Independent scholarship · Open accessThe monthly digest

Full manuscript · browser edition

Connections, Curvature, and Higher Geometry: From covariant differentiation to differential cohomology

Continuous, single-column reading. Equations, tables, proofs, references and appendices stay in document order.

Contents · 138 sections

Metrics and tangent geometry

A tangent vector describes the first-order motion of a point, and a metric assigns a length to that motion. This assignment gives norms on the fibers of the tangent bundle. Its smooth dependence on the base point is the input for connection and curvature. We begin with the way the assignment transforms under a change of coordinates, as in Tu’s treatment of metrics and curves [1, §§ 1–4].

The tangent map and pullback

For a smooth manifold MM, a tangent vector at pp can be defined as a derivation on germs of smooth functions at pp. In coordinates x1,…,xnx^1,\dots,x^n, any such derivation is ∑ivi∂i∣p\sum_i v^i\partial_i|_p, where vi=v(xi)v^i=v(x^i). To see the uniqueness, expand a function near pp as its value at pp plus ∑i(xi−xi(p))ai,ai(p)=∂if(p).\sum_i (x^i-x^i(p))a_i, \qquad a_i(p)=\partial_i f(p). One obtains the expansion by integrating the gradient along coordinate line segments. The product rule then gives v(f)=∑iviai(p)v(f)=\sum_i v^ia_i(p). This identifies TpMT_pM with Rn\mathbb R^n after a chart is chosen. If f:M→Nf:M\to N, set (dfpv)(h)=v(h∘f)(df_pv)(h)=v(h\circ f). It is a derivation at f(p)f(p), and this definition proves the chain rule d(g∘f)p=dgf(p) dfpd(g\circ f)_p=dg_{f(p)}\,df_p directly.

Definition 1.1. A Riemannian metric on MM is a smooth section gg of Sym⁡2T∗M\operatorname{Sym}^2T^*M such that gpg_p is positive definite for every p∈Mp\in M. Write ∣v∣g=gp(v,v)|v|_g=\sqrt{g_p(v,v)} for v∈TpMv\in T_pM.

A metric has the local expression gij dxi dxjg_{ij}\,dx^i\,dx^j, with summation understood and the product denoting the corresponding symmetric tensor. If x=x(y)x=x(y), the matrix in yy-coordinates equals JTgJJ^{\mathsf T}gJ, where J=(∂xi/∂ya)J=(\partial x^i/\partial y^a). This formula follows by evaluating gg on the vectors ∂/∂ya\partial/\partial y^a. In particular, positive definiteness does not depend on the chart.

Proposition 1.2. For every smooth f:M→Nf:M\to N and metric hh on NN, the symmetric tensor (f∗h)p(v,w)=hf(p)(dfpv,dfpw)(f^*h)_p(v,w)=h_{f(p)}(df_pv,df_pw) is positive semidefinite. It is positive definite at pp exactly when dfpdf_p is injective. If j:N→Qj:N\to Q and hh is a metric on QQ, then (j∘f)∗h=f∗(j∗h)(j\circ f)^*h=f^*(j^*h).

Proof. The squared length of vv under the tensor is ∣dfpv∣h2|df_pv|_h^2, whose zero set is ker⁡dfp\ker df_p. Smoothness follows from the coordinate formula. The formula for composites follows by applying the chain rule to both inputs. ◻

For an immersion M→NM\to N, the metric is the restriction of the ambient inner product to the immersed tangent directions. For a submersion with positive-dimensional fibers, its pullback metric vanishes along the vertical tangent bundle. This distinction will matter when taking quotient atlases. A tensor that is positive definite transversely to the fibers can encode a metric on a quotient even though it is degenerate on the atlas.

Existence by convex gluing

We use Hausdorff second-countable smooth manifolds, so smooth partitions of unity exist. On a coordinate chart UiU_i choose its Euclidean metric gig_i. Take a locally finite partition of unity {ϕi}\{\phi_i\} subordinate to the charts and put g=∑iϕigi.g=\sum_i \phi_i g_i . Each product extends smoothly by zero outside the chart because the support of ϕi\phi_i lies in UiU_i. The sum is smooth by local finiteness. At each pp and for each nonzero v∈TpMv\in T_pM, every summand in gp(v,v)g_p(v,v) is nonnegative, and at least one is positive since the coefficients sum to one. Thus gg is a metric. The same argument for Euclidean forms in bundle frames gives a metric on any real vector bundle over MM.

For any two metrics g0,g1g_0,g_1, the tensor gt=(1−t)g0+tg1g_t=(1-t)g_0+t g_1 is positive definite for 0≤t≤10\leq t\leq1. Hence every two metrics are joined by a smooth family. This interpolation is a statement about the convex space of metrics, with the underlying manifold and points fixed. Geodesics and curvature may vary along such an interpolation.

Example 1.3 (Rotational coordinates). The map (r,θ)↦(rcos⁡θ,rsin⁡θ)(r,\theta)\mapsto(r\cos\theta,r\sin\theta) has Euclidean pullback dx2+dy2=dr2+r2dθ2dx^2+dy^2=dr^2+r^2d\theta^2 on any angular chart with r>0r>0. This follows by squaring the two differentials and canceling the cross terms. The determinant of the metric matrix is r2r^2. At r=0r=0, the angular vector maps to zero, so the tensor written in these variables ceases to be positive definite. The degeneracy reflects the parametrization at the origin.

Length, speed, and oriented turning

For a piecewise smooth curve γ:[a,b]→M\gamma:[a,b]\to M, the metric defines its length by Lg(γ)=∫ab∣γ˙(t)∣g dt.L_g(\gamma)=\int_a^b |\dot\gamma(t)|_g\,dt . Length is invariant under increasing changes of parameter by the change-of-variables formula. Reversing the parameter leaves length unchanged because the norm is even. On a regular smooth segment, the length coordinate s(t)=∫at∣γ˙(u)∣g dus(t)=\int_a^t|\dot\gamma(u)|_g\,du has positive derivative. The inverse function theorem gives a smooth unit-speed parametrization on that segment.

In the standard oriented plane, let JJ be rotation by +π/2+\pi/2, write T=dγ/dsT=d\gamma/ds, and take the normal JTJT. Differentiation of ⟨T,T⟩=1\langle T,T\rangle=1 shows that dT/dsdT/ds is perpendicular to TT. Hence dTds=κ JT.\frac{dT}{ds}=\kappa\,JT . The real number κ\kappa is the signed curvature. Its value for an arbitrary regular parameter tt is κ(t)=det⁡(γ˙(t),γ¨(t))∣γ˙(t)∣3.\kappa(t)= \frac{\det(\dot\gamma(t),\ddot\gamma(t))} {|\dot\gamma(t)|^3}. Indeed γ˙=vT\dot\gamma=vT, with v=ds/dtv=ds/dt, so γ¨=v˙T+v2κJT\ddot\gamma=\dot v T+v^2\kappa JT. The determinant of T,JTT,JT equals one. The term involving v˙\dot v drops out, and the result is v3κv^3\kappa. If a local real angle θ\theta satisfies T=(cos⁡θ,sin⁡θ)T=(\cos\theta,\sin\theta), then κ=dθ/ds\kappa=d\theta/ds. The oriented tangent angle therefore records a one-dimensional form of curvature.

Example 1.4 (An ellipse and its osculating radius). For the counterclockwise ellipse γ(t)=(acos⁡t,bsin⁡t)\gamma(t)=(a\cos t,b\sin t), a,b>0a,b>0, the determinant in (1) is abab, and κ(t)=ab(a2sin⁡2t+b2cos⁡2t)3/2.\kappa(t)= \frac{ab}{(a^2\sin^2t+b^2\cos^2t)^{3/2}} . At (a,0)(a,0) the curvature is a/b2a/b^2. Its reciprocal b2/ab^2/a is the radius of the oriented circle tangent there with the same curvature. If a=ba=b, the signed curvature is 1/a1/a along the entire loop. Reversing the direction of traversal negates κ\kappa, while the norm of its curvature vector stays the same.

A tangent-vector field along a plane curve is an ordinary map into R2\mathbb R^2, once the ambient parallel coordinates have been chosen. A field along a curve in MM is instead a section of γ∗TM\gamma^*TM. The vectors above distinct points then belong to distinct vector spaces. Chapter [ch:levi] constructs a derivative of such sections by choosing a connection; the metric provides a canonical connection for the tangent bundle.

Volume and coordinate changes

On an oriented nn-manifold, the metric gives the positive form dVg=det⁡(gij) dx1∧⋯∧dxndV_g=\sqrt{\det(g_{ij})}\,dx^1\wedge\cdots\wedge dx^n in an oriented chart. Here det⁡(JTgJ)=(det⁡J)2det⁡g\det(J^{\mathsf T}gJ)=(\det J)^2\det g. Consequently the formula transforms as an nn-form under orientation-preserving changes of chart. Forgetting orientation gives the well-defined density ∣dVg∣|dV_g|. For a map ff that is an orientation-preserving local diffeomorphism, the same computation gives dVf∗h=f∗dVhdV_{f^*h}=f^*dV_h.

This construction depends on the metric algebraically at each point. If g′=e2ugg'=e^{2u}g for a smooth real function uu, the length of a tangent vector is multiplied by eue^u and dVg′=enudVgdV_{g'}=e^{nu}dV_g. Sectional curvature will depend on derivatives of uu as well. The different orders of dependence, first for connections and second for curvature, reappear in the local description by differential forms.

The length metric on each connected component is defined by dg(p,q)=inf⁡γLg(γ)d_g(p,q)=\inf_\gamma L_g(\gamma) over piecewise smooth paths from pp to qq. In a relatively compact coordinate ball about pp, the positive metric matrix admits uniform bounds cI≤(gij)≤CIcI\leq(g_{ij})\leq CI for some 0<c≤C0<c\leq C. Paths staying in the chart obey the same two-sided bounds for lengths up to factors c,C\sqrt c,\sqrt C. Paths leaving a smaller ball incur a positive lower length bound before their first exit. Coordinate line segments supply the local upper estimate in the smaller ball. These facts prove nondegeneracy and show that dgd_g induces the manifold topology. Finding length minimizers is the next question. Its local equation is a connection equation, as treated in Chapter [ch:geodesics].

Levi-Civita connection and torsion

A connection furnishes a derivative of tangent-vector fields. It compares vectors above adjacent points by specifying an infinitesimal rule with a product law for multiplying sections by functions. The construction recovers the directional derivative on Euclidean space and gives an intrinsic rule when the tangent bundle is presented only by local coordinates. We use the curvature convention R(X,Y)=[∇X,∇Y]−∇[X,Y]R(X,Y)=[\nabla_X,\nabla_Y]-\nabla_{[X,Y]} throughout this chapter and Chapter [ch:intrinsic].

Affine connections and the derivative of a section

Definition 1.1. An affine connection on MM is an R\mathbb R-bilinear operation (X,Y)↦∇XY(X,Y)\mapsto\nabla_XY on smooth vector fields such that ∇fXY=f∇XY,∇X(fY)=X(f)Y+f∇XY.\nabla_{fX}Y=f\nabla_XY,\qquad \nabla_X(fY)=X(f)Y+f\nabla_XY . Its torsion is the operation T(X,Y)=∇XY−∇YX−[X,Y].T(X,Y)=\nabla_XY-\nabla_YX-[X,Y].

These identities make all expressions local. For example, if YY vanishes near a given point pp, choose a cutoff that is one near pp with support in the vanishing neighborhood. Applying the product rule to the zero field obtained by multiplying by this cutoff gives ∇XY∣p=0\nabla_XY|_p=0. Smooth local formulas for ∇\nabla therefore restrict to smaller coordinate domains. In a local frame e1,…,ene_1,\ldots,e_n, put ∇eiej=Γkijek\nabla_{e_i}e_j=\Gamma^k{}_{ij}e_k. Expanding the connection rules gives ∇XY=∑k(X(Yk)+∑i,jXiYjΓkij)ek.\nabla_XY=\sum_k\left(X(Y^k)+ \sum_{i,j}X^iY^j\Gamma^k{}_{ij}\right)e_k . The coefficients are arbitrary smooth functions for a local connection. A second connection differs from the first by a section of T∗M⊗T∗M⊗TMT^*M\otimes T^*M\otimes TM, since the terms differentiating a multiplier cancel. For torsion the bracket satisfies [X,fY]=X(f)Y+f[X,Y][X,fY]=X(f)Y+f[X,Y]. The same cancellation proves C∞(M)C^\infty(M)-bilinearity and skew symmetry. In a coordinate frame Tkij=Γkij−ΓkjiT^k{}_{ij}=\Gamma^k{}_{ij}-\Gamma^k{}_{ji}. In a general frame there is also the contribution from [ei,ej][e_i,e_j].

Torsion can be changed without changing the equation for a curve. For any affine connection, put ∇X∘Y=∇XY−12T(X,Y)\nabla^\circ_XY=\nabla_XY-\tfrac12T(X,Y). The subtracted term is a tensor, so the Leibniz rule for the connection is preserved. The antisymmetric part of that tensor has exactly the size required to cancel TT. Thus ∇∘\nabla^\circ is torsion-free. Since T(v,v)=0T(v,v)=0, the two connections have the same affine geodesics. This construction uses the antisymmetry of torsion and requires no metric compatibility.

The metric and symmetry determine the connection

Theorem 1.2 (Levi-Civita). For a Riemannian metric gg, exactly one affine connection satisfies Xg(Y,Z)=g(∇XY,Z)+g(Y,∇XZ),T(X,Y)=0.Xg(Y,Z)=g(\nabla_XY,Z)+g(Y,\nabla_XZ), \qquad T(X,Y)=0.

Proof. Write the compatibility identity in the cyclic orders (X,Y,Z)(X,Y,Z), (Y,Z,X)(Y,Z,X), and (Z,X,Y)(Z,X,Y). Add the first two and subtract the third. Torsion-freeness replaces each reversed derivative by the given derivative and a bracket. The result is the Koszul formula 2g(∇XY,Z)=Xg(Y,Z)+Yg(Z,X)−Zg(X,Y)−g(X,[Y,Z])+g(Y,[Z,X])+g(Z,[X,Y]).\begin{aligned} 2g(\nabla_XY,Z) &= Xg(Y,Z)+Yg(Z,X)-Zg(X,Y)\notag\\ &\quad -g(X,[Y,Z])+g(Y,[Z,X])+g(Z,[X,Y]). \end{aligned} Nondegeneracy gives uniqueness. To see existence, denote the right side by K(X,Y,Z)K(X,Y,Z). The product rule and the bracket identity show that K(X,Y,fZ)=fK(X,Y,Z)K(X,Y,fZ)=fK(X,Y,Z): the coefficients of X(f)X(f) and Y(f)Y(f) from differentiating the metric evaluations cancel the corresponding bracket terms. The same expansion gives K(fX,Y,Z)=fK(X,Y,Z)K(fX,Y,Z)=fK(X,Y,Z) and K(X,fY,Z)=fK(X,Y,Z)+2X(f)g(Y,Z)K(X,fY,Z)=fK(X,Y,Z)+2X(f)g(Y,Z). Thus (2) defines a smooth affine connection using gg to recover a vector from a covector. Subtracting K(Y,X,Z)K(Y,X,Z) from K(X,Y,Z)K(X,Y,Z) leaves 2g([X,Y],Z)2g([X,Y],Z), which proves zero torsion. Adding K(X,Y,Z)K(X,Y,Z) to K(X,Z,Y)K(X,Z,Y) leaves 2Xg(Y,Z)2Xg(Y,Z), which proves metric compatibility. ◻

With coordinate fields all brackets vanish. Inverting the metric matrix gives the explicit formula Γkij=12∑ℓgkℓ(∂igℓj+∂jgℓi−∂ℓgij).\Gamma^k{}_{ij}=\frac12\sum_\ell g^{k\ell} \bigl(\partial_i g_{\ell j}+\partial_j g_{\ell i} -\partial_\ell g_{ij}\bigr). In particular, the connection depends on the first derivatives of the metric. Formula (2) applies also to orthonormal frames with nonzero brackets, where the derivative terms of the metric evaluations are then zero. These are two complementary ways to compute the same connection [1, §§ 6,13].

Proposition 1.3. Let f:(M,g)→(N,h)f:(M,g)\to(N,h) be an isometric local diffeomorphism. If X,YX,Y on MM correspond locally to df(X),df(Y)df(X),df(Y) on NN, then df(∇XgY)=∇df(X)hdf(Y)df(\nabla^g_XY)=\nabla^h_{df(X)}df(Y) there.

Proof. The derivative of a local diffeomorphism preserves the vector-field bracket because the conjugated derivations have conjugated commutators. Transport the connection on NN to MM in any neighborhood where ff is a diffeomorphism. The transported connection has zero torsion. It is compatible with f∗h=gf^*h=g by differentiation of the pulled-back inner products. Uniqueness gives the asserted identity. The local computations agree on overlaps. ◻

There is a closely related rule for an isometric immersion into Euclidean space. Decompose the ambient derivative DXYD_XY into its tangential and normal parts. The tangential projection is compatible with the induced metric because normal fields are perpendicular to tangent fields. Its torsion is the tangential projection of DXY−DYX−[X,Y]=0D_XY-D_YX-[X,Y]=0. Thus the tangential projection is the Levi-Civita connection. The normal projection defines the second fundamental form used in Chapter [ch:intrinsic].

Covariant differentiation along a path

Let γ:I→M\gamma:I\to M and let V(t)∈Tγ(t)MV(t)\in T_{\gamma(t)}M be a smooth section of the pullback tangent bundle. In coordinates around a segment of γ\gamma define (∇tV)k=dVkdt+∑i,jΓkij(γ)γ˙iVj.(\nabla_t V)^k =\frac{d V^k}{dt}+ \sum_{i,j}\Gamma^k{}_{ij}(\gamma)\dot\gamma^i V^j . This can also be derived locally by expressing V=∑jVj(t)ej(γ(t))V=\sum_jV^j(t)e_j(\gamma(t)) in any frame and applying the product rule with ∇t(ej∘γ)=(∇ej)γ(t)(γ˙(t))\nabla_t(e_j\circ\gamma)=(\nabla e_j)_{\gamma(t)}(\dot\gamma(t)). If a frame is changed by a matrix B(γ(t))B(\gamma(t)), differentiating BB by the ordinary chain rule gives the same change as applying ∇γ˙\nabla_{\dot\gamma} to the new frame fields. Hence the result is independent of the frame. Formula (4) also applies when the path crosses itself or is momentarily stationary. It requires only the pulled-back section and no extension as a field on MM.

A parallel section has ∇tV=0\nabla_t V=0. This is a linear differential equation in VV, so an initial vector determines a unique solution over the path interval. The solution operators compose on concatenation, and reversing the path gives the inverse. Here γ12∗γ01\gamma_{12}*\gamma_{01} denotes first γ01\gamma_{01} and then γ12\gamma_{12}. image For a metric connection, compatibility gives ddtg(V,W)=g(∇tV,W)+g(V,∇tW).\frac{d}{dt} g(V,W) =g(\nabla_tV,W)+g(V,\nabla_tW). In particular, parallel transport preserves the Euclidean structures on the initial and terminal fibers. Its dependence on the full path encodes curvature, as developed in Chapter [ch:geodesics] and in the subsequent treatment of holonomy.

A geodesic with affine parameter satisfies ∇tγ˙=0\nabla_t\dot\gamma=0, or γ¨k+Γkij(γ)γ˙iγ˙j=0\ddot\gamma^k+\Gamma^k{}_{ij}(\gamma)\dot\gamma^i\dot\gamma^j=0 in coordinates. Local existence and uniqueness follow by converting the equation into a smooth first-order system on TMTM. Metric compatibility makes the speed constant. The equation is invariant under affine reparametrization t↦at+bt\mapsto at+b. A smooth reparametrization with a nonzero second derivative introduces a tangential acceleration proportional to that second derivative.

Example 1.4 (A circle in a flat metric). For the metric dr2+r2dθ2dr^2+r^2d\theta^2 in Chapter [ch:metric], the nonzero coordinate coefficients are Γrθθ=−r,Γθrθ=Γθθr=r−1.\Gamma^r{}_{\theta\theta}=-r,\qquad \Gamma^\theta{}_{r\theta}=\Gamma^\theta{}_{\theta r}=r^{-1}. The unit field eθ=r−1∂θe_\theta=r^{-1}\partial_\theta satisfies ∇eθeθ=−r−1∂r\nabla_{e_\theta}e_\theta=-r^{-1}\partial_r on a circle of radius rr. Its acceleration has length 1/r1/r and points toward the center. This agrees with (1). Ambient Euclidean geodesics remain straight lines. In polar coordinates these same lines satisfy an equation with nonzero connection coefficients, since the frame varies with the point.

A connection on a Lie group

Let GG carry a bi-invariant Riemannian metric, such as the metric induced by the negative trace pairing on the skew-Hermitian matrices of a unitary group. Identify the Lie algebra with the left-invariant fields. Invariance of the inner product at the identity under the adjoint action gives g([X,Y],Z)+g(Y,[X,Z])=0g([X,Y],Z)+g(Y,[X,Z])=0 on such fields, by differentiating the adjoint action along a one-parameter subgroup. In the Koszul formula their pairwise inner products are constant, and interchanging arguments in the bracket terms using the displayed identity gives ∇XY=12[X,Y]\nabla_XY=\tfrac12[X,Y]. A one-parameter subgroup γ(t)=exp⁡(tξ)\gamma(t)=\exp(t\xi) therefore has ∇tγ˙=0\nabla_t\dot\gamma=0. In left-translated coordinates v(t)∈gv(t)\in\mathfrak g, a field along this curve is parallel when v˙(t)+12[ξ,v(t)]=0.\dot v(t)+\tfrac12[\xi,v(t)]=0. Its solution is v(t)=exp⁡(−t ad⁡ξ/2)v(0)v(t)=\exp(-t\,\operatorname{ad}_\xi/2)v(0). The first exponential above is the Lie group exponential, and the one in this formula is the exponential of a linear operator. The two exponentials describe the moving base point and the moving vector, respectively. The operator is skew for the invariant inner product, so transport is orthogonal. In the next chapter the same connection gives curvature through a single Lie bracket.

The connection is the structure governing motion of a single vector along a single curve. The intrinsic effect of commuting two such infinitesimal motions is the curvature tensor. Its definition subtracts the derivative along the commutator of the underlying vector fields. This subtraction is essential when the fields come from a noncoordinate frame.

Intrinsic and extrinsic curvature

The acceleration of a curve in Euclidean space can be resolved into components tangent and normal to an immersed submanifold. The tangential component is governed by the induced metric. In dimension two the normal component also determines the intrinsic Gaussian curvature through a quadratic identity. We derive that identity from flatness of the ambient derivative and the splitting of the ambient tangent bundle. This gives the tensor form of the calculation underlying Tu’s treatment [1, §§ 5,8–9,12].

The curvature of a connection

The curvature of ∇\nabla is R(X,Y)Z=∇X∇YZ−∇Y∇XZ−∇[X,Y]Z.R(X,Y)Z=\nabla_X\nabla_Y Z-\nabla_Y\nabla_X Z -\nabla_{[X,Y]}Z . It is C∞(M)C^\infty(M)-linear in all three fields. In the first input, the term −Y(f)∇XZ-Y(f)\nabla_XZ arising from fXfX cancels the contribution of the bracket [fX,Y]=f[X,Y]−Y(f)X[fX,Y]=f[X,Y]-Y(f)X. Skew symmetry then handles the second input. In the third input, the terms differentiating ff once cancel between the two compositions of ∇\nabla, and the scalar second derivatives cancel against [X,Y](f)[X,Y](f). Thus RR is a two-form with values in endomorphisms of TMTM.

For a metric connection, applying XY−YX−[X,Y]XY-YX-[X,Y] to the scalar g(Z,W)g(Z,W) and expanding metric compatibility gives g(R(X,Y)Z,W)=−g(Z,R(X,Y)W).g(R(X,Y)Z,W)=-g(Z,R(X,Y)W). The mixed terms from differentiating both fields cancel in pairs. For a torsion-free connection, the cyclic sum in X,Y,ZX,Y,Z of (5) is zero. To check this at a point, use coordinate fields, so that all brackets vanish. Pairing terms with the same outer derivative leaves ∇X(∇YZ−∇ZY)\nabla_X(\nabla_Y Z-\nabla_Z Y) and its cyclic permutations, each zero by symmetry of the Christoffel symbols. Tensoriality extends the statement to all fields.

For Levi-Civita curvature there is also pair exchange. Use the components Rijkℓ=g(R(∂i,∂j)∂k,∂ℓ).R_{ijk\ell}=g(R(\partial_i,\partial_j)\partial_k,\partial_\ell). At a point pp, start with a chart xx centered there and introduce yy by xa=ya−12Γabc(p)ybyc.x^a=y^a-\tfrac12\Gamma^a{}_{bc}(p)y^by^c . This defines new coordinates near pp by the inverse function theorem. The second derivatives of x(y)x(y) cancel the old Christoffel symbols at pp. Torsion-freeness ensures they are symmetric in b,cb,c. In the new chart Γ(p)=0\Gamma(p)=0, and metric compatibility also gives ∂igjk(p)=0\partial_i g_{jk}(p)=0. Evaluating (3) and its first derivatives at pp gives 2Rijkℓ=∂i∂kgℓj−∂i∂ℓgjk−∂j∂kgℓi+∂j∂ℓgik.\begin{split} 2R_{ijk\ell} &=\partial_i\partial_k g_{\ell j} -\partial_i\partial_\ell g_{jk}\\ &\quad-\partial_j\partial_k g_{\ell i} +\partial_j\partial_\ell g_{ik}. \end{split} Since mixed derivatives commute and gg is symmetric, this expression equals 2Rkℓij2R_{k\ell ij}. These identities characterize the symmetries of the metric curvature tensor [2].

For a two-plane Π⊆TpM\Pi\subseteq T_pM, choose a spanning pair X,YX,Y. Its sectional curvature is K(Π)=g(R(X,Y)Y,X)∣X∣g2∣Y∣g2−g(X,Y)2.K(\Pi)= \frac{g(R(X,Y)Y,X)} {|X|_g^2|Y|_g^2-g(X,Y)^2}. In the numerator, skew symmetry in each pair gives a quadratic expression on X∧YX\wedge Y. Both numerator and denominator therefore scale by the square of the determinant of a change of basis of Π\Pi. This proves independence of the spanning pair. On a surface there is only one plane at each point, giving its Gaussian curvature.

Example 1.1 (Bi-invariant metrics). In the Lie group example of Chapter [ch:levi], let X,Y,ZX,Y,Z be left-invariant fields. The commutator of their covariant derivatives acts on ZZ by 14[[X,Y],Z]\tfrac14[[X,Y],Z], since the Jacobi identity rewrites the difference of the two nested brackets. The last term of (5) acts by 12[[X,Y],Z]\tfrac12[[X,Y],Z]. Thus R(X,Y)Z=−14[[X,Y],Z].R(X,Y)Z=-\tfrac14[[X,Y],Z]. Taking Z=YZ=Y, pairing with XX, and using skew-adjointness of the bracket gives g(R(X,Y)Y,X)=14∣[X,Y]∣2g(R(X,Y)Y,X)=\tfrac14|[X,Y]|^2. Positive sectional curvature in a plane therefore occurs exactly when the two directions fail to commute. On a torus all brackets vanish, and this metric is flat.

The Gauss and Codazzi identities

Let MnM^n be immersed in Rn+1\mathbb R^{n+1} with induced metric. Work locally with a smooth unit normal NN and write DD for the ambient Euclidean derivative. Differentiating ⟨N,N⟩=1\langle N,N\rangle=1 shows that DXND_XN is tangent. Define the shape operator S(X)=−DXNS(X)=-D_XN and the second fundamental form II(X,Y)=g(SX,Y)\mathrm{II}(X,Y)=g(SX,Y). Differentiating ⟨Y,N⟩=0\langle Y,N\rangle=0 yields DXY=∇XY+II(X,Y)N.D_XY=\nabla_XY+\mathrm{II}(X,Y)N . Here ∇\nabla is Levi-Civita by Chapter [ch:levi]. Taking the normal component of DXY−DYX=[X,Y]D_XY-D_YX=[X,Y] proves that II\mathrm{II} is symmetric. This decomposition can also be recorded by the splitting image The derivative in the ambient bundle does not preserve either summand in general, and SS and II\mathrm{II} measure its off-diagonal terms.

Theorem 1.2 (Gauss and Codazzi for a Euclidean hypersurface). One has R(X,Y)Z=II(Y,Z)SX−II(X,Z)SY,(∇XII)(Y,Z)=(∇YII)(X,Z).\begin{aligned} R(X,Y)Z &=\mathrm{II}(Y,Z)SX-\mathrm{II}(X,Z)SY,\\ (\nabla_X\mathrm{II})(Y,Z)&=(\nabla_Y\mathrm{II})(X,Z). \end{aligned}

Proof. Insert the decomposition of DYZD_YZ into DXDYZD_XD_YZ, subtract the counterpart with X,YX,Y interchanged, and subtract D[X,Y]ZD_{[X,Y]}Z. Flatness of the Euclidean derivative makes the result zero. Its tangential component is R(X,Y)Z−II(Y,Z)SX+II(X,Z)SYR(X,Y)Z-\mathrm{II}(Y,Z)SX+\mathrm{II}(X,Z)SY. Its normal component is the coefficient XII(Y,Z)−YII(X,Z)+II(X,∇YZ)−II(Y,∇XZ)−II([X,Y],Z).\begin{split} X\mathrm{II}(Y,Z)-Y\mathrm{II}(X,Z) &+\mathrm{II}(X,\nabla_YZ)-\mathrm{II}(Y,\nabla_XZ)\\ &-\mathrm{II}([X,Y],Z). \end{split} Use [X,Y]=∇XY−∇YX[X,Y]=\nabla_XY-\nabla_YX and the product rule for the covariant derivative of a covariant tensor. The coefficient is the difference of the two derivatives in the second identity. ◻

When n=2n=2, choosing an orthonormal basis gives K=det⁡SK=\det S. Changing NN to −N-N negates SS while preserving its determinant in dimension two. Because the right-hand definition of KK uses only the metric and its Levi-Civita connection, two isometric surfaces have equal Gaussian curvature at corresponding points. This is the intrinsic content of the Theorema Egregium.

Example 1.3 (Graph, cylinder, and sphere). For a graph z=f(x,y)z=f(x,y) in R3\mathbb R^3, put w=(1+∣∇f∣2)1/2w=(1+|\nabla f|^2)^{1/2} and use the upward normal N=(−fx,−fy,1)/wN=(-f_x,-f_y,1)/w. Its first fundamental form has determinant w2w^2. The second fundamental form has matrix (fij/w)(f_{ij}/w). Thus K=det⁡(fij)/w4K=\det(f_{ij})/w^4. Taking f=xyf=xy gives negative curvature at every point of its graph. On the cylinder of radius aa, the two eigenvalues of SS with outward normal are −1/a-1/a in the circular direction and 00 along its axis. Hence K=0K=0. The local parametrization (s,t)↦(acos⁡(s/a),asin⁡(s/a),t)(s,t)\mapsto(a\cos(s/a),a\sin(s/a),t) exhibits the same result by pulling the metric back to ds2+dt2ds^2+dt^2. On the sphere of radius aa one has S=−a−1id⁡S=-a^{-1}\operatorname{id}, so K=a−2K=a^{-2}.

A curvature form on an oriented surface

Take a positive orthonormal frame e1,e2e_1,e_2 and define a one-form ω\omega by ∇e1=ωe2\nabla e_1=\omega e_2. Metric compatibility gives ∇e2=−ωe1\nabla e_2=-\omega e_1. Insert these two equations into (5). The product terms in ω(X)ω(Y)\omega(X)\omega(Y) cancel, yielding R(X,Y)e1=(dω)(X,Y)e2,dω=−K dVg.R(X,Y)e_1=(d\omega)(X,Y)e_2,\qquad d\omega=-K\,dV_g . The minus sign follows from K=g(R(e1,e2)e2,e1)K=g(R(e_1,e_2)e_2,e_1). For the positive coframe (θ1,θ2)(\theta^1,\theta^2), the first structure equations are dθ1=ω∧θ2d\theta^1=\omega\wedge\theta^2 and dθ2=−ω∧θ1d\theta^2=-\omega\wedge\theta^1. They follow by evaluating dθid\theta^i on two fields and replacing their bracket by the difference of the covariant derivatives. Rotating the frame through an angle ϕ\phi changes ω\omega to ω+dϕ\omega+d\phi, by differentiation of the rotation. The curvature form is thus unchanged. When local angles have only circle-valued transitions, the differentials dϕd\phi still agree for different real lifts. In Chapter [ch:geodesics], integration of this identity turns those frame changes into a topological integer.

Example 1.4 (The effect of conformally scaling a plane metric). Take g=e2u(dx2+dy2)g=e^{2u}(dx^2+dy^2) and the positive coframe (eudx,eudy)(e^u dx,e^u dy). The first structure equations give ω=−uy dx+ux dy\omega=-u_y\,dx+u_x\,dy. Consequently K=−e−2u(uxx+uyy).K=-e^{-2u}(u_{xx}+u_{yy}). For u=−log⁡yu=-\log y on the upper half-plane, this is −1-1. For u=log⁡(2a2)−log⁡(a2+x2+y2)u=\log(2a^2)-\log(a^2+x^2+y^2), the Laplacian is −4a2/(a2+x2+y2)2-4a^2/(a^2+x^2+y^2)^2, giving K=a−2K=a^{-2}. The latter metric is the usual stereographic expression for the round sphere of radius aa. More generally, on a conformal chart the curvature is controlled by a scalar second-order equation. The sign of its Laplacian determines the sign of the Gaussian curvature pointwise.

Geodesics, Jacobi fields, and surfaces

The covariant derivative turns geodesic motion into a differential equation. A variation of one geodesic through other geodesics satisfies a new, linear equation containing curvature. On surfaces the same curvature has a global integral constrained by the topology. The local motion and the global identity thus give two ways of detecting the curvature defined in Chapter [ch:intrinsic]. We use Tu’s conventions for the geodesic equation and his presentation of the surface theorem as a comparison [1, §§ 13–17].

Energy variation and the exponential map

Let Γ(s,t)\Gamma(s,t) be a smooth variation of a path γ(t)=Γ(0,t)\gamma(t)=\Gamma(0,t) on [a,b][a,b]. Put T(s,t)=∂tΓ(s,t)T(s,t)=\partial_t\Gamma(s,t) and V(s,t)=∂sΓ(s,t)V(s,t)=\partial_s\Gamma(s,t). The energy of the path t↦Γ(s,t)t\mapsto\Gamma(s,t) is E(s)=12∫abgΓ(s,t)(T(s,t),T(s,t)) dt,E(s)=\frac12\int_a^b g_{\Gamma(s,t)}\bigl(T(s,t),T(s,t)\bigr)\,dt , with ss held fixed during integration. Both fields are sections of Γ∗TM\Gamma^*TM. The torsion-free identity ∇sT=∇tV\nabla_sT=\nabla_tV holds because the mixed partials of Γ\Gamma agree and the Christoffel contributions are symmetric. Therefore E′(s)=[g(V,T)]ab−∫abg(V,∇tT) dt.E'(s)=\bigl[g(V,T)\bigr]_a^b- \int_a^b g(V,\nabla_tT)\,dt . Here metric compatibility first replaces the derivative of the squared speed by 2g(∇sT,T)2g(\nabla_sT,T), and then gives the integration-by-parts formula. For variations fixing both endpoints, the boundary term vanishes. Every smooth field supported in a small segment of the interior serves as a variation vector there, using coordinate displacement multiplied by a cutoff. Thus E′(0)=0E'(0)=0 for every fixed-endpoint variation implies the geodesic equation along γ\gamma by the fundamental lemma of the calculus of variations. The reverse implication follows from (6).

Let γv\gamma_v start at pp with velocity vv. Put exp⁡p(v)=γv(1)\exp_p(v)=\gamma_v(1) when the solution exists on [0,1][0,1]. The solution depends smoothly on the initial conditions. The rescaling relation γtv(1)=γv(t)\gamma_{tv}(1)=\gamma_v(t) on its domain gives (dexp⁡p)0=id⁡TpM(d\exp_p)_0=\operatorname{id}_{T_pM}, under the canonical identification of the tangent space of TpMT_pM at zero. Thus exp⁡p\exp_p maps a sufficiently small open ball diffeomorphically onto a normal neighborhood. The coordinates from a linear orthonormal identification TpM≃RnT_pM\simeq\mathbb R^n then have radial lines as geodesics.

Proposition 1.1 (Radial orthogonality). At a point vv in a normal ball and for w∈TpMw\in T_pM, one has gexp⁡p(v)((dexp⁡p)vv,(dexp⁡p)vw)=gp(v,w).g_{\exp_p(v)}\big((d\exp_p)_v v,(d\exp_p)_v w\big) =g_p(v,w).

Proof. Use Γ(s,t)=exp⁡p(t(v+sw))\Gamma(s,t)=\exp_p(t(v+sw)) for small ss and 0≤t≤10\leq t\leq1. Along a radial geodesic ∇tT=0\nabla_tT=0, so ∂tg(V,T)=g(∇sT,T)=12∂s∣v+sw∣gp2.\partial_t g(V,T)=g(\nabla_sT,T) =\tfrac12\partial_s |v+sw|_{g_p}^2. As V(s,0)=0V(s,0)=0, integration to t=1t=1 and evaluation at s=0s=0 give the formula. The vector T(s,1)T(s,1) at that endpoint is (dexp⁡p)v+sw(v+sw)(d\exp_p)_{v+sw}(v+sw). ◻

Consequently radial directions have unit length under the radial distance coordinate and are perpendicular to variations tangent to its spheres. For a path in the normal neighborhood the speed is at least the absolute derivative of the radius ∣exp⁡p−1(γ(t))∣|\exp_p^{-1}(\gamma(t))| wherever the radius is positive. Integrating, with limiting endpoints at zero allowed, proves minimality of the radial segment among paths in the neighborhood. If the endpoint has radius rr smaller than the radius RR of a compactly contained normal ball, any path leaving that ball has initial length at least RR by the same estimate up to its first exit. Thus the radial segment minimizes among all paths to the sufficiently near endpoint. These arguments are local and do not assume geodesic completeness.

The linearized geodesic equation

Suppose now that every t↦Γ(s,t)t\mapsto\Gamma(s,t) is geodesic. The curvature commutator on Γ∗TM\Gamma^*TM gives ∇t2V=∇t∇sT=∇s∇tT+R(T,V)T=−R(V,T)T.\nabla_t^2 V=\nabla_t\nabla_s T =\nabla_s\nabla_t T+R(T,V)T =-R(V,T)T . At s=0s=0, the field J=V∣s=0J=V|_{s=0} thus satisfies ∇t2J+R(J,γ˙)γ˙=0.\nabla_t^2 J+R(J,\dot\gamma)\dot\gamma=0 . Conversely, the linear equation has unique solutions prescribed by J(0),∇tJ(0)J(0),\nabla_t J(0). Choose a smooth curve of initial positions with derivative J(0)J(0), and vary the initial velocities along it with covariant derivative ∇tJ(0)\nabla_t J(0). To construct the varied velocities, use parallel transport along the curve of positions and add the desired first-order term in that trivialization. Smooth dependence of the geodesic flow gives a variation, at least locally on the geodesic interval, whose variation field has those two initial values. Uniqueness of the linear equation identifies it with JJ. This exhibits Jacobi fields as the linear response of geodesic motion [2].

For a unit-speed geodesic in constant sectional curvature KK, the operator J↦R(J,γ˙)γ˙J\mapsto R(J,\dot\gamma)\dot\gamma on the orthogonal complement of γ˙\dot\gamma is Kid⁡K\operatorname{id}. Indeed it is self-adjoint by the curvature symmetries, and its quadratic form on perpendicular vectors is K∣J∣2K|J|^2. In a parallel orthonormal frame perpendicular to γ˙\dot\gamma, a field with J(0)=0J(0)=0 has coefficients proportional to sK(t)={sin⁡(Kt)/K,K>0,t,K=0,sinh⁡(−Kt)/−K,K<0.s_K(t)= \begin{cases} \sin(\sqrt Kt)/\sqrt K,&K>0,\\ t,&K=0,\\ \sinh(\sqrt{-K}t)/\sqrt{-K},&K<0 . \end{cases} The tangential component with zero initial value is linear in tt. On the round nn-sphere of radius aa with n≥2n\geq2, the first positive time for a nonzero such field to vanish is πa\pi a, so the antipode is the first conjugate point. Varying the initial velocity in a geodesic with initial point held fixed identifies these vanishing fields with the kernel of the derivative of the exponential map at that time. For K≤0K\leq0 the displayed solutions give no positive conjugate time in a constant-curvature model.

Second variation and local stability

Let γ\gamma be a geodesic and suppose the variation fixes both endpoints. Differentiate (6) at s=0s=0. The term involving ∇sV\nabla_sV vanishes there because ∇tT=0\nabla_tT=0, so E′′(0)=−∫abg(V,∇s∇tT) dt.E''(0)=-\int_a^b g(V,\nabla_s\nabla_tT)\,dt . The curvature commutator and ∇sT=∇tV\nabla_sT=\nabla_tV give ∇s∇tT=∇t2V+R(V,T)T\nabla_s\nabla_tT=\nabla_t^2V+R(V,T)T at s=0s=0. Integration by parts, using V(a)=V(b)=0V(a)=V(b)=0, yields the quadratic form I(V,V)=∫ab(∣∇tV∣2−g(R(V,T)T,V)) dt.I(V,V)=\int_a^b \bigl(|\nabla_t V|^2-g(R(V,T)T,V)\bigr)\,dt. Its bilinear polarization is symmetric by the metric curvature identities. A Jacobi field JJ vanishing at the endpoints pairs to zero with every such VV, by (7) and integration by parts. The converse follows by testing against fields supported on interior segments. Thus conjugate endpoint data are detected by the kernel of the second variation.

For a unit-speed segment of length LL on the round sphere, choose a perpendicular parallel unit field ee, and set V(t)=sin⁡(πt/L)e(t)V(t)=\sin(\pi t/L)e(t) for 0≤t≤L0\leq t\leq L. The two terms in II integrate explicitly to I(V,V)=L2((π/L)2−a−2).I(V,V)=\frac L2\bigl((\pi/L)^2-a^{-2}\bigr). The variation (s,t)↦exp⁡γ(t)(sV(t))(s,t)\mapsto\exp_{\gamma(t)}(sV(t)) is defined for sufficiently small ss, by compactness of the path interval and the local existence theorem. When L>πaL>\pi a the quadratic form is negative on VV, giving a decrease in energy through that variation. At L=πaL=\pi a, the displayed field is the conjugate Jacobi field already computed.

Example 1.2 (The upper half-plane). On y>0y>0, the metric g=y−2(dx2+dy2)g=y^{-2}(dx^2+dy^2) has oriented orthonormal frame e1=y∂xe_1=y\partial_x, e2=y∂ye_2=y\partial_y. Its bracket is [e1,e2]=−e1[e_1,e_2]=-e_1, so the Koszul formula yields ∇e1e1=e2\nabla_{e_1}e_1=e_2 and ∇e2e1=0\nabla_{e_2}e_1=0. Hence ω=dx/y\omega=dx/y and dω=dx∧dy/y2=dVgd\omega=dx\wedge dy/y^2=dV_g, giving K=−1K=-1. Curves t↦x0+iy0ett\mapsto x_0+iy_0e^t have unit velocity e2e_2 and are geodesics. For a,b,c,d∈Ra,b,c,d\in\mathbb R with ad−bc>0ad-bc>0, the map z↦(az+b)/(cz+d)z\mapsto (az+b)/(cz+d) multiplies the imaginary part and the norm of the Euclidean differential by the same factor (ad−bc)/∣cz+d∣2(ad-bc)/|cz+d|^2. It therefore preserves gg. The images of the imaginary axis under z↦x0+r(z−1)/(z+1)z\mapsto x_0+r(z-1)/(z+1), r>0r>0, are the semicircles centered at x0x_0. These are geodesics as well, by the naturality result of Chapter [ch:levi].

From boundary turning to Gauss–Bonnet

For a unit oriented boundary tangent T=cos⁡θ e1+sin⁡θ e2T=\cos\theta\,e_1+\sin\theta\,e_2 on a smooth arc, define kg=g(∇TT,JT)k_g=g(\nabla_TT,JT), where JJ is positive rotation by π/2\pi/2. In a local positive frame the derivative of the expression for TT gives kg=dθ/ds+ω(T)k_g=d\theta/ds+\omega(T). On a positively oriented embedded triangular disk DD in an oriented chart, the boundary tangent, with signed exterior turns inserted at its three corners, turns by 2π2\pi. One may compute this using the planar turning theorem in the chart, since a positive frame change extending over a disk contributes zero winding. The turning identity and Stokes’ theorem, with dω=−K dVgd\omega=-K\,dV_g, give ∫DK dVg+∫∂Dkg ds+∑corners(π−α)=2π,\int_D K\,dV_g+\int_{\partial D} k_g\,ds +\sum_{\mathrm{corners}}(\pi-\alpha)=2\pi , where α\alpha is the interior angle at that corner [1, § 17.4].

For a compact oriented surface without boundary, choose a smooth triangulation fine enough that each triangle lies in a coordinate neighborhood. Sum (8) over triangles. The contributions along a shared edge cancel: reversing its orientation negates kgk_g and preserves dsds. Interior angles around each vertex sum to 2π2\pi, because the incident tangent sectors fill its tangent circle. As each triangle has three corners, the sum of exterior turns is 3πF−2πV3\pi F-2\pi V, where F,VF,V count faces and vertices. With 3F=2E3F=2E for the edges, this yields ∫MK dVg=2π(V−E+F)=2πχ(M).\int_M K\,dV_g=2\pi(V-E+F)=2\pi\chi(M). For the round sphere the calculation gives a−2(4πa2)=4πa^{-2}(4\pi a^2)=4\pi. The independence from aa expresses a general fact: a curvature form can vary with the metric while the cohomology class it represents stays fixed. This is the basic surface example of the characteristic-class construction developed later in the book.

Vector bundles and their descent

A frame identifies vectors over nearby points with a single model vector space. Two frames identify that space in different ways, and their change of coordinates is part of the geometry. The descent description retains the change of coordinates and the equations it satisfies. It thus applies to sections, to bundle maps and to connections. We use smooth manifolds that are Hausdorff and second countable, with real or complex vector bundles of finite rank. The local frame and section descriptions follow the setting of Tu [1, §7].

Pullback, local frames, and sections

A rank-rr vector bundle over MM is a smooth map p ⁣:E→Mp\colon E\to M with a vector space structure on each fiber and local product identifications E∣U≃U×Kr,K=R or C,E|_U \simeq U\times \mathbb K^r,\qquad \mathbb K=\mathbb R\ \text{or}\ \mathbb C, which are linear on fibers. A frame on UU is an ordered tuple of sections whose values are a basis at every point of UU. Write eie_i for the row of basis vectors in a frame on UiU_i, so eivie_i v_i denotes the vector with column of coordinates viv_i. On an overlap put ej=eigije_j=e_i g_{ij}. Then gij ⁣:Ui∩Uj⟶GLr(K),vi=gijvj.g_{ij}\colon U_i\cap U_j\longrightarrow \mathrm{GL}_r(\mathbb K), \qquad v_i=g_{ij}v_j. The matrix entries of gijg_{ij} are smooth because a bundle chart smoothly extracts the coordinates of the smooth sections of the second frame. On a triple overlap, gijgjk=gikg_{ij}g_{jk}=g_{ik}, by expressing eke_k in frame eie_i in two ways.

For f ⁣:N→Mf\colon N\to M the pullback bundle is f∗E={(y,e)∈N×E∣f(y)=p(e)}.f^*E=\{(y,e)\in N\times E\mid f(y)=p(e)\}. It is smooth and locally trivial since the inverse images of bundle charts give product charts for it. The commutative square image has the pullback property in smooth manifolds. A smooth map into the upper left corner consists of maps into NN and EE satisfying the compatibility at MM. Smoothness of the resulting map can be checked in the product charts above, so no assumption that arbitrary fiber products of manifolds exist is needed. On frames, pullback replaces gijg_{ij} by gij∘fg_{ij}\circ f. For a further map h ⁣:P→Nh\colon P\to N, the comparison h∗f∗E≃(fh)∗Eh^*f^*E\simeq(fh)^*E forgets the redundant intermediate base point. The comparisons for three successive maps agree since both delete the same intermediate points.

Sections over an open set UU form a C∞(U,K)C^\infty(U,\mathbb K)-module Γ(U,E)\Gamma(U,E). Their restrictions form a sheaf, because smoothness of a section can be tested in local frames. Concretely a section over MM is a family of smooth vi ⁣:Ui→Krv_i\colon U_i\to\mathbb K^r with vi=gijvjv_i=g_{ij}v_j. A bundle map F ⁣:E→E′F\colon E\to E' over MM is recorded on a shared trivializing cover by matrices FiF_i with Figij=gij′Fj.F_i g_{ij}=g'_{ij}F_j . This equation holds for rectangular matrices as well. It states that converting a vector to the ii-frame and applying the map agrees with applying the map in the jj-frame and converting the result.

Proposition 1.1. Every C∞(M,K)C^\infty(M,\mathbb K)-linear map Γ(M,E)→Γ(M,E′)\Gamma(M,E)\to\Gamma(M,E') arises from a unique bundle map.

Proof. Let TT be such a map. If a section ss vanishes near xx, choose a bump function χ\chi equal to one in a smaller neighborhood and supported where s=0s=0. Then 0=T(χs)=χT(s)0=T(\chi s)=\chi T(s), so T(s)(x)=0T(s)(x)=0. Thus TT depends at xx only on the germ of ss. Multiply a local frame by a cutoff equal to one on a smaller open set and extend it by zero. The resulting global sections have images under TT that supply the columns of a smooth local matrix. For any ss, write it in that frame near xx and extend its local coefficient functions by another cutoff. Germ locality and linearity over functions show that T(s)(x)T(s)(x) is the action of the local matrix on s(x)s(x). Every vector in ExE_x extends to a section by a local frame and a cutoff. This proves uniqueness and forces the local matrices to agree as bundle maps. ◻

Effective gluing

Theorem 1.2 (Descent for vector bundles). Let (Ui)(U_i) be an open cover of MM. Bundles over MM are equivalent to families of bundles Ei→UiE_i\to U_i with isomorphisms φij ⁣:Ej∣Ui∩Uj→Ei∣Ui∩Uj\varphi_{ij}\colon E_j|_{U_i\cap U_j}\to E_i|_{U_i\cap U_j} such that φii=1\varphi_{ii}=1 and φijφjk=φik\varphi_{ij}\varphi_{jk}=\varphi_{ik} on triple intersections. Maps of these families are bundle maps commuting with the isomorphisms.

Proof. Identify the point e∈Eje\in E_j over x∈Ui∩Ujx\in U_i\cap U_j with φij(e)∈Ei\varphi_{ij}(e)\in E_i. The cocycle law makes these identifications an equivalence relation since it ensures both inverses and closure under composition. The quotient is a set over MM, and over UiU_i every class has a unique representative in EiE_i. The quotient topology restricts to the original topology on EiE_i, since each overlap map is a homeomorphism. The transported smooth bundle charts agree on overlaps because the overlap isomorphisms are smooth. Different fibers can be separated by base neighborhoods, and two vectors in the same fiber lie in a common Hausdorff chart. Hence the total space is Hausdorff. A countable subcover of local charts exists because the base is second countable, and a countable union of their countable bases gives a countable base upstairs. The linear structures in the model charts glue under the linear isomorphisms.

A compatible family of maps is constant on equivalence classes after taking the target quotient, and so gives a unique bundle map. It is smooth on the covering charts. Restriction and gluing are inverse up to the canonical isomorphisms supplied by the identifications over UiU_i. ◻

For trivial local bundles this construction is the quotient by the matrix transitions gijg_{ij}. Replacing eie_i by eihie_i h_i changes the transition to hi−1gijhjh_i^{-1}g_{ij}h_j. Thus a cocycle is an object in a local presentation and its changes of frame are arrows. This matters when forming a moduli object. A bundle can possess many automorphisms even when its isomorphism class is fixed.

Example 1.3 (Two real line bundles on a circle). Use two open arcs covering S1S^1, with two components in their intersection. The transition can be +1+1 on both components or +1+1 on one and −1-1 on the other. A nowhere-zero section in the latter case would give a continuous nonzero function on each arc. Each function has one sign on its entire arc. The transition would require their signs to agree on one overlap component and to disagree on the other, which is impossible. The first case has the constant unit section, so the two gluings already differ as bundles. Their local frames present the same model fiber.

Example 1.4 (Winding in a clutching presentation). On the two disc-like charts of S2S^2, use a unitary line transition restricting on the equator to g(eiθ)=eimθg(e^{i\theta})=e^{im\theta}. A change of frame on either disc has zero winding on its boundary because it extends across the disc. Thus the winding of the transition is unchanged by any bundle isomorphism, written as changes of frame over the given charts. In particular a nonzero mm yields a nontrivial bundle. This topological obstruction will also be detectable by integration of curvature once a connection has been chosen.

Descent along a surjective submersion

Open inclusions suffice to construct each bundle, but descent under maps with local sections will be useful for quotient presentations. Let q ⁣:Y→Mq\colon Y\to M be a surjective submersion and write Y[j]=Y×M⋯×MYY^{[j]}=Y\times_M\cdots\times_M Y with jj factors. A descent datum is a bundle D→YD\to Y with compatible isomorphisms α(y1,y2) ⁣:Dy2⟶Dy1,α(y1,y2)α(y2,y3)=α(y1,y3).\alpha(y_1,y_2)\colon D_{y_2}\longrightarrow D_{y_1}, \qquad \alpha(y_1,y_2)\alpha(y_2,y_3)=\alpha(y_1,y_3). The first formula is a smooth bundle isomorphism on Y[2]Y^{[2]} and the second is an equality over Y[3]Y^{[3]}. A bundle pulled back from MM has this datum because both fibers above (y1,y2)(y_1,y_2) are copies of the same fiber over q(y1)q(y_1).

Conversely, the submersion theorem supplies a cover by open sets UiU_i with sections si ⁣:Ui→Ys_i\colon U_i\to Y of qq. Use Di=si∗DD_i=s_i^*D and the transition α(si(x),sj(x))\alpha(s_i(x),s_j(x)) in Theorem 1.2. The resulting EE is a bundle over MM. Its pullback to YY identifies with DD over q−1(Ui)q^{-1}(U_i) by applying α(y,si(q(y)))\alpha(y,s_i(q(y))) in the fiber. The cocycle law makes the identifications agree on overlaps and intertwine the given descent datum. A map commuting with α\alpha restricts to maps of the DiD_i that glue by the same theorem. This gives effectivity on objects and on maps. Using a common refinement compares two choices of local sections, so the isomorphisms furnished by this construction are canonical for the given descent datum.

Frames as a torsor

The fiber of the frame bundle Fr⁡(E)\operatorname{Fr}(E) at xx is the set of linear isomorphisms Kr→Ex\mathbb K^r\to E_x. Right composition by GLr(K)\mathrm{GL}_r(\mathbb K) acts freely and transitively. Here transitive means that any two frames are related by some change of basis, and free means that change of basis is unique. The quotient of Fr⁡(E)×Kr\operatorname{Fr}(E)\times\mathbb K^r by (eg,v)∼(e,gv)(e g,v)\sim(e,g v) recovers EE by [e,v]↦e(v)[e,v]\mapsto e(v), with the local charts above. An action of another group GG can replace the action of GLr\mathrm{GL}_r when a representation of GG on the model fiber is specified. Chapter [ch:principal] develops the connection on such a principal bundle and the connection it induces on the associated bundle.

Connections on vector bundles

Sections of a vector bundle live in different fibers at different points. A connection specifies how to take a derivative of a section while accounting for that variation. Tu develops the covariant derivative, its matrices and their curvature in §§10–11 and 22 [1]. The gluing theorem of Chapter [ch:bundles] allows the derivative to be specified entirely in local frames.

Leibniz rule and existence

Definition 1.1. A connection on a vector bundle E→ME\to M is an R\mathbb R-linear operator ∇ ⁣:Γ(E)→Ω1(M;E)\nabla\colon\Gamma(E)\to\Omega^1(M;E), complex-linear for a complex bundle, such that ∇(fs)=df⊗s+f∇s.\nabla(fs)=df\otimes s+f\nabla s. It acts on a vector field XX by contraction, ∇Xs=(∇s)(X)\nabla_Xs=(\nabla s)(X).

If ss vanishes near xx, multiply by a cutoff χ\chi supported in this neighborhood and equal to one near xx. Applying the Leibniz rule to χs=0\chi s=0 gives (∇s)(x)=0(\nabla s)(x)=0. This locality induces connection operators on restrictions to open sets. In a frame eie_i write ∇ei,b=∑aei,a(Ai)ab\nabla e_{i,b}=\sum_a e_{i,a}(A_i)_{ab}. The local formula on a section eivie_i v_i is then ∇(eivi)=ei(dvi+Aivi).\nabla(e_i v_i)=e_i(dv_i+A_i v_i). Every smooth matrix AiA_i of one-forms defines a connection on this product chart. Choose local product connections ∇(i)\nabla^{(i)} and a locally finite smooth partition of unity (ρi)(\rho_i) subordinate to their domains. The products ρi∇(i)s\rho_i\nabla^{(i)}s extend by zero and their sum is smooth. The coefficient of df⊗sdf\otimes s in its Leibniz identity is ∑iρi=1\sum_i\rho_i=1, which proves existence.

The difference of two connections satisfies (∇′−∇)(fs)=f(∇′−∇)(s)(\nabla'-\nabla)(fs)=f(\nabla'-\nabla)(s). The local matrix argument of Chapter [ch:bundles] therefore identifies it with a section of T∗M⊗End⁡ET^*M\otimes\operatorname{End}E, and adding any such section to a connection gives another. The set of connections is consequently an affine space modeled on Ω1(M;End⁡E)\Omega^1(M;\operatorname{End}E). For a fiber metric, the compatibility condition is d⟨s,t⟩=⟨∇s,t⟩+⟨s,∇t⟩.d\langle s,t\rangle =\langle\nabla s,t\rangle+\langle s,\nabla t\rangle. In an orthonormal frame this is Ai∗=−AiA_i^*=-A_i, with transpose in the real case and Hermitian adjoint in the complex case. Local orthonormal frames supply local metric connections. The partition of unity construction preserves this metric identity, and two metric connections differ by a one-form valued in skew-adjoint endomorphisms.

Changing the frame and pulling back

The frame ej=eigije_j=e_i g_{ij} gives the coordinate relation vi=gijvjv_i=g_{ij}v_j. Substituting into the Leibniz rule yields Aj=gij−1Aigij+gij−1dgij. A_j=g_{ij}^{-1}A_i g_{ij}+g_{ij}^{-1}dg_{ij}. Indeed (d+Ai)(gijvj)=gijdvj+(dgij+Aigij)vj(d+A_i)(g_{ij}v_j) =g_{ij}dv_j+(dg_{ij}+A_i g_{ij})v_j. The compatibility of local derivatives is the square image Conversely, matrices satisfying (9) give compatible operators in this square, hence a connection on the glued bundle by the sheaf property. The term gij−1dgijg_{ij}^{-1}dg_{ij} records differentiation of the frame itself. A bundle isomorphism which identifies two connections is accordingly governed by the same formula.

For f ⁣:N→Mf\colon N\to M, define the pullback connection by matrices f∗Aif^*A_i in the pullback frames. Pullback commutes with dd on functions, so these matrices satisfy (9) for the pulled-back transitions. This definition works for any smooth map and commutes with the canonical comparison for two successive pullbacks.

Curvature and the Bianchi identity

Extend ∇\nabla to EE-valued differential forms by ∇(α⊗s)=dα⊗s+(−1)kα∧∇s(α∈Ωk(M)).\nabla(\alpha\otimes s) =d\alpha\otimes s+(-1)^k\alpha\wedge\nabla s \qquad (\alpha\in\Omega^k(M)). The Leibniz rule gives the same result after moving a smooth function between the two factors. In a frame the extension is d+A∧d+A\wedge, where matrix products are combined with wedge products of entries. Applying it twice gives (d+A∧)2=F∧(d+A\wedge)^2=F\wedge, with F=dA+A∧A. F=dA+A\wedge A. The dd-terms hitting the argument cancel by the graded product rule. Thus the square is tensorial and defines an End⁡E\operatorname{End}E-valued two-form. In the jj-frame its matrix is Fj=gij−1FigijF_j=g_{ij}^{-1}F_i g_{ij} by applying the compatible operators twice. On vector fields the same square reads F(X,Y)s=∇X∇Ys−∇Y∇Xs−∇[X,Y]s.F(X,Y)s =\nabla_X\nabla_Ys-\nabla_Y\nabla_Xs-\nabla_{[X,Y]}s . The matrix formula shows that Ff∗∇=f∗F∇F_{f^*\nabla}=f^*F_\nabla.

The induced covariant derivative on endomorphism-valued kk-forms is Dη=dη+A∧η−(−1)kη∧A.D\eta=d\eta+A\wedge\eta-(-1)^k\eta\wedge A . Its Bianchi identity is DF=0DF=0. Expanding gives dF=dA∧A−A∧dAdF=dA\wedge A-A\wedge dA, while A∧F−F∧A=A∧dA−dA∧AA\wedge F-F\wedge A=A\wedge dA-dA\wedge A. The cubic terms cancel by associativity. The two-form FF therefore satisfies a differential constraint involving the same connection that produced it. In particular dtr⁡F=0d\operatorname{tr}F=0, since the graded commutator has zero trace. Closed traces of higher curvature products will appear in Chern–Weil theory.

Tensor operations and a matrix example

For bundles E,E′E,E' with connections, the connection on their tensor product is ∇(s⊗t)=(∇s)⊗t+s⊗(∇′t).\nabla(s\otimes t)=(\nabla s)\otimes t+s\otimes(\nabla' t). The Leibniz rules make this well-defined on the tensor product over functions. Its curvature is F⊗1+1⊗F′F\otimes 1+1\otimes F', as the mixed terms have opposite signs when the derivative is applied twice. The dual connection is fixed by requiring d(λ(s))=(∇∨λ)(s)+λ(∇s)d(\lambda(s))=(\nabla^\vee\lambda)(s)+\lambda(\nabla s). In dual frames its matrix is −AT-A^{\mathsf T} and its curvature is −FT-F^{\mathsf T}. The connection on the determinant ⋀rE\bigwedge^r E is obtained by differentiating each wedge factor. Only the diagonal entries contribute to the connection of the top exterior frame, so its local one-form is tr⁡A\operatorname{tr}A and its curvature is tr⁡F=dtr⁡A\operatorname{tr}F=d\operatorname{tr}A. For unitary connections we normalize the determinant curvature by i/(2π)i/(2\pi). The scalar resulting two-form is real.

Example 1.2 (Constant coefficients with nonzero curvature). On the trivial Hermitian bundle R2×C2\mathbb R^2\times\mathbb C^2, take A=X dx+Y dyA=X\,dx+Y\,dy with X=(i00−i),Y=(01−10).X=\begin{pmatrix}i&0\\0&-i\end{pmatrix},\qquad Y=\begin{pmatrix}0&1\\-1&0\end{pmatrix}. The entries of AA are constant and both matrices are skew-Hermitian. Its curvature is F=[X,Y] dx∧dy=(02i2i0)dx∧dy.F=[X,Y]\,dx\wedge dy =\begin{pmatrix}0&2i\\2i&0\end{pmatrix}dx\wedge dy. Thus the connection has curvature even though the determinant connection is the product connection. The commutator is also the leading difference between the two orders of transport along short coordinate segments. Chapter [ch:transport] makes this comparison explicit.

Normal frames and varying the connection

At a prescribed point x0x_0, choose coordinates xμx^\mu and write A=∑μAμ dxμA=\sum_\mu A_\mu\,dx^\mu. The frame change h(x)=exp⁡(−∑μ(xμ−x0μ)Aμ(x0))h(x)=\exp\Bigl(-\sum_\mu (x^\mu-x_0^\mu)A_\mu(x_0)\Bigr) has h(x0)=Ih(x_0)=I and (dh)x0=−Ax0(dh)_{x_0}=-A_{x_0}. Equation (9) then gives Ax0′=0A'_{x_0}=0. If the initial frame is orthonormal for a metric connection, each exponent is skew-adjoint and hh remains orthogonal or unitary. Such a frame simplifies the derivative at a point. The curvature there is dAx0′dA'_{x_0}, which can be nonzero. Vanishing on an entire neighborhood would imply F=0F=0 there by (10). This distinguishes pointwise normalization from a local parallel frame.

For a global one-form α\alpha valued in End⁡E\operatorname{End}E, the connection ∇+α\nabla+\alpha has curvature F∇+α=F∇+Dα+α∧α.F_{\nabla+\alpha}=F_\nabla+D\alpha+\alpha\wedge\alpha. The identity follows by expanding d(A+α)+(A+α)2d(A+\alpha)+(A+\alpha)^2 and collecting terms of each degree in α\alpha. For the path of connections ∇t=∇+tα\nabla_t=\nabla+t\alpha, it follows that ddtF∇t=Dtα\frac{d}{dt}F_{\nabla_t}=D_t\alpha, where DtD_t is the derivative induced by ∇t\nabla_t. The first variation of the curvature is linear in the direction, whereas the finite change includes the quadratic term. This distinction will produce a transgression between characteristic forms of two connections.

Example 1.3 (A projected connection). Let P ⁣:M→End⁡(CN)P\colon M\to\operatorname{End}(\mathbb C^N) be a smooth family of orthogonal projections of constant rank. The images form a vector bundle EE because projected fixed vectors give local frames wherever an appropriate minor is invertible. Define ∇s=P ds\nabla s=P\,ds. The relation Ps=sPs=s gives the Leibniz rule, and orthogonality shows that pairing with a vector in EE is unchanged when projecting. Hence the connection preserves the induced metric. Its extension to EE-valued forms is obtained by applying dd componentwise and projecting. Differentiating P2=PP^2=P gives P(dP)P=0P(dP)P=0. For a section ss one has ds=(dP)s+Pdsds=(dP)s+Pds, so ∇2s=P d(Pds)=P dP∧ds=P(dP∧dP)s.\nabla^2 s=P\,d(Pds) =P\,dP\wedge ds =P(dP\wedge dP)s . The projection can therefore produce a curved connection on a subbundle of a flat product. In the tautological line of Chapter [ch:principal], the projection is P=zz†P=zz^\dagger for a local unit column zz. The induced one-form in that frame is z†dzz^\dagger dz, which provides a second construction of the same connection.

Principal connections

A frame bundle is a bundle of choices of coordinates. Its fibers have a free transitive group action, so a local coordinate choice can be changed by a unique group element. Principal bundles keep this structure for any Lie group GG. Connections become invariant horizontal distributions and can then be applied to several representations of the same group. The constructions below correspond to Tu’s treatment of horizontal distributions, connection forms and associated bundles [1, §§27–31].

Right torsors and horizontal directions

A principal GG-bundle is a smooth locally trivial bundle π ⁣:P→M\pi\colon P\to M whose fibers are right GG-torsors. Local trivializations intertwine the given action with multiplication on the right of the second factor of U×GU\times G. Equivalently, with the bundle charts in place, the action gives the diffeomorphism P×G≃P×MPP\times G\simeq P\times_M P, (p,g)↦(p,pg)(p,g)\mapsto(p,pg). For each pp, the map g⟶VpP=ker⁡(dπp),ξ⟼ξp#=ddt∣0pexp⁡(tξ)\mathfrak g\longrightarrow V_pP=\ker(d\pi_p),\qquad \xi\longmapsto \xi^\#_p =\left.\frac{d}{dt}\right|_{0}p\exp(t\xi) is an isomorphism, as seen in a product chart. Right translation RhR_h sends ξp#\xi^\#_p to (Ad⁡h−1ξ)ph#(\operatorname{Ad}_{h^{-1}}\xi)^\#_{ph}. This follows from exp⁡(tξ)h=hexp⁡(tAd⁡h−1ξ)\exp(t\xi)h=h\exp(t\operatorname{Ad}_{h^{-1}}\xi).

A horizontal distribution is a smooth complement HH to VPVP in TPTP, invariant under right translation. The restriction dπ∣Hpd\pi|_{H_p} is an isomorphism onto Tπ(p)MT_{\pi(p)}M, so a tangent vector downstairs has one horizontal lift at each point of its fiber. Map any tangent vector vv to the Lie algebra element corresponding to its vertical component. This gives the connection form ω∈Ω1(P;g)\omega\in\Omega^1(P;\mathfrak g) satisfying ω(ξ#)=ξ,Rh∗ω=Ad⁡h−1ω. \omega(\xi^\#)=\xi,\qquad R_h^*\omega=\operatorname{Ad}_{h^{-1}}\omega . Conversely, (11) makes v↦(ω(v))#v\mapsto(\omega(v))^\# an equivariant projection onto VPVP. Its kernel is the invariant complement. Thus the form and the horizontal distribution determine each other.

Local potentials and curvature

A local section si ⁣:Ui→Ps_i\colon U_i\to P gives the potential Ai=si∗ωA_i=s_i^*\omega. In coordinates p=si(x)gp=s_i(x)g, ω=Ad⁡g−1π∗Ai+g−1dg. \omega=\operatorname{Ad}_{g^{-1}}\pi^*A_i+g^{-1}dg . For the base tangent this is equivariance with gg fixed. For the group tangent it is the left Maurer–Cartan form g−1dgg^{-1}dg, meaning the tangent translated back to the identity by left multiplication by g−1g^{-1}. On an overlap write sj=sigijs_j=s_i g_{ij}. Pulling (12) along sjs_j gives Aj=Ad⁡gij−1Ai+gij−1dgij.A_j=\operatorname{Ad}_{g_{ij}^{-1}}A_i +g_{ij}^{-1}dg_{ij}. Local forms with these transitions give a global connection by (12). To check compatibility, substitute g=gijg′g=g_{ij}g' and use the product rule for the Maurer–Cartan form. Local product connections can also be combined using a partition of unity pulled back from MM. Their weighted sum reproduces each vertical generator because the weights add to one, and equivariance survives because the weights are constant on the fibers. In particular every smooth principal bundle in our setting admits a connection.

Use the graded Lie bracket convention [ω,ω](v,w)=2[ω(v),ω(w)][\omega,\omega](v,w)=2[\omega(v),\omega(w)]. Define the curvature by Ω=dω+12[ω,ω].\Omega=d\omega+\tfrac12[\omega,\omega]. It is equivariant by (11). It is horizontal as well. Indeed the flow of ξ#\xi^\# is Rexp⁡(tξ)R_{\exp(t\xi)}. Differentiating equivariance gives Lξ#ω=−[ξ,ω]\mathcal L_{\xi^\#}\omega=-[\xi,\omega]. Cartan’s formula and ω(ξ#)=ξ\omega(\xi^\#)=\xi, which is constant on PP, give ιξ#dω=−[ξ,ω]\iota_{\xi^\#}d\omega=-[\xi,\omega]. The contraction of the bracket term is [ξ,ω][\xi,\omega], so the sum vanishes. In local sections the curvature is Fi=dAi+12[Ai,Ai]F_i=dA_i+\frac12[A_i,A_i]. In matrix representations this becomes dAi+Ai∧AidA_i+A_i\wedge A_i. The motion of horizontal directions is measured by the same form. For horizontal fields XH,YHX^H,Y^H, Ω(XH,YH)=−ω([XH,YH]). \Omega(X^H,Y^H)=-\omega([X^H,Y^H]). This is the exterior derivative on a pair of fields in ker⁡ω\ker\omega. Thus the horizontal subbundle is involutive precisely when Ω=0\Omega=0. This identifies flatness with an integrability condition before constructing any transport.

Flat gauges and smaller structure groups

When Ω=0\Omega=0, the Frobenius theorem supplies a local integral submanifold of HH through any point of PP. At that point the restriction of dπd\pi to its tangent space is an isomorphism. After shrinking, the restriction of π\pi is a diffeomorphism to an open neighborhood in MM. Its inverse gives a horizontal local section, so the potential in that section is zero. Conversely a zero potential gives zero curvature on its domain. If two horizontal sections coexist on a connected open set, then gij−1dgij=0g_{ij}^{-1}dg_{ij}=0 in their change of frame law. Thus they differ by a single constant group element. Flat bundles are consequently locally assembled by constant transitions, although their global gluing can be nontrivial.

A reduction of the structure group to a closed Lie subgroup K⊂GK\subset G is a principal KK-subbundle Q⊂PQ\subset P. The connection restricts to a principal KK-connection if ω(TQ)⊂k\omega(TQ)\subset\mathfrak k. This condition is equivalent to Hq⊂TqQH_q\subset T_qQ at every qq. If the image is in k\mathfrak k, subtracting the corresponding vertical KK-vector makes any tangent vector in TqQT_qQ horizontal. Since TqQT_qQ surjects onto Tπ(q)MT_{\pi(q)}M, it contains all horizontal lifts there. In the other direction, the decomposition of TqQT_qQ into its horizontal vectors and its vertical KK-directions shows that its image is in k\mathfrak k. The reproduction and equivariance axioms restrict as well. For instance, an orthogonal reduction of a real frame bundle arises from a metric, and requiring the connection to restrict to that reduction is precisely the skew matrix condition of Chapter [ch:vector-conn].

Associated bundles and pullback

For a smooth linear representation ρ ⁣:G→GL(W)\rho\colon G\to \mathrm{GL}(W), define E=P×GWE=P\times_G W by the relation (pg,w)∼(p,ρ(g)w)(pg,w)\sim(p,\rho(g)w). The chart on UiU_i identifies [si(x),w][s_i(x),w] with ww, giving transition ρ(gij)\rho(g_{ij}). A section is equivalently a smooth map u ⁣:P→Wu\colon P\to W such that u(pg)=ρ(g)−1u(p)u(pg)=\rho(g)^{-1}u(p). Restrict dudu to horizontal vectors to define a connection on EE. Equivariance of HH ensures this definition is independent of the representative of a point of EE. The Leibniz rule follows from differentiating (f∘π)u(f\circ\pi)u.

In local coordinates let vi=u∘siv_i=u\circ s_i. Subtracting (Ai(X))#(A_i(X))^\# from dsi(X)ds_i(X) gives the horizontal lift of XX at si(x)s_i(x). Differentiating uu in that vertical direction gives −ρ∗(Ai(X))vi-\rho_*(A_i(X))v_i. The local derivative is consequently dvi+ρ∗(Ai)vi.dv_i+\rho_*(A_i)v_i . As ρ∗\rho_* preserves brackets, the curvature is ρ∗(Fi)\rho_*(F_i). In the adjoint representation the horizontal equivariant curvature Ω\Omega thus gives a form with values in ad⁡P=P×Gg\operatorname{ad}P=P\times_G\mathfrak g. Here one uses equivariance to transport the value of Ω\Omega between representatives and horizontality to choose tangent lifts. The curvature endomorphism of the associated adjoint connection acts by the Lie bracket with the descended value of Ω\Omega.

A smooth map f ⁣:N→Mf\colon N\to M gives a pullback principal bundle and connection by pulling ω\omega back along the projection to PP. Fibers are mapped equivariantly, so the reproduction and equivariance conditions still hold. Associated bundles commute with this construction, as both sides identify [((y,p),w)][((y,p),w)] with (y,[p,w])(y,[p,w]). The local potential formula then shows that their induced connections agree. For the frame bundle of a vector bundle, using its defining representation recovers the original bundle connection from the matrices of Chapter [ch:vector-conn].

Example 1.1 (The unit circle bundle on the projective line). Give P=S3⊂C2P=S^3\subset\mathbb C^2 the right scalar action of U(1)U(1), so its quotient is CP1\mathbb{CP}^1. The form ω=z†dz\omega=z^\dagger dz is imaginary on S3S^3, since the derivative of z†z=1z^\dagger z=1 has zero real part. It evaluates to ξ\xi on zξz\xi, and unitary scalar multiplication preserves it. It therefore defines a principal connection. On the chart of lines spanned by (1,w)(1,w), take the unit section (1,w)/1+∣w∣2(1,w)/\sqrt{1+|w|^2}. The potential and curvature are A=wˉ dw−w dwˉ2(1+∣w∣2),F=dA=dwˉ∧dw(1+∣w∣2)2.A=\frac{\bar w\,dw-w\,d\bar w}{2(1+|w|^2)},\qquad F=dA=\frac{d\bar w\wedge dw}{(1+|w|^2)^2}. For example, differentiating the numerator gives 2 dwˉ∧dw2\,d\bar w\wedge dw, and differentiating the denominator subtracts the term proportional to ∣w∣2|w|^2. The associated line for the standard representation has fiber the given line in C2\mathbb C^2, hence is the tautological line. In the complex orientation, writing w=x+iyw=x+iy gives ∫CP1i2πF=−1π∫R2dx dy(1+x2+y2)2=−1.\int_{\mathbb{CP}^1}\frac{i}{2\pi}F =-\frac{1}{\pi} \int_{\mathbb R^2}\frac{dx\,dy}{(1+x^2+y^2)^2} =-1. The omitted point has zero area. In polar coordinates the real integral is 2π∫0∞r(1+r2)2 dr=π.2\pi\int_0^\infty \frac{r}{(1+r^2)^2}\,dr=\pi . The normalization records both the orientation and the choice of line in this principal bundle construction.

Paths, thin homotopy, and transport

A connection can be tested by transporting vectors and frames along paths. Composition of paths is then a composition of invertible maps of fibers. Curvature controls the dependence on a homotopy of paths. The appropriate functor for a general connection uses homotopies that sweep out no infinitesimal area.

Transport equations

Write a section along γ ⁣:[0,1]→M\gamma\colon[0,1]\to M in a frame as a column v(t)v(t). It is parallel when v˙(t)=−Aγ(t)(γ˙(t))v(t).\dot v(t)=-A_{\gamma(t)}(\dot\gamma(t))v(t). Existence and uniqueness for linear ordinary differential equations give a solution on each closed segment contained in a chart. Along chart changes the relation vi=gijvjv_i=g_{ij}v_j and equation (9) give the same equation in the new frame. Subdivision and uniqueness therefore give a well-defined transport isomorphism Tγ ⁣:Eγ(0)→Eγ(1)T_\gamma\colon E_{\gamma(0)}\to E_{\gamma(1)}. On a principal bundle, a lift γH(t)=si(γ(t))g(t)\gamma^H(t)=s_i(\gamma(t))g(t) is horizontal when g˙(t)=−(dRg(t))1Ai(γ˙(t)).\dot g(t)=-(dR_{g(t)})_1 A_i(\dot\gamma(t)). This is (12) evaluated on the lifted tangent. It has a solution over each compact chart segment. To see the extension in the group variable, take a short solution starting at 11 and right translate it to the required initial value. The Lie algebra input is bounded on the segment, so local existence near 11 gives a uniform short time for these restarts. For a matrix group the right-hand side is simply −Ai(γ˙(t))g(t)-A_i(\dot\gamma(t))g(t). The uniqueness of the horizontal lift gives equivariance, Tγ(pg)=Tγ(p)gT_\gamma(pg)=T_\gamma(p)g. In a representation, the horizontal lift induces exactly the parallel equation with potential ρ∗(Ai)\rho_*(A_i). This agrees with the frame interpretation in Tu [1, §29] and the principal parallel displacement of Kobayashi–Nomizu [3, Ch. II].

We use smooth paths with sitting instants, meaning constant segments near both endpoints. They can be joined smoothly at an endpoint. Ordinary changes of speed preserving the endpoint order leave transport unchanged by the chain rule. If γ1\gamma_1 ends where γ2\gamma_2 begins, uniqueness on the two pieces gives Tγ2⋆γ1=Tγ2∘Tγ1.T_{\gamma_2\star\gamma_1} =T_{\gamma_2}\circ T_{\gamma_1}. The constant path gives the identity. The transport of a reversed path is the inverse, since reversing the solution of the parallel equation gives the solution with reversed endpoints.

Thin homotopy and curvature variation

A thin homotopy of paths with fixed endpoints is a smooth map H(s,t)H(s,t) interpolating between them, with constant ends in the path variable and rank dH≤1dH\leq1 everywhere. Homotopies can be reparameterized to sit near their boundaries before concatenating them. Reparameterizations of a path are thin because their interpolation factors through the path interval. A path followed by its reverse contracts by shortening the portion of the path traversed in both directions. This homotopy likewise factors through an interval. These facts give associativity, units and inverses for the thin homotopy classes of sitting paths. Write P1(M)\mathcal P_1(M) for this groupoid, with the points of MM as objects [4, §2.1].

Proposition 1.1 (Thin invariance). Parallel transport of a principal connection is invariant under thin homotopies. For a flat connection it is invariant under all smooth homotopies relative to the endpoints.

Proof. Lift t↦H(s,t)t\mapsto H(s,t) horizontally, starting for every ss at the same point p0p_0 above the initial endpoint. Smooth dependence of the differential equation on parameters gives a smooth family p(s,t)p(s,t). On the square let as=ω(∂sp)a_s=\omega(\partial_s p) and at=ω(∂tp)a_t=\omega(\partial_t p). Then at=0a_t=0, and the structure equation gives (p∗Ω)(∂s,∂t)=∂sat−∂tas+[as,at]=−∂tas.(p^*\Omega)(\partial_s,\partial_t) =\partial_s a_t-\partial_t a_s+[a_s,a_t] =-\partial_t a_s. The curvature is horizontal, so its value on a pair of tangent vectors is determined by their projections to the base. Under a thin homotopy those projections are linearly dependent. Thus the left side vanishes. It also vanishes for an arbitrary homotopy if the connection is flat. It follows that asa_s is constant in tt. At t=0t=0 it is zero because the lifted initial point is fixed. At t=1t=1 the field ∂sp\partial_s p is vertical because the terminal base point is fixed, and as=0a_s=0 implies that this vertical vector is zero. Hence the lifted terminal point is independent of ss. ◻

It follows that transport assigns a functor P1(M)⟶G-Tor,x⟼Px.\mathcal P_1(M)\longrightarrow G\text{-}\mathrm{Tor}, \qquad x\longmapsto P_x . The target consists of right GG-torsors and their equivariant isomorphisms. A connection-preserving bundle map induces a natural transformation, since it carries horizontal lifts with specified initial values to the corresponding horizontal lifts. In particular the following square commutes for a path γ ⁣:x→y\gamma\colon x\to y and such a bundle map Φ\Phi. image

Small rectangles and global holonomy

Consider a vector bundle and a coordinate rectangle of side ε\varepsilon in the x,yx,y plane, beginning at a point taken as the origin and following +x,+y,−x,−y+x,+y,-x,-y in that order. Take a frame with A(0)=0A(0)=0, as in Chapter [ch:vector-conn]. The integral of AA along each edge is O(ε2)O(\varepsilon^2). The integral equation for parallel transport on an edge consequently gives I−∫edgeA+O(ε4)I-\int_{\mathrm{edge}}A+O(\varepsilon^4), with the bound from the second iterated integral and higher terms. Multiplying the four edge matrices gives T∂Rε=I−∮∂RεA+O(ε4)=I−ε2F0(∂x,∂y)+O(ε3).T_{\partial R_\varepsilon} = I-\oint_{\partial R_\varepsilon} A+O(\varepsilon^4) = I-\varepsilon^2 F_0(\partial_x,\partial_y) +O(\varepsilon^3). The last equality uses Stokes entry by entry and F0=(dA)0F_0=(dA)_0. The same leading formula holds in any frame at the base point, by conjugation. For constant A=X dx+Y dyA=X\,dx+Y\,dy one can also multiply the exact transports in reverse traversal order, T∂Rε=eεYeεXe−εYe−εX.T_{\partial R_\varepsilon} =e^{\varepsilon Y}e^{\varepsilon X} e^{-\varepsilon Y}e^{-\varepsilon X}. Their second-order term is −ε2[X,Y]-\varepsilon^2[X,Y], in agreement with curvature. In particular if vector transport descends to the ordinary fundamental groupoid, every small rectangle has trivial transport. Dividing by ε2\varepsilon^2 and taking the limit proves F=0F=0. For a general Lie group, use coordinates on the base centered at the initial point. The gauge h(x)=exp⁡(−∑μxμAμ(0))h(x)=\exp(-\sum_\mu x^\mu A_\mu(0)) makes A(0)=0A(0)=0 by the same first-derivative argument. Use a group chart at 11 with differential equal to the identity. Each edge solution starts at 11 and stays O(ε2)O(\varepsilon^2) from it in that chart, because its right-invariant driving term is O(ε)O(\varepsilon) on an edge of length ε\varepsilon. Translating its tangent back to the group chart changes the driving term by O(ε3)O(\varepsilon^3), so its displacement is −∫edgeA+O(ε4)-\int_{\mathrm{edge}}A+O(\varepsilon^4). Group multiplication likewise differs from addition by O(ε4)O(\varepsilon^4) for these displacements. This proves the same curvature limit in the Lie algebra. Thus flatness is exactly the condition for factoring through ordinary path homotopy. This formulation uses the usual smooth relative path model of the fundamental groupoid.

Example 1.2 (Holonomy distinguishes connections). For a unitary line connection A=−2πi aA=-2\pi i\,a in a global frame, the parallel equation integrates to Tγ=exp⁡(2πi∫γa)T_\gamma=\exp(2\pi i\int_\gamma a). On R2\mathbb R^2 take a=x dya=x\,dy. The loop around a rectangle of oriented area bb has holonomy e2πibe^{2\pi i b} even though it is contractible. Its phase is determined by the curvature integral. On S1=R/ZS^1=\mathbb R/\mathbb Z, take a=λ dθa=\lambda\,d\theta. This connection is flat and transport around a loop of winding mm multiplies by e2πiλme^{2\pi i\lambda m}. The gauge g=e−2πikθg=e^{-2\pi i k\theta} for integer kk replaces λ\lambda by λ+k\lambda+k. Conversely, two such constants with equal holonomy differ by an integer. Here curvature vanishes and the global holonomy still detects different connections up to gauge.

Flat transport and local systems

A functor from the ordinary fundamental groupoid to right GG-torsors determines a flat bundle by gluing. Choose a cover by coordinate balls UiU_i with distinguished points xix_i. On UiU_i use the product of UiU_i with the torsor assigned to xix_i, with zero potential in a chosen model frame. A path inside UiU_i from xix_i to xx identifies that model torsor with the torsor assigned to xx. All such paths have the same relative homotopy class. On the overlap, compose one of these identifications with the inverse of the other. The resulting transition is constant on each connected piece of the overlap. To verify constancy, join nearby points of the overlap by a path inside it and append it to both coordinate-ball paths. The induced map in the target groupoid cancels between the second identification and the inverse first one.

The transitions satisfy the cocycle law because they all use the torsor assigned to the point in MM. Gluing the local products uses the same quotient as in Theorem [thm:vector-descent], with multiplication on the GG-coordinate in place of linear maps. The local zero potentials agree under the constant transitions, so the bundle is flat. Its transport recovers the original functor by subdivision into the coordinate balls. Conversely, a flat connection has horizontal trivializations by Chapter [ch:principal], and the products above recover its transition maps. Natural transformations give compatible maps between these products, locally constant in the chosen frames, and therefore produce the connection-preserving isomorphisms. In a connected manifold, a choice of base point and a point in its torsor describes the resulting groupoid by representations of the fundamental group up to conjugation. The description without a chosen base point retains the conjugating maps as morphisms.

Recovering a connection from transport

The smooth information in transport is essential for recovery. In a locally chosen frame, suppose a family of short paths from xx to x+εvx+\varepsilon v is obtained by smoothly reparameterizing the straight segment to have sitting instants. Let U(ε)U(\varepsilon) be its arrow in GG, written in that frame at both endpoints. The transport equation yields Ax(v)=−U′(0)∈g.A_x(v)=-U'(0)\in\mathfrak g . Indeed the infinitesimal path velocity is proportional to vv, and the integral of the time derivative of the scalar reparameterization is one. The slope thus recovers the local one-form. Under a change of frame gg, a path arrow from xx to yy changes to g(y)−1Ug(x)g(y)^{-1}U g(x). Its derivative along a short segment recovers precisely the gauge law for the one-form.

The reconstruction theorem of Schreiber and Waldorf places this local argument in an equivalence of categories. A transport functor with BGBG-structure admits local trivializations by smooth BGBG-valued functors on a surjective submersion, with smooth descent data. Under those conditions it comes from a principal bundle with connection, and its natural morphisms arise from maps of the corresponding bundles with connection [4, Theorem 5.4]. The local hypotheses give the smooth dependence needed for the derivative and the transition data needed to glue the resulting forms. Thus the base groupoid and the smooth descent condition together retain both the local curvature and the global bundle.

Invariant polynomials and Chern–Weil forms

Curvature transforms by conjugation when a frame changes. This transformation law suggests a way to construct ordinary differential forms from it. A polynomial on the Lie algebra invariant under conjugation eliminates dependence on the frame. Closedness and connection independence then follow from two distinct differential identities. Tu develops this passage first for vector bundles and then for principal bundles [1, §§23,32].

Encoding a polynomial by polarization

Let GG be a Lie group with real Lie algebra g\mathfrak g. For a homogeneous polynomial p ⁣:g→Rp\colon\mathfrak g\to\mathbb R of degree mm, define P(X1,…,Xm)=1m!∂m∂t1⋯∂tmp(∑itiXi)∣t=0.P(X_1,\ldots,X_m)=\frac{1}{m!} \frac{\partial^m}{\partial t_1\cdots\partial t_m} p\left(\sum_i t_iX_i\right)\bigg|_{\boldsymbol t=0}. Then PP is symmetric and multilinear, and p(X)=P(X,…,X)p(X)=P(X,\ldots,X). The formula follows by extracting the coefficient of t1⋯tmt_1\cdots t_m, which also proves uniqueness. If pp is invariant under Ad⁡(G)\operatorname{Ad}(G), simultaneous application of Ad⁡g\operatorname{Ad}_g to all inputs of PP leaves it unchanged. Differentiating at g=exp⁡(sZ)g=\exp(sZ) gives ∑i=1mP(X1,…,[Z,Xi],…,Xm)=0.\sum_{i=1}^{m}P(X_1,\ldots,[Z,X_i],\ldots,X_m)=0 . For disconnected GG, the differentiated equation tests only the identity component. The statement about descent uses invariance under the whole group.

The notation P(β1,…,βm)P(\beta_1,\ldots,\beta_m) also applies to g\mathfrak g-valued differential forms. In a fixed basis, insert their coefficients in the displayed order and wedge the scalar forms. This description is independent of the basis by multilinearity. Permuting two inputs of degrees a,ba,b gives the sign (−1)ab(-1)^{ab}. In particular, degree-two forms can be interchanged inside PP. Write P(Fm)P(F^m) for P(F,…,F)P(F,\ldots,F). For example, when g\mathfrak g is a matrix Lie algebra and p(X)=tr⁡(X2)p(X)=\operatorname{tr}(X^2), one gets P(F2)=tr⁡(F∧F).P(F^2)=\operatorname{tr}(F\wedge F). The symmetric polarization is useful even for matrix polynomials. The inputs are multiplied as matrices with form coefficients before taking the trace.

Descent and the Bianchi identity

Let Q→XQ\to X be a smooth right principal GG-bundle. Give it a connection with local potentials AiA_i in sections sis_i, where sj=sigijs_j=s_i g_{ij}. Write Fi=dAi+12[Ai,Ai].F_i=dA_i+\tfrac12[A_i,A_i]. For matrix-valued forms the bracket is the graded commutator of their wedge products, so this formula reads dAi+Ai∧AidA_i+A_i\wedge A_i. If DA=d+[A,−]D_A=d+[A,-] is the covariant exterior derivative, then DAF=0D_AF=0. In a matrix Lie algebra the cancellation is explicit: dF=dA∧A−A∧dA,[A,F]=A∧dA−dA∧A.\begin{split} dF&=dA\wedge A-A\wedge dA,\\ [A,F]&=A\wedge dA-dA\wedge A. \end{split} The intrinsic Lie-algebra calculation replaces associativity of the matrix product by the graded Jacobi identity.

Theorem 1.1 (Chern–Weil forms). For every Ad⁡(G)\operatorname{Ad}(G)-invariant polynomial pp of homogeneous degree mm, the expressions P(Fim)P(F_i^m) patch to a closed 2m2m-form p(F)p(F) on XX. The cohomology class [p(F)]dR[p(F)]_{\mathrm{dR}} depends only on the principal bundle. It is natural under pullback, and the assignment from invariant polynomials to de Rham cohomology is a graded algebra homomorphism when polynomial degree is doubled.

Proof. Since Fj=Ad⁡gij−1FiF_j=\operatorname{Ad}_{g_{ij}^{-1}}F_i, invariance identifies the two expressions on each overlap. Equivalently the curvature on QQ is horizontal and equivariant. Applying PP gives a horizontal invariant form on QQ, which is the pullback of the form on XX. To see the uniqueness of such a descended form, evaluate on lifts of tangent vectors at a point in a fiber. Horizontality makes the lifts immaterial, and invariance makes the fiber point immaterial. Local sections give smoothness.

The derivative of P(Fim)P(F_i^m) is mP(dFi,Fim−1)mP(dF_i,F_i^{m-1}) because FiF_i has even degree. The coefficientwise version of (14), applied with the coefficients of AiA_i and with its scalar one-form placed first, gives mP([Ai,Fi],Fim−1)=0.mP([A_i,F_i],F_i^{m-1})=0. Consequently dP(Fim)=mP(DAiFi,Fim−1)=0dP(F_i^m)=mP(D_{A_i}F_i,F_i^{m-1})=0.

For independence take connections A0,A1A_0,A_1. Their difference on QQ is horizontal and equivariant, so their affine combination At=(1−t)A0+tA1A_t=(1-t)A_0+tA_1 satisfies the connection identities for every tt. Use AtA_t as a connection on the pullback of QQ to [0,1]×X[0,1]\times X, with zero potential in the tt-direction. Its curvature in frames constant in tt is F~=Ft+dt∧A˙t.\widetilde F=F_t+dt\wedge\dot A_t. Inserting it into PP and dropping terms with two copies of dtdt gives P(F~m)=P(Ftm)+dt∧mP(A˙t,Ftm−1).P(\widetilde F^m) = P(F_t^m)+dt\wedge mP(\dot A_t,F_t^{m-1}). Everything on the right descends. In particular A˙t\dot A_t transforms homogeneously since the transition functions have no tt-dependence. If a closed form is written αt+dt∧βt\alpha_t+dt\wedge\beta_t, closedness in the coefficient of dtdt gives ∂tαt=dXβt\partial_t\alpha_t=d_X\beta_t. Integrating this identity shows p(F1)−p(F0)=dX(m∫01P(A˙t,Ftm−1) dt).p(F_1)-p(F_0) =d_X\left(m\int_0^1 P(\dot A_t,F_t^{m-1})\,dt\right).

The proof for constant polynomials uses the constant form. Pullback commutes both with curvature and with applying a polynomial. Multiplicativity holds already at the level of forms because the entries being substituted have even degree and commute in the exterior algebra. These properties give the assertion about classes. ◻

The central gluing step of the proof can be expressed on an overlap by the diagram image Here each upper entry is a local section of the curvature-form bundle expressed in its own frame. The lower entry is a genuine scalar form.

Two distinctions visible on a torus

For U(1)U(1), use a real local potential aa with unitary connection matrix A=−2πi aA=-2\pi i\,a. Then F=−2πi daF=-2\pi i\,da and p(F)=iF/(2π)=dap(F)=iF/(2\pi)=da for the normalized degree-one polynomial. Take the trivial Hermitian line over R2/Z2\mathbb R^2/\mathbb Z^2 with a=sin⁡(2πx) dy.a=\sin(2\pi x)\,dy . Its normalized curvature is p(F)=2πcos⁡(2πx) dx∧dy.p(F)=2\pi\cos(2\pi x)\,dx\wedge dy . This form varies pointwise and is nonzero on open subsets, while its integral is zero. The trivial line admits the zero connection as well, and (15) exhibits the exact difference as dada. Thus a fixed characteristic class allows different local curvature forms.

On the same torus, a closed real one-form such as a=θ dxa=\theta\,dx gives curvature zero. A parallel section along a path solves ds−2πi a s=0ds-2\pi i\,a\,s=0, so a loop in the first circle direction transports by multiplication by exp⁡(2πiθ)\exp(2\pi i\theta). These flat connections share zero curvature and a trivial topological line. Their parallel transport distinguishes them when the parameters differ in R/Z\mathbb R/\mathbb Z. Differential characteristic classes will retain that information.

Change of group

For a Lie-group homomorphism ρ ⁣:G→H\rho\colon G\to H, extension of structure group constructs Q×GHQ\times_G H and its induced connection. Its local potentials and curvature are ρ∗Ai\rho_* A_i and ρ∗Fi\rho_*F_i, since ρ∗\rho_* preserves brackets. For an invariant polynomial qq on h\mathfrak h, the polynomial q∘ρ∗q\circ\rho_* on g\mathfrak g is invariant, and q(FQ×GH)=(q∘ρ∗)(FQ).q(F_{Q\times_G H})=(q\circ\rho_*)(F_Q). This identity on forms explains how a representation of a principal bundle leads to the characteristic classes of an associated vector bundle. Replacing a representation by a direct sum replaces its curvature by a block sum. The determinant and trace extract different algebraic expressions from that block sum, as the next chapter develops.

A degree-four matrix polynomial

For a unitary matrix connection let H=iF/(2π)H=iF/(2\pi). Eliminating the eigenvalues from the degree-two elementary symmetric polynomial gives c2(∇)=12((tr⁡H)2−tr⁡(H∧H))=tr⁡(F∧F)−tr⁡F∧tr⁡F8π2.c_2(\nabla)=\tfrac12\big((\operatorname{tr}H)^2-\operatorname{tr}(H\wedge H)\big) =\frac{\operatorname{tr}(F\wedge F)-\operatorname{tr}F\wedge\operatorname{tr}F}{8\pi^2}. The first equality is a polynomial identity, so it holds over the even scalar forms even where diagonalization would make no sense. On an SU(r)SU(r)-bundle the connection and curvature are traceless and only tr⁡(F∧F)/(8π2)\operatorname{tr}(F\wedge F)/(8\pi^2) remains. On a compact oriented four-manifold the normalization therefore turns an integral of a quadratic local expression into an integral topological number. Chapter [ch:characteristic] justifies the normalization through line bundles. Chapter [ch:chern-simons] compares two such local representatives by an explicit three-form.

Chern, Pontryagin, and Euler classes

The invariant polynomials of Chapter [ch:chern-weil] yield many closed forms. A normalization selects those corresponding to the named topological classes. The comparison uses a complex line, where the period can be computed, and a flag construction, where the cohomology pullback is injective. Integral classes also contain torsion, which the computation with real forms cannot see [1, §25].

Unitary curvature and Chern forms

Let EE be a complex vector bundle of rank rr with Hermitian metric and unitary connection. In a unitary frame its curvature FF is skew-Hermitian as an endomorphism-valued two-form. Define the total Chern form by c(∇)=det⁡(I+i2πF)=1+c1(∇)+⋯+cr(∇).c(\nabla)=\det\left(I+\frac{i}{2\pi}F\right) =1+c_1(\nabla)+\cdots+c_r(\nabla). The determinant is taken over the commutative algebra of scalar even-degree forms. Its homogeneous term containing jj curvature factors has degree 2j2j. Conjugation invariance allows Chapter [ch:chern-weil] to be applied. The coefficients are real because the matrix iFiF equals its conjugate transpose, and the determinant is unchanged by transpose. All higher Chern forms vanish by rank.

For two bundles, the direct-sum connection has block diagonal curvature. Consequently its Chern form is the product of the two original Chern forms. A metric on a short exact sequence of smooth complex bundles identifies the middle bundle with the direct sum of the ends by taking the orthogonal complement of the subbundle. Together with connection independence this yields the Whitney product identity for the real classes. The corresponding integral Whitney identity is the standard topological one [5, §21].

The Chern character uses an additive invariant polynomial in each degree: ch⁡(∇)=tr⁡exp⁡ ⁣(i2πF).\operatorname{ch}(\nabla)=\operatorname{tr} \exp\!\left(\frac{i}{2\pi}F\right). Only finitely many terms survive on a finite-dimensional base. The constant term is rr; the terms of degrees two and four are c1(∇)c_1(\nabla) and 12(c1(∇)2−2c2(∇))\tfrac12(c_1(\nabla)^2-2c_2(\nabla)). These formulas follow from the Newton identities over the algebra of forms. On E⊗E′E\otimes E', the induced curvature is F⊗I+I⊗F′F\otimes I+I\otimes F', with commuting summands. The exponential factors, and the trace of the tensor product is the product of the traces. This gives multiplicativity for the Chern character under tensor products and additivity under direct sums.

The normalization on a projective line

Let L=O(−1)→CP1L=\mathcal O(-1)\to\mathbb{CP}^1 be the tautological line with its metric induced from C2\mathbb C^2. On the affine chart [1:z][1:z] the holomorphic frame (1,z)(1,z) has squared norm h=1+∣z∣2h=1+|z|^2. The Chern connection in this frame has potential A=∂log⁡hA=\partial\log h. This formula follows from two conditions: its (0,1)(0,1) part is ordinary ∂ˉ\bar\partial in a holomorphic frame, and preserving the metric requires dh=h(A+A‾)dh=h(A+\overline A). Taking the (1,0)(1,0) part gives the claimed expression. Although the potential in this frame need not be imaginary, it becomes imaginary in the corresponding unitary frame. Its curvature is independent of the frame: A=z‾ dz1+∣z∣2,F=dz‾∧dz(1+∣z∣2)2.A=\frac{\overline z\,dz}{1+|z|^2}, \qquad F=\frac{d\overline z\wedge dz}{(1+|z|^2)^2}. The normalization integrates to ∫CP1iF2π=−1π∫R2dx dy(1+x2+y2)2=−1.\int_{\mathbb{CP}^1}\frac{iF}{2\pi} =-\frac1\pi \int_{\mathbb R^2} \frac{dx\,dy}{(1+x^2+y^2)^2} =-1 . The integral is computed in polar coordinates, since the omitted point has measure zero for a smooth top form. The bundle’s topological first Chern class has the same sign: the dual line O(1)\mathcal O(1) has a section defined by a nonzero linear functional on C2\mathbb C^2, with one positively oriented transverse zero. Its Euler number and hence its degree are 11. The duality of lines reverses the Chern class. One can also recover the sign by using the transition z−1z^{-1} from the frame (1,z)(1,z) to the frame (z−1,1)(z^{-1},1).

The standard comparison with integral Chern classes reduces to this calculation [5, Chapter IV]. For a line bundle, naturality applied to the universal line over projective space reduces normalization to the projective-line inclusion, which detects its degree-two generator. For a bundle EE, pass to the full flag bundle π ⁣:Flag⁡(E)→X\pi\colon\operatorname{Flag}(E)\to X. The successive projectivizations have cohomology free over the cohomology of their bases by the projective bundle theorem. The unit is one of the module generators, so each pullback is injective. In particular π∗\pi^* is injective in real cohomology. The pullback of EE has a flag with line-bundle quotients; choosing a metric splits the filtration as a direct sum of these lines. On the flag bundle, both the real image of the topological total Chern class and the curvature-defined total class are the product of 1+c11+c_1 of the line factors. Injectivity proves equality downstairs.

Real bundles and Pontryagin classes

For a real bundle VV, set pj(V)=(−1)jc2j(V⊗RC)∈H4j(X;Z).p_j(V)=(-1)^j c_{2j}(V\otimes_{\mathbb R}\mathbb C) \in H^{4j}(X;\mathbb Z). A metric connection on VV complexifies to a unitary connection, so the formula gives a Chern–Weil representative for the real image. In a real orthonormal frame FT=−FF^{\mathsf T}=-F, and transposing the matrix in det⁡(I+iF/(2π))\det(I+iF/(2\pi)) shows that this polynomial is even in FF. Thus the odd-index Chern forms of the complexification vanish. The integral odd-index Chern classes may still be two-torsion. Indeed conjugation identifies the complexification with its conjugate, and ck(E‾)=(−1)kck(E)c_k(\overline E)=(-1)^k c_k(E). It follows that the Pontryagin total class satisfies the product formula after inverting 22. Over the integers, products of two odd-index classes in the Whitney formula can contribute two-torsion [6, §15].

For an explicit torsion calculation, let λ\lambda be the tautological real line on RP4\mathbb{RP}^4 and let u=w1(λ)u=w_1(\lambda). The real bundle underlying λC\lambda_{\mathbb C} is λ⊕λ\lambda\oplus\lambda, whose second Stiefel–Whitney class is u2u^2. Thus a=c1(λC)a=c_1(\lambda_{\mathbb C}) reduces to u2u^2, so a2a^2 reduces to u4≠0u^4\ne0. For V=λ⊕λV=\lambda\oplus\lambda the integral p1(V)=−a2p_1(V)=-a^2 is nonzero, whereas p1(λ)=0p_1(\lambda)=0 by rank. The transition functions can be chosen to take values in {±1}\{\pm1\}, and the two line summands admit flat connections. Their real characteristic forms vanish. This example exhibits both the two-primary qualification on the Whitney identity and the failure of real forms to detect torsion.

Orientation and the Pfaffian

For an oriented real bundle of rank 2m2m, take an oriented orthonormal frame. Define the Euler form e(∇)=Pf⁡(F2π).e(\nabla)=\operatorname{Pf}\left(\frac{F}{2\pi}\right). We use the Pfaffian with Pf⁡(0a−a0)=a\operatorname{Pf}\left(\begin{smallmatrix}0&a\\-a&0\end{smallmatrix}\right)=a. One definition is the coefficient of the ordered top exterior basis vector in the mm-th divided power of ∑i<jFijei∧ej/(2π)\sum_{i<j}F_{ij}e_i\wedge e_j/(2\pi). Conjugation by an orthogonal matrix multiplies this coefficient by its determinant. The polynomial is therefore invariant under SO(2m)SO(2m), and reversing orientation changes its sign. On a complex line with its complex orientation, the skew matrix for multiplication by ii is (0−110)\left(\begin{smallmatrix}0&-1\\1&0\end{smallmatrix}\right). For F=i dαF=i\,d\alpha this gives e(∇)=−dα/(2π)=c1(∇)e(\nabla)=-d\alpha/(2\pi)=c_1(\nabla). The Pfaffian product on oriented block sums then implies equality of the Euler form and the top Chern form for the realification of a unitary bundle. This can be checked on block diagonal matrices with real scalar entries, which determine the restricted invariant polynomial on u(m)\mathfrak u(m) by unitary diagonalization. Polarization extends the equality to matrices of two-forms.

The comparison of the Pfaffian class with the real image of the integral Euler class is the Euler Chern–Weil theorem. The latter class is obtained by pulling the Thom class back along the zero section [1, §25]. In odd rank, negation of each fiber reverses orientation and fixes the zero section. It follows that the integral Euler class is killed by 22. Its real image is zero.

Take E=O(1)⊕O(1)E=\mathcal O(1)\oplus\mathcal O(1) on CP2\mathbb{CP}^2 and let h=c1(O(1))h=c_1(\mathcal O(1)). The top Chern class is h2h^2. Since (O(1))R⊗C≅O(1)⊕O(−1)(\mathcal O(1))_{\mathbb R}\otimes\mathbb C\cong\mathcal O(1)\oplus\mathcal O(-1), the first Pontryagin class of ERE_{\mathbb R} is 2h22h^2. Two independent linear functionals give hyperplanes intersecting in one positively oriented point, so ∫CP2h2=1\int_{\mathbb{CP}^2}h^2=1. The resulting Euler and Pontryagin numbers are 11 and 22, respectively, for the same oriented real bundle. They detect different invariant polynomials of its curvature.

Chern–Simons forms and coherent transgression

A form that witnesses the equality of two de Rham classes contains additional information. It records how the two curvature representatives are related, and it changes when the comparison is changed. Chern and Simons used this principle to construct secondary invariants [7]. Here we make the comparison depend on a simplex of connections, so comparisons between comparisons have their own differential identities.

An oriented path of connections

Let PP be an invariant symmetric mm-linear form with m≥1m\geq1, and let A0,A1A_0,A_1 be connections of a fixed principal bundle Q→XQ\to X. Set B=A1−A0B=A_1-A_0 and At=A0+tBA_t=A_0+tB. Define the relative transgression by T1(A0,A1)=m∫01P(B,Ftm−1) dt.T_1(A_0,A_1)=m\int_0^1P(B,F_t^{m-1})\,dt. Although written in a frame, the integrand is a global form because BB and FtF_t transform by simultaneous conjugation. Its exterior derivative is P(F1m)−P(F0m)P(F_1^m)-P(F_0^m) by (15). A second proof differentiates the curvature formula: F˙t=DAtB.\dot F_t= D_{A_t}B. The derivative of P(Ftm)P(F_t^m) is mP(DAtB,Ftm−1)mP(D_{A_t}B,F_t^{m-1}). Invariance converts covariant differentiation to exterior differentiation inside PP. Terms differentiating FtF_t vanish by Bianchi, so this derivative equals m dXP(B,Ftm−1)m\,d_X P(B,F_t^{m-1}).

For the quadratic polynomial p(X)=tr⁡(X2)p(X)=\operatorname{tr}(X^2) on a matrix Lie algebra, expand Ft=F0+tDA0B+t2B∧B.F_t=F_0+tD_{A_0}B+t^2 B\wedge B. Integration gives the three-form T1(A0,A1)=tr⁡ ⁣(2B∧F0+B∧DA0B+23B∧B∧B).T_1(A_0,A_1) =\operatorname{tr}\!\left(2B\wedge F_0 +B\wedge D_{A_0}B+\tfrac23 B\wedge B\wedge B\right). On a trivial bundle with the reference connection A0=0A_0=0, this reduces to the usual local expression tr⁡(A∧dA+23A∧3)\operatorname{tr}(A\wedge dA+\tfrac23 A^{\wedge 3}). The definition of T1T_1 also makes sense when neither connection has a global potential, since it uses a reference connection on the same bundle.

The sign from integration over a simplex

Orient the simplex with ordered vertices 0,…,q0,\ldots,q by coordinates t1,…,tqt_1,\ldots,t_q, with t0=1−∑i=1qtit_0=1-\sum_{i=1}^q t_i. On Δq×X\Delta^q\times X use the connection A~=∑i=0qtiAi\widetilde A=\sum_{i=0}^q t_i A_i in the XX directions. It has zero components in the simplex directions in local frames pulled back from XX. For a form ω\omega on the product, let IqωI_q\omega mean the integral of the component with simplex degree qq, with all simplex differentials written first. Its degree on XX is deg⁡(ω)−q\deg(\omega)-q. For a closed ω\omega, dXIqω=(−1)q+1I∂Δqω.d_X I_q\omega=(-1)^{q+1}I_{\partial\Delta^q}\omega. Indeed, write ω=∑pωp\omega=\sum_p\omega_p by simplex degree. In degree qq, the closedness equation is the sum of the simplex derivative of ωq−1\omega_{q-1} and (−1)q(-1)^q times the XX-differential of the coefficients of ωq\omega_q. Stokes on the parameter simplex gives (16).

Apply this to the closed form P(F~m)P(\widetilde F^m) and define, for q≥0q\geq0, Tq(A0,…,Aq)=(−1)q(q−1)/2IqP(F~m).T_q(A_0,\ldots,A_q) =(-1)^{q(q-1)/2} I_q P(\widetilde F^m). At q=0q=0 this is P(F0m)P(F_0^m). Each oriented face pulls the connection back to the affine family defined by the surviving vertices. Since (−1)q(q−1)/2(−1)q+1=(−1)(q−1)(q−2)/2(-1)^{q(q-1)/2}(-1)^{q+1}=(-1)^{(q-1)(q-2)/2}, the sign convention gives dXTq(A0,…,Aq)=∑i=0q(−1)iTq−1(A0,…,A^i,…,Aq).d_X T_q(A_0,\ldots,A_q) =\sum_{i=0}^q(-1)^iT_{q-1}(A_0,\ldots,\widehat A_i,\ldots,A_q). The hat omits the indicated connection. The case q=2q=2 is the relation T1(A0,A1)+T1(A1,A2)−T1(A0,A2)=dXT2(A0,A1,A2).T_1(A_0,A_1)+T_1(A_1,A_2)-T_1(A_0,A_2) = d_XT_2(A_0,A_1,A_2). It says that making the comparison in two steps and taking the direct comparison differ by an exact form. For a polynomial of degree mm, the form TqT_q vanishes for q>mq>m, since each curvature factor carries at most one simplex differential. The relations still hold at those levels because the boundary sum in (18) then vanishes.

A two-parameter calculation

For q=2q=2, put Bi=Ai−A0B_i=A_i-A_0 for i=1,2i=1,2. The curvature of the affine family consists of FtF_{\boldsymbol t} and the terms dti∧Bidt_i\wedge B_i. Each product of the two mixed curvature terms contributes a minus sign when dt2dt_2 is moved past B1B_1. The prefactor in (17) is also negative, so for m≥2m\geq2, T2(A0,A1,A2)=m(m−1)∫Δ2P(B1,B2,Ftm−2) dt1 dt2.T_2(A_0,A_1,A_2)=m(m-1)\int_{\Delta^2} P(B_1,B_2,F_{\boldsymbol t}^{m-2})\,dt_1\,dt_2 . If m=2m=2 the integral is over a coefficient independent of the parameter, and area⁡(Δ2)=1/2\operatorname{area}(\Delta^2)=1/2 in these coordinates. Thus T2=P(B1,B2)T_2=P(B_1,B_2). With p=c12p=c_1^2 and Ai=−2πi aiA_i=-2\pi i\,a_i for a unitary line, write bi=ai−a0b_i=a_i-a_0, which is a globally defined real one-form. The identity becomes T2=b1∧b2T_2=b_1\wedge b_2. Expanding the closed curvature expressions directly gives T1(A0,A1)+T1(A1,A2)−T1(A0,A2)=db1∧b2−b1∧db2.T_1(A_0,A_1)+T_1(A_1,A_2)-T_1(A_0,A_2) =db_1\wedge b_2-b_1\wedge db_2 . The two sides agree with d(b1∧b2)d(b_1\wedge b_2), confirming the orientation choice in a case where the two-form can be nonzero.

Gauge transformations and periods

Let XX be closed and oriented of dimension 2m−12m-1. Suppose P(Fm)P(F^m) is normalized so that it represents the real image of an integral class on every principal GG-bundle under consideration. A gauge automorphism u ⁣:Q→Qu\colon Q\to Q produces A1=u∗A0A_1=u^*A_0. Choose a path from A0A_0 to A1A_1 constant near the endpoints. Gluing the product bundle along (1,p)∼(0,u(p))(1,p)\sim(0,u(p)) gives a principal bundle on S1×XS^1\times X. The endpoint conditions make the family of connections descend smoothly. Orient S1×XS^1\times X with the parameter first. Integrating its characteristic form yields ∫XT1(A0,u∗A0)=∫S1×XP(F~m)∈Z.\int_X T_1(A_0,u^*A_0) = \int_{S^1\times X} P(\widetilde F^m) \in \mathbb Z . Here T1T_1 can be computed using the path with constant ends. A monotone reparametrization of the affine path gives one such path and leaves its integral unchanged. More generally the transgression for any two smooth paths with the same endpoints differs by an exact form, by Stokes applied to an interpolating square in the affine space of connections. This period identity uses both the closedness of XX and the integral normalization of the polynomial. A choice of an arbitrary multiple of a real polynomial changes the period group.

For P=c12P=c_1^2, the comparison of a line connection under a gauge map u ⁣:X→U(1)u\colon X\to U(1) can also be computed on XX itself. Use the normalized forms f=iF2π,β=i2πu−1du.f=\frac{iF}{2\pi},\qquad \beta=\frac{i}{2\pi}u^{-1}du . The comparison gives T1=2β∧fT_1=2\beta\wedge f. For X=S1×S2X=S^1\times S^2, take u(t,x)=e−2πitu(t,x)=e^{-2\pi i t} and let the line be pulled back from a degree-one line on S2S^2. Then β=dt\beta=dt and the period is 22. The mapping-torus argument explains its integrality without any choice of a global potential. Later, additive differential cohomology will package curvature and secondary period data in a single object.

A secondary value from a filling

Let X=∂ZX=\partial Z with its boundary orientation, and assume that a bundle with connection on XX extends over the compact oriented 2m2m-manifold ZZ. For the same integral normalization, define a number in R/Z\mathbb R/\mathbb Z by S(X,Q,A)=[∫ZP(Fm)](modZ).\mathcal S(X,Q,A)=\left[\int_Z P(F^m)\right]\pmod{\mathbb Z}. Two choices of connections on a fixed extension with equal boundary restriction yield the same integral: their difference is the integral of dT1dT_1, and the restricted T1T_1 is zero. By this argument, one may arrange that the connection is constant in the normal direction in a collar. Two different fillings can then be glued using the boundary bundle isomorphism, with the second filling oppositely oriented. The difference of integrals is the characteristic number of the glued closed manifold, an integer. This proves independence of the filling whenever at least one filling exists.

This construction determines a change formula. Suppose that both boundary connections A0,A1A_0,A_1 occur as ends of a cylinder on the same bundle and the first bounds with its connection. Attach the cylinder with its product orientation, with the interval first, to obtain a filling of the second. The two values then differ by ∫XT1(A0,A1)\int_X T_1(A_0,A_1) modulo integers. For the normalized first Chern form on a disk carrying a trivialized line connection A=−2πi aA=-2\pi i\,a, the filling rule gives exp⁡(2πi∫Dda)=exp⁡(2πi∫∂Da)\exp(2\pi i\int_D da)=\exp(2\pi i\int_{\partial D}a), the actual boundary holonomy by Stokes. A differential character extends this relationship to cycles without requiring a chosen filling.

The Weil algebra and equivariant curvature

A principal connection assigns forms to its potential and curvature, subject to the structure equation and Bianchi’s equation. The Weil algebra encodes those two equations before a bundle or a base has been chosen. To describe geometry with an extra group action, one combines this algebra with the forms of the space acted upon. The Cartan model results from eliminating the formal connection variables. The comparison is classical [8, Chapters 3–5].

Formal generators and their basic subalgebra

Let KK be a Lie group with finite-dimensional real Lie algebra k\mathfrak k. The Weil algebra as a graded commutative algebra is W(k)=Λ(k∗[1])⊗Sym⁡(k∗[2]).W(\mathfrak k)= \Lambda(\mathfrak k^*[1])\otimes \operatorname{Sym}(\mathfrak k^*[2]). Write θ\theta and uu for the tautological k\mathfrak k-valued expressions in its generators of degrees 11 and 22, respectively. Define the differential by dθ=u−12[θ,θ],du=−[θ,u].d\theta=u-\tfrac12[\theta,\theta],\qquad du=-[\theta,u]. The structure constants are those of k\mathfrak k, with graded signs from the exterior product. Extending dd by the Leibniz rule gives d2=0d^2=0 by the graded Jacobi identity. For example, d2u=−[u,u]+12[[θ,θ],u]−[θ,[θ,u]]=0.d^2u=-[u,u]+\tfrac12[[\theta,\theta],u]-[\theta,[\theta,u]]=0. The first term vanishes because uu has even degree, and Jacobi identifies the other two terms.

For ξ∈k\xi\in\mathfrak k the contraction ιξ\iota_\xi is the degree −1-1 derivation with ιξθ=ξ, ιξu=0\iota_\xi\theta=\xi,\ \iota_\xi u=0. Its commutator with dd is the infinitesimal coadjoint action on the generators, expressed as Lξθ=−[ξ,θ]L_\xi\theta=-[\xi,\theta] and Lξu=−[ξ,u]L_\xi u=-[\xi,u] in the tautological expressions. A form is horizontal if every contraction annihilates it, and basic if it is horizontal and invariant under KK. The horizontal subalgebra consists of polynomials in uu: the contractions are all derivatives with respect to the exterior generators. Thus W(k)bas=(Sym⁡(k∗))K,d∣bas=0.W(\mathfrak k)_{\mathrm{bas}} =\big(\operatorname{Sym}(\mathfrak k^*)\big)^K,\qquad d|_{\mathrm{bas}}=0 . The differential vanishes there by the infinitesimal invariance identity of Chapter [ch:chern-weil]. Full invariance under the group accounts for disconnected components.

The cohomology of the entire Weil algebra is R\mathbb R in degree zero. Indeed, replacing the degree-two generators uau^a by va=dθav^a=d\theta^a is an invertible polynomial change of variables. In the new generators dθa=vad\theta^a=v^a and dva=0dv^a=0. The degree −1-1 operator h=∑aθa∂/∂vah=\sum_a\theta^a\partial/\partial v^a has dh+hddh+hd equal to the operator that counts all factors of θ\theta and vv in a monomial. On the positive-count part, divide hh by the factor count to obtain a contraction. This explains how the basic algebra can carry nontrivial classes although the unqualified algebra is acyclic in positive degrees. Primitives in the whole algebra can fail the basic condition.

Given a right principal KK-bundle with connection, evaluation θ↦A, u↦F\theta\mapsto A,\ u\mapsto F is a morphism of differential graded algebras into forms on the total space. It preserves contractions by the reproduction property for AA and the horizontality of FF. It also preserves group actions. Its restriction to basic forms recovers the Chern–Weil homomorphism on representatives discussed in Chapter [ch:chern-weil].

Forms on a space with a group action

Let XX have a smooth right KK-action. Use ξX(x)=ddt(x⋅exp⁡(tξ))∣0\xi_X(x)=\frac{d}{dt}(x\cdot\exp(t\xi))|_0 for a fundamental vector field. On W(k)⊗Ω(X)W(\mathfrak k)\otimes\Omega(X) take diagonal contractions and the diagonal group action, using the graded tensor signs. The basic subcomplex is the Weil model. The alternative Cartan model is ΩK(X)=(Sym⁡(k∗)⊗Ω(X))K,(dKα)(ξ)=d(α(ξ))−ιξXα(ξ),\Omega_K(X) =\big(\operatorname{Sym}(\mathfrak k^*)\otimes\Omega(X)\big)^K, \qquad (d_K\alpha)(\xi)=d\big(\alpha(\xi)\big)-\iota_{\xi_X}\alpha(\xi), with polynomial generators in degree two. The equivariance convention is Rk∗α(ξ)=α(Ad⁡kξ).R_k^*\alpha(\xi)=\alpha(\operatorname{Ad}_k\xi). Cartan’s identity gives dK2α(ξ)=−LξXα(ξ)=0d_K^2\alpha(\xi)=-L_{\xi_X}\alpha(\xi)=0, since the one-parameter subgroup generated by ξ\xi fixes ξ\xi under the adjoint action.

Here is the comparison at the level of complexes. Choose dual bases eae_a of k\mathfrak k and θa\theta^a of degree-one generators. Put T=∑aθaιeaXT=\sum_a\theta^a\iota_{e_a}^X on the tensor product, with contractions on the second factor carrying the usual Koszul sign. This degree-zero derivation is nilpotent because it adds one exterior generator at each application. Its exponential is finite. One checks [ιξtot,T]=ιξX,[ιξX,T]=0.[\iota_\xi^{\mathrm{tot}},T]=\iota_\xi^X,\qquad [\iota_\xi^X,T]=0 . The first identity follows by differentiating θa\theta^a and the second uses anticommutation of contractions. It follows that exp⁡(−T)\exp(-T) sends elements independent of θ\theta to horizontally constrained elements, and its inverse sends horizontal elements back to ones independent of θ\theta. The operator TT is invariant by the pairing of a basis with its dual. Conjugating the differential and then setting θ=0\theta=0, its only terms that survive are the coefficient differential dXd_X and −∑auaιeaX-\sum_a u^a\iota_{e_a}^X. The latter comes from differentiating the first-order term of exp⁡(−T)\exp(-T), since the degree-two term of dθd\theta is uu. This proves the Cartan differential formula.

For compact KK the equivariant de Rham theorem identifies the cohomology of this complex with H∗(EK×KX;R)H^*(EK\times_K X;\mathbb R) [8, Chapters 1,4–5]. This comparison invokes the compact-group hypotheses. The algebraic identification of the Weil and Cartan complexes above is a separate calculation. A useful geometric evaluation is obtained from any right principal KK-bundle R→SR\to S with connection and an equivariant map R→XR\to X. Evaluate its Weil generators on the bundle and its XX-forms by pullback. The resulting forms on RR are basic and descend to SS. Thus an equivariant cohomology class yields ordinary cohomology classes for these bundles equipped with a map to XX. This pattern will reappear as a characteristic map on a quotient stack.

The moment term of an invariant connection

Let Q→XQ\to X be a right principal GG-bundle and suppose the right KK-action on XX lifts to QQ commuting with the GG-action. Choose a KK-invariant connection AA; when KK is compact one can average any connection using normalized Haar measure. For ξ∈k\xi\in\mathfrak k put μ(ξ)=ιξQA,FK(ξ)=F−μ(ξ).\mu(\xi)=\iota_{\xi_Q}A,\qquad F_K(\xi)=F-\mu(\xi). The term μ\mu is an equivariant g\mathfrak g-valued function on QQ, linear in the parameter, so both summands of FKF_K have total degree two. Our sign follows from Cartan’s identity and LξQA=0L_{\xi_Q}A=0: ιξQF=−dμ(ξ)+[μ(ξ),A]=−DAμ(ξ).\iota_{\xi_Q}F =-d\mu(\xi)+[\mu(\xi),A]=-D_A\mu(\xi). Consequently (DA−ιξQ)FK(ξ)=0(D_A-\iota_{\xi_Q})F_K(\xi)=0, by the ordinary Bianchi identity. Each P(FKm)P(F_K^m) is GG-basic, so it descends to an equivariant form on XX. Invariance of PP converts DAD_A to dd after polarization. Therefore the equivariant form is dKd_K-closed. Independence of the invariant connection follows by taking an affine path AtA_t and differentiating: ∂tFK,t=(DAt−ιξQ)A˙t.\partial_t F_{K,t}=(D_{A_t}-\iota_{\xi_Q})\dot A_t. The same polarization argument as for (15) expresses the derivative of P(FK,tm)P(F_{K,t}^m) as a Cartan coboundary.

For a concrete calculation take X=∗X=\ast and K=G=U(1)K=G=U(1). On Q=U(1)Q=U(1) use A=z−1dzA=z^{-1}dz, and lift the KK-action by z⋅k=zkmz\cdot k=z k^m for an integer mm. Write the real Lie-algebra coordinate as ξ=2πi s\xi=2\pi i\,s. Then F=0F=0 and μ(ξ)=2πi ms\mu(\xi)=2\pi i\,ms. The equivariant first Chern polynomial is i2π(F−μ(ξ))=ms.\frac{i}{2\pi}(F-\mu(\xi))=ms. The degree-two generator of the Cartan model of the point is normalized here by the weight-one action. Taking higher weights multiplies this generator even though the ordinary curvature on the point is zero. The moment term records the isotropy action that ordinary forms on the base cannot record.

Free actions and the sign of the universal generator

For a free compact KK-action with smooth principal quotient X→X/KX\to X/K, the projection EK×KX→X/KEK\times_KX\to X/K has contractible fiber and is a homotopy equivalence. The Cartan map can be evaluated using a connection on the principal quotient and the identity equivariant map X→XX\to X. In the Weil description its evaluation sends uu to the curvature and uses the horizontal part of the coefficient forms. The horizontal projection arises from exp⁡(−T)\exp(-T): its degree-one effect on a one-form α\alpha is α−∑aAaα((ea)X)\alpha-\sum_a A^a\alpha((e_a)_X), and multiplicativity gives the projection on higher forms. Changing the connection gives the same cohomology map by applying the path construction on the quotient times an interval.

For the Hopf circle bundle S3→CP1S^3\to\mathbb{CP}^1, a point over a line is a unit vector in that line. The right circle action is multiplication of the vector by the circle coordinate. Its associated line for the standard representation is O(−1)\mathcal O(-1), using (qg,v)∼(q,gv)(qg,v)\sim(q,gv). If AA is the corresponding circle connection, the variable ss in the preceding weight example evaluates to F/(2πi)=−iF/(2π)F/(2\pi i)=-iF/(2\pi). Its integral over CP1\mathbb{CP}^1 is therefore 11 by Chapter [ch:characteristic]. This also explains the sign for right actions in the quotient construction. A separate KK-action q⋅k=qkmq\cdot k=qk^m on a circle torsor over a point, pulled back along the principal Hopf bundle, gives the diagonal relation (rk,qkm)∼(r,q)(r k,qk^m)\sim(r,q). The associated character on the underlying Hopf bundle is k−mk^{-m}, so its first Chern number is mm, exactly the evaluation of msms.

The smooth site and its higher sheaves

A connection can be described in every coordinate chart of its base. The description includes transition functions and their equations. Keeping the transitions is essential. If one retains only the isomorphism class of the object on each chart, the information needed to assemble the object over the whole base can disappear. We now make the difference between these two procedures precise.

Test manifolds and local data

Let Mfd\mathbf{Mfd} denote a category of smooth Hausdorff second-countable finite-dimensional manifolds without boundary, in a fixed universe. We use open covers to define a Grothendieck topology. The later arguments about bundles also use surjective submersions. Such a map has local smooth sections by the submersion theorem, so it is a cover for the topology generated by open covers. When integrating a family, a boundary hypothesis will be stated explicitly.

For a manifold XX, put yX(U)=C∞(U,X)yX(U)=C^\infty(U,X). This is a sheaf of sets because smoothness is local on the source. If F\mathcal F is a presheaf, the action of a map v:V→Uv:V\to U goes in the direction F(U)→F(V)\mathcal F(U)\to\mathcal F(V). For differential forms it is the ordinary pullback v∗v^*. The sheaf Ωq\Omega^q assigns to UU its real qq-forms, and Ωclq\Omega_{\mathrm{cl}}^q assigns the kernel of d:Ωq(U)→Ωq+1(U)d:\Omega^q(U)\to\Omega^{q+1}(U). A family of forms on UiU_i glues uniquely when its two restrictions to each Ui∩UjU_i\cap U_j agree. Checking the exterior derivative locally shows that the glued form is closed when the given forms are closed.

On a cover {Ui→U}\{U_i\to U\}, write Ui0⋯ip=Ui0∩⋯∩UipU_{i_0\cdots i_p}=U_{i_0}\cap\cdots\cap U_{i_p}. The Čech nerve has at level pp the union of the Ui0⋯ipU_{i_0\cdots i_p}, with a separate copy for every tuple. Its face maps omit one index and its degeneracies repeat an index. If necessary we use formal coproducts of representable sheaves for this construction. At the first two levels this is the diagram ∐i,jyUij  ⇉  ∐iyUi  ⟶  yU.\coprod_{i,j}yU_{ij}\;\rightrightarrows\; \coprod_i yU_i\;\longrightarrow\;yU . For a sheaf of sets the limit test amounts to the equalizer at levels zero and one. The higher levels record identities already forced by equality of restrictions.

Descent with isomorphisms

A presheaf of groupoids has objects and their isomorphisms over each UU. We allow restriction functors associative up to specified coherent natural isomorphisms. This is how pullback of bundles behaves without choosing a strict model. The descent groupoid over a cover has as objects ai∈F(Ui)a_i\in\mathcal F(U_i) and isomorphisms ϕij:aj∣Uij→ai∣Uij\phi_{ij}:a_j|_{U_{ij}}\to a_i|_{U_{ij}}. We suppress the canonical pullback isomorphisms from the notation. The equations are ϕii=1,ϕijϕjk=ϕik.\phi_{ii}=1,\qquad \phi_{ij}\phi_{jk}=\phi_{ik}. A morphism is a family fi:ai→bif_i:a_i\to b_i with fiϕij=ψijfjf_i\phi_{ij}=\psi_{ij}f_j. Composition and inversion are componentwise. The restriction functor from F(U)\mathcal F(U) to this groupoid is fully faithful exactly when isomorphisms between global objects glue uniquely. It is essentially surjective exactly when compatible local objects can be glued, up to isomorphism. A stack satisfies both tests.

Proposition 1.1 (Vector bundles satisfy descent). Let Vr(U)\mathcal V_r(U) be the groupoid of real rank rr smooth vector bundles over UU and bundle isomorphisms over its identity. Then Vr\mathcal V_r is a stack.

Proof. Given descent data, take the disjoint union of the local total spaces EiE_i. Over a base point in UijU_{ij}, identify vjv_j with ϕij(vj)\phi_{ij}(v_j). The equations make this an equivalence relation. Every equivalence class over a point of UiU_i has exactly one representative in EiE_i. Give the quotient the topology for which the quotient charts from EiE_i over UiU_i are open, and use the vector bundle charts of EiE_i as smooth trivializations. The identifications on overlaps are smooth and linear, so they supply the change of coordinates. Points over different base points have disjoint neighborhoods by Hausdorffness of UU. Points over the same base point lie in one of the open bundle charts and are separated there. With a countable refinement of the base cover and countable trivializations this gives the usual second-countable total space. Applying a compatible family of local isomorphisms to each representative gives the unique global isomorphism. This verifies both parts of descent. ◻

This proof does not identify nonisomorphic gluing choices. For example, cover a circle by two connected arcs whose intersection has two components. Take trivial real lines on the two arcs. Gluing by the sign +1+1 on one component and −1-1 on the other produces the Möbius line. A change of local frames consists of two continuous nonzero functions, each with a constant sign on its arc. Their sign ratio is therefore the same on both components. It cannot convert the chosen transition to the identity on both. Consequently a presheaf keeping just isomorphism classes is not separated: the Möbius line and the trivial line give different classes on the circle and identical classes on all chart domains. The different transition signs are precisely the information missing from the local isomorphism classes.

From groupoids to spaces

The nerve of a groupoid is a space with no homotopy above degree one. Its points are objects, paths represent isomorphisms, and the composition law determines the higher simplices. The groupoid of descent data above models a homotopy limit of the Čech diagram of nerves. One way to see this is to read the data of a coherent cone over the diagram. Level zero specifies the aia_i. Level one supplies paths between their restrictions. A triangle filling in level two is the displayed composition equation, since the mapping spaces between fixed objects of a groupoid nerve are discrete. Higher coherence contains no new independent choices once the equations hold.

For sheaves of arbitrary spaces we use the corresponding homotopy-limit condition F(U)≃holim⁡[p]∈Δ∏i0,…,ipF(Ui0⋯ip).\mathcal F(U)\simeq \operatorname*{holim}_{[p]\in\mathbf\Delta} \prod_{i_0,\ldots,i_p}\mathcal F(U_{i_0\cdots i_p}). In this case an agreement may itself have nontrivial higher paths. A weakly associative system of transition maps can therefore carry additional information in the fillers and their compatibilities. We use H\mathcal H for this ∞\infty-category of smooth sheaves of spaces. It is an ∞\infty-topos by the general theory of sheaves of spaces on a site [9, Chapter 6]. In these chapters descent refers to cover descent unless a hypercover is specified. Degree-one descriptions only require groupoid descent. For the bounded differential complexes used later, the constructions can be checked on refinements by good covers of manifolds.

Representables remain sheaves, and the Yoneda equivalence has the form Map⁡H(yU,F)≃F(U).\operatorname{Map}_{\mathcal H}(yU,\mathcal F)\simeq \mathcal F(U). In particular, a natural operation on families can be written as a morphism of sheaves of spaces. Pullback along a map of test spaces is built into the morphism.

Example 1.2 (A homotopy fiber retains an identification). Let F:X→YF:\mathcal X\to\mathcal Y be a functor between groupoids and yy an object of Y\mathcal Y. A model of the homotopy fiber has objects (x,α:F(x)≅y)(x,\alpha:F(x)\cong y) and arrows u:x→x′u:x\to x' satisfying α′F(u)=α\alpha'F(u)=\alpha. The square of nerves is a homotopy pullback. To check the assertion, replace the target point by the groupoid of arrows with target yy, with morphisms the commuting triangles. This groupoid is contractible, and its source projection to Y\mathcal Y lifts isomorphisms. The ordinary pullback along that projection is precisely the indicated groupoid. An automorphism of xx survives in the fiber if F(u)=1F(u)=1. We will use this test to fix the underlying bundle while allowing its connection to vary.

Local presentations of a moduli object

A presentation by an atlas starts with a map yX→XyX\to\mathcal X admitting local lifts of points in the sense of sheaves. Its relation object is the homotopy fiber product yX×XhyXyX\times^h_{\mathcal X}yX. A point of this relation has two lifts and an identification between their images. Iterating the construction gives the Čech nerve of the atlas. When the map is an effective epimorphism, the geometric realization of that nerve is X\mathcal X. The local lifting condition is the ordinary criterion for effective epimorphisms of sheaves of spaces in terms of locally surjective maps on components. In the groupoid case the assertion follows directly by gluing the chosen lifts and the isomorphisms on their overlaps.

The atlas by itself supplies a collection of local representatives. Curvature and characteristic forms computed on those representatives must be compatible with the full relation object in order to define natural operations. This is one reason to distinguish basic forms on a principal bundle, differential forms on a chosen chart, and a cohomology class on the stack. The next chapter works out the relation object for torsor classification.

The classifying stack of smooth torsors

The familiar classifying space BGBG provides information about the isomorphism class of a topological GG-bundle. A smooth classifying stack records bundles, their isomorphisms, and smooth pullback along test maps. This refinement lets us express a construction such as extension of structure group by an actual map of moduli objects.

The stack BG\mathbf BG

Let GG be a Lie group, regarded as the sheaf U↦C∞(U,G)U\mapsto C^\infty(U,G). The one-object groupoid with automorphism sheaf GG gives a presheaf of groupoids PG\mathcal P_G. Its stackification is denoted BG\mathbf BG. Stackification permits local objects with descent isomorphisms to define global objects. Thus, for a cover {Ui}\{U_i\}, an object is represented by smooth functions gij:Uij→Gg_{ij}:U_{ij}\to G satisfying gii=1,gijgjk=gik.g_{ii}=1,\qquad g_{ij}g_{jk}=g_{ik}. A choice of right-principal local sections sis_i relates them by sj=sigijs_j=s_i g_{ij}. With this convention the overlap gluing identifies (x,h)j(x,h)_j with (x,gij(x)h)i(x,g_{ij}(x)h)_i.

Theorem 1.1 (Classification on a test manifold). For every UU, BG(U)\mathbf BG(U) is equivalent to the groupoid of smooth right principal GG-bundles over UU, and these equivalences commute with pullback up to the canonical isomorphisms.

Proof. A principal bundle is locally a product Ui×GU_i\times G. Choose local sections to express each change of frame by gijg_{ij}. Multiplying changes of frame establishes the cocycle identity. Conversely the indicated gluing of the local products produces a smooth locally trivial right GG-space with base UU. The equivalence relation commutes with the right action and a local product identifies each fiber with GG acting on the right. Given an equivariant isomorphism between such products over UiU_i, it is left multiplication by a smooth function on UiU_i. The compatibility condition on the overlaps is exactly the condition that these local maps respect the identifications defining the quotient. They therefore glue to the given global isomorphisms, and descriptions from different covers agree on a common refinement. Pullback replaces every local product and every transition function by their pullbacks. This identifies the stackification with the indicated groupoid on each test. ◻

The terminal smooth sheaf ∗* maps to BG\mathbf BG by picking the trivial torsor. Its relation object ∗×BGh∗≃G*\times^h_{\mathbf BG}*\simeq G is the sheaf of its automorphisms. On a manifold UU, an isomorphism between the two trivial right bundles has the form (x,h)↦(x,k(x)h)(x,h)\mapsto (x,k(x)h), for a unique k:U→Gk:U\to G. This computes the fiber product directly, including its functorial action on tests. Iteration gives the nerve with GnG^n at level nn, with faces given by multiplication and deletion and degeneracies by identities. In particular, the smooth group is recovered by looping the stack at the trivial torsor.

Fibers over a given torsor

A map yU→BGyU\to\mathbf BG classifying PP has a homotopy pullback square image For a test map v:V→Uv:V\to U, a point in the upper left consists of an isomorphism from the trivial bundle over VV to v∗Pv^*P. Evaluation at the identity picks a section of v∗Pv^*P, and that section extends equivariantly to the isomorphism. Such sections are the lifts V→PV\to P. There are no further automorphisms of a given identification in this fiber. Thus the square reconstructs the original bundle, including its projection, without a choice of coordinates. Right multiplication of the evaluation point corresponds to precomposing the isomorphism by the appropriate automorphism of the trivial right bundle.

For a homomorphism ρ:G→K\rho:G\to K, extension of structure group sends PP to P×GKP\times_G K, the quotient of P×KP\times K by (pg,k)∼(p,ρ(g)k)(p g,k)\sim(p,\rho(g)k). It gives a map BG→BK\mathbf BG\to\mathbf BK. The transition functions become ρ(gij)\rho(g_{ij}), which already proves compatibility with any base pullback and with iterated extension. A linear representation ρ:G→GL(V)\rho:G\to GL(V) works in the same way with VV in place of KK, giving the associated vector bundle P×GVP\times_G V. Its section over UU is an equivariant map P→VP\to V satisfying t(pg)=ρ(g)−1t(p)t(p g)=\rho(g)^{-1}t(p). Evaluation in the local sections sis_i gives t(sj)=ρ(gij)−1t(si)t(s_j)=\rho(g_{ij})^{-1}t(s_i), which is exactly the change of columns in the associated frames.

Proposition 1.2 (Reductions as sections). Let HH be a closed Lie subgroup of GG, and let PP be a smooth right GG-bundle over UU. The groupoid of HH-bundles together with an identification of their extended GG-bundle with PP is equivalent to the set of smooth sections of P/H→UP/H\to U, viewed as a discrete groupoid.

Proof. An identification Q×HG≅PQ\times_H G\cong P sends the canonical copy of QQ to an HH-subbundle of PP. Taking its cosets gives a section of P/HP/H. Conversely a section selects one right HH-coset in each fiber. Pulling back P→P/HP\to P/H along the section gives an HH-bundle. Here G→G/HG\to G/H is a locally trivial smooth HH-bundle, so the pullback gives exactly the indicated smooth subbundle. Multiplication extends it to PP. Constructions in both directions are inverse up to unique isomorphism, since a map between reductions is fixed by its required extension to PP. Thus no additional automorphisms remain when the identification is specified. ◻

For H=O(r)⊂GLr(R)H=O(r)\subset GL_r(\mathbb R), this is the familiar relation between a positive definite inner product and its orthonormal frame bundle. In a real vector bundle EE, a frame pp transports the Euclidean inner product onto the fiber. Replacing pp by php h, with h∈O(r)h\in O(r), leaves that inner product unchanged. Conversely frames defining the same inner product differ by an orthogonal matrix. Therefore P/HP/H is the bundle of positive definite forms and the section in the proposition is a metric. A metric is a reduction of structure group. A connection is different data and is treated by a separate moduli stack in Chapter [ch:bg-connection].

Stacks and underlying homotopy types

A classifying stack is functorial on every smooth test. Its set of connected components on a test gives isomorphism classes of torsors, but it also retains their gauge groups. For a bundle over a point all objects of BG(∗)\mathbf BG(*) are isomorphic, while its automorphisms are the elements of GG. Over a manifold the group of gauge automorphisms of PP is the group of sections of the conjugation bundle P×GGP\times_G G. Indeed an automorphism can be written p↦p u(p)p\mapsto p\,u(p), with u(pg)=g−1u(p)gu(p g)=g^{-1}u(p)g. This formula also exhibits the group law by pointwise multiplication in the conjugation bundle.

One can apply a realization or shape construction that forgets the smooth structure of the sheaf and obtains an ordinary homotopy type. For the Čech presentation above this recovers the classifying homotopy type of the topological Lie group, with the usual model obtained by realizing the nerve of the topological group. It answers a different question from evaluation on a fixed smooth test. A smooth-family map into BG\mathbf BG specifies a bundle over the product of the parameter manifold with the base, together with its descent data when no global trivialization is available. The same distinction will be needed for connections, differential cohomology, and higher bundles.

Connections as a smooth moduli stack

To a bundle with connection we can assign its transport, its curvature, or a characteristic form. Each assignment is compatible with pullback, but they retain different amounts of the original data. A stack makes the domain of these assignments explicit. Its automorphisms are the gauge transformations that preserve the given connection.

Gluing connective data

In a local right-principal section, the connective data for a Lie group GG are a form A∈Ω1(U;g)A\in\Omega^1(U;\mathfrak g). Define Ag=Ad⁡g−1A+g∗θL,A^g=\operatorname{Ad}_{g^{-1}}A+g^*\theta_L , where θL\theta_L is the left Maurer–Cartan form on GG. For a matrix group g∗θL=g−1dgg^*\theta_L=g^{-1}dg. The product rule for ghgh gives (Ag)h=Agh(A^g)^h=A^{gh}. Thus there is an action groupoid of these forms and smooth gauge changes on each UU. Its stackification is written BG∇\mathbf BG_\nabla.

An object of BG∇(U)\mathbf BG_\nabla(U) has local data (gij,Ai)(g_{ij},A_i) satisfying gijgjk=gik,Aj=Aigij.g_{ij}g_{jk}=g_{ik},\qquad A_j=A_i^{g_{ij}} . Replacing frames by sikis_i k_i sends AiA_i to AikiA_i^{k_i} and sends gijg_{ij} to ki−1gijkjk_i^{-1}g_{ij}k_j. When the local descriptions come from different covers one first restricts to a common refinement. This presentation of the connection stack is used in higher Chern–Weil theory [11, 10].

Theorem 1.1 (Classification with connection). For a test manifold UU, BG∇(U)\mathbf BG_\nabla(U) is equivalent to the groupoid of smooth principal GG-bundles with connection and connection-preserving isomorphisms.

Proof. A global connection ω\omega on a right bundle PP gives Ai=si∗ωA_i=s_i^*\omega. Right equivariance for a constant group element and reproduction of fundamental vector fields for the derivative of a varying one give Aj=AigijA_j=A_i^{g_{ij}}. Conversely, glue the local trivial bundles using gijg_{ij}, as in Theorem [thm:bg-classifies]. On the local product with group coordinate hh set ωi=Ad⁡h−1Ai+θL(h).\omega_i=\operatorname{Ad}_{h^{-1}}A_i+\theta_L(h). The form reproduces vertical fundamental vectors and has the required behavior under right translation. The chart change from jj to ii replaces hh by gij(x)hg_{ij}(x)h. Pulling ωi\omega_i back along it gives Ad⁡h−1Aigij+θL(h)\operatorname{Ad}_{h^{-1}}A_i^{g_{ij}}+\theta_L(h), which is ωj\omega_j. Hence the forms glue to a connection. Local gauge isomorphisms between two presentations preserve the local connection forms by the corresponding formula for AgA^g, and they glue by the compatibility of the underlying bundle maps. All operations commute with pullback along maps of test manifolds. ◻

The pointwise curvature of the local connection is Fi=dAi+12[Ai,Ai]F_i=dA_i+\frac12[A_i,A_i], using the graded bracket on forms. The change of frame gives Fj=Ad⁡gij−1FiF_j=\operatorname{Ad}_{g_{ij}^{-1}}F_i by the Maurer–Cartan equation. Consequently the FiF_i glue to a two-form with values in ad⁡P=P×Gg\operatorname{ad}P=P\times_G\mathfrak g. One cannot in general use a single g\mathfrak g-valued form on the whole base without choosing a gauge. An invariant polynomial on g\mathfrak g eliminates this dependence and produces a scalar form.

Forgetting the connection

The map u:BG∇→BGu:\mathbf BG_\nabla\to\mathbf BG forgets the connective data. Take its homotopy fiber on UU over a bundle PP. A point consists of (Q,ω)(Q,\omega) and an isomorphism Q≅PQ\cong P. Transport of ω\omega across the isomorphism gives a connection on PP. Every connection arises this way. An isomorphism over a prescribed identification of the underlying bundle is forced to be the identity after transport. Thus the fiber on UU is equivalent to the discrete groupoid of connections on PP. When nonempty it is an affine space under Ω1(U;ad⁡P)\Omega^1(U;\operatorname{ad}P): the difference of two principal connections is horizontal and equivariant, and adding an adjoint-valued one-form preserves the defining conditions. In the present manifold setting, local connections patch by a partition of unity, so the affine space is nonempty.

Discreteness here refers to automorphisms in a groupoid over a fixed test. On mathematical families it still has smooth variation. A map y(S×U)→BG∇y(S\times U)\to\mathbf BG_\nabla specifies a connection on the whole product. In a chosen trivialization it may include one-form components in the parameter directions of SS. A family of connections only along fibers of S×U→SS\times U\to S uses relative one-forms and a relative exterior derivative. A relative connection by itself gives no chosen components along SS. The difference can be seen on U=∗U=*: a connection along the point fibers is empty data, whereas an ordinary connection on SS has nontrivial curvature in general.

Extension of structure group induces a square image for a Lie group homomorphism ρ:G→K\rho:G\to K. In local data the upper map uses ρ(gij)\rho(g_{ij}) and the Lie algebra map dρ(Ai)d\rho(A_i). The naturality of θL\theta_L under ρ\rho verifies the gauge law, and dρ(Fi)d\rho(F_i) gives the resulting curvature. For a representation this recovers the connection on the associated vector bundle.

Infinitesimal gauge data

Choose a model PP for the underlying bundle and a connection AA. In the first-order problem below we retain gauge transformations of this model that reduce to the identity. In the marked fiber of Section 1.2, the identification of the underlying bundle removes these transformations and the variations form Ω1(U;ad⁡P)\Omega^1(U;\operatorname{ad}P) alone. A change is A+ϵaA+\epsilon a, where ϵ2=0\epsilon^2=0 and a∈Ω1(U;ad⁡P)a\in\Omega^1(U;\operatorname{ad}P). An infinitesimal change of frame with parameter ϵξ\epsilon\xi, for ξ∈Ω0(U;ad⁡P)\xi\in\Omega^0(U;\operatorname{ad}P), adds dAξd_A\xi to aa, with dAξ=dξ+[A,ξ].d_A\xi=d\xi+[A,\xi]. This follows by expanding the gauge formula to first order. It is independent of the chosen matrix realization because the derivative of the adjoint action is the Lie bracket. The first-order gauge action groupoid is therefore encoded by Ω0(U;ad⁡P)→dAΩ1(U;ad⁡P).\Omega^0(U;\operatorname{ad}P)\xrightarrow{d_A} \Omega^1(U;\operatorname{ad}P). Place the two terms in degrees −1,0-1,0. Infinitesimal automorphisms lie in the kernel, and isomorphism classes of changes form the cokernel.

The derivative of curvature is dAad_Aa. It follows by expanding F(A+ϵa)=F(A)+ϵ(da+[A,a])F(A+\epsilon a)=F(A)+\epsilon(da+[A,a]). A direct calculation with the graded Jacobi identity gives dA2η=[F(A),η]d_A^2\eta=[F(A),\eta]. At a flat connection, imposing flatness to first order requires dAa=0d_Aa=0, and the two arrows extend to the ordinary covariant de Rham complex. At a curved connection the linearized curvature of a gauge change is [F(A),ξ][F(A),\xi], which is the infinitesimal conjugation law. The flat deformation complex thus has a flatness hypothesis as part of its definition.

Example 1.2 (A flat circle connection on the two-torus). On U=R2/Z2U=\mathbb R^2/\mathbb Z^2 consider A=−2πi(r dx+s dy)A=-2\pi i(r\,dx+s\,dy) on the trivial circle bundle. Its curvature vanishes. The two fundamental loops have holonomies exp⁡(2πir)\exp(2\pi i r) and exp⁡(2πis)\exp(2\pi i s). A gauge change g=exp⁡(−2πi(mx+ny))g=\exp(-2\pi i(mx+ny)) for m,n∈Zm,n\in\mathbb Z adds the integer periods (m,n)(m,n). Conversely the logarithmic derivative of any smooth circle-valued gauge has integral periods in this normalization. Hence two members of this family are gauge equivalent exactly when their two periods differ by integers. Constant circle gauges preserve every member. This example exhibits curvature, holonomy and automorphisms as three separate pieces of data of the connection stack.

Characteristic maps and their differential lifts

Classical Chern–Weil theory produces a closed form and identifies its cohomology class. The two constructions, the form itself and its cohomology class, have different domains of naturality. There is a map on bundles with connection that produces the form. Integrality concerns the corresponding bundle after forgetting the connection. The space of differential refinements is a way to keep both together with their comparison.

The characteristic form as a natural operation

Let PmP_m be an invariant symmetric mm-linear form on g\mathfrak g. Applying it to the curvatures of a connection gives ωPm=Pm(F,…,F)∈Ω2m(U).\omega_{P_m}=P_m(F,\ldots,F)\in\Omega^{2m}(U). Invariance implies that the local expressions agree under changes of frame. Their exterior derivative is the sum of the terms obtained by replacing one input by dAFd_AF. To see why covariant differentiation can replace ordinary differentiation in this expression, differentiate the invariance identity along a one-parameter subgroup. The resulting sum with [ξ,F][\xi,F] in each position vanishes. In local coordinates this cancels the terms [A,F][A,F] in the graded derivative. The Bianchi identity then proves dωPm=0d\omega_{P_m}=0. This repeats the defining calculation in a form suitable for descent.

The assignment commutes with smooth pullback, since exterior differentiation and substitution into PmP_m commute with pullback. A connection-preserving isomorphism fixes the resulting form, so this is a morphism BG∇⟶Ωcl2m\mathbf BG_\nabla\longrightarrow\Omega_{\mathrm{cl}}^{2m} of smooth sheaves of spaces. Its values are actual closed forms. If two connections on the same bundle are joined by their affine interpolation, the transgression calculation of Chapter [ch:chern-simons] gives an exact difference between their characteristic forms. Equality of classes here uses the interpolation, whereas equality of forms would require further hypotheses.

Representing the real comparison

For n≥1n\geq 1, write K(C,n)K(C,n) for the nn-fold delooping of the locally constant sheaf of an abelian group CC. On a test manifold its space of sections represents cohomology with those coefficients, with π0\pi_0 given by Hn(U;C)H^n(U;C). One model uses the sheafified cochain complex of CC placed in degree −n-n, followed by the Dold–Kan construction. Equivalently, cover UU by a good cover, apply the cochain construction on its nerve, and keep the homotopies between cocycles. This computes the same space because the good cover is acyclic for the locally constant coefficients.

The Poincaré lemma resolves the locally constant reals by differential forms. In fact Ω0→dΩ1→d⋯→dΩn−1→dΩcln\Omega^0 \xrightarrow{d}\Omega^1\xrightarrow{d}\cdots \xrightarrow{d}\Omega^{n-1}\xrightarrow{d}\Omega_{\mathrm{cl}}^n has only degree-zero local cohomology, the locally constant functions, when displayed starting in cohomological degree zero. Placing its rightmost term in degree zero instead realizes K(R,n)K(\mathbb R,n), after Dold–Kan and sheafification. Inclusion of the rightmost term gives the morphism dR:Ωcln→K(R,n)\mathrm{dR}:\Omega_{\mathrm{cl}}^n\to K(\mathbb R,n). On connected components over UU it sends a closed form to its de Rham class. The comparison uses sheaves because locally exact potentials need not be globally exact.

The map real:K(Z,n)→K(R,n)\mathrm{real}:K(\mathbb Z,n)\to K(\mathbb R,n) comes from changing coefficients. Form the pullback image All pullbacks in this square are homotopy pullbacks. Sections of H^n\widehat{\mathbf H}^{n} therefore include an integral cocycle, a closed form, and a path comparing their real images. The path can be represented by a cochain with precisely the prescribed differential. Replacing a path by a homotopic one, or altering the chosen cocycle representatives with their compatibilities, gives the usual equivalence in this space. Its set of components is the degree-nn differential cohomology group modeled in Chapter [ch:diff-cohomology]. This is the discrete-coefficient specialization of the differential refinement constructions in [13, 14].

Proposition 1.1 (A lift supplies the real comparison). Let XX be a smooth sheaf of spaces and let c:X→K(Z,n)c:X\to K(\mathbb Z,n) and ω:X→Ωcln\omega:X\to\Omega_{\mathrm{cl}}^n be given. The space of lifts to H^n\widehat{\mathbf H}^{n} with these components fixed is the space of paths from real∘c\mathrm{real}\circ c to dR∘ω\mathrm{dR}\circ\omega in Map⁡(X,K(R,n))\operatorname{Map}(X,K(\mathbb R,n)).

Proof. The universal property of the homotopy pullback identifies maps into it with the homotopy pullback of the three mapping spaces. Taking the fiber over the pair (c,ω)(c,\omega) gives a model with exactly the indicated path as its remaining datum. ◻

For a normalized Chern polynomial, the integral component is the ordinary characteristic map of the underlying bundle and the form component is its Chern–Weil form. A differential characteristic map records a natural comparison in the above sense. Existence of a refinement on each individual test does not itself produce the coherent map on all tests. The classical differential refinements have concrete cocycle constructions [12, 11]. The circle case gives an elementary illustration.

The differential first Chern class in coordinates

Take a unitary frame of a complex line over each UiU_i. Use Ai=−2πi aiA_i=-2\pi i\,a_i and, on a good cover, write gij=exp⁡(−2πi λij)g_{ij}=\exp(-2\pi i\,\lambda_{ij}). The connection law gives aj−ai=dλija_j-a_i=d\lambda_{ij}, and δλ\delta\lambda is integer-valued. The ordinary exponential sequence with exponential t↦exp⁡(2πit)t\mapsto\exp(2\pi i t) computes the integral class from the logarithms −λij-\lambda_{ij}. With those signs put n=−δλ,ω=dai.n=-\delta\lambda,\qquad \omega=da_i . The daida_i agree on overlaps. For total Čech degree pp and form degree qq, use the total differential D=δ+(−1)pdD=\delta+(-1)^p d. Then D(a+λ)=ω−n.D(a+\lambda)=\omega-n . Thus the cochain a+λa+\lambda gives the real comparison between the two displayed representatives. The curvature is exactly iF/(2π)iF/(2\pi), since F=dAi=−2πi daiF=dA_i=-2\pi i\,da_i.

For completeness, a lift of the comparison on a good cover is represented by nn in integral Čech degree two, a closed global ω\omega, and total real degree-one data a+λa+\lambda with the displayed differential. Reading its bidegrees gives dai=ω, δa=dλ, δλ=−nda_i=\omega,\ \delta a=d\lambda,\ \delta\lambda=-n. Consequently the exponentials reconstruct the cocycle of a circle bundle and the aia_i reconstruct its connection. Changing a local frame by exp⁡(−2πiμi)\exp(-2\pi i\mu_i) replaces aia_i by ai+dμia_i+d\mu_i and λ\lambda by λ+δμ+k\lambda+\delta\mu+k, for an integral overlap cochain kk. The comparison cochain changes by Dμ+kD\mu+k, and nn by −δk-\delta k, giving the equivalent triple. Changing μ\mu by an integral local function ℓ\ell and kk by −δℓ-\delta\ell gives the same gauge transformation. These changes of representatives recover the Čech equivalences, retaining the circle-valued gauge automorphisms. Locally refinements of covers merely restrict the data. This yields the identification of the degree-two differential cocycle stack with BU(1)∇\mathbf BU(1)_\nabla, sending a line and connection to its first differential Chern class.

In particular a flat line on a circle may have nontrivial holonomy although ω\omega and nn can both be taken zero. It is the comparison, modulo integral changes, that remembers the flat class. For a nontrivial bundle the same comparison also records how the curvature periods agree with the integral class. Higher differential cohomology replaces these overlap logarithms by longer strings of cochains, as the next part develops.

Differential cohomology as a pullback

A characteristic form has a real cohomology class. An integral characteristic class can contain torsion, and a connection can have nontrivial holonomy even when both that form and that class vanish. Differential cohomology records all three phenomena. We develop it in ordinary integral degrees, using smooth singular chains to specify exactly how differential forms and holonomy fit together.

Characters on cycles

Write Ck(M;Z)C_k(M;\mathbb Z) for smooth singular chains, Zk(M;Z)=ker⁡∂Z_k(M;\mathbb Z)=\ker\partial, and Ck(M;A)=Hom⁡(Ck(M;Z),A)C^k(M;A)=\operatorname{Hom}(C_k(M;\mathbb Z),A). Smooth means smooth up to the faces of each simplex. The cochain differential is (δa)(c)=a(∂c)(\delta a)(c)=a(\partial c). Integration gives a cochain map J:Ω∙(M)→C∙(M;R)J:\Omega^\bullet(M)\to C^\bullet(M;\mathbb R). The smooth singular and ordinary singular cohomologies agree for manifolds. Write ΩZq(M)\Omega^q_{\mathbb Z}(M) for closed qq-forms whose integrals on smooth integral qq-cycles lie in Z\mathbb Z.

Definition 1.1. A degree q≥1q\geq 1 differential character on MM is a homomorphism h:Zq−1(M;Z)→R/Zh:Z_{q-1}(M;\mathbb Z)\to\mathbb R/\mathbb Z for which there is a qq-form ω\omega satisfying h(∂c)=∫cω(modZ)(c∈Cq(M;Z)).h(\partial c)=\int_c\omega\pmod{\mathbb Z} \qquad(c\in C_q(M;\mathbb Z)). The group under pointwise addition is denoted H^q(M;Z)\widehat H^q(M;\mathbb Z).

Pullback along f:N→Mf:N\to M sends hh to h∘f∗h\circ f_* on cycles and sends ω\omega to f∗ωf^*\omega. Both conclusions follow from the corresponding statements for smooth singular chains and forms. This is the contravariant functoriality of differential characters [12].

Lemma 1.2. The form ω\omega in (19) is unique and belongs to ΩZq(M)\Omega^q_{\mathbb Z}(M).

Proof. The difference of two such forms integrates to an integer over every smooth qq-simplex. Take a shrinking affine simplex in a chart, with edge vectors tv1,…,tvqt v_1,\ldots,t v_q. The difference of the integrals tends to zero, so it is zero for small tt. The limit after division by tq/q!t^q/q! is evaluation of the difference of the forms on v1,…,vqv_1,\ldots,v_q. This proves uniqueness. Replacing cc in (19) by a cycle proves integral periods. If bb is a smooth (q+1)(q+1)-simplex, the same equation for c=∂bc=\partial b gives ∫bdω∈Z\int_b d\omega\in\mathbb Z by Stokes. The shrinking-simplex argument now proves dω=0d\omega=0. ◻

Write R(h)=ωR(h)=\omega for the curvature. When q=1q=1, a character is precisely a smooth map f:M→R/Zf:M\to\mathbb R/\mathbb Z, regarded as a homomorphism on the free group of zero-chains. Indeed the boundary of a path gives the change in ff as the integral of the curvature. On a small coordinate ball this expresses ff as a constant plus a smooth real primitive of the curvature, reduced modulo Z\mathbb Z. Conversely, the normalized differential of a circle-valued function satisfies the boundary rule on paths.

Differential cocycles and integral classes

A real cochain h~∈Cq−1(M;R)\widetilde h\in C^{q-1}(M;\mathbb R) lifting hh on cycles exists. The cycles and boundaries in each degree are free abelian as subgroups of free abelian groups. The surjection Cq−1→Bq−2C_{q-1}\to B_{q-2} therefore splits, and one may lift the values on a basis of Zq−1Z_{q-1} and extend by zero on a complement. In degree zero the complement can be taken to be zero. The equation for curvature implies that c=Jω−δh~c=J\omega-\delta\widetilde h takes integral values. Since dω=0d\omega=0, it satisfies δc=0\delta c=0.

Proposition 1.3. A character is represented by a triple (c,h~,ω)(c,\widetilde h,\omega) with c∈Cq(M;Z)c\in C^q(M;\mathbb Z), δc=0\delta c=0, ω∈Ωq(M)\omega\in\Omega^q(M), dω=0d\omega=0, and δh~=Jω−c.\delta\widetilde h=J\omega-c . Two triples represent the same character precisely when they differ by (δb,−b−δu,0),b∈Cq−1(M;Z),u∈Cq−2(M;R).(\delta b,-b-\delta u,0), \quad b\in C^{q-1}(M;\mathbb Z),\quad u\in C^{q-2}(M;\mathbb R). In particular I(h)=[c]∈Hq(M;Z)I(h)=[c]\in H^q(M;\mathbb Z) is well defined and its realification is [ω]dR[\omega]_{\mathrm{dR}}.

Proof. A triple defines a character by evaluation of h~\widetilde h on cycles modulo integers. The defining cochain equation gives (19). Terms of the form (20) evaluate to zero on cycles and leave the form unchanged. Conversely, two lifts of the same character differ on the cycle group by an integral homomorphism. Subtract a cochain extending that integral homomorphism using the splitting above. The remaining real difference vanishes on cycles, hence factors through Bq−2B_{q-2}. Extend its real functional from Bq−2B_{q-2} to Cq−2C_{q-2}, using divisibility of R\mathbb R. Thus the lift difference is an integral cochain plus a real coboundary, with the change in cc then forced by the equation. Realifying the equation and applying de Rham’s theorem gives the last assertion. ◻

Let a(η)a(\eta) be the character z↦∫zη mod Zz\mapsto\int_z\eta\bmod\mathbb Z for a (q−1)(q-1)-form η\eta. Its representative is (0,Jη,dη)(0,J\eta,d\eta). A flat class ξ∈Hq−1(M;R/Z)\xi\in H^{q-1}(M;\mathbb R/\mathbb Z) defines a character j(ξ)j(\xi) by evaluation on cycles. Under the convention δh~=Jω−c\delta\widetilde h=J\omega-c, the integral class of j(ξ)j(\xi) is the negative of the Bockstein calculated by lifting ξ\xi to a real cochain and taking its coboundary. The sign follows directly by taking ω=0\omega=0.

Theorem 1.4. There are natural short exact sequences 0⟶Hq−1(M;R/Z)→jH^q(M;Z)→RΩZq(M)⟶0,0⟶Ωq−1(M)/ΩZq−1(M)→aH^q(M;Z)→IHq(M;Z)⟶0.\begin{aligned} 0&\longrightarrow H^{q-1}(M;\mathbb R/\mathbb Z) \xrightarrow{j}\widehat H^q(M;\mathbb Z) \xrightarrow{R}\Omega_{\mathbb Z}^q(M)\longrightarrow 0,\\ 0&\longrightarrow \Omega^{q-1}(M)/\Omega^{q-1}_{\mathbb Z}(M) \xrightarrow{a}\widehat H^q(M;\mathbb Z) \xrightarrow{I} H^q(M;\mathbb Z)\longrightarrow 0. \end{aligned} Here ΩZ0\Omega^0_{\mathbb Z} consists of locally constant integer-valued functions.

Proof. A flat character factors through Hq−1(M;Z)H_{q-1}(M;\mathbb Z), and every such homomorphism is a unique cohomology class by the universal coefficient theorem with injective coefficients R/Z\mathbb R/\mathbb Z. For an integral-period form ω\omega, the integer homomorphism on Hq(M;Z)H_q(M;\mathbb Z) given by periods lifts to integral cohomology by the universal coefficient theorem. Choose a cocycle cc representing that lift. The real cochain Jω−cJ\omega-c has zero real class and is a coboundary. This produces a character with curvature ω\omega.

The condition a(η)=0a(\eta)=0 says that the integrals of η\eta on all (q−1)(q-1)-cycles are integers. Integrating over boundaries and applying the shrinking-simplex argument gives dη=0d\eta=0. Thus its kernel is the indicated subgroup. If I(h)=0I(h)=0, the cochain representative can be chosen with c=0c=0. Then the curvature is exact by de Rham’s theorem, so write ω=dη\omega=d\eta. The real cochain h~−Jη\widetilde h-J\eta is closed and can be represented up to coboundary by JβJ\beta for a closed (q−1)(q-1)-form β\beta. Evaluating on cycles gives h=a(η+β)h=a(\eta+\beta). Finally, for any integral cocycle cc one can choose a closed ω\omega representing its real class, then solve δh~=Jω−c\delta\widetilde h=J\omega-c. This proves surjectivity of II. ◻

For example, on S1=R/ZS^1=\mathbb R/\mathbb Z the degree-two group is H1(S1;R/Z)≅R/ZH^1(S^1;\mathbb R/\mathbb Z)\cong\mathbb R/\mathbb Z. All its classes have R=I=0R=I=0. The form a0=θ dxa_0=\theta\,dx on the trivial line defines the character a(a0)a(a_0), whose value on the positively oriented circle is θ mod Z\theta\bmod\mathbb Z. This is the phase of the flat connection with matrix −2πia0-2\pi i a_0.

The space of cocycles

Fix qq and use the form complex truncated by setting degrees less than qq to zero. The data in a triple are cochains in the homotopy pullback \includegraphics\includegraphics{hcs-diagram-2a8a53f1a12448a0259cdcf09565c953a3a08d7133f6029de675b3c60c7daf3f.png} The lower map is realification. Concretely the pullback complex has degree-nn terms (c,h~,ω)(c,\widetilde h,\omega) with respective degrees n,n−1,nn,n-1,n and differential (c,h~,ω)⟼(δc,Jω−c−δh~,dω).(c,\widetilde h,\omega)\longmapsto (\delta c,J\omega-c-\delta\widetilde h,d\omega). The middle component makes the comparison of the two real images part of the data. Its second differential vanishes because JJ commutes with the differentials. At degree qq, cocycles and coboundaries are precisely those above. Dropping the comparison component before passing to cohomology would retain only the pair (I,R)(I,R), which already loses the circle example. This mapping-cone model is the ordinary integral case of differential cocycles [13].

The space attached to (21) is obtained by shifting degree qq to degree zero and taking the associated infinite loop space. Its connected components are H^q(M;Z)\widehat H^q(M;\mathbb Z). For i>0i>0 its homotopy groups are Hq−i(D(q)(M))H^{q-i}(D(q)(M)). In degrees less than qq the form term does not enter the cohomology calculation, leaving the homotopy fiber of Z→R\mathbb Z\to\mathbb R on cochains. Its cohomology in degree rr is Hr−1(M;R/Z)H^{r-1}(M;\mathbb R/\mathbb Z). Thus degree two has connection-preserving gauge automorphisms H0(M;R/Z)H^0(M;\mathbb R/\mathbb Z) at each object, and degree three has H1(M;R/Z)H^1(M;\mathbb R/\mathbb Z) and H0(M;R/Z)H^0(M;\mathbb R/\mathbb Z) in the first two homotopy degrees. The groups of components alone do not specify these automorphisms.

For a manifold of parameters SS, evaluating the cocycle space on S×MS\times M gives families with forms in all directions of the product, including the parameter directions. Restrictions at points of SS give individual characters. A collection of such pointwise restrictions can miss curvature components involving tangent vectors to SS. The presheaf of spaces defined by (21) satisfies descent. Smooth singular cochains satisfy homotopy descent for covers by subdivision, and each sheaf of forms is fine on manifolds. These give descent for the constituents and hence for their homotopy pullback. This use of sheaves of complexes separates the family object from evaluation in a single degree [14].

Products and differential characteristic classes

The Chern–Weil construction gives a closed form from a connection. To refine an integral characteristic class, that form must be supplied with its integral class and a comparison between them in real cochains. The comparison participates in products and remembers the secondary change under variation of the connection.

Multiplication at the cochain level

Use JJ and the cocycle model from Chapter [ch:diff-cohomology]. Fix the usual singular cup product, for instance the Alexander–Whitney product, and a natural real cochain homotopy BB between integrating the wedge product and taking the cup product of the integrations. For forms ω,η\omega,\eta with ω\omega of degree pp, the sign convention is δB(ω,η)+B(dω,η)+(−1)pB(ω,dη)=J(ω∧η)−Jω⌣Jη.\begin{aligned} \delta B(\omega,\eta)+B(d\omega,\eta) +(-1)^p B(\omega,d\eta) &=J(\omega\wedge\eta)-J\omega\smile J\eta . \end{aligned} Such a homotopy is a form of the multiplicative de Rham comparison. The construction and its coherent choices are treated in the differential cochain setting by Hopkins and Singer [13, §3.2].

Take cocycles x=(c,h,ω)x=(c,h,\omega) of degree pp and y=(c′,h′,η)y=(c',h',\eta) of degree qq. Define their product by the triple (c⌣c′, h⌣Jη+(−1)pc⌣h′+B(ω,η), ω∧η).\left(c\smile c',\ h\smile J\eta+(-1)^p c\smile h'+B(\omega,\eta),\ \omega\wedge\eta\right).

Proposition 1.1. Formula (23) defines a product on differential cohomology classes. Its integral class and curvature are the products of the corresponding integral classes and curvatures. For a form α\alpha of degree p−1p-1, a(α)⌣y=a(α∧R(y)).a(\alpha)\smile y=a\bigl(\alpha\wedge R(y)\bigr).

Proof. The coboundary of the middle component of (23) is (Jω−c)⌣Jη+c⌣(Jη−c′)+J(ω∧η)−Jω⌣Jη=J(ω∧η)−c⌣c′.\begin{aligned} &(J\omega-c)\smile J\eta +c\smile(J\eta-c') +J(\omega\wedge\eta)-J\omega\smile J\eta\\ &\hspace{2em}=J(\omega\wedge\eta)-c\smile c' . \end{aligned} This is the cocycle equation. The first and last components give the claimed product formulas for II and RR.

For invariance under changing xx, replace (c,h)(c,h) by (c+δb,h−b−δu)(c+\delta b,h-b-\delta u). The change in the first component is δ(b⌣c′)\delta(b\smile c'). If ΔH\Delta H is the change in the middle component, the coboundary rule for the cup product gives ΔH+b⌣c′=(−1)pδ(b⌣h′)−δ(u⌣Jη).\Delta H+b\smile c' =(-1)^p\delta(b\smile h')-\delta(u\smile J\eta). Thus the change is an equivalence of the form (20). Changing yy by (δb′,−b′−δu′,0)(\delta b',-b'-\delta u',0) instead changes the first component by δ((−1)pc⌣b′)\delta((-1)^p c\smile b') and the middle component by −(−1)pc⌣b′−δ(c⌣u′)-(-1)^p c\smile b'-\delta(c\smile u'). This too is a cocycle equivalence.

For (24), substitute c=0c=0, h=Jαh=J\alpha, and ω=dα\omega=d\alpha. Equation (22) says that the resulting middle component and J(α∧η)J(\alpha\wedge\eta) differ by a coboundary. The first component is zero in both triples and the last component is d(α∧η)d(\alpha\wedge\eta). ◻

The coherent multiplicative comparison makes the operation associative and graded commutative on classes [13, §3.2]. The correction BB has a concrete purpose. The two cochain products in (22) agree in real cohomology, and the term BB supplies the comparison before passing to cohomology. Consequently multiplication preserves both the integral class and the curvature while controlling the character on cycles. Omitting this correction generally breaks the cocycle equation.

The cylinder and changes of connection

Lemma 1.2 (Homotopy formula). For x∈H^q([0,1]×M;Z)x\in\widehat H^q([0,1]\times M;\mathbb Z) and the inclusions i0,i1i_0,i_1, i1∗x−i0∗x=a(∫[0,1]R(x)).i_1^*x-i_0^*x =a\left(\int_{[0,1]} R(x)\right). Here the interval direction is first in the integral. The smooth cochain and character constructions admit manifolds with boundary in this calculation.

Proof. For a cycle zz, triangulate the prism [0,1]×z[0,1]\times z so that its boundary is (i1)∗z−(i0)∗z(i_1)_*z-(i_0)_*z. This is the chain cross product with the interval as the first factor. The boundary formula for the cross product gives the two endpoint terms, since ∂z=0\partial z=0. Apply (19) and integrate first over the interval. The equality on every cycle proves the formula. ◻

For two connections on one bundle, a smooth path of connections gives a connection on the pullback bundle over the cylinder. The curvature of a degree-2j2j differential characteristic class there is the Chern–Weil form of the cylinder connection. Its interval integral is the transgression of that form. The homotopy formula says that the secondary change of the differential characteristic class is exactly the character aa of that transgression. A constant curvature form on each slice does not determine the interval components of the cylinder curvature.

Characteristic refinements for unitary bundles

Let GG be a Lie group with finitely many components, and let PP be a principal GG-bundle with connection ∇\nabla. Suppose an invariant polynomial pp of degree jj and a class ξ∈H2j(BG;Z)\xi\in H^{2j}(BG;\mathbb Z) have the same real universal Chern–Weil class. The Cheeger–Simons characteristic refinement for this compatible pair is natural under pullback and satisfies R(ξ^(P,∇))=p(F∇),I(ξ^(P,∇))=ξ(P).R\bigl(\widehat{\xi}(P,\nabla)\bigr)=p(F_\nabla), \qquad I\bigl(\widehat{\xi}(P,\nabla)\bigr)=\xi(P). Existence in this setting is the characteristic-class construction of Cheeger and Simons [12, Theorem 2.2]. We use its specialization to G=U(r)G=U(r), with the polynomial obtained as the degree-jj coefficient of the determinant det⁡(1+t iF∇2π).\det\left(1+t\,\frac{iF_\nabla}{2\pi}\right). It refines the Chern class cj(E)c_j(E) of the associated Hermitian vector bundle. We denote the differential class by c^j(E,∇)\widehat c_j(E,\nabla). Compatibility between the determinant form and the real integral Chern class is the classical input [1].

Theorem 1.3 (Uniqueness for the unitary construction). On smooth manifolds, there is at most one natural rule on Hermitian rank-rr bundles with unitary connection whose integral class is cjc_j and whose curvature is the corresponding determinant coefficient.

Proof. Subtract two such rules. Their difference has zero curvature on every manifold, including the cylinder. Interpolate two unitary connections by their affine path on the cylinder. The homotopy formula applied to the difference shows that it is independent of the connection. Naturality then reduces it to a rule on Hermitian bundles.

First consider a bundle admitting a finite trivializing cover. Using a smooth partition of unity subordinate to that cover, it embeds isometrically into a finite-dimensional trivial Hermitian bundle. For example, choose smooth functions ρi\rho_i supported in local trivializations with ∑iρi2=1\sum_i\rho_i^2=1, and send a vector to the tuple of its coordinates weighted by ρi\rho_i. Such functions are obtained by normalizing the squares of a finite partition of unity. The image gives a smooth classifying map into a finite complex Grassmannian, and the bundle is pulled back from its tautological bundle. Since the Grassmannian has only even-dimensional cells, its H2j−1(−;R/Z)H^{2j-1}(-;\mathbb R/\mathbb Z) vanishes. The curvature exact sequence then forces the difference for its tautological bundle to be zero. Naturality forces the same for every bundle admitting a finite trivializing cover.

For an arbitrary base, the image of each smooth singular cycle is compact. Choose an open neighborhood of that image formed from finitely many trivializing opens. Naturality and the finite-cover argument show that the difference evaluates to zero on that cycle. A character with zero value on every cycle is zero. ◻

Write c^=1+∑j≥1c^j\widehat c=1+\sum_{j\geq1}\widehat c_j, extending degree zero by H0(M;Z)H^0(M;\mathbb Z). The proof also gives the differential Whitney sum formula by applying the same argument on products of Grassmannians. Both candidate total classes have the same curvature because the block-diagonal curvature of a direct sum gives a product of determinant forms. Their integral classes agree by the ordinary Whitney sum theorem. Each positive-degree coefficient of their difference is flat and becomes zero on a product of the finite Grassmannians used for the two bundles. This proves c^(E⊕E′,∇⊕∇′)=c^(E,∇)⌣c^(E′,∇′).\widehat c(E\oplus E',\nabla\oplus\nabla') =\widehat c(E,\nabla)\smile\widehat c(E',\nabla').

First Chern class as holonomy

For a unitary line, take the real local one-form aia_i such that the frame connection is Ai=−2πiaiA_i=-2\pi i a_i. Parallel transport in a frame on an interval multiplies by exp⁡(2πi∫ai)\exp(2\pi i\int a_i). Along a loop, include the changes of frame where the path crosses charts. The abelian holonomy defines a character hh on smooth one-cycles. One can choose frames at the vertices of the one-chain and multiply the transport scalars of its oriented edges with their integer powers. Changing a vertex frame contributes opposite powers on incoming and outgoing edges, so the result is unchanged. On a smooth two-simplex the pulled back bundle is trivial. Stokes in a frame on that simplex gives the boundary rule with curvature dai=iF/(2π)da_i=iF/(2\pi).

The induced integral class is c1c_1. In local frames sj=sigijs_j=s_i g_{ij}, write gij=exp⁡(−2πiλij)g_{ij}=\exp(-2\pi i\lambda_{ij}). Then aj−ai=dλija_j-a_i=d\lambda_{ij} and the integral representative for c1c_1 is −δλ-\delta\lambda using the exponential t↦exp⁡(2πit)t\mapsto\exp(2\pi it). With D=δ+(−1)pdD=\delta+(-1)^p d in Čech degree pp, one has D(a+λ)=da+δλD(a+\lambda)=da+\delta\lambda, exhibiting the real comparison with curvature. The higher-holonomy interpretation of c^1\widehat c_1 thus recovers the actual phase of parallel transport. For a flat line its curvature is zero, but that phase can still detect either a nonzero torsion integral class or a continuous holonomy parameter.

Integration, holonomy, and transgression

Integrating along an oriented fiber takes a differential cocycle on a total space to a cocycle on the space of parameters. The integration acts before curvature is taken. This lets the output retain the holonomy on the fibers even when integrating the curvature alone would lose it.

Coherent transfers of smooth chains

Let π:E→M\pi:E\to M be a smooth proper submersion with closed fibers of dimension dd, equipped with an orientation of the vertical tangent bundle. One may also take an oriented smooth fiber bundle with a closed fiber as the starting point. For a smooth simplex σ:Δk→M\sigma:\Delta^k\to M, put Eσ=Δk×MEE_\sigma=\Delta^k\times_M E, a compact smooth manifold with corners. Orient it with the simplex direction first and the vertical direction second. Proper local triviality of the oriented family and the contractibility of the simplex give EσE_\sigma the homotopy type of its fiber. The same conclusion follows by pulling back a smooth fiber bundle and taking a trivialization over the simplex. Homology here may be calculated by smooth singular chains, including for the manifolds with corners that occur in this construction.

Let EσiE_{\sigma_i} be the pullback over face ii and ji:Eσi→Eσj_i:E_{\sigma_i}\to E_\sigma its inclusion. We choose, for each simplex, a smooth relative fundamental chain sσs_\sigma of degree k+dk+d, with prescribed boundary ∂sσ=∑i=0k(−1)i(ji)∗sσi(k>0).\partial s_\sigma=\sum_{i=0}^k(-1)^i(j_i)_*s_{\sigma_i} \quad(k>0). For k=0k=0, choose an oriented smooth fundamental cycle of the fiber. Suppose lower degrees have been chosen. The right side of (25) is a cycle in ∂Eσ\partial E_\sigma because the contributions from codimension-two faces cancel. It represents the boundary fundamental class. Indeed on the interior of each facet it has the corresponding local orientation, and those interiors meet every component of the boundary. The boundary of any relative fundamental representative also represents this class. Their difference bounds a chain in ∂Eσ\partial E_\sigma. Adding such a chain to the representative gives the desired boundary. This argument requires orientations of the entire family, since the face choices have to agree under every identification of fibers.

Compose sσs_\sigma with the projection Eσ→EE_\sigma\to E and extend linearly. The resulting degree-dd transfer TT on smooth singular chains satisfies ∂T=T∂\partial T=T\partial in the base-first convention. For a (k+d)(k+d)-form ω\omega, relative fundamental chains compute its integral on EσE_\sigma after pullback. This holds for every top form on EσE_\sigma. A boundary change integrates to zero by Stokes, and a chain supported in the boundary has zero integral by dimension. Fiberwise Fubini therefore gives ∫T(σ)ω=∫Eσω∣Eσ=∫σπ!ω,\int_{T(\sigma)}\omega =\int_{E_\sigma}\omega|_{E_\sigma} =\int_{\sigma}\pi_!\omega , where the first term uses the projection and the middle term uses its pullback. Here π!\pi_! on forms is defined by evaluating first on horizontal lifts of base directions, then on the vertical directions. Changing a lift by a vertical vector gives a vanishing term by the fiber dimension, so this definition needs no horizontal connection. It gives, in particular, π!(π∗α∧β)=α∧π!β\pi_!(\pi^*\alpha\wedge\beta)=\alpha\wedge\pi_!\beta.

Lemma 1.1. Two choices T,T′T,T' in this construction are chain homotopic by a homotopy KK for which K(σ)K(\sigma) factors through EσE_\sigma.

Proof. Induct on kk, keeping the homotopy values as chains in the pullbacks before projecting to EE. In degree zero the difference of the two fundamental cycles bounds in the fiber. Suppose chains kσik_{\sigma_i} giving the homotopy have been chosen on the faces. In EσE_\sigma, the chain sσ′−sσ−∑i(−1)i(ji)∗kσis'_\sigma-s_\sigma-\sum_i(-1)^i(j_i)_*k_{\sigma_i} is a cycle because of (25) and the identities on the faces. For k>0k>0 its degree is k+d>dk+d>d, so it bounds in EσE_\sigma by the fiber homotopy type. Let kσk_\sigma be a bounding chain there. Projecting to EE gives T′−T=∂K+K∂T'-T=\partial K+K\partial. In each degree K(σ)K(\sigma) thus factors through a smooth space of dimension one less than the degree of that chain, a property stronger than the existence of an abstract chain homotopy. ◻

Pushforward of differential characters

Suppose q>dq>d and h∈H^q(E;Z)h\in\widehat H^q(E;\mathbb Z). For a smooth integral (q−d−1)(q-d-1)-cycle zz of MM, define (π^!h)(z)=h(T(z)).\bigl(\widehat\pi_!h\bigr)(z)=h(T(z)). By the chain equation T(z)T(z) is a cycle. The boundary of a (q−d)(q-d)-chain cc is sent to the boundary of T(c)T(c). The character boundary rule and (26) give (π^!h)(∂c)=∫T(c)R(h)=∫cπ!R(h)(modZ).\bigl(\widehat\pi_!h\bigr)(\partial c) =\int_{T(c)}R(h) =\int_c\pi_!R(h)\pmod{\mathbb Z}. Thus the pushforward is a character. For different choices of transfer its change on zz is h(∂K(z))h(\partial K(z)). The integral of R(h)R(h) on K(z)K(z) vanishes because each summand factors through a pullback of dimension q−1q-1. The character is consequently independent of the fundamental chains used to define it.

In an oriented pullback square image the bundles over a simplex of M′M' are canonically the bundles over the composite simplex of MM. Using corresponding fundamental chains and applying independence of choices gives g∗π^!(h)=π^!′(g~∗h).g^*\widehat\pi_!(h)=\widehat\pi'_!(\widetilde g^*h). The curvature and integral class formulas are R(π^!h)=π!R(h),I(π^!h)=π!I(h).R(\widehat\pi_!h)=\pi_!R(h), \qquad I(\widehat\pi_!h)=\pi_!I(h). For the second formula, use a triple (c,h~,ω)(c,\widetilde h,\omega) on EE and precompose the two cochains with TT. Equation (26) identifies the resulting triple with curvature π!ω\pi_!\omega. The integral cochain operation is the ordinary oriented transfer dual to the relative fundamental chains of the bundle. This is a smooth singular description of the integral Gysin map. Fiber integration for characters has several equivalent constructions [15, §7].

A useful special case is h=a(η)h=a(\eta). By (26) in degree q−1q-1, π^!a(η)=a(π!η)\widehat\pi_!a(\eta)=a(\pi_!\eta). For a product M×F→MM\times F\to M, the transfer can be chosen as the chain cross product with a smooth fundamental cycle of FF, with the base cycle written first. A relative fundamental chain of a fiber with boundary would instead contribute an additional term to ∂T(z)\partial T(z). The definition with absolute cycles above uses closed fibers. Relative characters describe integration when boundary data, such as a trivialization, are specified [15, §§7.2 and 8.3].

Loop families

Let BB be a finite-dimensional smooth parameter manifold and let Γ:B×S1→M\Gamma:B\times S^1\to M be a smooth family of loops. For every h∈H^q(M;Z)h\in\widehat H^q(M;\mathbb Z) with q≥2q\geq2, set τΓ(h)=π^!Γ∗h∈H^q−1(B;Z),π:B×S1⟶B.\tau_\Gamma(h)=\widehat\pi_!\Gamma^*h \in\widehat H^{q-1}(B;\mathbb Z), \qquad \pi:B\times S^1\longrightarrow B. For q=2q=2, this character is the smooth S1S^1-valued holonomy function on the family. Evaluating it at bb evaluates hh on the oriented loop Γ(b,−)\Gamma(b,-). The differential of the function is described in real angular units by the fiber integral of Γ∗R(h)\Gamma^*R(h), including its component with one base direction and one circle direction. The formula explains how holonomies vary in a family even when each individual loop sees only the phase.

For q=3q=3, the output is a degree-two class, so it is represented by a line with connection on BB. Evaluate on a loop b:S1→Bb:S^1\to B. Equation (27) identifies the resulting phase with hh evaluated on the mapped torus S1×S1⟶M,(s,t)⟼Γ(b(s),t).S^1\times S^1\longrightarrow M,\qquad (s,t)\longmapsto\Gamma(b(s),t). The orientation of the torus is ss first, tt second. This characterization lets one use loop-space transgression without choosing a particular calculus on the whole infinite-dimensional free loop space. Each smooth family gives a well-defined pullback datum, and the construction is natural in the parameter manifold.

Example 1.2. On M=R2/Z2M=\mathbb R^2/\mathbb Z^2 take h=a(θ dx∧dy)h=a(\theta\,dx\wedge dy) of degree three with θ∈R\theta\in\mathbb R. Let B=S1B=S^1 and Γ(s,t)=(s,t)\Gamma(s,t)=(s,t). The resulting degree-two class is a(θ ds)a(\theta\,ds). Its curvature is zero, and its phase on the positive generator is θ mod Z\theta\bmod\mathbb Z. Transgressing the curvature alone would give zero for every θ\theta, whereas transgressing the character distinguishes the phases.

Bundle gerbes and surface phases

A unitary line with connection has one-form potentials on charts and circle-valued gluing functions on double intersections. A bundle gerbe with connection and curving has two-form potentials on charts, one-form connection data on double intersections, and circle-valued coherence on triple intersections. Its invariant is a degree-three character, so the boundary rule compares its phase on a surface with the integral of a three-form over a filling.

The Deligne complex and its signs

At the level of sheaves, the real cochain sheaf in the pullback model of Chapter [ch:diff-cohomology] can be computed by the de Rham complex, using the Poincaré lemma. Write Z‾\underline{\mathbb Z} for the locally constant integer sheaf. The homotopy pullback then identifies with the homotopy fiber of Z‾→Ω<q\underline{\mathbb Z}\to\Omega^{<q}, since Ω<q\Omega^{<q} is the quotient by the subcomplex Ω≥q\Omega^{\ge q}. A signed presentation of that fiber is the Deligne complex Z‾ ⟶ Ω0 →d ⋯ →d Ωq−1,\underline{\mathbb Z}\ \longrightarrow\ \Omega^0\ \xrightarrow{d}\ \cdots\ \xrightarrow{d}\ \Omega^{q-1}, with Z‾\underline{\mathbb Z} in degree zero and the first map the inclusion. The sign convention can be read directly from the middle component −c−δh~-c-\delta\widetilde h of the fiber differential. In internal degree n>0n>0, multiply that component’s cochain variable by (−1)n(-1)^n to obtain the displayed positive differentials. In degree qq the curvature map on the top form is therefore (−1)qd(-1)^q d. Projection onto Z‾\underline{\mathbb Z} gives the integral class. The hypercohomology in degree qq is H^q(M;Z)\widehat H^q(M;\mathbb Z), and the shifted complex retains the higher equivalences between cocycles [13, 14].

Choose a good open cover, so that every nonempty finite intersection is contractible and the Čech complex computes the required hypercohomology. Use δ\delta for the alternating Čech differential. In Čech degree pp the total differential is D=δ+(−1)pdDelD=\delta+(-1)^p d_{\rm Del}, where dDeld_{\rm Del} includes the first map of the Deligne complex. In total degree three write the entries in order of increasing internal degree as (n,λ,−A,−B).(n,\lambda,-A,-B). The entries have Čech degrees 3,2,1,03,2,1,0, respectively, with nn integer-valued, λ\lambda real-valued, AA a one-form cochain and BB a two-form cochain. The cocycle conditions become δn=0,δλ=n,δA=dλ,δB=dA.\delta n=0,\qquad \delta\lambda=n,\qquad \delta A=d\lambda,\qquad \delta B=dA . The curvings BiB_i give a global three-form H=dBiH=dB_i, because the difference on a double intersection is the differential of AijA_{ij}. This is the curvature of the class. For comparison, D(B+A−λ)=H−nD(B+A-\lambda)=H-n in the full real Čech–de Rham complex with D=δ+(−1)pdD=\delta+(-1)^p d. Thus the real image of [n][n] is [H][H]. The integer cochain remains part of the cocycle even when it represents torsion.

For an integer two-cochain mm, a real one-cochain μ\mu, and a zero-cochain Λ\Lambda of one-forms, adding a total coboundary with components (m,μ,−Λ)(m,\mu,-\Lambda) in degree two gives B′=B+dΛ,A′=A+δΛ+dμ,λ′=λ+δμ+m,n′=n+δm.\begin{aligned} B'&=B+d\Lambda,& A'&=A+\delta\Lambda+d\mu,\\ \lambda'&=\lambda+\delta\mu+m,& n'&=n+\delta m . \end{aligned} These signs follow by applying DD to each component. They preserve (28) and HH. Higher gauge relations between the changes are obtained from the preceding degrees of the same total complex. This explains why the gerbe cocycle belongs to a space with nontrivial higher homotopy groups.

In degree two the analogous entries are (n,λ,a)(n,\lambda,a), and the equations are δa=dλ\delta a=d\lambda and n=−δλn=-\delta\lambda. Exponentiating with gij=exp⁡(−2πiλij)g_{ij}=\exp(-2\pi i\lambda_{ij}) gives precisely the frame changes for a line with Ai=−2πiaiA_i=-2\pi i a_i. Thus the Deligne model identifies the integral component of its holonomy character with the integral Chern cocycle, including its possible torsion. This also verifies the normalization in Chapter [ch:diff-products].

Geometric realization by bundle gerbes

Take a surjective submersion p:Y→Mp:Y\to M possessing local sections. Denote its repeated fiber products by Y[k]Y^{[k]}. A unitary bundle gerbe has a Hermitian line L→Y[2]L\to Y^{[2]}, with multiplication over Y[3]Y^{[3]} Ly1,y2⊗Ly2,y3⟶Ly1,y3.L_{y_1,y_2}\otimes L_{y_2,y_3}\longrightarrow L_{y_1,y_3}. The map is a unitary isomorphism and the two composites on Y[4]Y^{[4]} agree. This is the line-bundle realization of the gerbe descent condition [16]. A connection is compatible when multiplication preserves the tensor product connection. Its normalized curvature FRL=iFL/(2π)F^L_{\mathbb R}=iF_L/(2\pi) therefore has zero alternating sum of pullbacks to Y[3]Y^{[3]}. A curving is a two-form BB on YY with FRL=pr⁡2∗B−pr⁡1∗B.F^L_{\mathbb R}=\operatorname{pr}_2^*B-\operatorname{pr}_1^*B .

A curving exists for a compatible line connection on a paracompact base. Choose local sections sis_i and a subordinate smooth partition of unity ρi\rho_i on MM. On p−1(Ui)p^{-1}(U_i) let Bi′B_i' be the pullback of FRLF^L_{\mathbb R} by y↦(si(p(y)),y)y\mapsto(s_i(p(y)),y). The cocycle identity for FRLF^L_{\mathbb R} gives Bi′(y2)−Bi′(y1)=FRL(y1,y2)B_i'(y_2)-B_i'(y_1)=F^L_{\mathbb R}(y_1,y_2) as pullback maps on forms. Hence B=∑ip∗ρiBi′B=\sum_i p^*\rho_i B_i', with terms extended by zero outside their supports, is a curving. Taking dd gives equality of the two pullbacks of dBdB. A form with this equality descends along pp, since pullback to any two local sections gives the same local form on the overlap. Thus dB=p∗HdB=p^*H.

Choose local sections of YY on a good cover and unit frames of the induced lines on double intersections. Write the connection form in each line frame as −2πiAij-2\pi i A_{ij}. With the frame multiplication scalar exp⁡(−2πiλijk)\exp(-2\pi i\lambda_{ijk}), compatibility of multiplication with the connection gives δA=dλ\delta A=d\lambda. Associativity gives δλ∈Z\delta\lambda\in\mathbb Z. The curving identity gives δB=dA\delta B=dA for Bi=si∗BB_i=s_i^*B. In these conventions the integral Dixmier–Douady cocycle is the Bockstein of the inverse frame multiplication scalars, which is δλ\delta\lambda for the positive exponential. This assigns the sign so that the integral class realifies to the normalized three-curvature.

Conversely, on a good cover, (28) constructs a bundle gerbe with Y=∐iUiY=\coprod_i U_i. Put a trivial line with connection −2πiAij-2\pi i A_{ij} on each double intersection. Use the exponential of −2πiλ-2\pi i\lambda for multiplication. The equations are exactly the compatibility, associativity and curving conditions. Coboundaries describe changes of local representatives and tensoring by the corresponding trivialization data. Up to refinement and stable equivalence with connection, this gives the gerbe classes described by degree-three differential cohomology. For the underlying topological gerbe the equivalence relation is stable isomorphism, and the class is the signed Dixmier–Douady class above [28]. Surface phases are evaluations of the associated differential character on smooth two-cycles.

A three-sphere and its equator

Let HH be a closed three-form on an oriented S3S^3 with ∫S3H=N∈Z\int_{S^3} H=N\in\mathbb Z. Cover S3S^3 by U+U_+ and U−U_-, the complements of the south and north pole. Choose two-forms B+B_+ and B−B_- there with dB±=HdB_\pm=H, since each open is contractible. The difference K=B−−B+K=B_--B_+ is closed on the overlap. Orient the equator as the boundary of the southern closed hemisphere. Stokes on the two hemispheres gives ∫S2K=∫S3H=N.\int_{S^2} K=\int_{S^3} H=N. The overlap has the homotopy type of that S2S^2, so KK has integral periods. Choose a line LL on the overlap with c1c_1 pairing to NN on the equator. Its normalized curvature can be chosen to equal KK, because any initial normalized curvature represents the same real class and can be adjusted by an exact form. Use LL on the ordered overlap (+,−)(+,-), its dual on (−,+)(-,+), and the trivial lines on the equal-index overlaps. Evaluation and tensor product multiplication give the associative gerbe multiplication. The two curvings yield exactly HH, and the integral class of the gerbe has value NN on the fundamental class of S3S^3.

Every integral two-cycle zz on S3S^3 bounds a three-chain CC, so the gerbe phase is exp⁡ ⁣(2πi∫CH).\exp\!\left(2\pi i\int_C H\right). A different filling leaves the phase unchanged because periods are integral. If HH is the round volume form normalized to have integral NN, its integral over either hemisphere is N/2N/2. The equatorial phase is (−1)N(-1)^N. This calculation uses both the orientation of the fiberwise boundary and the period convention. A global curving on S3S^3 could not give this nonzero integral of HH for N≠0N\ne0, since the integral of its exterior derivative would vanish by Stokes.

Strict two-groups and crossed modules

An ordinary gauge transformation changes a frame. In higher gauge theory there are also transformations between frame changes. A useful local model is a groupoid whose objects and arrows have group structures and whose structure maps respect multiplication. It has a particularly explicit description by two groups, a boundary homomorphism, and an action. The distinction between an arrow and its boundary will become the distinction between a surface transport and the path transports on its edges.

The groupoid of higher arrows

Definition 1.1 (Crossed module). A crossed module of Lie groups consists of a smooth homomorphism ∂:H→G\partial:H\to G and a smooth left action g▹hg\triangleright h of GG on HH by group automorphisms, such that ∂(g▹h)=g ∂(h) g−1,∂(h)▹k=hkh−1.\begin{aligned} \partial(g\triangleright h)&=g\,\partial(h)\,g^{-1},\\ \partial(h)\triangleright k&=hkh^{-1}. \end{aligned} The second identity is the Peiffer identity.

The associated strict Lie two-group has objects g∈Gg\in G. An arrow (h,g)(h,g) starts at gg and ends at ∂(h)g\partial(h)g. Composition in the groupoid is multiplication of the HH-labels in chronological order from right to left: (k,∂(h)g)∘(h,g)=(kh,g).(k,\partial(h)g)\circ(h,g)=(kh,g). The identity of gg is (1,g)(1,g) and the inverse for groupoid composition is (h−1,∂(h)g)(h^{-1},\partial(h)g). There is a second multiplication, the tensor product, with (h,g)⊗(k,ℓ)=(h(g▹k),gℓ).(h,g)\otimes(k,\ell) =\bigl(h(g\triangleright k),g\ell\bigr). The object group multiplies by g⊗ℓ=gℓg\otimes\ell=g\ell. Equation (29) verifies the target of the tensor product: ∂(h(g▹k))gℓ=(∂(h)g)(∂(k)ℓ).\partial\bigl(h(g\triangleright k)\bigr)g\ell =(\partial(h)g)(\partial(k)\ell). The group law on arrows is the semidirect product H⋊GH\rtimes G. Its inverse uses action by g−1g^{-1}, whereas the inverse of an arrow in the groupoid uses the target ∂(h)g\partial(h)g.

Tensor and groupoid composition must satisfy interchange. If the upper arrows in two composable pairs carry labels h′h' and k′k' and the lower arrows carry hh and kk, respectively, multiplying each pair first gives the label h′h(g▹(k′k)).h'h\bigl(g\triangleright(k'k)\bigr). Tensoring at each level first gives the label h′((∂(h)g)▹k′)h(g▹k).h'\bigl((\partial(h)g)\triangleright k'\bigr) h(g\triangleright k). Since (∂(h)g)▹k′=h(g▹k′)h−1(\partial(h)g)\triangleright k' =h(g\triangleright k')h^{-1}, the two labels agree. This calculation explains why equivariance and the Peiffer identity need separate hypotheses.

A strict monoidal groupoid with such an arrow group can also be read as a one-object strict two-category. Its one-morphisms are GG, its two-morphisms are the arrows just described, and horizontal composition is the tensor product. In this reading, when one path runs after another, the corresponding one-morphism is multiplied on the left. This agrees with successive parallel transport.

Recovering a crossed module from a group object

Proposition 1.2. The crossed module construction and passage to the kernel of the source map are inverse up to natural isomorphism. They identify crossed modules of Lie groups with internal groupoids in Lie groups.

Proof. Let G1⇉G0\mathcal G_1\rightrightarrows\mathcal G_0 be an internal groupoid with source ss, target tt, and unit uu. Put G=G0G=\mathcal G_0, H=ker⁡sH=\ker s, and ∂=t∣H\partial=t|_H. The split surjection ss is a submersion and its kernel is a Lie subgroup. Conjugation by u(g)u(g) defines the action on HH and applying tt verifies (29).

To prove the Peiffer identity, use that groupoid composition is a group homomorphism on composable pairs. If a∈ker⁡sa\in\ker s and b∈ker⁡tb\in\ker t, then the composable pairs (a,1)(a,1) and (1,b)(1,b) commute in the group of composable pairs. Their composites are aa and bb, so aa commutes with bb in G1\mathcal G_1. For h∈Hh\in H, the element h−1u(∂(h))h^{-1}u(\partial(h)) belongs to ker⁡t\ker t. Its commutation with k∈Hk\in H is exactly u(∂(h))ku(∂(h))−1=hkh−1.u(\partial(h))ku(\partial(h))^{-1}=hkh^{-1}.

The map (h,g)↦hu(g)(h,g)\mapsto hu(g) identifies H⋊GH\rtimes G with G1\mathcal G_1, with inverse a↦(au(s(a))−1,s(a))a\mapsto \bigl(au(s(a))^{-1},s(a)\bigr). This identification respects source and target. For a composable pair with lower arrow hu(g)hu(g), the upper arrow has the form ku(∂(h)g)k u(\partial(h)g). Factor the pair as (k,1)⋅(u(∂(h)g),hu(g))(k,1)\cdot \bigl(u(\partial(h)g),hu(g)\bigr) in G1×GG1\mathcal G_1\times_G\mathcal G_1. Applying the composition homomorphism gives khu(g)khu(g). This is the composition formula of the crossed module groupoid. The constructions on homomorphisms follow by restricting to HH and applying the same identification. ◻

The result addresses a strict model. A weak two-group has coherent associators, and questions about smooth presentations must keep track of how it is presented. In particular a choice of crossed module as a smooth model is extra structure. Baez and Lauda develop the corresponding algebraic and internal notions and the role of coherence [17].

The nerve and the two homotopy groups

The groupoid nerve has an explicit nn-simplex. It consists of nn consecutive HH-labels and an initial object: (h1,…,hn;g) ⟷ g→h1∂(h1)g→h2∂(h2h1)g⟶⋯ .(h_1,\ldots,h_n;g) \ \longleftrightarrow\ g\xrightarrow{h_1}\partial(h_1)g \xrightarrow{h_2}\partial(h_2h_1)g \longrightarrow\cdots . In degree two its middle face replaces h1,h2h_1,h_2 by h2h1h_2h_1. The first face adjusts the initial object by ∂(h1)\partial(h_1). Degeneracies insert identity arrows. Since all structure maps of the groupoid respect group multiplication, the nerve is a simplicial group. This gives a nerve description of the same strict model, including the operation responsible for the triangular faces.

For the next interpretation take the underlying discrete groups. The natural components of the groupoid are cosets of ∂(H)\partial(H), which is normal by (29). At each object the automorphism group is ker⁡∂\ker\partial. A groupoid can be replaced, up to equivalence of categories, by one representative of each component together with all of its automorphisms. Indeed inclusion of that subcategory is fully faithful and essentially surjective. The groupoid’s nerve therefore has components of the form B(ker⁡∂)B(\ker\partial), with contractible universal cover. Delooping the simplicial group gives a connected two-type with π1≅G/∂(H),π2≅ker⁡∂.\pi_1\cong G/\partial(H),\qquad \pi_2\cong\ker\partial. Here the delooping can be defined by the simplicial classifying construction, whose loop space is the realization of the simplicial group. Passing instead to smooth sheaves or keeping the topology of the Lie groups changes that interpretation. For example, the classifying space of the topological circle has its first nonzero homotopy group in degree two, while the classifying space of the discrete circle has fundamental group the circle regarded as an abstract group.

If a∈ker⁡∂a\in\ker\partial, then (30) gives aka−1=kaka^{-1}=k for every k∈Hk\in H. Thus ker⁡∂\ker\partial is abelian, and the action of ∂(H)\partial(H) on it is trivial. This supplies the G/∂(H)G/\partial(H)-action on the second homotopy group. The action is visible in the tensor formula as conjugation of a two-cell by the one-cell gg.

Models and their geometric interpretations

Example 1.3 (Ordinary groups and abelian two-groups). For H={1}H=\{1\} the associated groupoid is discrete on GG, and its delooping gives BGBG in the discrete interpretation. If G={1}G=\{1\} and HH is abelian, the groupoid has a single object and all its arrows are elements of HH. The resulting two-category models a twice-delooped abelian group. When H=U(1)H=U(1) is kept as a smooth group, the corresponding smooth bundle theory is the gerbe theory of Chapter [ch:gerbes].

Example 1.4 (An inner action). Let HH be a Lie group, and take GG to be a Lie group of smooth automorphisms of HH containing its inner automorphisms, with smooth evaluation and inner-automorphism map. The map h↦Ad⁡hh\mapsto \operatorname{Ad}_h with the evaluation action gives a crossed module since αAd⁡hα−1=Ad⁡α(h)\alpha\operatorname{Ad}_h\alpha^{-1}=\operatorname{Ad}_{\alpha(h)}. Its kernel is the center Z(H)Z(H), and its quotient of objects is the chosen group of automorphisms modulo inner automorphisms. A choice of finite-dimensional Lie group GG is part of this example. For instance H=U(n)H=U(n) and G=PU(n)G=PU(n) with conjugation work. This model keeps central two-automorphisms even when every object belongs to a single component.

Example 1.5 (An abelian kernel with a nontrivial action). Let VV be a finite-dimensional real representation of GG, with H=(V,+)H=(V,+) and trivial boundary. The Peiffer identity holds because VV is abelian. There are no arrows between distinct gg’s, but tensoring with gg acts on the automorphism arrows by the given representation. This provides a model in which the two-form connection takes values in an abelian group and still depends on transport by a nonabelian one-form connection.

Two-connections and their curvatures

A strict two-group assigns a boundary in GG to each arrow label in HH. On a small surface, the path transport around the boundary is controlled by the curvature of an ordinary connection. A two-form suitable for surface transport must have that curvature as its infinitesimal boundary. This yields a compatibility condition between the one- and two-form data.

Forms for a differential crossed module

Differentiate the crossed module of Chapter [ch:crossed-modules]. Write h,g\mathfrak h,\mathfrak g for the Lie algebras, ∂∗\partial_* for the differential of the boundary, and X▹bX\triangleright b for the induced action of g\mathfrak g on h\mathfrak h by derivations. The differentiated identities are ∂∗(X▹b)=[X,∂∗b],(∂∗a)▹b=[a,b].\partial_*(X\triangleright b)=[X,\partial_*b], \qquad (\partial_*a)\triangleright b=[a,b]. They extend to forms by combining wedge of scalar forms with the indicated bracket or action on coefficients. In particular, brackets of forms are graded antisymmetric. When both inputs have degree one, the factor 12\tfrac12 in 12[A,A]\tfrac12[A,A] accounts for the two cross terms.

On a trivialization over UU, a pair of potentials consists of A∈Ω1(U;g)A\in\Omega^1(U;\mathfrak g) and B∈Ω2(U;h)B\in\Omega^2(U;\mathfrak h). Write DA=d+A▹(−)D_A=d+A\triangleright(-) on h\mathfrak h-valued forms and DA=d+[A,−]D_A=d+[A,-] on g\mathfrak g-valued forms. Our conventions are FA=dA+12[A,A],K=FA−∂∗B,H=DAB.\begin{aligned} F_A&=dA+\tfrac12[A,A],\\ K&=F_A-\partial_* B,\\ \mathcal H&=D_AB. \end{aligned} The form KK is the fake curvature and H\mathcal H is the three-form curvature. We reserve the term fake-flat for K=0K=0. In the path two-functor model this condition is part of the local correspondence; a pair of forms can also be considered without imposing it [18, 19].

Proposition 1.1 (Higher Bianchi identities). For a pair of potentials with conventions (31)–(33), DAK=−∂∗H,DAH=K▹B.D_A K=-\partial_*\mathcal H,\qquad D_A\mathcal H=K\triangleright B. If K=0K=0, then H\mathcal H takes values in ker⁡∂∗\ker\partial_* and is DAD_A-closed there. The operator DAD_A is flat on ker⁡∂∗\ker\partial_*.

Proof. Equivariance of ∂∗\partial_* gives DA∂∗B=∂∗DABD_A\partial_*B=\partial_*D_AB. The ordinary Bianchi identity DAFA=0D_AF_A=0, applied to (32), gives the first formula. On an h\mathfrak h-valued form uu, expansion gives DA2u=dA▹u+A▹(A▹u)=FA▹u.D_A^2u =dA\triangleright u+A\triangleright(A\triangleright u) =F_A\triangleright u. The two terms involving dudu cancel, and the representation identity makes A▹(A▹u)A\triangleright(A\triangleright u) equal to 12[A,A]▹u\tfrac12[A,A]\triangleright u. In particular, DAH=K▹B+(∂∗B)▹B=K▹B+[B,B].D_A\mathcal H =K\triangleright B+(\partial_*B)\triangleright B =K\triangleright B+[B,B]. Here [B,B]=0[B,B]=0 because BB has even degree and the coefficients are over R\mathbb R. When K=0K=0, the first identity gives ∂∗H=0\partial_*\mathcal H=0. The differentiated Peiffer identity makes ker⁡∂∗\ker\partial_* central in h\mathfrak h, and equivariance makes it stable under the GG-action. On that subspace DA2u=(∂∗B)▹u=[B,u]=0D_A^2u=(\partial_*B)\triangleright u=[B,u]=0. ◻

Even under fake flatness DAD_A-closedness is a statement about the flat local system on the kernel, rather than ordinary closedness of each coefficient in an arbitrary frame. In a parallel local frame for that system it becomes ordinary closedness. The distinction is meaningful when the G/∂(H)G/\partial(H)-action on ker⁡∂∗\ker\partial_* is nontrivial.

Two elementary changes of potentials

The ordinary frame change by g:U→Gg:U\to G acts by Ag=Ad⁡(g−1)A+g−1dg,Bg=g−1▹B.A^g=\operatorname{Ad}(g^{-1})A+g^{-1}dg, \qquad B^g=g^{-1}\triangleright B. Differentiating the representation sends g−1dgg^{-1}dg to the logarithmic derivative of the matrix acting on h\mathfrak h. This cancels the derivative of g−1g^{-1} in DAg(g−1▹B)D_{A^g}(g^{-1}\triangleright B). It follows that KK transforms by Ad⁡(g−1)\operatorname{Ad}(g^{-1}) and H\mathcal H by g−1▹(−)g^{-1}\triangleright(-).

There is also a one-form change. For a∈Ω1(U;h)a\in\Omega^1(U;\mathfrak h), put A′=A+∂∗a,B′=B+DAa+12[a,a].A'=A+\partial_*a,\qquad B'=B+D_Aa+\tfrac12[a,a]. Expanding the ordinary curvature gives FA′=FA+∂∗(DAa+12[a,a]),K′=K.F_{A'}=F_A+\partial_*\bigl(D_Aa+\tfrac12[a,a]\bigr), \qquad K'=K. By the Peiffer identity DA′=DA+[a,−]D_{A'}=D_A+[a,-]. Use the graded derivation rule DA[a,a]=−2[a,DAa]D_A[a,a]=-2[a,D_Aa] and the graded Jacobi identity [a,[a,a]]=0[a,[a,a]]=0. Expanding the new three-form curvature leaves H′=H+FA▹a+[a,B]=H+K▹a.\mathcal H'=\mathcal H+F_A\triangleright a+[a,B] =\mathcal H+K\triangleright a. In the last equality the terms [B,a][B,a] and [a,B][a,B] cancel. Thus the one-form change preserves H\mathcal H in the fake-flat case. These formulas are local. Gluing a two-bundle requires transition data and coherence on overlaps in addition to these changes of representatives.

The flat quotient connection

Suppose the image N=∂(H)N=\partial(H) is a closed Lie subgroup of GG with Lie algebra ∂∗h\partial_*\mathfrak h. Equivariance makes it normal. Write Q=G/NQ=G/N, and use qq for the quotient map and q∗q_* for its differential. In a fake-flat model Fq∗A=q∗FA=q∗∂∗B=0.F_{q_*A}=q_*F_A=q_*\partial_*B=0. The equality uses the fact that q∗q_* preserves brackets. The change A↦A+∂∗aA\mapsto A+\partial_*a has no effect on q∗Aq_*A.

This calculation also gives a descent statement with explicit hypotheses. Suppose local fake-flat pairs have changes of frame gijg_{ij} on double overlaps, and hijkh_{ijk} on triple overlaps with gijgjk=∂(hijk)gikg_{ij}g_{jk}=\partial(h_{ijk})g_{ik}. Suppose the local one-forms satisfy Aj=Aigij+∂∗aijA_j=A_i^{g_{ij}}+\partial_*a_{ij} for one-forms aija_{ij} valued in h\mathfrak h. Then q(gij)q(gjk)=q(gik)q(g_{ij})q(g_{jk})=q(g_{ik}), so the quotient transition functions glue an ordinary principal QQ-bundle. The projected forms satisfy the ordinary gauge relation with these transition functions. They define a flat connection on that bundle. This is the part of the higher connection visible after identifying two path transports joined by a two-arrow. A smooth Lie group quotient was an assumption in this argument. At the Lie-algebra level the projection to g/∂∗h\mathfrak g/\partial_*\mathfrak h and its local flatness still make sense.

Local models and transport compatibility

Example 1.2 (Abelian curving). Let GG be trivial and H=U(1)H=U(1). In real coordinates on the Lie algebra, the two-form potential bb has three-form curvature dbdb. A one-form change is b↦b+dλb\mapsto b+d\lambda. On R3\mathbb R^3, take b=x dy∧dzb=x\,dy\wedge dz. Then db=dx∧dy∧dzdb=dx\wedge dy\wedge dz, so a fake-flat potential can have nonzero curvature. The leading transport on a surface comes from integrating bb, and its change along a three-dimensional deformation comes from integrating dbdb.

Example 1.3 (A twisted abelian kernel). For the crossed module V→GV\to G with zero boundary from Chapter [ch:crossed-modules], fake flatness forces FA=0F_A=0. The form BB is VV-valued and its curvature is DABD_AB. Take G=GL(V)G=GL(V) and U=(R/Z)×S2U=(\mathbb R/\mathbb Z)\times S^2. Let T∈End⁡(V)T\in\operatorname{End}(V), A=T dθA=T\,d\theta, and B=v ωB=v\,\omega, with θ\theta the circle coordinate, v∈Vv\in V, and ω\omega an area form on S2S^2 pulled back to UU. Then FA=0F_A=0 and H=Tv dθ∧ω.\mathcal H=Tv\,d\theta\wedge\omega. The local system can have monodromy exp⁡(−T)\exp(-T) around the circle. The three-form uses the action even though its coefficients lie in an abelian Lie algebra.

Example 1.4 (Identity boundary). For ∂=id⁡G\partial=\operatorname{id}_G with conjugation action, fake flatness sets B=FAB=F_A, hence H=DAFA=0\mathcal H=D_AF_A=0. The underlying groupoid has a unique arrow between any pair of objects. It records a comparison between transports but has no nonidentity automorphism at a fixed transport.

For the relevance of KK to surface integration, consider a small coordinate rectangle oriented by dx∧dydx\wedge dy. Its bottom and right edges give a path γ0\gamma_0 from the lower left to the upper right. The left and top edges give a parallel path γ1\gamma_1. Use UγU_\gamma for parallel transport with equation U˙=−A(γ˙)U\dot U=-A(\dot\gamma)U. Comparing the two endpoint operators gives Uγ1Uγ0−1U_{\gamma_1}U_{\gamma_0}^{-1}. This product corresponds to the clockwise boundary based at the common endpoint. On a square of side length ϵ\epsilon, the standard curvature expansion, in a smooth frame, reads Uγ1Uγ0−1=1+ϵ2FA(∂x,∂y)+O(ϵ3)U_{\gamma_1}U_{\gamma_0}^{-1} =1+\epsilon^2 F_A(\partial_x,\partial_y)+O(\epsilon^3) in coordinates at the identity. One can see the coefficient by choosing a frame with A=0A=0 at the lower left. The line integral around the rectangle gives the exterior derivative term to order two, and the bracket term at that point is zero in the chosen frame. Covariance gives the formula in the original frame. A surface arrow (h,Uγ0)(h,U_{\gamma_0}) has target Uγ1U_{\gamma_1} exactly when ∂(h)=Uγ1Uγ0−1\partial(h)=U_{\gamma_1}U_{\gamma_0}^{-1}. The coefficient of area in hh is B(∂x,∂y)B(\partial_x,\partial_y). Equating coefficients gives K=0K=0. The next chapter describes the global composition law for such arrows.

Thin homotopy and surface transport

A bundle with connection sends a path to an isomorphism between fibers. A higher connection sends a deformation between two paths to a comparison between their isomorphisms. The parameterization of the deformation should have no effect on this comparison. The rank conditions defining thin homotopy express the precise independence needed here. A homotopy whose image sweeps out a three-dimensional volume can change the comparison, and the three-form curvature measures that effect.

The path two-groupoid

Take paths with sitting instants, so a path is constant on collars of its endpoints. One-morphisms from xx to yy are paths modulo endpoint-preserving smooth homotopies of differential rank at most one. A bigon from γ0\gamma_0 to γ1\gamma_1 is a smooth map Σ(t,s)\Sigma(t,s) on the square. The parameter tt runs along a path and ss varies it: Σ(t,0)=γ0(t),Σ(t,1)=γ1(t),Σ(0,s)=x,Σ(1,s)=y.\Sigma(t,0)=\gamma_0(t),\qquad \Sigma(t,1)=\gamma_1(t),\qquad \Sigma(0,s)=x,\qquad \Sigma(1,s)=y. Require collars on which the map is independent of the coordinate normal to each side. We orient its integration domain by dt∧dsdt\wedge ds, so its boundary is γ0−γ1\gamma_0-\gamma_1 after discarding the constant edges. Two bigons are equivalent if they are joined by a collared smooth homotopy of rank at most two, with fixed end objects and with rank at most one on the two side faces that vary the boundary paths. The two-morphisms are these classes of bigons.

Vertical composition stacks squares in ss. If boundary paths have only the same thin class, a thin bigon joins the chosen representatives before stacking. Horizontal composition uses concatenation in tt. The collars make concatenation smooth. Different choices of rescaling on the concatenated intervals give homotopic parameterizations through the same surface. Such a homotopy has rank at most two because it factors through that surface, and its restrictions to the moving path faces have rank at most one. In the full quotient these compositions give the strict path two-groupoid Π2thin(M)\Pi_2^{\mathrm{thin}}(M). In particular the independence of the connecting thin bigon and the interchange law are part of the path two-groupoid construction in [19, §2.1]. Thin equivalence of two bigons is more restrictive than ordinary homotopy, even when the boundary paths agree pointwise.

Smooth two-functors and local potentials

Let G\mathcal G be the finite-dimensional strict Lie two-group defined by a smooth crossed module. Its one-object two-category is denoted by BGB\mathcal G in this chapter. A two-functor T:Π2thin(U)→BGT:\Pi_2^{\mathrm{thin}}(U)\to B\mathcal G sends the paths to elements of GG. It associates to a bigon Σ\Sigma with source γ0\gamma_0 an arrow (hΣ,T(γ0))(h_\Sigma,T(\gamma_0)) as in Chapter [ch:crossed-modules]. In particular ∂(hΣ)T(γ0)=T(γ1).\partial(h_\Sigma)T(\gamma_0)=T(\gamma_1). For successive vertical bigons with labels h,h′h,h', the composite has label h′hh'h. For horizontal composition, if Σ1\Sigma_1 runs before Σ2\Sigma_2 and their sources have transports g1,g2g_1,g_2, the label is h2(g2▹h1).h_2(g_2\triangleright h_1). The action accounts for transporting the first surface label to the frame used by the second. Diagrammatically, the comparison belongs to the hom-groupoid of the target two-category: image Smoothness here means smoothness on smooth families in the path and bigon quotients with their quotient diffeologies. Use the smooth structure of GG and H⋊GH\rtimes G on the values. Thus a set-theoretic two-functor need not have local differential-form data.

With a fixed trivialization on UU, Schreiber and Waldorf prove that these smooth two-functors correspond to pairs (A,B)∈Ω1(U;g)×Ω2(U;h)(A,B)\in\Omega^1(U;\mathfrak g)\times\Omega^2(U;\mathfrak h) satisfying FA=∂∗BF_A=\partial_*B, in the conventions for orientation and infinitesimal extraction used there [19, Theorem 2.21]. One extracts the one-form from short paths and the two-form from shrinking rectangles. The converse uses ordinary differential equations for path transport and a second integration for the bigon label. The boundary calculation of Chapter [ch:2connections] is the infinitesimal necessity of the equation in our conventions. The theorem supplies sufficiency and reconstructs the two-functor. Global transport for a nontrivial higher bundle uses local trivializations and their higher transformation data. The global formulation with transport two-functors and descent is developed in [20].

Abelian transport from a two-form

The abelian case has a direct construction. Take G={1}G=\{1\} and H=U(1)H=U(1). Let b∈Ω2(M)b\in\Omega^2(M) be a global curving for a trivial gerbe. All path transports are the sole GG-value, and the surface assignment is hb(Σ)=exp⁡ ⁣(2πi∫[0,1]2Σ∗b).h_b(\Sigma)=\exp\!\left(2\pi i\int_{[0,1]^2}\Sigma^*b\right). For either form of concatenation, the integral over the new square is the sum of the integrals over the two old squares after orientation-preserving changes of variable. Thus the assignment respects both compositions. The identity bigon factors through a path, so its integral is zero. Smoothness on parameter families follows by differentiating an integral over a fixed compact domain.

For thin invariance take a bigon homotopy VV on the cube. At every point rank⁡(dV)≤2\operatorname{rank}(dV)\le 2, so the pullback of dbdb is zero. Two side faces are constant and the other two have rank at most one, so their pullbacks of bb are zero. Using the cube orientation with the deformation parameter first and Stokes’ theorem gives ∫Σ′∗b−∫Σ∗b=∫V∗db=0.\int \Sigma'^*b-\int\Sigma^*b=\int V^*db=0. This proves that the surface formula descends to the thin quotient even if dbdb is nonzero. For an unrestricted homotopy with fixed boundary paths the same calculation gives hb(Σ′)hb(Σ)=exp⁡ ⁣(2πi∫V∗db).\frac{h_b(\Sigma')}{h_b(\Sigma)} =\exp\!\left(2\pi i\int V^*db\right). In particular the local curvature is recoverable from infinitesimal three-dimensional changes of holonomy.

If b′=b+dλb'=b+d\lambda for a real one-form λ\lambda, then hb′(Σ)hb(Σ)=exp⁡ ⁣(2πi(∫γ0λ−∫γ1λ)).\frac{h_{b'}(\Sigma)}{h_b(\Sigma)} =\exp\!\left(2\pi i\bigl( \textstyle\int_{\gamma_0}\lambda- \textstyle\int_{\gamma_1}\lambda\bigr)\right). The ratio is a boundary comparison for an open bigon and is one for a closed surface. On smooth singular two-cycles zz, the rule z↦∫zb mod Zz\mapsto\int_z b\bmod\mathbb Z is a degree-three differential character. For any smooth singular three-chain cc, its value on ∂c\partial c is ∫cdb mod Z\int_c db\bmod\mathbb Z. The underlying integral class is zero for this globally defined potential. A gerbe with nonzero integral class uses locally defined curvings with transition terms, as in Chapter [ch:gerbes].

For example, on R3\mathbb R^3 take b=12x dy∧dzb=\tfrac12 x\,dy\wedge dz and let cc be the unit cube with its standard orientation. The curvature integral is 1/21/2. Equivalently the face at x=1x=1 contributes 1/21/2 with outward orientation while the other faces contribute zero. Its surface holonomy is −1-1 even though ∂c\partial c is a boundary. This is consistent with the differential character law and with the invariance of holonomy under thin deformations.

Transgression and orientation

For a global bb, a smooth family of loops carries the transgressed one-form (τb)γ(V)=∫S1bγ(t)(γ˙(t),V(t)) dt.(\tau b)_\gamma(V) =\int_{S^1} b_{\gamma(t)}(\dot\gamma(t),V(t))\,dt . Here the variation vector VV is along the loop. The formula defines a form on families without putting a choice of infinite-dimensional manifold structure on the free loop space. A path of loops makes a cylinder. With the loop tangent ordered first, the integral of τb\tau b over the path equals the integral of bb on this cylinder. Consequently the trivial line with connection form −2πi τb-2\pi i\,\tau b has precisely the cylinder transport coming from hbh_b.

The sign in its curvature can be checked on any finite-dimensional parameter space SS. Write the pulled-back two-form on S1×SS^1\times S as dt∧α(t)+β(t)dt\wedge\alpha(t)+\beta(t), with α,β\alpha,\beta forms on SS. The coefficient of dtdt in its exterior derivative is ∂tβ−dSα\partial_t\beta-d_S\alpha. Integrating in tt, the derivative of a periodic function integrates to zero. Thus d τb=−τ(db)d\,\tau b=-\tau(db) for this fiber-first convention. One can instead choose a convention that reverses both transgression and cylinder orientation. The formula on a trivial gerbe explains why transgression on a global gerbe produces a line bundle with connection on loop families and why the resulting curvature is a transgressed three-form.

Weil algebras and higher characteristic maps

For an ordinary Lie group a characteristic form is obtained by evaluating an invariant polynomial on curvature. Chapter [ch:weil] expressed that evaluation in the Weil algebra. A higher analogue starts from a graded algebra whose differential encodes brackets of several arities. The characteristic form, the Chern–Simons element, and the infinitesimal cocycle then appear as three parts of one calculation. Passing from that calculation to a differential characteristic class requires global integration and descent.

The algebra of connections and curvatures

Work over R\mathbb R with cohomological degrees. For the local models here, a finite-dimensional L∞L_\infty-algebra is presented by its Chevalley–Eilenberg algebra CE⁡=(Sym⁡(V),q)\operatorname{CE}=\bigl(\operatorname{Sym}(V),q\bigr) where VV is a finite graded space in positive degrees, Sym⁡\operatorname{Sym} is graded-commutative, and qq is an augmentation-preserving derivation of degree one with q2=0q^2=0. The homogeneous components of qq in each word length are dual to the brackets. For an ordinary Lie algebra the generators have degree one and qq is quadratic. A degree-two generator can record a higher abelian direction. We only use free graded algebras with polynomial differential, so no completion is needed.

Write sVsV for a copy of VV raised by one degree. As a graded algebra, the Weil algebra is W=Sym⁡(V⊕sV).W=\operatorname{Sym}(V\oplus sV). Extend s(v)=svs(v)=sv, s(sv)=0s(sv)=0 as a derivation of degree one. The differential is determined by dWv=qv+sv,dW(sv)=−s(qv).d_W v=qv+sv,\qquad d_W(sv)=-s(qv). On the unshifted algebra, dWd_W is q+sq+s. The identity dWs+sdW=0d_Ws+sd_W=0 holds on the generators and hence everywhere. On vv, dW2v=s(qv)−s(qv)=0d_W^2v=s(qv)-s(qv)=0. On svsv it is s dWqv=s2qv=0s\,d_Wqv=s^2qv=0, so this is indeed a differential graded algebra. The map j∗:W→CE⁡j^*:W\to\operatorname{CE} kills sVsV and is a chain map. It restricts the algebra of connections to the corresponding flat infinitesimal data.

Proposition 1.1 (Acyclicity of the Weil algebra). The cohomology of WW is R\mathbb R in degree zero and zero in positive degrees.

Proof. Use vv and t(v)=sv+qvt(v)=sv+qv as algebra generators; the change is invertible because qvqv belongs to the unshifted algebra. In these coordinates dWv=t(v)d_W v=t(v) and dWt(v)=0d_W t(v)=0. Define the degree-minus-one derivation hh by h(v)=0h(v)=0 and h(t(v))=vh(t(v))=v. The derivation dWh+hdWd_Wh+hd_W acts as the identity on generators, so on monomials of word length rr it acts as multiplication by rr. The differential preserves this word length. On positive word length define the contraction by h/rh/r. It contracts the augmentation ideal, leaving the constants. ◻

A map of differential graded algebras Φ:W→Ω∙(U)\Phi:W\to\Omega^\bullet(U) evaluates vv as a potential of degree ∣v∣|v|, and svsv as its raw curvature of degree ∣v∣+1|v|+1. It is specified by the potentials: the shifted values must be dΦ(v)−Φ(qv)d\Phi(v)-\Phi(qv), and the equation for dW(sv)d_W(sv) then follows from dW2=0d_W^2=0. For a Lie algebra, write AA for the collection of degree-one generators and FF for their shifts. The universal identities are exactly dWA=F−12[A,A],dWF=−[A,F].d_W A=F-\tfrac12[A,A],\qquad d_WF=-[A,F]. They recover the structure equation and Bianchi identity after evaluation. A map that annihilates the shifted generators factors through CE⁡\operatorname{CE}.

An invariant form and its two primitives

The subalgebra generated by sVsV is the algebra of horizontal polynomials. Call a horizontal polynomial PP invariant if dWP=0d_W P=0. For ordinary reductive Lie algebras this gives the usual invariant polynomials on their curvature variables [11, Remark 4.1.17]. This elementary definition suffices for the examples below.

When PP has positive degree n+1n+1, choose C∈WnC\in W^n with dWC=Pd_W C=P using acyclicity. Since j∗P=0j^*P=0, μ=j∗C\mu=j^*C is a cocycle of degree nn. Two choices of CC differ by a closed element. If n>0n>0, acyclicity gives an exact difference, so the class of μ\mu is independent of that choice. This is the transgression relation in [11, Definition 4.1.20]. On UU, evaluation gives dΦ(C)=Φ(P).d\Phi(C)=\Phi(P). Under flat evaluation it gives Φ(C)=Φ(μ)\Phi(C)=\Phi(\mu), using the factorization through the Chevalley–Eilenberg algebra. Thus Φ(C)\Phi(C) is a curvature primitive on a chart and μ\mu records its flat restriction.

The diagram separates algebraic restriction from form evaluation: image The dashed factorization is defined when Φ(sV)=0\Phi(sV)=0. The algebra gives a local real form for PP even without that factorization. Gluing this form and refining it to an integral differential class requires the global structure group, compatible descent data, and an integral normalization. For example, if Φ(P)\Phi(P) is the curvature of a differential character, its integrals on integral cycles must be integers. The local Weil calculation by itself has no such period condition.

Example 1.2 (The degree-three abelian case). Take CE⁡=R[b]\operatorname{CE}=\mathbb R[b] with ∣b∣=2|b|=2 and qb=0qb=0. The Weil algebra adds h=sbh=sb with ∣h∣=3|h|=3 and dWb=h, dWh=0d_Wb=h,\ d_Wh=0. Thus P=h, C=b, μ=bP=h,\ C=b,\ \mu=b. A map of Weil algebras into forms assigns B=Φ(b)B=\Phi(b) and H=Φ(h)=dB\mathcal H=\Phi(h)=dB. Taking the smooth circle as the integrated coefficient group gives the curving of a gerbe on a trivializing open set. Integral Čech descent data can change its underlying degree-three class while leaving the same equations on each open set. The holonomy and curvature law for a single global BB were computed in Chapter [ch:surface-transport].

The string transgression

Let g\mathfrak g be an ordinary Lie algebra with an invariant symmetric bilinear form QQ. Its degree-four characteristic polynomial is P=Q(F,F)P=Q(F,F). The connection interpolation At=tAA_t=tA in the ordinary Weil algebra has curvature Ft=tF+12(t2−t)[A,A].F_t=tF+\tfrac12(t^2-t)[A,A]. Differentiating the polynomial with respect to tt gives 2dWQ(A,Ft)2d_WQ(A,F_t) by invariance of QQ and the Bianchi identity for AtA_t. Integrating from 00 to 11 gives C=Q(A,F)−16Q(A,[A,A]),dWC=P.C=Q(A,F)-\tfrac16Q(A,[A,A]),\qquad d_W C=P . In these signs, μ=−16Q(A,[A,A])\mu=-\tfrac16Q(A,[A,A]) is the Chevalley–Eilenberg restriction. Reversing the sign of the new abelian generator gives the alternate convention for the corresponding central extension.

Adjoin a degree-two generator bb to the Chevalley–Eilenberg algebra with qb=−μqb=-\mu. This is a differential because qμ=0q\mu=0. In its Weil algebra, write h=sbh=sb. Then dWb=h−μd_Wb=h-\mu, dWh=sμd_Wh=s\mu, and dWμ=sμd_W\mu=s\mu. The modified element K=h+C−μ\mathcal K=h+C-\mu satisfies j∗K=0j^*\mathcal K=0 and dWK=Pd_W\mathcal K=P. After evaluation this reads H=dB+Φ(C),dH=Φ(P).H=dB+\Phi(C),\qquad dH=\Phi(P). The new three-form HH uses a Chern–Simons term. This Bianchi equation is different from the fake-flat strict crossed module equation in Chapter [ch:2connections]. It arises from an extension whose Chevalley–Eilenberg differential has the cubic term μ\mu.

For the spin group in stable rank, take QQ normalized so PP represents the real image of the integral class λ=12p1\lambda=\tfrac12 p_1. The topological string lift is a lift through the homotopy fiber of BSpin⁡→ λ K(Z,4).B\operatorname{Spin}\xrightarrow{\ \lambda\ } K(\mathbb Z,4). By the universal property of a homotopy fiber, a lift of a fixed spin classifying map is a chosen nullhomotopy of its composite with λ\lambda. Consequently existence is equivalent to vanishing of the integral class on the base. Vanishing of the de Rham image alone leaves a possible torsion obstruction. At the differential level, if the underlying integral class vanishes, the differential character is represented by a three-form HH modulo closed forms of integral periods, and its curvature is dHdH. A geometric differential string structure keeps track of a choice of trivialization and its coherences. Treating it as a nullhomotopy to the zero differential character would impose zero curvature and would miss the three-form in this Bianchi identity.

Fiorenza, Schreiber, and Stasheff construct the global differential refinement of the fractional Pontryagin map at the level of smooth cocycle spaces and describe the corresponding differential trivializations [11, §§5.3 and 6]. Their construction uses integration of the infinitesimal model and compatible periods. The local equations above explain which forms such a refinement must recover. The global characteristic map also retains the integral class and the higher automorphisms that the form equations leave undetermined.

The string obstruction and its differential lifts

A spin connection has a curvature polynomial of degree four. The same polynomial represents a distinguished integral class, whose vanishing is a lifting problem. A primitive of the curvature form can exist even when the integral obstruction survives. Conversely an integral lift can exist while a particular differential refinement is nonzero. Keeping track of the spaces in which those conditions live separates the choices.

The topological fiber

Let E→ME\to M be an oriented Euclidean real vector bundle of rank n≥5n\geq 5 over a smooth manifold, with a specified spin lift PP of its oriented orthonormal frame bundle. A spin lift supplies a classifying map cP:M→BSpin⁡(n)c_P:M\to B\operatorname{Spin}(n), where the domain denotes the underlying homotopy type. The universal integral class λ∈H4(BSpin⁡(n);Z)\lambda\in H^4(B\operatorname{Spin}(n);\mathbb Z) satisfies 2λ=p12\lambda=p_1 pulled back along BSpin⁡(n)→BSO⁡(n)B\operatorname{Spin}(n)\to B\operatorname{SO}(n). We use this universal class to define the half-Pontryagin class of PP. An equation 2x=p1(E)2x=p_1(E) in the cohomology of a given MM would not define it uniquely when H4(M;Z)H^4(M;\mathbb Z) has two-torsion. The integral assertion is part of the spin classification [21, Section 5].

Represent λ\lambda by a map to K(Z,4)K(\mathbb Z,4). Define the topological classifying space for string lifts by the homotopy pullback image The right-hand arrow chooses the zero class. A lift is taken over cPc_P, so the choice includes the comparison with that specified spin bundle.

Proposition 1.1 (Lifts of a fixed spin bundle). The space of lifts of cPc_P is nonempty precisely when λ(P)=0\lambda(P)=0. When it is nonempty, its components form a torsor for H3(M;Z)H^3(M;\mathbb Z). After choosing a lift as basepoint, its homotopy groups in degrees 11 through 33 are H3−j(M;Z)H^{3-j}(M;\mathbb Z) in degree jj, and the higher groups vanish.

Proof. Mapping out of MM preserves limits in the category of spaces. The lift space is therefore the space of paths from λcP\lambda c_P to the constant map inside Map⁡(M,K(Z,4))\operatorname{Map}(M,K(\mathbb Z,4)). The description of connected components by cohomology gives the obstruction. When a path exists, composing with it identifies every other path with a loop at the constant map, up to coherent homotopy. Looping commutes with taking this mapping space, giving Map⁡(M,K(Z,3))\operatorname{Map}(M,K(\mathbb Z,3)). Its components are H3(M;Z)H^3(M;\mathbb Z), and the action by path composition is simply transitive on components. For j≥1j\geq 1, a based jj-sphere of maps to K(Z,3)K(\mathbb Z,3) has adjoint a map from Sj∧M+S^j\wedge M_+ to K(Z,3)K(\mathbb Z,3), where M+M_+ adds a disjoint basepoint. The suspension isomorphism gives the formula. ◻

The proof distinguishes classification from existence. For instance, the trivial spin bundle over S3S^3 has infinitely many lift classes, indexed by Z\mathbb Z after choosing one of them. Its spin curvature may be identically zero throughout this family. The same proof yields automorphisms and higher automorphisms, since a choice of lift class identifies a component without eliminating its higher homotopy.

Normalization and refinement

Take the curvature F=dA+A∧AF=dA+A\wedge A of a spin connection in the real defining representation of so(n)\mathfrak{so}(n). In our Chern normalization, the real form for λ\lambda is λ(F)=−116π2tr⁡(F∧F).\lambda(F)=-\frac{1}{16\pi^2}\operatorname{tr}(F\wedge F). Indeed the degree-four part of det⁡(1+iF/(2π))\det(1+iF/(2\pi)) for the complexification of EE is tr⁡(F∧F)/(8π2)\operatorname{tr}(F\wedge F)/(8\pi^2), since tr⁡F=0\operatorname{tr}F=0. The relation p1(E)=−c2(E⊗C)p_1(E)=-c_2(E\otimes\mathbb C) supplies the minus sign. Trace invariance and the Bianchi identity make this form closed and independent of the frame. The universal integral normalization specifies which refinement is meant.

The differential characteristic class λ^(P,A)\widehat\lambda(P,A) has integral class λ(P)\lambda(P) and curvature λ(F)\lambda(F). It can be realized by the Chern–Simons bundle 22-gerbe of the spin bundle, equipped with its induced connection [22, Theorems 1.1.3 and 1.3.3]. A geometric string structure in this model consists of a trivialization of that bundle 22-gerbe and compatible connective data. Such data produce a 33-form HH with dH=λ(F)dH=\lambda(F). Tensoring the trivialization by a bundle gerbe with connection adds that gerbe’s curvature to HH. Waldorf’s existence theorem gives compatible connections on every topological trivialization [22, Theorem 1.3.4 and Section 3.2].

The exact sequence of differential cohomology makes the role of HH explicit. Write ΩZ3(M)\Omega^3_{\mathbb Z}(M) for closed forms with integral periods. We have Ω3(M)/ΩZ3(M)→ a H^4(M;Z)→ I H4(M;Z)⟶0.\Omega^3(M)/\Omega^3_{\mathbb Z}(M) \xrightarrow{\ a\ } \widehat H^4(M;\mathbb Z) \xrightarrow{\ I\ } H^4(M;\mathbb Z)\longrightarrow 0 . Here a(H)a(H) evaluates on a 33-cycle by integration modulo Z\mathbb Z. If λ(P)=0\lambda(P)=0, the refinement has the form a(H)a(H). Its curvature is dHdH, which may be nonzero. Changing HH by a closed form of integral periods keeps a(H)a(H) fixed. The exact sequence describes the image of the differential class as a form modulo periods. A specified trivialization includes gluing data beyond that image. In particular the requirement λ^(P,A)=0\widehat\lambda(P,A)=0 is stronger than existence of a geometric string structure, which allows a topologically trivial 22-gerbe carrying a nonzero global curving.

The difference of two such compatible trivializations, for the same spin connection, is a bundle gerbe with connection. One way to see this is to compare their local descent cochains. Their coboundaries represent the same obstruction cocycle, so the quotient of their data satisfies the unobstructed gerbe descent equations. Conversely multiplication by such a cocycle changes the trivialization and preserves its relation to the obstruction cocycle. On isomorphism classes this is the torsor action of H^3(M;Z)\widehat H^3(M;\mathbb Z). Taking underlying integral classes recovers the H3H^3-action in Proposition 1.1, and taking curvature recovers addition to HH. In the 22-gerbe description these operations are tensor product and inverse of gerbes with connection [22, Corollary 1.3.2 and Section 3.2].

Nonzero integral and differential obstructions

The quaternionic Hopf bundle QQ on S4=HP1S^4=\mathbb H P^1 has complex rank two and structure group Sp⁡(1)=SU⁡(2)\operatorname{Sp}(1)=\operatorname{SU}(2). Its unit sphere bundle is S7→S4S^7\to S^4, so the Gysin sequence identifies its Euler class with a generator of H4(S4;Z)H^4(S^4;\mathbb Z), up to orientation. This Euler class is c2(Q)c_2(Q) for the complex orientation. The real five-plane bundle QR⊕RQ_{\mathbb R}\oplus\mathbb R has a spin lift. The group SU⁡(2)\operatorname{SU}(2) is simply connected, so its orthogonal representation on H⊕R\mathbb H\oplus\mathbb R lifts to Spin⁡(5)\operatorname{Spin}(5). For a complex two-plane bundle, p1(QR)=c1(Q)2−2c2(Q)=−2c2(Q)p_1(Q_{\mathbb R})=c_1(Q)^2-2c_2(Q)=-2c_2(Q), as follows from the formal Chern roots. Here H4(S4;Z)H^4(S^4;\mathbb Z) is torsion-free, so the spin λ\lambda is −c2(Q)-c_2(Q). Proposition 1.1 then obstructs a string structure.

There is also a flat model with a torsion obstruction. Let Y=S5/μmY=S^5/\mu_m, where the mm-th roots of unity act by common scalar multiplication on the unit sphere in C3\mathbb C^3 and m≥2m\geq 2. The map p:Y→CP2p:Y\to\mathbb C P^2 is the unit circle bundle of L⊗mL^{\otimes m}, with LL the tautological line. Indeed the fiberwise tensor-power map on the unit circle bundle of LL is surjective with kernel μm\mu_m. Put x=c1(L)x=c_1(L). The Gysin sequence for pp contains H2(CP2;Z)→ ⌣mx H4(CP2;Z)→ p∗ H4(Y;Z)⟶0.H^2(\mathbb C P^2;\mathbb Z)\xrightarrow{\ \smile mx\ } H^4(\mathbb C P^2;\mathbb Z)\xrightarrow{\ p^*\ } H^4(Y;\mathbb Z)\longrightarrow 0 . The final zero is H3(CP2;Z)H^3(\mathbb C P^2;\mathbb Z), and the first two groups are generated by xx and x2x^2. It follows that u=p∗(x2)u=p^*(x^2) has order exactly mm. The line p∗Lp^*L admits a flat connection because its pullback to S5S^5 has the unit frame q↦qq\mapsto q, and the deck group changes this frame by constant phases. The rank-two bundle Q=p∗L⊕(p∗L)−1Q=p^*L\oplus(p^*L)^{-1} carries the direct sum of the flat connection and its dual. Its determinant reduction has structure group SU⁡(2)\operatorname{SU}(2) and c2(Q)=−uc_2(Q)=-u. For the associated spin five-plane bundle the identity λ=−c2\lambda=-c_2 follows by naturality from BSU⁡(2)B\operatorname{SU}(2), whose integral H4H^4 is free, by the complexification computation above. Thus λ=u\lambda=u while F=0F=0. This obstruction comes entirely from the integral gluing data.

For the other direction take the trivial spin bundle on T3=(R/Z)3T^3=(\mathbb R/\mathbb Z)^3 and the connection A=tXαA=tX\alpha in a global frame. Here XX is a nonzero real skew-symmetric matrix from spin(n)\mathfrak{spin}(n), t∈Rt\in\mathbb R, and α=cos⁡(2πz) dx+sin⁡(2πz) dy.\alpha=\cos(2\pi z)\,dx+\sin(2\pi z)\,dy. The transgression from the trivial connection to AA is −t2tr⁡(X2)α∧dα/(16π2)-t^2\operatorname{tr}(X^2)\alpha\wedge d\alpha/(16\pi^2). This follows by inserting sAsA into the curvature polynomial and integrating its derivative for 0≤s≤10\leq s\leq 1. The cubic term vanishes because AA has just one form factor. Direct differentiation gives α∧dα=−2π dx∧dy∧dz\alpha\wedge d\alpha=-2\pi\,dx\wedge dy\wedge dz, so the transgression integral is a nonzero multiple of t2t^2. The differential characteristic class changes by aa of this transgression. For small nonzero tt the resulting degree-four differential character has nonzero value on T3T^3, while its integral class and its 44-form curvature both vanish. It records the comparison of two spin connections on the same topologically trivial bundle. A geometric trivialization can retain that comparison as a curving.

Connections on quotient stacks

A group action carries information at each point about its stabilizer. Taking an orbit space discards those groups. A quotient stack keeps them as automorphisms, and characteristic classes can detect representations of those automorphisms even when the base orbit space is a point.

The action groupoid and its test objects

Let a Lie group KK act smoothly from the left on XX. The action groupoid has objects x∈Xx\in X and arrows x→kxx\to kx for k∈Kk\in K. Given arrows labelled kk and ll in that order, their composite is labelled lklk. Its nerve in degree qq is Kq×XK^q\times X. A sequence of composable arrows is determined by its initial point and its qq group elements, and the inner face maps multiply adjacent elements in their composition order. The outer faces forget the first or last vertex. In particular the first face also moves the initial point.

The quotient stack [X/K][X/K] is described on a test manifold TT by pairs (P,u)(P,u) with P→TP\to T a right principal KK-bundle and u:P→Xu:P\to X a smooth map satisfying u(pk)=k−1u(p)u(pk)=k^{-1}u(p). Arrows are principal bundle isomorphisms commuting with the maps to XX. A frame pi:Ti→Pp_i:T_i\to P gives a map xi:Ti→Xx_i:T_i\to X. On TijT_{ij} write pj=pikijp_j=p_i k_{ij}, so xj=kij−1xix_j=k_{ij}^{-1}x_i. On triples kijkjk=kikk_{ij}k_{jk}=k_{ik}. Conversely this local data glue the trivial principal bundles by their transition functions, and equivariance of the local maps makes them glue to uu. The same argument on isomorphisms identifies the arrows. This exhibits the passage from the action groupoid to its stack as descent for locally trivial bundles.

Taking KK trivial recovers XX. Taking X=∗X=* gives BKBK, because the map uu is then forced. For a free proper action with smooth quotient and principal bundle X→X/KX\to X/K, this stack is represented by X/KX/K. Indeed the composition P→X→X/KP\to X\to X/K descends to TT, and the resulting map to X/KX/K makes PP isomorphic to its pullback of XX, with the right action on XX obtained by inversion. This follows fiberwise since an equivariant map between torsors is an isomorphism, and local sections give smoothness of the inverse. In general a test map ∗→[X/K]*\to[X/K] determined by x∈Xx\in X has automorphisms identified with its stabilizer. The projection to the orbit set sends each such automorphism to the identity.

Bundles and the condition on a connection

Let GG be another Lie group. A principal GG-bundle on [X/K][X/K] amounts to a KK-equivariant principal GG-bundle Q→XQ\to X. Pullback to the atlas X→[X/K]X\to[X/K] gives QQ. The chosen descent isomorphism on K×XK\times X acts in the fiber over xx by Qx→QkxQ_x\to Q_{kx}. The cocycle condition on K2×XK^2\times X says this action respects multiplication. Conversely these isomorphisms give the bundle over any test (P,u)(P,u), namely (u∗Q)/K→T(u^*Q)/K\to T. The right action on u∗Qu^*Q sends (p,q)(p,q) to (pk,k−1q)(pk,k^{-1}q). Its quotient is a principal GG-bundle, as can be checked after a local trivialization of PP. This also proves the assertion for morphisms.

A connection form ω\omega on QQ descends with equality on the atlas nerve exactly when b∗ω=pr⁡2∗ω,b:K×Q⟶Q,b(k,q)=kq.b^*\omega=\operatorname{pr}_2^*\omega,\qquad b:K\times Q\longrightarrow Q,\quad b(k,q)=kq. Evaluating on the QQ-directions at each fixed kk requires KK-invariance. Evaluating on vectors tangent to KK at the identity requires ω(ξQ)=0\omega(\xi_Q)=0, where ξQ(q)=ddt∣0exp⁡(tξ)q\xi_Q(q)=\frac{d}{dt}|_0\exp(t\xi)q. Those conditions imply the equality on all tangent directions by their behavior under translations. Thus a KK-invariant connection can fail this descent condition. An invariant connection still has a role in computing equivariant characteristic forms. The groupoid formulation and the comparison with the usual equivariant Chern–Weil map are developed in [23].

For a KK-invariant connection write μξ=ω(ξQ)\mu_\xi=\omega(\xi_Q), an equivariant g\mathfrak g-valued function. If FF is the ordinary curvature and Dω=d+[ω,−]D_\omega=d+[\omega,-], then ιξQF=−Dωμξ.\iota_{\xi_Q}F=-D_\omega\mu_\xi . In fact Cartan’s formula gives ιξQdω=−dμξ\iota_{\xi_Q}d\omega=-d\mu_\xi, and contraction of 12[ω,ω]\tfrac12[\omega,\omega] is [μξ,ω][\mu_\xi,\omega]. Consequently F−μξF-\mu_\xi satisfies (Dω−ιξQ)(F−μξ)=0(D_\omega-\iota_{\xi_Q})(F-\mu_\xi)=0, using the Bianchi identity and the fact that contraction of a function vanishes. For an invariant polynomial, evaluating on this equivariant curvature gives a closed expression in the Cartan complex with differential d−ιξXd-\iota_{\xi_X}. The calculation comes from the derivation rule for the invariant polynomial, which kills the adjoint-bracket terms. The ξ\xi-polynomial degree is counted as twice its ordinary degree. For compact KK the Cartan comparison computes the real cohomology of the Borel construction, or equivalently the simplicial de Rham model of the action groupoid. This use of an equivariant complex differs from the complex of invariant horizontal forms on XX.

An example is X=∗X=*, K=U(1)K=U(1), and Q=U(1)Q=U(1) with left action kq=krqkq=k^rq for some integer rr. The connection q−1dqq^{-1}dq is invariant, but μξ=irξ\mu_\xi=ir\xi for the generator parameterization exp⁡(itξ)\exp(i t\xi). The ordinary curvature is zero. The equivariant Chern polynomial in degree two gives rξ/(2π)r\xi/(2\pi). Thus the real first Chern class over BU(1)BU(1) can be nonzero, whereas the positive-degree ordinary forms on the atlas point vanish.

A flat torsion line on a stack

Let Γ=Z/m\Gamma=\mathbb Z/m with m≥2m\geq 2. A complex line on BΓB\Gamma is supplied by the unitary character χ(j)=exp⁡(−2πij/m)\chi(j)=\exp(-2\pi i j/m). A trivial connection on the line over the point is preserved by each constant unitary isomorphism. Since Γ\Gamma is discrete, the horizontal condition has no infinitesimal group directions to impose, so this is a flat line with connection on the stack. Its holonomy on a generator of the stabilizer is nontrivial.

Compute its integral class using 0⟶Z⟶R→ t↦e2πit U(1)⟶0.0\longrightarrow\mathbb Z\longrightarrow\mathbb R \xrightarrow{\ t\mapsto e^{2\pi i t}\ } U(1) \longrightarrow 0 . For homogeneous real group cochains in positive degree, averaging over the first group entry gives a contracting homotopy: insert a first entry hh, sum over hh, and divide by mm. The alternating deletion differential and this homotopy have anticommutator the identity in positive degree because the term deleting hh is the original cochain and the other terms cancel in pairs. Invariance is preserved by reindexing the average. Thus Hq(BΓ;R)=0H^q(B\Gamma;\mathbb R)=0 for q>0q>0. The connecting homomorphism gives an isomorphism Hom⁡(Γ,U(1))=H1(BΓ;U(1))→ ∼ H2(BΓ;Z),\operatorname{Hom}(\Gamma,U(1)) =H^1(B\Gamma;U(1))\xrightarrow{\ \sim\ }H^2(B\Gamma;\mathbb Z), where the coefficients on the left are constant coefficients on the realized groupoid. Lifting the character to the real cochain j↦−j/mj\mapsto -j/m for 0≤j<m0\le j<m gives the integer cocycle (j,k)↦−⌊(j+k)/m⌋(j,k)\mapsto-\lfloor(j+k)/m\rfloor. This is the Chern cocycle in the chosen exponential normalization. The character has order mm, so its first Chern class generates a cyclic group of order mm. The curvature and the coarse space are both unable to show this class. The stack retains the discrete descent arrows, and the differential class retains their phases.

The same mechanism occurs on a quotient with a nontrivial atlas. Let Γ=Z/m\Gamma=\mathbb Z/m rotate CP1\mathbb C P^1 by j⋅[a:b]=[a:e2πij/mb]j\cdot[a:b]=[a:e^{2\pi i j/m}b]. The corresponding unitary diagonal matrices act on the ambient C2\mathbb C^2 and carry each tautological line to the tautological line at its image. The projection connection of Chapter [ch:synthesis] is equivariant because these matrices preserve both the trivial differential and orthogonal projection. Since the group is finite, this is a descent connection on [CP1/Γ][\mathbb C P^1/\Gamma]. At [1:0][1:0] the action on the fiber is trivial, and at [0:1][0:1] the action is the inverse of χ\chi above. Each fixed point defines a map BΓ→[CP1/Γ]B\Gamma\to[\mathbb C P^1/\Gamma]. The restrictions of the equivariant first Chern class along those two maps are respectively zero and a generator of H2(BΓ;Z)H^2(B\Gamma;\mathbb Z). Its restriction to the atlas is the first Chern class of the tautological line on the sphere, which has Chern number −1-1. These are three different tests of the same characteristic class. If the tautological equivariant bundle were pulled back from the orbit space, each stabilizer would act trivially on its fiber, because the pullback from a space acts on fibers by the identity along arrows mapping to identities. The action at the second fixed point disproves that descent to the coarse orbit space.

Derived intersections and Atiyah classes

The first infinitesimal neighborhood of a diagonal gives a local description of connections. A derived intersection of that diagonal with itself retains the relations among its equations. This chapter works in smooth algebraic geometry over a characteristic-zero field kk, using the algebraic differential dd. Smooth real connections and algebraic connections have different existence properties. The comparison rests on the common Leibniz identity.

Intersecting schemes through a derived tensor product

For affine schemes mapping to Spec⁡R\operatorname{Spec} R, let their coordinate algebras be BB and CC. The ordinary pullback has coordinate algebra B⊗RCB\otimes_R C. In derived algebraic geometry one uses B⊗RLCB\otimes_R^{\mathbf L}C, the homotopy pushout of the corresponding commutative algebras. We use cohomological degrees, so H−j(B⊗RLC)=Tor⁡jR(B,C),j≥0.H^{-j}(B\otimes_R^{\mathbf L}C) =\operatorname{Tor}^R_j(B,C),\qquad j\geq 0. The homotopy pushout can be modeled by commutative differential graded algebras in nonpositive degrees in characteristic zero [24, Section 2]. The formula is understood locally on schemes when the rings vary from chart to chart.

Consider two embeddings of the line {y=0}\{y=0\} in Ak2\mathbb A^2_k, so R=k[x,y]R=k[x,y] and B=C=R/(y)B=C=R/(y). The algebra obtained by adjoining one odd generator ee of degree −1-1 to RR with de=yde=y resolves BB. Indeed its underlying complex is R→yRR\xrightarrow{y}R, with zero kernel and cokernel BB. As a free extension in graded commutative algebras this is a Koszul model for the derived base change. Tensoring with CC gives a differential of zero, and the resulting algebra is k[x]⊕k[x]ek[x]\oplus k[x]e, with e2=0e^2=0. The ordinary intersection is just the line, while H−1H^{-1} records a normal equation that occurs on both sides.

For a different second embedding take the parabola y=x2y=x^2. On R/(y−x2)≅k[x]R/(y-x^2)\cong k[x], the same differential is multiplication by x2x^2, hence it is injective. The degree-zero algebra is k[x]/(x2)k[x]/(x^2), with no higher Tor. Here failure of transversality appears as a nonreduced ordinary intersection. The two examples distinguish nilpotents in degree zero from negative-degree homology.

The diagonal and its self-intersection

Let XX be smooth, separated and of finite type over kk. Its diagonal Δ:X→X×kX\Delta:X\to X\times_k X is a regular closed immersion. Writing I\mathcal I for its ideal, the conormal I/I2\mathcal I/\mathcal I^2 is ΩX/k1\Omega^1_{X/k}. The identification takes the class of p2∗f−p1∗fp_2^*f-p_1^*f to dfdf. Indeed modulo I2\mathcal I^2 the difference obeys Leibniz, and on an étale coordinate neighborhood the coordinate differences generate the conormal freely. The same statement also follows from the universal derivation of a commutative algebra by applying its defining property to the multiplication map.

Proposition 1.1 (Tor sheaves along the diagonal). The homology sheaves on the derived self-intersection X×X×kXhXX\times^{h}_{X\times_k X}X satisfy H−j(OX×X×kXhX)≅⋀jΩX/k1.\mathcal H^{-j}(\mathcal O_{X\times^{h}_{X\times_k X}X}) \cong \bigwedge\nolimits^j\Omega^1_{X/k}. These identifications respect the graded homology algebra.

Proof. Take local generators f1,…,fdf_1,\ldots,f_d of the diagonal ideal forming a regular sequence. The Koszul algebra adjoins odd generators eie_i with dei=fide_i=f_i. It resolves the diagonal, as shown by induction on the length of the sequence. Passing from d−1d-1 to dd takes the cone of multiplication by fdf_d on the preceding Koszul complex, with the shift that puts the new exterior generator in degree −1-1. The preceding homology is the quotient by the first d−1d-1 generators. The regular-sequence hypothesis makes the multiplication injective on that quotient, so the cone has only the requisite degree-zero homology. This completes the induction. Pulling back along the diagonal kills the fif_i, leaving the exterior algebra with zero differential. Under a change of generators, the induced map in degree −1-1 is the change-of-basis map of I/I2\mathcal I/\mathcal I^2. Multiplication extends the identification to the exterior powers and makes it independent of the local resolution. ◻

The loop-space expression here is categorical. Cotensoring by S1≃∗∐S0h∗S^1\simeq *\coprod^h_{S^0}* turns the homotopy pushout into the pullback of the two diagonal maps. A degree-zero scheme of points would miss the derived directions of this pullback. The global Hochschild–Kostant–Rosenberg comparison for the smooth diagonal gives the corresponding splitting as a derived object. For other embeddings a local Koszul model by itself only determines the homology sheaves. The extension condition for a global splitting is treated in [26, Theorem 1].

First jets and the obstruction to a connection

For a vector bundle EE on XX, let ZZ be the first infinitesimal neighborhood of the diagonal, with OZ=OX×X/I2\mathcal O_Z=\mathcal O_{X\times X}/\mathcal I^2. The two projections of ZZ to XX are still distinct. Define J1E=p1∗p2∗EJ^1E=p_{1*}p_2^*E, regarded as an OX\mathcal O_X-module by the first projection. There is a sequence 0⟶ΩX/k1⊗E⟶J1E⟶E⟶0.0\longrightarrow\Omega^1_{X/k}\otimes E \longrightarrow J^1E\longrightarrow E\longrightarrow 0. After trivializing EE, this is the exact sequence coming from reduction OZ→OΔ\mathcal O_Z\to\mathcal O_\Delta; the conormal module multiplies by coefficients restricted to the diagonal. Thus the sequence is exact independently of the chosen frame. Its extension class At⁡(E)∈Ext⁡OX1(E,ΩX/k1⊗E)\operatorname{At}(E)\in \operatorname{Ext}^1_{\mathcal O_X}(E,\Omega^1_{X/k}\otimes E) uses the first jet convention. Atiyah’s obstruction construction treats the holomorphic counterpart [25, Sections 2–4].

The jet operator jj pulls a section along p2p_2. On OZ\mathcal O_Z, multiplication by p2∗fp_2^*f differs from multiplication by p1∗fp_1^*f by dfdf in the conormal ideal. Thus j(fs)=fj(s)+df⊗sj(fs)=fj(s)+df\otimes s. If σ:E→J1E\sigma:E\to J^1E is an OX\mathcal O_X-linear splitting, j−σj-\sigma takes values in the kernel and satisfies the connection Leibniz rule. Conversely a connection ∇\nabla gives the splitting j−∇j-\nabla. The extension class vanishes exactly when a connection of this kind exists. Its vanishing alone makes no integrability assertion. Flatness imposes the separate curvature equation.

Here is a coordinate account of the class. Take local frames with ej=eigije_j=e_i g_{ij} and in frame ii take the connection ∇(i)\nabla^{(i)} for which eie_i is parallel. Write σi=j−∇(i)\sigma_i=j-\nabla^{(i)}. On a double overlap σj−σi=∇(i)−∇(j).\sigma_j-\sigma_i=\nabla^{(i)}-\nabla^{(j)}. In the jj-frame this difference is gij−1dgijg_{ij}^{-1}dg_{ij}. The overlap differences form a Čech cocycle of Ω1⊗End⁡E\Omega^1\otimes\operatorname{End}E, and changing the splittings adds a coboundary. This is the same inhomogeneous term that enters the connection descent law.

Proposition 1.2 (Tensor products and determinant). With the jet convention above, the class of E⊗GE\otimes G is the sum of At⁡(E)⊗1G\operatorname{At}(E)\otimes 1_G and 1E⊗At⁡(G)1_E\otimes\operatorname{At}(G), with the tensor factors reordered so that both take values in Ω1⊗E⊗G\Omega^1\otimes E\otimes G. The class of det⁡E\det E is the trace of At⁡(E)\operatorname{At}(E).

Proof. On a common trivializing cover, the tensor connection built from the chosen local connections satisfies the Leibniz rule on s⊗ts\otimes t by differentiating each factor in turn. Taking the difference between the two local tensor connections on an overlap leaves the sum of the differences in each input, since all terms are OX\mathcal O_X-linear after subtraction. This proves the tensor formula on the extension cocycles. For the determinant use the local connection induced on the top exterior power. In a wedge of a frame, replacing its ii-th vector by a multiple of any other frame element gives zero, so only the diagonal entries of the difference matrix survive. Their sum gives the trace formula. ◻

On Pk1\mathbb P^1_k, the tautological bundle has frames (1,z)(1,z) and (w,1)(w,1) with w=1/zw=1/z, giving transition g=z−1g=z^{-1}. Its jet cocycle is −dz/z-dz/z. In the two standard charts, H1(P1,Ω1)=k[z,z−1] dzk[z] dz+z−2k[z−1] dz.H^1(\mathbb P^1,\Omega^1) =\frac{k[z,z^{-1}]\,dz} {k[z]\,dz+z^{-2}k[z^{-1}]\,dz}. This quotient follows by rewriting regular forms in the ww-chart using dw=−z−2dzdw=-z^{-2}dz. The class of dz/zdz/z survives, so the tautological bundle has no algebraic connection. Over C\mathbb C the same underlying smooth line has a Hermitian connection, computed explicitly in Chapter [ch:synthesis]. The change of coefficient sheaf accounts for the different answers.

From a Hopf connection to surface holonomy

The Hopf fibration gives a sequence of explicit geometric models. Its line bundle over the two-sphere has a curvature form with nonzero period. On the total space this form admits a global primitive. The primitive and the curvature together give an integral 33-form, leading to surface holonomy on the total space and ordinary line holonomy on its loop space.

The tautological connection and its periods

Let L→CP1L\to\mathbb C P^1 be the tautological Hermitian line. It is a subbundle of CP1×C2\mathbb C P^1\times\mathbb C^2. Projecting the trivial differential orthogonally to LL gives a metric connection. For a section ss, define ∇s=πLds\nabla s=\pi_L ds. The Leibniz identity follows from πLs=s\pi_Ls=s, and metric compatibility follows because the perpendicular component of dsds pairs to zero with any section of LL.

In the affine zz-chart use the unit frame s0(z)=(1,z)1+∣z∣2.s_0(z)=\frac{(1,z)}{\sqrt{1+|z|^2}}. The connection matrix is A0=s0∗ds0A_0=s_0^*ds_0, namely A0=zˉ dz−z dzˉ2(1+∣z∣2)=i r21+r2 dθ,z=reiθ.A_0=\frac{\bar z\,dz-z\,d\bar z}{2(1+|z|^2)} =i\,\frac{r^2}{1+r^2}\,d\theta , \qquad z=re^{i\theta}. The polar expression holds away from the origin. The Cartesian formula is smooth there. On the chart w=1/zw=1/z take s1=(w,1)/1+∣w∣2s_1=(w,1)/\sqrt{1+|w|^2}. On the overlap s1=s0e−iθs_1=s_0 e^{-i\theta}, so the two matrices obey A1=A0−i dθA_1=A_0-i\,d\theta. This is the transformation law from Chapter [ch:principal].

Set ω=iF/(2π)\omega=iF/(2\pi). On the zz-chart, ω=−1πr(1+r2)2 dr∧dθ,∫CP1ω=−1.\omega=-\frac{1}{\pi}\frac{r}{(1+r^2)^2}\,dr\wedge d\theta, \qquad \int_{\mathbb C P^1}\omega=-1. For the integral one first integrates over r<Rr<R and sends RR to infinity. The missing point has measure zero for the smooth form in the other chart. With the complex orientation this gives the first Chern number of LL [1, 5]. Its sign also follows by using Stokes on the two chart disks, since A0−A1=i dθA_0-A_1=i\,d\theta on their common boundary oriented from the zz-disk.

On a positively oriented zz-circle of radius RR, the parallel transport equation for the coefficient in the frame is v˙=−A0(γ˙)v\dot v=-A_0(\dot\gamma)v, so Hol⁡L(γ)=exp⁡ ⁣(−2πiR21+R2).\operatorname{Hol}_L(\gamma)= \exp\!\left(-2\pi i\frac{R^2}{1+R^2}\right). The integral of ω\omega over the disk it bounds is −R2/(1+R2)-R^2/(1+R^2). Choosing the complementary filling with the same boundary changes that integral by an integer. The exponent therefore comes from a differential character with curvature ω\omega, not just from a choice of potential on the chart.

An integral 3-form on the total space

The unit circle bundle p:S3→CP1p:S^3\to\mathbb C P^1 of LL has global real 11-form α=q∗dq/(2πi)\alpha=q^*dq/(2\pi i) for a point q∈S3q\in S^3 viewed as a unit complex column vector. Its value on a positively oriented fiber integrates to one. In the frame q=s0(z)eiϕq=s_0(z)e^{i\phi} the form is dϕ/(2π)+A0/(2πi)d\phi/(2\pi)+A_0/(2\pi i). Consequently dα=−p∗ω,η=α∧p∗(−ω),∫S3η=1.d\alpha=-p^*\omega,\qquad \eta=\alpha\wedge p^*(-\omega),\qquad \int_{S^3}\eta=1. The orientation here uses the positive circle and complex base orientations. The integral follows by fiber integration: on each trivialization only the dϕd\phi term contributes, and the local base integrals glue by a partition of unity. Also dη=p∗((−ω)∧(−ω))=0d\eta=p^*((-\omega)\wedge(-\omega))=0. Thus η\eta represents the integral generator of H3(S3;Z)H^3(S^3;\mathbb Z) in real cohomology.

One can construct the associated degree-three differential character directly. A smooth singular 22-cycle zz in S3S^3 bounds a smooth singular 33-chain DD. Put h(z)=∫Dη mod Zh(z)=\int_D\eta\bmod\mathbb Z. The difference between two choices of DD is a 33-cycle, and integrality of η\eta gives independence. Linearity follows by adding fillings. Taking z=∂Dz=\partial D gives exactly the character boundary identity with curvature η\eta. The integral class of hh maps to the de Rham class of η\eta; here the degree-three cohomology is free, so this class is the generator. This character can also be represented by a bundle gerbe with connection.

To see the overlap geometry, cover S3S^3 by the complements of the two poles, with local 22-forms BN,BSB_N,B_S satisfying dBN=dBS=ηdB_N=dB_S=\eta. Orient the equator as the boundary of the northern hemisphere. Stokes gives ∫S2(BN−BS)=1\int_{S^2}(B_N-B_S)=1. On the overlap this difference is a closed 22-form of integral periods because the equator generates its H2H_2. It can be realized as a line curvature there by the differential cohomology curvature sequence. For the convention in which the overlap curvature is BS−BNB_S-B_N, take the dual line. Use the reverse overlap line as its dual and the identity lines on repeated charts. Evaluation of the line against its dual supplies the product on alternating overlaps. Associativity reduces to the tensor identities for a line and its dual, which can be checked in a unit frame. This is the connective descent presentation whose curvature is η\eta. The difference of local 22-forms carries the same integer that the original Hopf curvature detected.

Evaluating families of loops

Let LS3LS^3 consist of smooth loops, with a smooth family defined by its adjoint smooth map to S3S^3. For a smooth singular 11-cycle cc of families of loops, write ev⁡:LS3×S1→S3\operatorname{ev}:LS^3\times S^1\to S^3 for evaluation. Then the transgressed holonomy of cc is (τh)(c)=h(ev⁡∗(c×[S1])).(\tau h)(c)=h\big(\operatorname{ev}_*(c\times[S^1])\big). A smooth singular fundamental cycle of S1S^1 is used in this expression. Replacing it by a homologous cycle changes the argument by a boundary evaluated through the one-dimensional circle factor. The lifted 33-chain has rank at most two under evaluation, so its curvature integral vanishes. This proves independence of that choice.

For a smooth 22-chain CC of loop families, the boundary of C×[S1]C\times[S^1] is ∂C×[S1]\partial C\times[S^1]. The character property gives (τh)(∂C)=∫Cτη mod Z,(τη)γ(V,W)=∫01ηγ(t)(γ˙(t),V(t),W(t)) dt.(\tau h)(\partial C)= \int_C\tau\eta\bmod\mathbb Z,\qquad (\tau\eta)_\gamma(V,W) =\int_0^1\eta_{\gamma(t)}(\dot\gamma(t),V(t),W(t))\,dt . In the orientation of C×S1C\times S^1, moving the loop tangent from the final slot to the first makes two interchanges, explaining the sign. The gerbe-to-loop-space transgression theorem realizes this degree-two character as a line bundle with connection. Waldorf’s convention uses inverse surface holonomy [27, Proposition 4.3.2]. The dual of his transgressed bundle realizes the positive evaluation convention above. Its fusion structure gives compatible products along triples of paths with common endpoints. The conditions on this structure for the inverse regression theorem are given in [27, Theorem A].

More explicitly, for paths γi\gamma_i from xx to yy that are constant near their endpoints, let ℓij\ell_{ij} traverse γi\gamma_i and then γj\gamma_j backwards. The line fibers TℓijT_{\ell_{ij}} of the transgression admit fusion maps Tℓ12⊗Tℓ23⟶Tℓ13.T_{\ell_{12}}\otimes T_{\ell_{23}}\longrightarrow T_{\ell_{13}} . For four paths the two composites of fusion maps agree. The maps express the cancellation of the common path segment in the surface description. A bare line bundle over the loop space does not supply the fusion maps or their compatibility with transport. The cited theorem uses fusion bundles with connections satisfying the stated superficiality and compatibility conditions, and takes the homotopy category on the gerbe side.

The transgressed curvature has a nonzero integral period. For this, choose an embedded oriented 33-ball in S2×S1S^2\times S^1 avoiding both S2×{t0}S^2\times\{t_0\} and {s0}×S1\{s_0\}\times S^1. Collapse its complement to a point in S3S^3, using a smooth representative f:S2×S1→S3f:S^2\times S^1\to S^3 constant near the boundary of the ball. Such a representative is obtained in polar coordinates by mapping the radius to a polar angle from zero to π\pi, with a smooth constant collar at π\pi and a linear radial behavior near zero. Choosing the orientation of the chart gives degree +1+1. The adjoint family S2→LS3S^2\to LS^3 takes values in based loops since f(s,t0)f(s,t_0) is constant. Its curvature period is ∫S2(adj⁡f)∗(τη)=∫S2×S1f∗η=1.\int_{S^2}(\operatorname{adj}f)^*(\tau\eta) =\int_{S^2\times S^1}f^*\eta=1. Thus this higher holonomy produces a line bundle with nontrivial first Chern class on loop space.

The examples identify the roles of the data at each step. A local matrix measures parallel transport, an integral class controls gluing, and a differential character compares fillings of cycles. Changes of presentation preserve these comparisons. Increasing the dimension of the cycle changes the descent data that carry the holonomy, and transgression converts one such comparison into another by evaluation on a family of loops.

Notation and comparison conventions

Bundles, frames, and curvature

For a principal bundle P→XP\to X the group acts on the right. Local sections are written sis_i, with sj=sigijs_j=s_i g_{ij} on overlaps. Matrix group connection forms transform as Aj=gij−1Aigij+gij−1dgij,Fj=gij−1Figij.A_j=g_{ij}^{-1}A_i g_{ij}+g_{ij}^{-1}dg_{ij}, \qquad F_j=g_{ij}^{-1}F_i g_{ij}. A column representing a section of the associated vector bundle changes by vj=gij−1viv_j=g_{ij}^{-1}v_i for the defining representation. The local covariant derivative on the column is dvi+Aividv_i+A_i v_i. When the Lie algebra-valued form is expressed in matrices, F=dA+A∧AF=dA+A\wedge A. In bracket notation the last term is 12[A,A]\tfrac12[A,A]. On an ordinary tangent bundle we take R(X,Y)Z=∇X∇YZ−∇Y∇XZ−∇[X,Y]Z.R(X,Y)Z=\nabla_X\nabla_Y Z-\nabla_Y\nabla_X Z-\nabla_{[X,Y]}Z. Sectional curvature uses ⟨R(X,Y)Y,X⟩\langle R(X,Y)Y,X\rangle divided by ⟨X,X⟩⟨Y,Y⟩−⟨X,Y⟩2\langle X,X\rangle\langle Y,Y\rangle-\langle X,Y\rangle^2.

For a unitary line A=−2πi aA=-2\pi i\,a, the real first Chern form is iF/(2π)=daiF/(2\pi)=da, and in a local frame parallel transport along a path is multiplication by exp⁡(2πi∫a)\exp(2\pi i\int a). The curvature on a general principal bundle is an adjoint-valued form on the base. Invariant polynomials give scalar differential forms. For a unitary vector bundle the total Chern form is det⁡(I+iF/(2π))\det(I+iF/(2\pi)). Changing the orientation of a real even-rank bundle changes the sign of its Euler class.

Differentials, local cochains, and transport

The Čech differential δ\delta is the alternating sum of the restrictions obtained by omitting an index. In a Čech complex with pp the Čech degree and qq the form degree, the total differential is D=δ+(−1)pdD=\delta+(-1)^p d on bidegree (p,q)(p,q). In degree two, if gij=exp⁡(−2πiλij)g_{ij}=\exp(-2\pi i\lambda_{ij}), then aj−ai=dλija_j-a_i=d\lambda_{ij}, and the integral first Chern cocycle for the positive exponential sequence is −δλ-\delta\lambda. The comparison D(a+λ)=da+δλD(a+\lambda)=da+\delta\lambda verifies the sign in real cohomology. The gerbe chapter specifies its multiplication and cocycle conventions separately.

Manifolds used as smooth tests are Hausdorff, second countable, and finite dimensional. Smooth singular simplices, paths with sitting instants, and fiber integration are considered with the boundary, orientation, and support conventions in the relevant chapters. In the path transport formulas, composing two paths means traversal of the first path followed by the second, with the corresponding linear maps composed in the usual function order.

Moduli and differential coefficients

BG\mathbf BG is the smooth stack of right GG-torsors, and BG∇\mathbf BG_\nabla includes their connections. A smooth test XX is identified with its representable sheaf when this causes no ambiguity. A superscript h{\mathrm h} on a pullback or a limit emphasizes the use of paths and their coherences. The pullback in Chapter [ch:characteristic-maps] uses locally constant coefficient sheaves K(Z,n)K(\mathbb Z,n) and K(R,n)K(\mathbb R,n). In differential cohomology II denotes the underlying integral class and RR denotes curvature. Where RR acts on a vector via a tangent connection it denotes the Riemannian curvature tensor. The section and page links in Appendix [app:index] point to the main treatments.

Authorship.

The manuscript has an anonymous byline. A language model assisted with drafting, mathematical review, typesetting, and revision under the submitter’s direction.

Index of concepts and constructions

Select a page number to open the corresponding treatment.

Concept Construction Page
Concept Construction Page
Adjoint bundle Lie algebra bundle carrying curvature and covariant differentiation. ,
Associated bundle Turns a torsor and a representation into a vector bundle. ,
Atiyah class The extension class of the first jet sequence.
Basic form Horizontal invariant form that descends to the base. ,
Bianchi identity Covariant differential of curvature and its higher analogue. ,
Cartan model Equivariant forms and the contraction term in their differential.
Characteristic form Closed form obtained by evaluating an invariant polynomial on curvature. ,
Čech descent Local objects, transition maps, and higher comparisons. ,
Chern classes Classes represented over the reals by the normalized determinant form. ,
Chern–Simons form Transgresses the difference of characteristic forms.
Christoffel symbols Local coefficients of the affine connection.
Classifying stack BG\mathbf BG for torsors, BG∇\mathbf BG_\nabla for torsors with connection. ,
Crossed module Two groups, an action, and Peiffer identities presenting a strict 22-group.
Curvature Obstruction detected by the square of the covariant differential and small loops. ,
Curving Local two-form involved in a gerbe connection.
Derived intersection Homotopy pullback computed locally by a derived tensor product.
Differential character Evaluates smooth cycles with a Stokes law for its curvature.
Differential refinement An integral class, a form, and a comparison of their real images. ,
Dixmier–Douady class Underlying degree-three characteristic class of a circle gerbe.
Euler class Top class of an oriented real bundle, with a Pfaffian form in even rank.
Fake curvature Difference between ordinary curvature and the boundary of the two-form in a strict two-connection.
Fiber integration Pushforward reducing cohomological degree by fiber dimension.
Gauss and Codazzi equations Compare ambient and intrinsic derivatives of a submanifold.
Gauss–Bonnet Identifies the total Gaussian curvature with 2π2\pi times the Euler characteristic.
Gauge transformation Changes a frame or acts as an isomorphism in the connection stack. ,
Gerbe holonomy Surface evaluation encoded by local two-forms and overlap data. ,
Holonomy Transport around a closed path or higher cycle. ,
Homotopy pullback A pair of maps with their specified compatibility path. ,
Jacobi field First variation of a geodesic through geodesics.
Levi–Civita connection The metric-compatible tangent connection with zero torsion.
Pontryagin classes Even Chern classes of the complexification with the conventional signs. ,
Principal connection An equivariant horizontal splitting, equivalently a normalized Lie algebra-valued one-form.
Quotient stack Remembers inertia and descent along an action groupoid.
Reduction An HH-subbundle of a GG-bundle, or a section of its G/HG/H-bundle.
Sectional curvature Scalar curvature associated to an unoriented tangent two-plane.
Smooth site Manifolds as tests, with gluing over open covers.
String structure A lift across the homotopy fiber of 12p1\tfrac12p_1.
Surface transport Two-dimensional transport on bigons with their thin relations.
Thin homotopy A rank-constrained homotopy respected by transport. ,
Transgression Integrates along a parameter family or along the evaluation circle. ,
Weil algebra Universal algebra for connection and curvature generators. ,

References

  1. Loring W. Tu. Differential Geometry. Connections, Curvature, and Characteristic Classes. Graduate Texts in Mathematics 275, Springer, 2017. https://link.springer.com/book/10.1007/978-3-319-55084-8.

  2. John M. Lee. Introduction to Riemannian Manifolds. 2nd ed., Graduate Texts in Mathematics 176, Springer, 2018. https://link.springer.com/book/10.1007/978-3-319-91755-9.

  3. Shoshichi Kobayashi and Katsumi Nomizu. Foundations of Differential Geometry, Vol. I. Wiley, 1963. https://www.wiley.com/en-us/foundations-of-differential-geometry-volume-1-p-9780471157335.

  4. Urs Schreiber and Konrad Waldorf. Parallel Transport and Functors. Journal of Homotopy and Related Structures 4(1), 187–244, 2009. https://arxiv.org/abs/0705.0452.

  5. Raoul Bott and Loring W. Tu. Differential Forms in Algebraic Topology. Graduate Texts in Mathematics 82, Springer, 1982. https://link.springer.com/book/10.1007/978-1-4757-3951-0.

  6. John W. Milnor and James D. Stasheff. Characteristic Classes. Annals of Mathematics Studies 76, Princeton University Press, 1974. https://press.princeton.edu/books/paperback/9780691081229/characteristic-classes.

  7. Shiing-Shen Chern and James Simons. Characteristic Forms and Geometric Invariants. Annals of Mathematics 99(1), 48–69, 1974. https://annals.math.princeton.edu/1974/99-1/p03.

  8. Victor W. Guillemin and Shlomo Sternberg. Supersymmetry and Equivariant de Rham Theory. Springer, 1999. https://link.springer.com/book/10.1007/978-3-662-03992-2.

  9. Jacob Lurie. Higher Topos Theory. Annals of Mathematics Studies 170, Princeton University Press, 2009. https://www.math.ias.edu/~lurie/papers/HTT.pdf.

  10. Urs Schreiber. Differential cohomology in a cohesive infinity-topos. 2013. https://arxiv.org/abs/1310.7930.

  11. Domenico Fiorenza, Urs Schreiber, and Jim Stasheff. Čech cocycles for differential characteristic classes: An infinity-Lie theoretic construction. Advances in Theoretical and Mathematical Physics 16(1), 149–250, 2012. https://arxiv.org/abs/1011.4735.

  12. Jeff Cheeger and James Simons. Differential characters and geometric invariants. In Geometry and Topology, Lecture Notes in Mathematics 1167, 50–80, Springer, 1985. https://math.mit.edu/juvitop/pastseminars/notes_2019_Fall/cheeger-simons.pdf.

  13. Michael J. Hopkins and Isadore M. Singer. Quadratic functions in geometry, topology, and M-theory. Journal of Differential Geometry 70(3), 329–452, 2005. https://arxiv.org/abs/math/0211216.

  14. Ulrich Bunke, Thomas Nikolaus, and Michael Völkl. Differential cohomology theories as sheaves of spectra. Journal of Homotopy and Related Structures 11, 1–66, 2016. https://arxiv.org/abs/1311.3188.

  15. Christian Bär and Christian Becker. Differential Characters and Geometric Chains. https://arxiv.org/abs/1303.6457.

  16. Michael K. Murray. Bundle gerbes. Journal of the London Mathematical Society 54(2), 403–416, 1996. https://arxiv.org/abs/dg-ga/9407015.

  17. John C. Baez and Aaron D. Lauda. Higher-dimensional algebra V: 2-groups. Theory and Applications of Categories 12, 423–491, 2004. https://arxiv.org/abs/math/0307200.

  18. John C. Baez and Urs Schreiber. Higher gauge theory: 2-connections on 2-bundles. 2004. https://arxiv.org/abs/hep-th/0412325.

  19. Urs Schreiber and Konrad Waldorf. Smooth functors vs. differential forms. Homology, Homotopy and Applications 13(1), 143–203, 2011. https://arxiv.org/abs/0802.0663.

  20. Urs Schreiber and Konrad Waldorf. Connections on non-abelian gerbes and their holonomy. Theory and Applications of Categories 28, 476–540, 2013. https://arxiv.org/abs/0808.1923.

  21. Stephan Stolz and Peter Teichner. What is an elliptic object? In Topology, Geometry and Quantum Field Theory, London Mathematical Society Lecture Note Series 308, 247–343, Cambridge University Press, 2004. https://doi.org/10.1017/CBO9780511526398.013.

  22. Konrad Waldorf. String connections and Chern–Simons theory. Transactions of the American Mathematical Society 365, 4393–4432, 2013. https://arxiv.org/abs/0906.0117.

  23. Camille Laurent-Gengoux, Jean-Louis Tu, and Ping Xu. Chern–Weil map for principal bundles over groupoids. Mathematische Zeitschrift 255, 451–491, 2007. https://arxiv.org/abs/math/0401420.

  24. Bertrand Toën. Derived algebraic geometry. EMS Surveys in Mathematical Sciences 1, 153–240, 2014. https://arxiv.org/abs/1401.1044.

  25. Michael F. Atiyah. Complex analytic connections in fibre bundles. Transactions of the American Mathematical Society 85, 181–207, 1957. https://doi.org/10.1090/S0002-9947-1957-0086359-5.

  26. Dima Arinkin and Andrei Căldăraru. When is the self-intersection of a subvariety a fibration? Advances in Mathematics 231, 815–842, 2012. https://arxiv.org/abs/1007.1671.

  27. Konrad Waldorf. Transgression to loop spaces and its inverse, II: Gerbes and fusion bundles with connection. Asian Journal of Mathematics 20, 59–116, 2016. https://arxiv.org/abs/1004.0031.

  28. Michael K. Murray and Daniel Stevenson. Bundle gerbes: Stable isomorphism and local theory. Journal of the London Mathematical Society 62, 925–937, 2000. https://arxiv.org/abs/math/9908135.

Complete flattened LaTeX source
\documentclass[11pt,oneside,openany]{book}
\usepackage[T1]{fontenc}

\usepackage{lmodern}

\usepackage[
  paperwidth=7in,paperheight=10in,left=0.87in,right=0.87in,
  top=0.90in,bottom=0.86in,headheight=15pt,headsep=14pt,footskip=25pt
]{geometry}

\usepackage{microtype}

\usepackage{mathtools,amssymb,amsthm}

\usepackage{xcolor}

\usepackage{tikz-cd}

\tikzcdset{
  labels={/tikz/auto=false,description,/tikz/inner sep=0.8pt},
  phantom/.append style={labels={/tikz/fill=none}}
}
\usepackage{enumitem}

\usepackage{array,longtable}

\usepackage{caption}

\usepackage{titlesec}

\usepackage{fancyhdr}

\usepackage{etoolbox}

\usepackage{xurl}

\usepackage{hyperref}

\usepackage{bookmark}


\definecolor{bookink}{HTML}{182C2B}
\definecolor{bookaccent}{HTML}{13615A}
\definecolor{bookmuted}{HTML}{546967}
\definecolor{bookrule}{HTML}{A3BBB7}

\hypersetup{
  pdftitle={Connections, Curvature, and Higher Geometry},
  pdfauthor={Anonymous},
  pdfsubject={Differential geometry, characteristic classes, differential cohomology, and higher gauge theory},
  colorlinks=true,linkcolor=bookaccent,citecolor=bookaccent,urlcolor=bookaccent,
  bookmarksdepth=2
}
\AtBeginDocument{\color{bookink}}
\linespread{1.055}
\setlength{\parskip}{2pt plus 1pt}
\setlength{\parindent}{1em}
\setlength{\emergencystretch}{1.2em}
\raggedbottom
\clubpenalty=3000
\widowpenalty=3000
\allowdisplaybreaks[1]

\titleformat{\chapter}[display]
  {\normalfont\huge\bfseries\raggedright}
  {\sffamily\small\bfseries\color{bookaccent}\MakeUppercase{\chaptertitlename}\ \thechapter}
  {8pt}{}
  [\vspace{4pt}{\color{bookrule}\titlerule}]
\titlespacing*{\chapter}{0pt}{12pt}{20pt}
\titleformat{\section}{\Large\bfseries\raggedright}
  {\color{bookaccent}\thesection}{0.65em}{}
\titlespacing*{\section}{0pt}{15pt plus 2pt minus 1pt}{6pt}
\titleformat{\subsection}{\large\bfseries\raggedright}
  {\color{bookaccent}\thesubsection}{0.65em}{}
\captionsetup{font=small,labelfont={bf,color=bookaccent},skip=6pt}

\pagestyle{fancy}
\fancyhf{}
\fancyhead[L]{\small\color{bookmuted}\nouppercase{\leftmark}}
\fancyfoot[R]{\sffamily\small\color{bookmuted}\thepage}
\renewcommand{\headrulewidth}{0pt}
\renewcommand{\chaptermark}[1]{\markboth{#1}{}}
\fancypagestyle{plain}{  \fancyhf{}  \fancyfoot[R]{\sffamily\small\color{bookmuted}\thepage}  \renewcommand{\headrulewidth}{0pt}}

\theoremstyle{plain}
\newtheorem{theorem}{Theorem}[chapter]
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{example}[theorem]{Example}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{Remark}

\DeclareMathOperator{\Hom}{Hom}
\DeclareMathOperator{\Ob}{Ob}
\DeclareMathOperator{\id}{id}
\DeclareMathOperator{\Id}{Id}
\DeclareMathOperator{\Nat}{Nat}
\DeclareMathOperator{\tr}{tr}
\DeclareMathOperator{\Ad}{Ad}
\DeclareMathOperator{\ad}{ad}
\DeclareMathOperator{\Hol}{Hol}
\DeclareMathOperator{\Map}{Map}
\DeclareMathOperator{\Pf}{Pf}
\DeclareMathOperator{\Aut}{Aut}
\DeclareMathOperator{\Tor}{Tor}
\DeclareMathOperator{\Ext}{Ext}
\DeclareMathOperator{\diag}{diag}
\newcommand{\Set}{\mathbf{Set}}
\newcommand{\FinSet}{\mathbf{FinSet}}
\newcommand{\Cat}{\mathcal{C}}
\newcommand{\bookindexpage}[1]{\hyperref[#1]{\pageref*{#1}}}

\setcounter{tocdepth}{1}

\makeatletter
\let\book@pendingpart\@empty
\newcommand{\book@partheading}[1]{  {\parindent\z@\raggedright\sffamily\large\bfseries\color{bookaccent}#1\par}  \nobreak
}
\newcommand{\book@numberedpart}[2]{  \refstepcounter{part}  \addcontentsline{toc}{part}{\thepart\hspace{1em}#1}  \book@partheading{\partname~\thepart\quad #2}}
\newcommand{\book@unnumberedpart}[1]{  \phantomsection
  \addcontentsline{toc}{part}{#1}  \book@partheading{#1}}
\NewDocumentCommand{\bookpart}{s o m}{  \IfBooleanTF{#1}
    {\gdef\book@pendingpart{\book@unnumberedpart{#3}}}
    {\IfNoValueTF{#2}
      {\gdef\book@pendingpart{\book@numberedpart{#3}{#3}}}
      {\gdef\book@pendingpart{\book@numberedpart{#2}{#3}}}}}
\pretocmd{\@chapter}{\book@pendingpart\global\let\book@pendingpart\@empty}
  {}
  {\PackageError{book}{Cannot attach part heading to chapter}{}}
\makeatother

\begin{document}
\hypersetup{pageanchor=false}
\begin{titlepage}
\thispagestyle{empty}
\noindent\textcolor{bookaccent}{\rule{36mm}{1.3pt}}
\vspace*{0.145\textheight}

\noindent{\fontsize{30}{37}\selectfont\bfseries
  Connections, Curvature,\\[0.12em] and Higher Geometry\par}
\vspace{1.3em}
\noindent{\Large\color{bookmuted}From covariant differentiation\\ to differential cohomology\par}
\vspace{2.25em}
\noindent{\large Anonymous\par}
\end{titlepage}
\hypersetup{pageanchor=true}
\frontmatter
\begingroup
  \fontsize{10}{10.25}\selectfont
  \setlength{\parskip}{0pt}
  \makeatletter
  \patchcmd{\l@chapter}{\penalty\@highpenalty}{\nobreak}{}{}
  \patchcmd{\l@chapter}{\vskip 1.0em \@plus\p@}{\vskip .92em \@plus\p@}{}{}
  \makeatother
  \tableofcontents
\endgroup

\mainmatter
\bookpart{Metric and tangent geometry}
\chapter{Metrics and tangent geometry}
\label{ch:metric}

A tangent vector describes the first-order motion of a point, and a metric assigns a length to that motion. This assignment gives norms on the fibers of the tangent bundle. Its smooth dependence on the base point is the input for connection and curvature. We begin with the way the assignment transforms under a change of coordinates, as in Tu's treatment of metrics and curves \cite[\S\S\,1--4]{tu}.

\section{The tangent map and pullback}

For a smooth manifold \(M\), a tangent vector at \(p\) can be defined as a derivation on germs of smooth functions at \(p\). In coordinates \(x^1,\dots,x^n\), any such derivation is
\(\sum_i v^i\partial_i|_p\), where \(v^i=v(x^i)\). To see the uniqueness, expand a function near \(p\) as its value at \(p\) plus
\[
 \sum_i (x^i-x^i(p))a_i,
 \qquad a_i(p)=\partial_i f(p).
\]
One obtains the expansion by integrating the gradient along coordinate line segments. The product rule then gives \(v(f)=\sum_i v^ia_i(p)\). This identifies \(T_pM\) with \(\mathbb R^n\) after a chart is chosen. If \(f:M\to N\), set \((df_pv)(h)=v(h\circ f)\). It is a derivation at \(f(p)\), and this definition proves the chain rule
\(d(g\circ f)_p=dg_{f(p)}\,df_p\) directly.

\begin{definition}
A Riemannian metric on \(M\) is a smooth section \(g\) of
\(\operatorname{Sym}^2T^*M\) such that \(g_p\) is positive definite for every \(p\in M\). Write \(|v|_g=\sqrt{g_p(v,v)}\) for \(v\in T_pM\).
\end{definition}

A metric has the local expression \(g_{ij}\,dx^i\,dx^j\), with summation understood and the product denoting the corresponding symmetric tensor. If \(x=x(y)\), the matrix in \(y\)-coordinates equals \(J^{\mathsf T}gJ\), where \(J=(\partial x^i/\partial y^a)\). This formula follows by evaluating \(g\) on the vectors \(\partial/\partial y^a\). In particular, positive definiteness does not depend on the chart.

\begin{proposition}
For every smooth \(f:M\to N\) and metric \(h\) on \(N\), the symmetric tensor
\[
 (f^*h)_p(v,w)=h_{f(p)}(df_pv,df_pw)
\]
is positive semidefinite. It is positive definite at \(p\) exactly when \(df_p\) is injective. If \(j:N\to Q\) and \(h\) is a metric on \(Q\), then \((j\circ f)^*h=f^*(j^*h)\).
\end{proposition}

\begin{proof}
The squared length of \(v\) under the tensor is \(|df_pv|_h^2\), whose zero set is \(\ker df_p\). Smoothness follows from the coordinate formula. The formula for composites follows by applying the chain rule to both inputs.
\end{proof}

For an immersion \(M\to N\), the metric is the restriction of the ambient inner product to the immersed tangent directions. For a submersion with positive-dimensional fibers, its pullback metric vanishes along the vertical tangent bundle. This distinction will matter when taking quotient atlases. A tensor that is positive definite transversely to the fibers can encode a metric on a quotient even though it is degenerate on the atlas.

\section{Existence by convex gluing}

We use Hausdorff second-countable smooth manifolds, so smooth partitions of unity exist. On a coordinate chart \(U_i\) choose its Euclidean metric \(g_i\). Take a locally finite partition of unity \(\{\phi_i\}\) subordinate to the charts and put
\[
 g=\sum_i \phi_i g_i .
\]
Each product extends smoothly by zero outside the chart because the support of \(\phi_i\) lies in \(U_i\). The sum is smooth by local finiteness. At each \(p\) and for each nonzero \(v\in T_pM\), every summand in \(g_p(v,v)\) is nonnegative, and at least one is positive since the coefficients sum to one. Thus \(g\) is a metric. The same argument for Euclidean forms in bundle frames gives a metric on any real vector bundle over \(M\).

For any two metrics \(g_0,g_1\), the tensor \(g_t=(1-t)g_0+t g_1\) is positive definite for \(0\leq t\leq1\). Hence every two metrics are joined by a smooth family. This interpolation is a statement about the convex space of metrics, with the underlying manifold and points fixed. Geodesics and curvature may vary along such an interpolation.

\begin{example}[Rotational coordinates]
The map \((r,\theta)\mapsto(r\cos\theta,r\sin\theta)\) has Euclidean pullback
\[
 dx^2+dy^2=dr^2+r^2d\theta^2
\]
on any angular chart with \(r>0\). This follows by squaring the two differentials and canceling the cross terms. The determinant of the metric matrix is \(r^2\). At \(r=0\), the angular vector maps to zero, so the tensor written in these variables ceases to be positive definite. The degeneracy reflects the parametrization at the origin.
\end{example}

\section{Length, speed, and oriented turning}

For a piecewise smooth curve \(\gamma:[a,b]\to M\), the metric defines its length by
\[
 L_g(\gamma)=\int_a^b |\dot\gamma(t)|_g\,dt .
\]
Length is invariant under increasing changes of parameter by the change-of-variables formula. Reversing the parameter leaves length unchanged because the norm is even. On a regular smooth segment, the length coordinate
\[
 s(t)=\int_a^t|\dot\gamma(u)|_g\,du
\]
has positive derivative. The inverse function theorem gives a smooth unit-speed parametrization on that segment.

In the standard oriented plane, let \(J\) be rotation by \(+\pi/2\), write \(T=d\gamma/ds\), and take the normal \(JT\). Differentiation of \(\langle T,T\rangle=1\) shows that \(dT/ds\) is perpendicular to \(T\). Hence
\[
 \frac{dT}{ds}=\kappa\,JT .
\]
The real number \(\kappa\) is the signed curvature. Its value for an arbitrary regular parameter \(t\) is
\begin{equation}
 \kappa(t)=
 \frac{\det(\dot\gamma(t),\ddot\gamma(t))}
      {|\dot\gamma(t)|^3}.
 \label{eq:metric-signed-curvature}
\end{equation}
Indeed \(\dot\gamma=vT\), with \(v=ds/dt\), so
\(\ddot\gamma=\dot v T+v^2\kappa JT\). The determinant of \(T,JT\) equals one. The term involving \(\dot v\) drops out, and the result is \(v^3\kappa\).
If a local real angle \(\theta\) satisfies \(T=(\cos\theta,\sin\theta)\), then \(\kappa=d\theta/ds\). The oriented tangent angle therefore records a one-dimensional form of curvature.

\begin{example}[An ellipse and its osculating radius]
For the counterclockwise ellipse \(\gamma(t)=(a\cos t,b\sin t)\), \(a,b>0\), the determinant in \eqref{eq:metric-signed-curvature} is \(ab\), and
\[
 \kappa(t)=
 \frac{ab}{(a^2\sin^2t+b^2\cos^2t)^{3/2}} .
\]
At \((a,0)\) the curvature is \(a/b^2\). Its reciprocal \(b^2/a\) is the radius of the oriented circle tangent there with the same curvature. If \(a=b\), the signed curvature is \(1/a\) along the entire loop. Reversing the direction of traversal negates \(\kappa\), while the norm of its curvature vector stays the same.
\end{example}

A tangent-vector field along a plane curve is an ordinary map into \(\mathbb R^2\), once the ambient parallel coordinates have been chosen. A field along a curve in \(M\) is instead a section of \(\gamma^*TM\). The vectors above distinct points then belong to distinct vector spaces. Chapter~\ref{ch:levi} constructs a derivative of such sections by choosing a connection; the metric provides a canonical connection for the tangent bundle.

\section{Volume and coordinate changes}

On an oriented \(n\)-manifold, the metric gives the positive form
\[
 dV_g=\sqrt{\det(g_{ij})}\,dx^1\wedge\cdots\wedge dx^n
\]
in an oriented chart. Here \(\det(J^{\mathsf T}gJ)=(\det J)^2\det g\). Consequently the formula transforms as an \(n\)-form under orientation-preserving changes of chart. Forgetting orientation gives the well-defined density \(|dV_g|\). For a map \(f\) that is an orientation-preserving local diffeomorphism, the same computation gives \(dV_{f^*h}=f^*dV_h\).

This construction depends on the metric algebraically at each point. If \(g'=e^{2u}g\) for a smooth real function \(u\), the length of a tangent vector is multiplied by \(e^u\) and \(dV_{g'}=e^{nu}dV_g\). Sectional curvature will depend on derivatives of \(u\) as well. The different orders of dependence, first for connections and second for curvature, reappear in the local description by differential forms.

The length metric on each connected component is defined by
\(d_g(p,q)=\inf_\gamma L_g(\gamma)\) over piecewise smooth paths from \(p\) to \(q\). In a relatively compact coordinate ball about \(p\), the positive metric matrix admits uniform bounds \(cI\leq(g_{ij})\leq CI\) for some \(0<c\leq C\). Paths staying in the chart obey the same two-sided bounds for lengths up to factors \(\sqrt c,\sqrt C\). Paths leaving a smaller ball incur a positive lower length bound before their first exit. Coordinate line segments supply the local upper estimate in the smaller ball. These facts prove nondegeneracy and show that \(d_g\) induces the manifold topology. Finding length minimizers is the next question. Its local equation is a connection equation, as treated in Chapter~\ref{ch:geodesics}.
\chapter{Levi-Civita connection and torsion}
\label{ch:levi}

A connection furnishes a derivative of tangent-vector fields. It compares vectors above adjacent points by specifying an infinitesimal rule with a product law for multiplying sections by functions. The construction recovers the directional derivative on Euclidean space and gives an intrinsic rule when the tangent bundle is presented only by local coordinates. We use the curvature convention \(R(X,Y)=[\nabla_X,\nabla_Y]-\nabla_{[X,Y]}\) throughout this chapter and Chapter~\ref{ch:intrinsic}.

\section{Affine connections and the derivative of a section}

\begin{definition}
An affine connection on \(M\) is an \(\mathbb R\)-bilinear operation \((X,Y)\mapsto\nabla_XY\) on smooth vector fields such that
\[
 \nabla_{fX}Y=f\nabla_XY,\qquad
 \nabla_X(fY)=X(f)Y+f\nabla_XY .
\]
Its torsion is the operation
\[
 T(X,Y)=\nabla_XY-\nabla_YX-[X,Y].
\]
\end{definition}

These identities make all expressions local. For example, if \(Y\) vanishes near a given point \(p\), choose a cutoff that is one near \(p\) with support in the vanishing neighborhood. Applying the product rule to the zero field obtained by multiplying by this cutoff gives \(\nabla_XY|_p=0\). Smooth local formulas for \(\nabla\) therefore restrict to smaller coordinate domains. In a local frame \(e_1,\ldots,e_n\), put
\(\nabla_{e_i}e_j=\Gamma^k{}_{ij}e_k\). Expanding the connection rules gives
\[
 \nabla_XY=\sum_k\left(X(Y^k)+
             \sum_{i,j}X^iY^j\Gamma^k{}_{ij}\right)e_k .
\]
The coefficients are arbitrary smooth functions for a local connection. A second connection differs from the first by a section of \(T^*M\otimes T^*M\otimes TM\), since the terms differentiating a multiplier cancel. For torsion the bracket satisfies \([X,fY]=X(f)Y+f[X,Y]\). The same cancellation proves \(C^\infty(M)\)-bilinearity and skew symmetry. In a coordinate frame \(T^k{}_{ij}=\Gamma^k{}_{ij}-\Gamma^k{}_{ji}\). In a general frame there is also the contribution from \([e_i,e_j]\).

Torsion can be changed without changing the equation for a curve.
For any affine connection, put
\(\nabla^\circ_XY=\nabla_XY-\tfrac12T(X,Y)\). The subtracted term
is a tensor, so the Leibniz rule for the connection is preserved.
The antisymmetric part of that tensor has exactly the size required
to cancel \(T\). Thus \(\nabla^\circ\) is torsion-free. Since
\(T(v,v)=0\), the two connections have the same affine geodesics.
This construction uses the antisymmetry of torsion and requires no
metric compatibility.

\section{The metric and symmetry determine the connection}

\begin{theorem}[Levi-Civita]
For a Riemannian metric \(g\), exactly one affine connection satisfies
\[
 Xg(Y,Z)=g(\nabla_XY,Z)+g(Y,\nabla_XZ),
 \qquad T(X,Y)=0.
\]
\end{theorem}

\begin{proof}
Write the compatibility identity in the cyclic orders \((X,Y,Z)\), \((Y,Z,X)\), and \((Z,X,Y)\). Add the first two and subtract the third. Torsion-freeness replaces each reversed derivative by the given derivative and a bracket. The result is the Koszul formula
\begin{align}
 2g(\nabla_XY,Z)
 &= Xg(Y,Z)+Yg(Z,X)-Zg(X,Y)\notag\\
 &\quad -g(X,[Y,Z])+g(Y,[Z,X])+g(Z,[X,Y]).
 \label{eq:levi-koszul}
\end{align}
Nondegeneracy gives uniqueness. To see existence, denote the right side by \(K(X,Y,Z)\). The product rule and the bracket identity show that \(K(X,Y,fZ)=fK(X,Y,Z)\): the coefficients of \(X(f)\) and \(Y(f)\) from differentiating the metric evaluations cancel the corresponding bracket terms. The same expansion gives \(K(fX,Y,Z)=fK(X,Y,Z)\) and \(K(X,fY,Z)=fK(X,Y,Z)+2X(f)g(Y,Z)\). Thus \eqref{eq:levi-koszul} defines a smooth affine connection using \(g\) to recover a vector from a covector. Subtracting \(K(Y,X,Z)\) from \(K(X,Y,Z)\) leaves \(2g([X,Y],Z)\), which proves zero torsion. Adding \(K(X,Y,Z)\) to \(K(X,Z,Y)\) leaves \(2Xg(Y,Z)\), which proves metric compatibility.
\end{proof}

With coordinate fields all brackets vanish. Inverting the metric matrix gives the explicit formula
\begin{equation}
 \Gamma^k{}_{ij}=\frac12\sum_\ell g^{k\ell}
     \bigl(\partial_i g_{\ell j}+\partial_j g_{\ell i}
                 -\partial_\ell g_{ij}\bigr).
 \label{eq:levi-christoffel}
\end{equation}
In particular, the connection depends on the first derivatives of the metric. Formula \eqref{eq:levi-koszul} applies also to orthonormal frames with nonzero brackets, where the derivative terms of the metric evaluations are then zero. These are two complementary ways to compute the same connection \cite[\S\S\,6,13]{tu}.

\begin{proposition}
Let \(f:(M,g)\to(N,h)\) be an isometric local diffeomorphism. If \(X,Y\) on \(M\) correspond locally to \(df(X),df(Y)\) on \(N\), then \(df(\nabla^g_XY)=\nabla^h_{df(X)}df(Y)\) there.
\end{proposition}

\begin{proof}
The derivative of a local diffeomorphism preserves the vector-field bracket because the conjugated derivations have conjugated commutators. Transport the connection on \(N\) to \(M\) in any neighborhood where \(f\) is a diffeomorphism. The transported connection has zero torsion. It is compatible with \(f^*h=g\) by differentiation of the pulled-back inner products. Uniqueness gives the asserted identity. The local computations agree on overlaps.
\end{proof}

There is a closely related rule for an isometric immersion into Euclidean space. Decompose the ambient derivative \(D_XY\) into its tangential and normal parts. The tangential projection is compatible with the induced metric because normal fields are perpendicular to tangent fields. Its torsion is the tangential projection of \(D_XY-D_YX-[X,Y]=0\). Thus the tangential projection is the Levi-Civita connection. The normal projection defines the second fundamental form used in Chapter~\ref{ch:intrinsic}.

\section{Covariant differentiation along a path}

Let \(\gamma:I\to M\) and let \(V(t)\in T_{\gamma(t)}M\) be a smooth section of the pullback tangent bundle. In coordinates around a segment of \(\gamma\) define
\begin{equation}
 (\nabla_t V)^k
 =\frac{d V^k}{dt}+
       \sum_{i,j}\Gamma^k{}_{ij}(\gamma)\dot\gamma^i V^j .
 \label{eq:levi-path}
\end{equation}
This can also be derived locally by expressing \(V=\sum_jV^j(t)e_j(\gamma(t))\) in any frame and applying the product rule with
\(\nabla_t(e_j\circ\gamma)=(\nabla e_j)_{\gamma(t)}(\dot\gamma(t))\). If a frame is changed by a matrix \(B(\gamma(t))\), differentiating \(B\) by the ordinary chain rule gives the same change as applying \(\nabla_{\dot\gamma}\) to the new frame fields. Hence the result is independent of the frame. Formula \eqref{eq:levi-path} also applies when the path crosses itself or is momentarily stationary. It requires only the pulled-back section and no extension as a field on \(M\).

A parallel section has \(\nabla_t V=0\). This is a linear differential equation in \(V\), so an initial vector determines a unique solution over the path interval. The solution operators compose on concatenation, and reversing the path gives the inverse. Here \(\gamma_{12}*\gamma_{01}\) denotes first \(\gamma_{01}\) and then \(\gamma_{12}\).
\[
 \begin{tikzcd}[column sep=3.3em]
 T_{p_0}M\arrow[r,"P_{\gamma_{01}}"]\arrow[rr,bend right=25,"{P_{\gamma_{12}*\gamma_{01}}}"']&
 T_{p_1}M\arrow[r,"P_{\gamma_{12}}"]&
 T_{p_2}M
 \end{tikzcd}
\]
For a metric connection, compatibility gives
\[
 \frac{d}{dt} g(V,W)
   =g(\nabla_tV,W)+g(V,\nabla_tW).
\]
In particular, parallel transport preserves the Euclidean structures on the initial and terminal fibers. Its dependence on the full path encodes curvature, as developed in Chapter~\ref{ch:geodesics} and in the subsequent treatment of holonomy.

A geodesic with affine parameter satisfies \(\nabla_t\dot\gamma=0\), or
\(\ddot\gamma^k+\Gamma^k{}_{ij}(\gamma)\dot\gamma^i\dot\gamma^j=0\)
in coordinates. Local existence and uniqueness follow by converting the equation into a smooth first-order system on \(TM\). Metric compatibility makes the speed constant. The equation is invariant under affine reparametrization \(t\mapsto at+b\). A smooth reparametrization with a nonzero second derivative introduces a tangential acceleration proportional to that second derivative.

\begin{example}[A circle in a flat metric]
For the metric \(dr^2+r^2d\theta^2\) in Chapter~\ref{ch:metric}, the nonzero coordinate coefficients are
\[
 \Gamma^r{}_{\theta\theta}=-r,\qquad
 \Gamma^\theta{}_{r\theta}=\Gamma^\theta{}_{\theta r}=r^{-1}.
\]
The unit field \(e_\theta=r^{-1}\partial_\theta\) satisfies
\(\nabla_{e_\theta}e_\theta=-r^{-1}\partial_r\) on a circle of radius \(r\). Its acceleration has length \(1/r\) and points toward the center. This agrees with \eqref{eq:metric-signed-curvature}. Ambient Euclidean geodesics remain straight lines. In polar coordinates these same lines satisfy an equation with nonzero connection coefficients, since the frame varies with the point.
\end{example}

\section{A connection on a Lie group}

Let \(G\) carry a bi-invariant Riemannian metric, such as the metric induced by the negative trace pairing on the skew-Hermitian matrices of a unitary group. Identify the Lie algebra with the left-invariant fields. Invariance of the inner product at the identity under the adjoint action gives
\[
 g([X,Y],Z)+g(Y,[X,Z])=0
\]
on such fields, by differentiating the adjoint action along a one-parameter subgroup. In the Koszul formula their pairwise inner products are constant, and interchanging arguments in the bracket terms using the displayed identity gives \(\nabla_XY=\tfrac12[X,Y]\). A one-parameter subgroup \(\gamma(t)=\exp(t\xi)\) therefore has \(\nabla_t\dot\gamma=0\). In left-translated coordinates \(v(t)\in\mathfrak g\), a field along this curve is parallel when
\[
 \dot v(t)+\tfrac12[\xi,v(t)]=0.
\]
Its solution is \(v(t)=\exp(-t\,\operatorname{ad}_\xi/2)v(0)\). The first exponential above is the Lie group exponential, and the one in this formula is the exponential of a linear operator. The two exponentials describe the moving base point and the moving vector, respectively. The operator is skew for the invariant inner product, so transport is orthogonal. In the next chapter the same connection gives curvature through a single Lie bracket.

The connection is the structure governing motion of a single vector along a single curve. The intrinsic effect of commuting two such infinitesimal motions is the curvature tensor. Its definition subtracts the derivative along the commutator of the underlying vector fields. This subtraction is essential when the fields come from a noncoordinate frame.
\chapter{Intrinsic and extrinsic curvature}
\label{ch:intrinsic}

The acceleration of a curve in Euclidean space can be resolved into components tangent and normal to an immersed submanifold. The tangential component is governed by the induced metric. In dimension two the normal component also determines the intrinsic Gaussian curvature through a quadratic identity. We derive that identity from flatness of the ambient derivative and the splitting of the ambient tangent bundle. This gives the tensor form of the calculation underlying Tu's treatment \cite[\S\S\,5,8--9,12]{tu}.

\section{The curvature of a connection}

The curvature of \(\nabla\) is
\begin{equation}
 R(X,Y)Z=\nabla_X\nabla_Y Z-\nabla_Y\nabla_X Z
                   -\nabla_{[X,Y]}Z .
 \label{eq:intrinsic-curvature}
\end{equation}
It is \(C^\infty(M)\)-linear in all three fields. In the first input, the term \(-Y(f)\nabla_XZ\) arising from \(fX\) cancels the contribution of the bracket \([fX,Y]=f[X,Y]-Y(f)X\). Skew symmetry then handles the second input. In the third input, the terms differentiating \(f\) once cancel between the two compositions of \(\nabla\), and the scalar second derivatives cancel against \([X,Y](f)\). Thus \(R\) is a two-form with values in endomorphisms of \(TM\).

For a metric connection, applying \(XY-YX-[X,Y]\) to the scalar \(g(Z,W)\) and expanding metric compatibility gives
\[
 g(R(X,Y)Z,W)=-g(Z,R(X,Y)W).
\]
The mixed terms from differentiating both fields cancel in pairs. For a torsion-free connection, the cyclic sum in \(X,Y,Z\) of \eqref{eq:intrinsic-curvature} is zero. To check this at a point, use coordinate fields, so that all brackets vanish. Pairing terms with the same outer derivative leaves
\(\nabla_X(\nabla_Y Z-\nabla_Z Y)\) and its cyclic permutations, each zero by symmetry of the Christoffel symbols. Tensoriality extends the statement to all fields.

For Levi-Civita curvature there is also pair exchange. Use the components
\[
 R_{ijk\ell}=g(R(\partial_i,\partial_j)\partial_k,\partial_\ell).
\]
At a point \(p\), start with a chart \(x\) centered there and introduce \(y\) by
\[
 x^a=y^a-\tfrac12\Gamma^a{}_{bc}(p)y^by^c .
\]
This defines new coordinates near \(p\) by the inverse function theorem. The second derivatives of \(x(y)\) cancel the old Christoffel symbols at \(p\). Torsion-freeness ensures they are symmetric in \(b,c\). In the new chart \(\Gamma(p)=0\), and metric compatibility also gives \(\partial_i g_{jk}(p)=0\). Evaluating \eqref{eq:levi-christoffel} and its first derivatives at \(p\) gives
\[
 \begin{split}
 2R_{ijk\ell}
 &=\partial_i\partial_k g_{\ell j}
   -\partial_i\partial_\ell g_{jk}\\
 &\quad-\partial_j\partial_k g_{\ell i}
   +\partial_j\partial_\ell g_{ik}.
 \end{split}
\]
Since mixed derivatives commute and \(g\) is symmetric, this expression equals \(2R_{k\ell ij}\). These identities characterize the symmetries of the metric curvature tensor \cite{lee-riemannian}.

For a two-plane \(\Pi\subseteq T_pM\), choose a spanning pair \(X,Y\). Its sectional curvature is
\[
 K(\Pi)=
 \frac{g(R(X,Y)Y,X)}
 {|X|_g^2|Y|_g^2-g(X,Y)^2}.
\]
In the numerator, skew symmetry in each pair gives a quadratic expression on \(X\wedge Y\). Both numerator and denominator therefore scale by the square of the determinant of a change of basis of \(\Pi\). This proves independence of the spanning pair. On a surface there is only one plane at each point, giving its Gaussian curvature.

\begin{example}[Bi-invariant metrics]
In the Lie group example of Chapter~\ref{ch:levi}, let \(X,Y,Z\) be left-invariant fields. The commutator of their covariant derivatives acts on \(Z\) by \(\tfrac14[[X,Y],Z]\), since the Jacobi identity rewrites the difference of the two nested brackets. The last term of \eqref{eq:intrinsic-curvature} acts by \(\tfrac12[[X,Y],Z]\). Thus
\[
 R(X,Y)Z=-\tfrac14[[X,Y],Z].
\]
Taking \(Z=Y\), pairing with \(X\), and using skew-adjointness of the bracket gives \(g(R(X,Y)Y,X)=\tfrac14|[X,Y]|^2\). Positive sectional curvature in a plane therefore occurs exactly when the two directions fail to commute. On a torus all brackets vanish, and this metric is flat.
\end{example}

\section{The Gauss and Codazzi identities}

Let \(M^n\) be immersed in \(\mathbb R^{n+1}\) with induced metric. Work locally with a smooth unit normal \(N\) and write \(D\) for the ambient Euclidean derivative. Differentiating \(\langle N,N\rangle=1\) shows that \(D_XN\) is tangent. Define the shape operator \(S(X)=-D_XN\) and the second fundamental form \(\mathrm{II}(X,Y)=g(SX,Y)\). Differentiating \(\langle Y,N\rangle=0\) yields
\[
 D_XY=\nabla_XY+\mathrm{II}(X,Y)N .
\]
Here \(\nabla\) is Levi-Civita by Chapter~\ref{ch:levi}. Taking the normal component of \(D_XY-D_YX=[X,Y]\) proves that \(\mathrm{II}\) is symmetric. This decomposition can also be recorded by the splitting
\[
 \begin{tikzcd}[column sep=3em]
 TM\oplus\mathbb RN \arrow[r,"\simeq"]
      \arrow[d,"\operatorname{pr}_{TM}"']&
      M\times\mathbb R^{n+1}\arrow[dl,"\operatorname{pr}_{TM}"]\\
 TM .
 \end{tikzcd}
\]
The derivative in the ambient bundle does not preserve either summand in general, and \(S\) and \(\mathrm{II}\) measure its off-diagonal terms.

\begin{theorem}[Gauss and Codazzi for a Euclidean hypersurface]
One has
\begin{align*}
 R(X,Y)Z
   &=\mathrm{II}(Y,Z)SX-\mathrm{II}(X,Z)SY,\\
 (\nabla_X\mathrm{II})(Y,Z)&=(\nabla_Y\mathrm{II})(X,Z).
\end{align*}
\end{theorem}

\begin{proof}
Insert the decomposition of \(D_YZ\) into \(D_XD_YZ\), subtract the counterpart with \(X,Y\) interchanged, and subtract \(D_{[X,Y]}Z\). Flatness of the Euclidean derivative makes the result zero. Its tangential component is \(R(X,Y)Z-\mathrm{II}(Y,Z)SX+\mathrm{II}(X,Z)SY\). Its normal component is the coefficient
\[
 \begin{split}
 X\mathrm{II}(Y,Z)-Y\mathrm{II}(X,Z)
 &+\mathrm{II}(X,\nabla_YZ)-\mathrm{II}(Y,\nabla_XZ)\\
 &-\mathrm{II}([X,Y],Z).
 \end{split}
\]
Use \([X,Y]=\nabla_XY-\nabla_YX\) and the product rule for the covariant derivative of a covariant tensor. The coefficient is the difference of the two derivatives in the second identity.
\end{proof}

When \(n=2\), choosing an orthonormal basis gives \(K=\det S\). Changing \(N\) to \(-N\) negates \(S\) while preserving its determinant in dimension two. Because the right-hand definition of \(K\) uses only the metric and its Levi-Civita connection, two isometric surfaces have equal Gaussian curvature at corresponding points. This is the intrinsic content of the Theorema Egregium.

\begin{example}[Graph, cylinder, and sphere]
For a graph \(z=f(x,y)\) in \(\mathbb R^3\), put \(w=(1+|\nabla f|^2)^{1/2}\) and use the upward normal \(N=(-f_x,-f_y,1)/w\). Its first fundamental form has determinant \(w^2\). The second fundamental form has matrix \((f_{ij}/w)\). Thus \(K=\det(f_{ij})/w^4\). Taking \(f=xy\) gives negative curvature at every point of its graph. On the cylinder of radius \(a\), the two eigenvalues of \(S\) with outward normal are \(-1/a\) in the circular direction and \(0\) along its axis. Hence \(K=0\). The local parametrization \((s,t)\mapsto(a\cos(s/a),a\sin(s/a),t)\) exhibits the same result by pulling the metric back to \(ds^2+dt^2\). On the sphere of radius \(a\) one has \(S=-a^{-1}\operatorname{id}\), so \(K=a^{-2}\).
\end{example}

\section{A curvature form on an oriented surface}

Take a positive orthonormal frame \(e_1,e_2\) and define a one-form \(\omega\) by \(\nabla e_1=\omega e_2\). Metric compatibility gives \(\nabla e_2=-\omega e_1\). Insert these two equations into \eqref{eq:intrinsic-curvature}. The product terms in \(\omega(X)\omega(Y)\) cancel, yielding
\[
 R(X,Y)e_1=(d\omega)(X,Y)e_2,\qquad
 d\omega=-K\,dV_g .
\]
The minus sign follows from \(K=g(R(e_1,e_2)e_2,e_1)\). For the positive coframe \((\theta^1,\theta^2)\), the first structure equations are
\(d\theta^1=\omega\wedge\theta^2\) and
\(d\theta^2=-\omega\wedge\theta^1\). They follow by evaluating \(d\theta^i\) on two fields and replacing their bracket by the difference of the covariant derivatives. Rotating the frame through an angle \(\phi\) changes \(\omega\) to \(\omega+d\phi\), by differentiation of the rotation. The curvature form is thus unchanged. When local angles have only circle-valued transitions, the differentials \(d\phi\) still agree for different real lifts. In Chapter~\ref{ch:geodesics}, integration of this identity turns those frame changes into a topological integer.

\begin{example}[The effect of conformally scaling a plane metric]
Take \(g=e^{2u}(dx^2+dy^2)\) and the positive coframe \((e^u dx,e^u dy)\). The first structure equations give
\(\omega=-u_y\,dx+u_x\,dy\). Consequently
\[
 K=-e^{-2u}(u_{xx}+u_{yy}).
\]
For \(u=-\log y\) on the upper half-plane, this is \(-1\). For \(u=\log(2a^2)-\log(a^2+x^2+y^2)\), the Laplacian is
\(-4a^2/(a^2+x^2+y^2)^2\), giving \(K=a^{-2}\). The latter metric is the usual stereographic expression for the round sphere of radius \(a\). More generally, on a conformal chart the curvature is controlled by a scalar second-order equation. The sign of its Laplacian determines the sign of the Gaussian curvature pointwise.
\end{example}
\chapter{Geodesics, Jacobi fields, and surfaces}
\label{ch:geodesics}

The covariant derivative turns geodesic motion into a differential equation. A variation of one geodesic through other geodesics satisfies a new, linear equation containing curvature. On surfaces the same curvature has a global integral constrained by the topology. The local motion and the global identity thus give two ways of detecting the curvature defined in Chapter~\ref{ch:intrinsic}. We use Tu's conventions for the geodesic equation and his presentation of the surface theorem as a comparison \cite[\S\S\,13--17]{tu}.

\section{Energy variation and the exponential map}

Let \(\Gamma(s,t)\) be a smooth variation of a path \(\gamma(t)=\Gamma(0,t)\) on \([a,b]\). Put \(T(s,t)=\partial_t\Gamma(s,t)\) and \(V(s,t)=\partial_s\Gamma(s,t)\). The energy of the path \(t\mapsto\Gamma(s,t)\) is
\[
 E(s)=\frac12\int_a^b
    g_{\Gamma(s,t)}\bigl(T(s,t),T(s,t)\bigr)\,dt ,
\]
with \(s\) held fixed during integration. Both fields are sections of \(\Gamma^*TM\). The torsion-free identity \(\nabla_sT=\nabla_tV\) holds because the mixed partials of \(\Gamma\) agree and the Christoffel contributions are symmetric. Therefore
\begin{equation}
 E'(s)=\bigl[g(V,T)\bigr]_a^b-
                     \int_a^b g(V,\nabla_tT)\,dt .
 \label{eq:geo-first-var}
\end{equation}
Here metric compatibility first replaces the derivative of the squared speed by \(2g(\nabla_sT,T)\), and then gives the integration-by-parts formula. For variations fixing both endpoints, the boundary term vanishes. Every smooth field supported in a small segment of the interior serves as a variation vector there, using coordinate displacement multiplied by a cutoff. Thus \(E'(0)=0\) for every fixed-endpoint variation implies the geodesic equation along \(\gamma\) by the fundamental lemma of the calculus of variations. The reverse implication follows from \eqref{eq:geo-first-var}.

Let \(\gamma_v\) start at \(p\) with velocity \(v\). Put \(\exp_p(v)=\gamma_v(1)\) when the solution exists on \([0,1]\). The solution depends smoothly on the initial conditions. The rescaling relation \(\gamma_{tv}(1)=\gamma_v(t)\) on its domain gives
\((d\exp_p)_0=\operatorname{id}_{T_pM}\), under the canonical identification of the tangent space of \(T_pM\) at zero. Thus \(\exp_p\) maps a sufficiently small open ball diffeomorphically onto a normal neighborhood. The coordinates from a linear orthonormal identification \(T_pM\simeq\mathbb R^n\) then have radial lines as geodesics.

\begin{proposition}[Radial orthogonality]
At a point \(v\) in a normal ball and for \(w\in T_pM\), one has
\[
 g_{\exp_p(v)}\big((d\exp_p)_v v,(d\exp_p)_v w\big)
      =g_p(v,w).
\]
\end{proposition}

\begin{proof}
Use \(\Gamma(s,t)=\exp_p(t(v+sw))\) for small \(s\) and \(0\leq t\leq1\). Along a radial geodesic \(\nabla_tT=0\), so
\[
 \partial_t g(V,T)=g(\nabla_sT,T)
            =\tfrac12\partial_s |v+sw|_{g_p}^2.
\]
As \(V(s,0)=0\), integration to \(t=1\) and evaluation at \(s=0\) give the formula. The vector \(T(s,1)\) at that endpoint is \((d\exp_p)_{v+sw}(v+sw)\).
\end{proof}

Consequently radial directions have unit length under the radial distance coordinate and are perpendicular to variations tangent to its spheres. For a path in the normal neighborhood the speed is at least the absolute derivative of the radius \(|\exp_p^{-1}(\gamma(t))|\) wherever the radius is positive. Integrating, with limiting endpoints at zero allowed, proves minimality of the radial segment among paths in the neighborhood. If the endpoint has radius \(r\) smaller than the radius \(R\) of a compactly contained normal ball, any path leaving that ball has initial length at least \(R\) by the same estimate up to its first exit. Thus the radial segment minimizes among all paths to the sufficiently near endpoint. These arguments are local and do not assume geodesic completeness.

\section{The linearized geodesic equation}

Suppose now that every \(t\mapsto\Gamma(s,t)\) is geodesic. The curvature commutator on \(\Gamma^*TM\) gives
\[
 \nabla_t^2 V=\nabla_t\nabla_s T
       =\nabla_s\nabla_t T+R(T,V)T
       =-R(V,T)T .
\]
At \(s=0\), the field \(J=V|_{s=0}\) thus satisfies
\begin{equation}
 \nabla_t^2 J+R(J,\dot\gamma)\dot\gamma=0 .
 \label{eq:geo-jacobi}
\end{equation}
Conversely, the linear equation has unique solutions prescribed by \(J(0),\nabla_t J(0)\). Choose a smooth curve of initial positions with derivative \(J(0)\), and vary the initial velocities along it with covariant derivative \(\nabla_t J(0)\). To construct the varied velocities, use parallel transport along the curve of positions and add the desired first-order term in that trivialization. Smooth dependence of the geodesic flow gives a variation, at least locally on the geodesic interval, whose variation field has those two initial values. Uniqueness of the linear equation identifies it with \(J\). This exhibits Jacobi fields as the linear response of geodesic motion \cite{lee-riemannian}.

For a unit-speed geodesic in constant sectional curvature \(K\), the operator \(J\mapsto R(J,\dot\gamma)\dot\gamma\) on the orthogonal complement of \(\dot\gamma\) is \(K\operatorname{id}\). Indeed it is self-adjoint by the curvature symmetries, and its quadratic form on perpendicular vectors is \(K|J|^2\). In a parallel orthonormal frame perpendicular to \(\dot\gamma\), a field with \(J(0)=0\) has coefficients proportional to
\[
 s_K(t)=
 \begin{cases}
 \sin(\sqrt Kt)/\sqrt K,&K>0,\\
 t,&K=0,\\
 \sinh(\sqrt{-K}t)/\sqrt{-K},&K<0 .
 \end{cases}
\]
The tangential component with zero initial value is linear in \(t\). On the round \(n\)-sphere of radius \(a\) with \(n\geq2\), the first positive time for a nonzero such field to vanish is \(\pi a\), so the antipode is the first conjugate point. Varying the initial velocity in a geodesic with initial point held fixed identifies these vanishing fields with the kernel of the derivative of the exponential map at that time. For \(K\leq0\) the displayed solutions give no positive conjugate time in a constant-curvature model.

\section{Second variation and local stability}

Let \(\gamma\) be a geodesic and suppose the variation fixes both endpoints. Differentiate \eqref{eq:geo-first-var} at \(s=0\). The term involving \(\nabla_sV\) vanishes there because \(\nabla_tT=0\), so
\[
 E''(0)=-\int_a^b g(V,\nabla_s\nabla_tT)\,dt .
\]
The curvature commutator and \(\nabla_sT=\nabla_tV\) give
\(\nabla_s\nabla_tT=\nabla_t^2V+R(V,T)T\) at \(s=0\).
Integration by parts, using \(V(a)=V(b)=0\), yields the quadratic form
\[
 I(V,V)=\int_a^b
   \bigl(|\nabla_t V|^2-g(R(V,T)T,V)\bigr)\,dt.
\]
Its bilinear polarization is symmetric by the metric curvature identities. A Jacobi field \(J\) vanishing at the endpoints pairs to zero with every such \(V\), by \eqref{eq:geo-jacobi} and integration by parts. The converse follows by testing against fields supported on interior segments. Thus conjugate endpoint data are detected by the kernel of the second variation.

For a unit-speed segment of length \(L\) on the round sphere, choose a perpendicular parallel unit field \(e\), and set \(V(t)=\sin(\pi t/L)e(t)\) for \(0\leq t\leq L\). The two terms in \(I\) integrate explicitly to
\[
 I(V,V)=\frac L2\bigl((\pi/L)^2-a^{-2}\bigr).
\]
The variation \((s,t)\mapsto\exp_{\gamma(t)}(sV(t))\) is defined for sufficiently small \(s\), by compactness of the path interval and the local existence theorem. When \(L>\pi a\) the quadratic form is negative on \(V\), giving a decrease in energy through that variation. At \(L=\pi a\), the displayed field is the conjugate Jacobi field already computed.

\begin{example}[The upper half-plane]
On \(y>0\), the metric \(g=y^{-2}(dx^2+dy^2)\) has oriented orthonormal frame \(e_1=y\partial_x\), \(e_2=y\partial_y\). Its bracket is \([e_1,e_2]=-e_1\), so the Koszul formula yields \(\nabla_{e_1}e_1=e_2\) and \(\nabla_{e_2}e_1=0\). Hence \(\omega=dx/y\) and \(d\omega=dx\wedge dy/y^2=dV_g\), giving \(K=-1\). Curves \(t\mapsto x_0+iy_0e^t\) have unit velocity \(e_2\) and are geodesics. For \(a,b,c,d\in\mathbb R\) with \(ad-bc>0\), the map \(z\mapsto (az+b)/(cz+d)\) multiplies the imaginary part and the norm of the Euclidean differential by the same factor \((ad-bc)/|cz+d|^2\). It therefore preserves \(g\). The images of the imaginary axis under \(z\mapsto x_0+r(z-1)/(z+1)\), \(r>0\), are the semicircles centered at \(x_0\). These are geodesics as well, by the naturality result of Chapter~\ref{ch:levi}.
\end{example}

\section{From boundary turning to Gauss--Bonnet}

For a unit oriented boundary tangent \(T=\cos\theta\,e_1+\sin\theta\,e_2\) on a smooth arc, define \(k_g=g(\nabla_TT,JT)\), where \(J\) is positive rotation by \(\pi/2\). In a local positive frame the derivative of the expression for \(T\) gives
\(k_g=d\theta/ds+\omega(T)\). On a positively oriented embedded triangular disk \(D\) in an oriented chart, the boundary tangent, with signed exterior turns inserted at its three corners, turns by \(2\pi\). One may compute this using the planar turning theorem in the chart, since a positive frame change extending over a disk contributes zero winding. The turning identity and Stokes' theorem, with \(d\omega=-K\,dV_g\), give
\begin{equation}
 \int_D K\,dV_g+\int_{\partial D} k_g\,ds
             +\sum_{\mathrm{corners}}(\pi-\alpha)=2\pi ,
 \label{eq:geo-polygon}
\end{equation}
where \(\alpha\) is the interior angle at that corner \cite[\S\,17.4]{tu}.

For a compact oriented surface without boundary, choose a smooth triangulation fine enough that each triangle lies in a coordinate neighborhood. Sum \eqref{eq:geo-polygon} over triangles. The contributions along a shared edge cancel: reversing its orientation negates \(k_g\) and preserves \(ds\). Interior angles around each vertex sum to \(2\pi\), because the incident tangent sectors fill its tangent circle. As each triangle has three corners, the sum of exterior turns is \(3\pi F-2\pi V\), where \(F,V\) count faces and vertices. With \(3F=2E\) for the edges, this yields
\[
 \int_M K\,dV_g=2\pi(V-E+F)=2\pi\chi(M).
\]
For the round sphere the calculation gives \(a^{-2}(4\pi a^2)=4\pi\). The independence from \(a\) expresses a general fact: a curvature form can vary with the metric while the cohomology class it represents stays fixed. This is the basic surface example of the characteristic-class construction developed later in the book.

\bookpart{Bundles, connections, and transport}
\chapter{Vector bundles and their descent}
\label{ch:bundles}

A frame identifies vectors over nearby points with a single model
vector space. Two frames identify that space in different ways, and
their change of coordinates is part of the geometry. The descent
description retains the change of coordinates and the equations it
satisfies. It thus applies to sections, to bundle maps and to
connections. We use smooth manifolds that are Hausdorff and second
countable, with real or complex vector bundles of finite rank. The
local frame and section descriptions follow the setting of
Tu~\cite[\S 7]{tu}.

\section{Pullback, local frames, and sections}

A rank-\(r\) vector bundle over \(M\) is a smooth map
\(p\colon E\to M\) with a vector space structure on each fiber and
local product identifications
\[
    E|_U \simeq U\times \mathbb K^r,\qquad
    \mathbb K=\mathbb R\ \text{or}\ \mathbb C,
\]
which are linear on fibers. A frame on \(U\) is an ordered tuple of
sections whose values are a basis at every point of \(U\). Write
\(e_i\) for the row of basis vectors in a frame on \(U_i\), so
\(e_i v_i\) denotes the vector with column of coordinates \(v_i\).
On an overlap put \(e_j=e_i g_{ij}\). Then
\[
 g_{ij}\colon U_i\cap U_j\longrightarrow \mathrm{GL}_r(\mathbb K),
 \qquad v_i=g_{ij}v_j.
\]
The matrix entries of \(g_{ij}\) are smooth because a bundle chart
smoothly extracts the coordinates of the smooth sections of the
second frame. On a triple overlap,
\(g_{ij}g_{jk}=g_{ik}\), by expressing \(e_k\) in frame \(e_i\)
in two ways.

For \(f\colon N\to M\) the pullback bundle is
\[
 f^*E=\{(y,e)\in N\times E\mid f(y)=p(e)\}.
\]
It is smooth and locally trivial since the inverse images of bundle
charts give product charts for it. The commutative square
\[
\begin{tikzcd}
  f^*E \arrow[r] \arrow[d] & E\arrow[d,"p"]\\
  N \arrow[r,"f"] & M
\end{tikzcd}
\]
has the pullback property in smooth manifolds. A smooth map into the
upper left corner consists of maps into \(N\) and \(E\) satisfying
the compatibility at \(M\). Smoothness of the resulting map can be
checked in the product charts above, so no assumption that arbitrary
fiber products of manifolds exist is needed. On frames, pullback
replaces \(g_{ij}\) by \(g_{ij}\circ f\). For a further map
\(h\colon P\to N\), the comparison \(h^*f^*E\simeq(fh)^*E\)
forgets the redundant intermediate base point. The comparisons for
three successive maps agree since both delete the same intermediate
points.

Sections over an open set \(U\) form a
\(C^\infty(U,\mathbb K)\)-module \(\Gamma(U,E)\).
Their restrictions form a sheaf, because smoothness of a section
can be tested in local frames. Concretely a section over \(M\) is
a family of smooth \(v_i\colon U_i\to\mathbb K^r\) with
\(v_i=g_{ij}v_j\). A bundle map \(F\colon E\to E'\) over \(M\)
is recorded on a shared trivializing cover by matrices \(F_i\) with
\[
            F_i g_{ij}=g'_{ij}F_j .
\]
This equation holds for rectangular matrices as well. It states
that converting a vector to the \(i\)-frame and applying the map
agrees with applying the map in the \(j\)-frame and converting the
result.

\begin{proposition}
Every \(C^\infty(M,\mathbb K)\)-linear map
\(\Gamma(M,E)\to\Gamma(M,E')\) arises from a unique bundle map.
\end{proposition}

\begin{proof}
Let \(T\) be such a map. If a section \(s\) vanishes near \(x\), choose
a bump function \(\chi\) equal to one in a smaller neighborhood and
supported where \(s=0\). Then \(0=T(\chi s)=\chi T(s)\), so
\(T(s)(x)=0\). Thus \(T\) depends at \(x\) only on the germ of \(s\).
Multiply a local frame by a cutoff equal to one on a smaller open
set and extend it by zero. The resulting global sections have
images under \(T\) that supply the columns of a smooth local
matrix. For any \(s\), write it in that frame near \(x\) and extend
its local coefficient functions by another cutoff. Germ locality
and linearity over functions show that \(T(s)(x)\) is the action of
the local matrix on \(s(x)\). Every vector in \(E_x\) extends to a
section by a local frame and a cutoff. This proves uniqueness and
forces the local matrices to agree as bundle maps.
\end{proof}

\section{Effective gluing}

\begin{theorem}[Descent for vector bundles]
\label{thm:vector-descent}
Let \((U_i)\) be an open cover of \(M\). Bundles over \(M\) are
equivalent to families of bundles \(E_i\to U_i\) with isomorphisms
\(\varphi_{ij}\colon E_j|_{U_i\cap U_j}\to E_i|_{U_i\cap U_j}\)
such that \(\varphi_{ii}=1\) and
\(\varphi_{ij}\varphi_{jk}=\varphi_{ik}\) on triple intersections.
Maps of these families are bundle maps commuting with the
isomorphisms.
\end{theorem}

\begin{proof}
Identify the point \(e\in E_j\) over \(x\in U_i\cap U_j\) with
\(\varphi_{ij}(e)\in E_i\). The cocycle law makes these
identifications an equivalence relation since it ensures both
inverses and closure under composition. The quotient is a set over
\(M\), and over \(U_i\) every class has a unique representative in
\(E_i\). The quotient topology restricts to the original topology
on \(E_i\), since each overlap map is a homeomorphism. The
transported smooth bundle charts agree on overlaps because the
overlap isomorphisms are smooth. Different fibers can be separated
by base neighborhoods, and two vectors in the same fiber lie in a
common Hausdorff chart. Hence the total space is Hausdorff. A countable subcover of local charts exists because the base
is second countable, and a countable union of their countable bases
gives a countable base upstairs. The linear structures in the model
charts glue under the linear isomorphisms.

A compatible family of maps is constant on equivalence classes
after taking the target quotient, and so gives a unique bundle map.
It is smooth on the covering charts. Restriction and gluing are
inverse up to the canonical isomorphisms supplied by the
identifications over \(U_i\).
\end{proof}

For trivial local bundles this construction is the quotient by the
matrix transitions \(g_{ij}\). Replacing \(e_i\) by \(e_i h_i\)
changes the transition to \(h_i^{-1}g_{ij}h_j\). Thus a cocycle
is an object in a local presentation and its changes of frame are
arrows. This matters when forming a moduli object. A bundle can
possess many automorphisms even when its isomorphism class is
fixed.

\begin{example}[Two real line bundles on a circle]
Use two open arcs covering \(S^1\), with two components in their
intersection. The transition can be \(+1\) on both components or
\(+1\) on one and \(-1\) on the other. A nowhere-zero section in
the latter case would give a continuous nonzero function on each
arc. Each function has one sign on its entire arc. The transition
would require their signs to agree on one overlap component and
to disagree on the other, which is impossible. The first case
has the constant unit section, so the two gluings already differ
as bundles. Their local frames present the same model fiber.
\end{example}

\begin{example}[Winding in a clutching presentation]
On the two disc-like charts of \(S^2\), use a unitary line
transition restricting on the equator to \(g(e^{i\theta})=e^{im\theta}\).
A change of frame on either disc has zero winding on its boundary
because it extends across the disc. Thus the winding of the
transition is unchanged by any bundle isomorphism, written as
changes of frame over the given charts. In particular a nonzero
\(m\) yields a nontrivial bundle. This topological obstruction will
also be detectable by integration of curvature once a connection
has been chosen.
\end{example}

\section{Descent along a surjective submersion}

Open inclusions suffice to construct each bundle, but descent
under maps with local sections will be useful for quotient
presentations. Let \(q\colon Y\to M\) be a surjective submersion
and write \(Y^{[j]}=Y\times_M\cdots\times_M Y\) with \(j\) factors.
A descent datum is a bundle \(D\to Y\) with compatible
isomorphisms
\[
 \alpha(y_1,y_2)\colon D_{y_2}\longrightarrow D_{y_1},
 \qquad
 \alpha(y_1,y_2)\alpha(y_2,y_3)=\alpha(y_1,y_3).
\]
The first formula is a smooth bundle isomorphism on \(Y^{[2]}\)
and the second is an equality over \(Y^{[3]}\). A bundle pulled
back from \(M\) has this datum because both fibers above
\((y_1,y_2)\) are copies of the same fiber over \(q(y_1)\).

Conversely, the submersion theorem supplies a cover by open sets
\(U_i\) with sections \(s_i\colon U_i\to Y\) of \(q\).
Use \(D_i=s_i^*D\) and the transition
\(\alpha(s_i(x),s_j(x))\) in Theorem~\ref{thm:vector-descent}.
The resulting \(E\) is a bundle over \(M\). Its pullback to
\(Y\) identifies with \(D\) over \(q^{-1}(U_i)\) by applying
\(\alpha(y,s_i(q(y)))\) in the fiber. The cocycle law makes
the identifications agree on overlaps and intertwine the
given descent datum. A map commuting with \(\alpha\)
restricts to maps of the \(D_i\) that glue by the same theorem.
This gives effectivity on objects and on maps. Using a common
refinement compares two choices of local sections, so the
isomorphisms furnished by this construction are canonical
for the given descent datum.

\section{Frames as a torsor}

The fiber of the frame bundle \(\operatorname{Fr}(E)\) at \(x\) is
the set of linear isomorphisms \(\mathbb K^r\to E_x\).
Right composition by \(\mathrm{GL}_r(\mathbb K)\) acts freely
and transitively. Here \emph{transitive} means that any two frames
are related by some change of basis, and \emph{free} means that
change of basis is unique. The quotient of
\(\operatorname{Fr}(E)\times\mathbb K^r\) by
\[
               (e g,v)\sim(e,g v)
\]
recovers \(E\) by \([e,v]\mapsto e(v)\), with the local charts
above. An action of another group \(G\) can replace the action of
\(\mathrm{GL}_r\) when a representation of \(G\) on the model
fiber is specified. Chapter~\ref{ch:principal} develops the
connection on such a principal bundle and the connection it induces
on the associated bundle.
\chapter{Connections on vector bundles}
\label{ch:vector-conn}

Sections of a vector bundle live in different fibers at different
points. A connection specifies how to take a derivative of a
section while accounting for that variation. Tu develops the
covariant derivative, its matrices and their curvature in
\S\S 10--11 and 22~\cite{tu}. The gluing theorem of
Chapter~\ref{ch:bundles} allows the derivative to be specified
entirely in local frames.

\section{Leibniz rule and existence}

\begin{definition}
A connection on a vector bundle \(E\to M\) is an
\(\mathbb R\)-linear operator
\(\nabla\colon\Gamma(E)\to\Omega^1(M;E)\), complex-linear for a
complex bundle, such that
\[
              \nabla(fs)=df\otimes s+f\nabla s.
\]
It acts on a vector field \(X\) by contraction,
\(\nabla_Xs=(\nabla s)(X)\).
\end{definition}

If \(s\) vanishes near \(x\), multiply by a cutoff \(\chi\) supported
in this neighborhood and equal to one near \(x\). Applying the
Leibniz rule to \(\chi s=0\) gives \((\nabla s)(x)=0\).
This locality induces connection operators on restrictions to open
sets. In a frame \(e_i\) write
\(\nabla e_{i,b}=\sum_a e_{i,a}(A_i)_{ab}\).
The local formula on a section \(e_i v_i\) is then
\[
                 \nabla(e_i v_i)=e_i(dv_i+A_i v_i).
\]
Every smooth matrix \(A_i\) of one-forms defines a connection
on this product chart. Choose local product connections
\(\nabla^{(i)}\) and a locally finite smooth partition of unity
\((\rho_i)\) subordinate to their domains. The products
\(\rho_i\nabla^{(i)}s\) extend by zero and their sum is smooth.
The coefficient of \(df\otimes s\) in its Leibniz identity is
\(\sum_i\rho_i=1\), which proves existence.

The difference of two connections satisfies
\((\nabla'-\nabla)(fs)=f(\nabla'-\nabla)(s)\).
The local matrix argument of Chapter~\ref{ch:bundles} therefore
identifies it with a section of \(T^*M\otimes\operatorname{End}E\),
and adding any such section to a connection gives another.
The set of connections is consequently an affine space modeled on
\(\Omega^1(M;\operatorname{End}E)\). For a fiber metric, the
compatibility condition is
\[
        d\langle s,t\rangle
        =\langle\nabla s,t\rangle+\langle s,\nabla t\rangle.
\]
In an orthonormal frame this is \(A_i^*=-A_i\), with transpose
in the real case and Hermitian adjoint in the complex case.
Local orthonormal frames supply local metric connections.
The partition of unity construction preserves this metric
identity, and two metric connections differ by a one-form
valued in skew-adjoint endomorphisms.

\section{Changing the frame and pulling back}

The frame \(e_j=e_i g_{ij}\) gives the coordinate relation
\(v_i=g_{ij}v_j\). Substituting into the Leibniz rule yields
\begin{equation}
\label{eq:vector-gauge}
 A_j=g_{ij}^{-1}A_i g_{ij}+g_{ij}^{-1}dg_{ij}.
\end{equation}
Indeed
\((d+A_i)(g_{ij}v_j)
=g_{ij}dv_j+(dg_{ij}+A_i g_{ij})v_j\).
The compatibility of local derivatives is the square
\[
\begin{tikzcd}[column sep=large]
 C^\infty(U_{ij},\mathbb K^r)
     \arrow[r,"d+A_j"]\arrow[d,"g_{ij}"'] &
 \Omega^1(U_{ij};\mathbb K^r)\arrow[d,"g_{ij}"]\\
 C^\infty(U_{ij},\mathbb K^r)\arrow[r,"d+A_i"] &
 \Omega^1(U_{ij};\mathbb K^r).
\end{tikzcd}
\]
Conversely, matrices satisfying \eqref{eq:vector-gauge} give
compatible operators in this square, hence a connection on the
glued bundle by the sheaf property. The term \(g_{ij}^{-1}dg_{ij}\)
records differentiation of the frame itself. A bundle isomorphism
which identifies two connections is accordingly governed by
the same formula.

For \(f\colon N\to M\), define the pullback connection by matrices
\(f^*A_i\) in the pullback frames. Pullback commutes with \(d\)
on functions, so these matrices satisfy \eqref{eq:vector-gauge}
for the pulled-back transitions. This definition works for any
smooth map and commutes with the canonical comparison for two
successive pullbacks.

\section{Curvature and the Bianchi identity}

Extend \(\nabla\) to \(E\)-valued differential forms by
\[
 \nabla(\alpha\otimes s)
     =d\alpha\otimes s+(-1)^k\alpha\wedge\nabla s
         \qquad (\alpha\in\Omega^k(M)).
\]
The Leibniz rule gives the same result after moving a smooth
function between the two factors. In a frame the extension is
\(d+A\wedge\), where matrix products are combined with wedge
products of entries. Applying it twice gives
\((d+A\wedge)^2=F\wedge\), with
\begin{equation}
\label{eq:vector-curvature}
             F=dA+A\wedge A.
\end{equation}
The \(d\)-terms hitting the argument cancel by the graded product
rule. Thus the square is tensorial and defines an
\(\operatorname{End}E\)-valued two-form. In the \(j\)-frame its
matrix is \(F_j=g_{ij}^{-1}F_i g_{ij}\) by applying the compatible
operators twice. On vector fields the same square reads
\[
  F(X,Y)s
   =\nabla_X\nabla_Ys-\nabla_Y\nabla_Xs-\nabla_{[X,Y]}s .
\]
The matrix formula shows that \(F_{f^*\nabla}=f^*F_\nabla\).

The induced covariant derivative on endomorphism-valued \(k\)-forms is
\[
       D\eta=d\eta+A\wedge\eta-(-1)^k\eta\wedge A .
\]
Its Bianchi identity is \(DF=0\). Expanding gives
\(dF=dA\wedge A-A\wedge dA\), while
\(A\wedge F-F\wedge A=A\wedge dA-dA\wedge A\).
The cubic terms cancel by associativity. The two-form \(F\)
therefore satisfies a differential constraint involving the same
connection that produced it. In particular
\(d\operatorname{tr}F=0\), since the graded commutator has zero
trace. Closed traces of higher curvature products will appear in
Chern--Weil theory.

\section{Tensor operations and a matrix example}

For bundles \(E,E'\) with connections, the connection on their
tensor product is
\[
 \nabla(s\otimes t)=(\nabla s)\otimes t+s\otimes(\nabla' t).
\]
The Leibniz rules make this well-defined on the tensor product
over functions. Its curvature is \(F\otimes 1+1\otimes F'\),
as the mixed terms have opposite signs when the derivative is
applied twice. The dual connection is fixed by requiring
\(d(\lambda(s))=(\nabla^\vee\lambda)(s)+\lambda(\nabla s)\).
In dual frames its matrix is \(-A^{\mathsf T}\) and its
curvature is \(-F^{\mathsf T}\). The connection on the determinant
\(\bigwedge^r E\) is obtained by differentiating each wedge factor.
Only the diagonal entries contribute to the connection of the
top exterior frame, so its local one-form is
\(\operatorname{tr}A\) and its curvature is
\(\operatorname{tr}F=d\operatorname{tr}A\).
For unitary connections we normalize the determinant curvature
by \(i/(2\pi)\). The scalar resulting two-form is real.

\begin{example}[Constant coefficients with nonzero curvature]
On the trivial Hermitian bundle \(\mathbb R^2\times\mathbb C^2\),
take \(A=X\,dx+Y\,dy\) with
\[
 X=\begin{pmatrix}i&0\\0&-i\end{pmatrix},\qquad
 Y=\begin{pmatrix}0&1\\-1&0\end{pmatrix}.
\]
The entries of \(A\) are constant and both matrices are
skew-Hermitian. Its curvature is
\[
 F=[X,Y]\,dx\wedge dy
   =\begin{pmatrix}0&2i\\2i&0\end{pmatrix}dx\wedge dy.
\]
Thus the connection has curvature even though the determinant
connection is the product connection. The commutator is also
the leading difference between the two orders of transport along
short coordinate segments. Chapter~\ref{ch:transport} makes this
comparison explicit.
\end{example}

\section{Normal frames and varying the connection}

At a prescribed point \(x_0\), choose coordinates \(x^\mu\)
and write \(A=\sum_\mu A_\mu\,dx^\mu\). The frame change
\[
 h(x)=\exp\Bigl(-\sum_\mu
      (x^\mu-x_0^\mu)A_\mu(x_0)\Bigr)
\]
has \(h(x_0)=I\) and \((dh)_{x_0}=-A_{x_0}\).
Equation~\eqref{eq:vector-gauge} then gives \(A'_{x_0}=0\).
If the initial frame is orthonormal for a metric connection,
each exponent is skew-adjoint and \(h\) remains orthogonal or
unitary. Such a frame simplifies the derivative at a point.
The curvature there is \(dA'_{x_0}\), which can be nonzero.
Vanishing on an entire neighborhood would imply \(F=0\)
there by \eqref{eq:vector-curvature}. This distinguishes
pointwise normalization from a local parallel frame.

For a global one-form \(\alpha\) valued in \(\operatorname{End}E\),
the connection \(\nabla+\alpha\) has curvature
\[
              F_{\nabla+\alpha}=F_\nabla+D\alpha+\alpha\wedge\alpha.
\]
The identity follows by expanding \(d(A+\alpha)+(A+\alpha)^2\)
and collecting terms of each degree in \(\alpha\).
For the path of connections \(\nabla_t=\nabla+t\alpha\), it follows
that \(\frac{d}{dt}F_{\nabla_t}=D_t\alpha\), where \(D_t\) is
the derivative induced by \(\nabla_t\).
The first variation of the curvature is linear in the direction,
whereas the finite change includes the quadratic term.
This distinction will produce a transgression between characteristic
forms of two connections.


\begin{example}[A projected connection]
Let \(P\colon M\to\operatorname{End}(\mathbb C^N)\) be a smooth
family of orthogonal projections of constant rank. The images
form a vector bundle \(E\) because projected fixed vectors give
local frames wherever an appropriate minor is invertible.
Define \(\nabla s=P\,ds\). The relation \(Ps=s\) gives the
Leibniz rule, and orthogonality shows that pairing with a vector
in \(E\) is unchanged when projecting. Hence the connection
preserves the induced metric. Its extension to \(E\)-valued
forms is obtained by applying \(d\) componentwise and projecting.
Differentiating \(P^2=P\) gives \(P(dP)P=0\).
For a section \(s\) one has \(ds=(dP)s+Pds\), so
\[
 \nabla^2 s=P\,d(Pds)
            =P\,dP\wedge ds
            =P(dP\wedge dP)s .
\]
The projection can therefore produce a curved connection on
a subbundle of a flat product. In the tautological line
of Chapter~\ref{ch:principal}, the projection is
\(P=zz^\dagger\) for a local unit column \(z\). The induced
one-form in that frame is \(z^\dagger dz\), which provides
a second construction of the same connection.
\end{example}
\chapter{Principal connections}
\label{ch:principal}

A frame bundle is a bundle of choices of coordinates. Its fibers
have a free transitive group action, so a local coordinate choice
can be changed by a unique group element. Principal bundles keep
this structure for any Lie group \(G\). Connections become
invariant horizontal distributions and can then be applied to
several representations of the same group. The constructions below
correspond to Tu's treatment of horizontal distributions,
connection forms and associated bundles~\cite[\S\S 27--31]{tu}.

\section{Right torsors and horizontal directions}

A principal \(G\)-bundle is a smooth locally trivial bundle
\(\pi\colon P\to M\) whose fibers are right \(G\)-torsors. Local
trivializations intertwine the given action with multiplication
on the right of the second factor of \(U\times G\).
Equivalently, with the bundle charts in place, the action gives
the diffeomorphism \(P\times G\simeq P\times_M P\),
\((p,g)\mapsto(p,pg)\). For each \(p\), the map
\[
 \mathfrak g\longrightarrow V_pP=\ker(d\pi_p),\qquad
 \xi\longmapsto \xi^\#_p
      =\left.\frac{d}{dt}\right|_{0}p\exp(t\xi)
\]
is an isomorphism, as seen in a product chart. Right translation
\(R_h\) sends \(\xi^\#_p\) to
\((\operatorname{Ad}_{h^{-1}}\xi)^\#_{ph}\). This follows from
\(\exp(t\xi)h=h\exp(t\operatorname{Ad}_{h^{-1}}\xi)\).

A horizontal distribution is a smooth complement \(H\) to \(VP\)
in \(TP\), invariant under right translation. The restriction
\(d\pi|_{H_p}\) is an isomorphism onto \(T_{\pi(p)}M\), so a tangent
vector downstairs has one horizontal lift at each point of its
fiber. Map any tangent vector \(v\) to the Lie algebra element
corresponding to its vertical component. This gives the connection
form \(\omega\in\Omega^1(P;\mathfrak g)\) satisfying
\begin{equation}
\label{eq:principal-form}
 \omega(\xi^\#)=\xi,\qquad
 R_h^*\omega=\operatorname{Ad}_{h^{-1}}\omega .
\end{equation}
Conversely, \eqref{eq:principal-form} makes
\(v\mapsto(\omega(v))^\#\) an equivariant projection onto \(VP\).
Its kernel is the invariant complement. Thus the form and the
horizontal distribution determine each other.

\section{Local potentials and curvature}

A local section \(s_i\colon U_i\to P\) gives the potential
\(A_i=s_i^*\omega\). In coordinates \(p=s_i(x)g\),
\begin{equation}
\label{eq:principal-local}
   \omega=\operatorname{Ad}_{g^{-1}}\pi^*A_i+g^{-1}dg .
\end{equation}
For the base tangent this is equivariance with \(g\) fixed.
For the group tangent it is the left Maurer--Cartan form
\(g^{-1}dg\), meaning the tangent translated back to the identity
by left multiplication by \(g^{-1}\). On an overlap write
\(s_j=s_i g_{ij}\). Pulling \eqref{eq:principal-local} along
\(s_j\) gives
\[
             A_j=\operatorname{Ad}_{g_{ij}^{-1}}A_i
                           +g_{ij}^{-1}dg_{ij}.
\]
Local forms with these transitions give a global connection by
\eqref{eq:principal-local}. To check compatibility, substitute
\(g=g_{ij}g'\) and use the product rule for the Maurer--Cartan
form. Local product connections can also be combined using a
partition of unity pulled back from \(M\). Their weighted sum
reproduces each vertical generator because the weights add to one,
and equivariance survives because the weights are constant on
the fibers. In particular every smooth principal bundle in our
setting admits a connection.

Use the graded Lie bracket convention
\([\omega,\omega](v,w)=2[\omega(v),\omega(w)]\).
Define the curvature by
\[
                    \Omega=d\omega+\tfrac12[\omega,\omega].
\]
It is equivariant by \eqref{eq:principal-form}. It is horizontal
as well. Indeed the flow of \(\xi^\#\) is \(R_{\exp(t\xi)}\).
Differentiating equivariance gives
\(\mathcal L_{\xi^\#}\omega=-[\xi,\omega]\).
Cartan's formula and \(\omega(\xi^\#)=\xi\), which is constant
on \(P\), give
\(\iota_{\xi^\#}d\omega=-[\xi,\omega]\). The contraction of the
bracket term is \([\xi,\omega]\), so the sum vanishes. In local
sections the curvature is \(F_i=dA_i+\frac12[A_i,A_i]\).
In matrix representations this becomes \(dA_i+A_i\wedge A_i\).
The motion of horizontal directions is measured by the same form.
For horizontal fields \(X^H,Y^H\),
\begin{equation}
\label{eq:curvature-vertical-bracket}
             \Omega(X^H,Y^H)=-\omega([X^H,Y^H]).
\end{equation}
This is the exterior derivative on a pair of fields in
\(\ker\omega\). Thus the horizontal subbundle is involutive
precisely when \(\Omega=0\). This identifies flatness with an
integrability condition before constructing any transport.

\section{Flat gauges and smaller structure groups}

When \(\Omega=0\), the Frobenius theorem supplies a local integral
submanifold of \(H\) through any point of \(P\). At that point
the restriction of \(d\pi\) to its tangent space is an isomorphism.
After shrinking, the restriction of \(\pi\) is a diffeomorphism
to an open neighborhood in \(M\). Its inverse gives a horizontal
local section, so the potential in that section is zero.
Conversely a zero potential gives zero curvature on its domain.
If two horizontal sections coexist on a connected open set,
then \(g_{ij}^{-1}dg_{ij}=0\) in their change of frame law.
Thus they differ by a single constant group element. Flat
bundles are consequently locally assembled by constant
transitions, although their global gluing can be nontrivial.

A reduction of the structure group to a closed Lie subgroup
\(K\subset G\) is a principal \(K\)-subbundle \(Q\subset P\).
The connection restricts to a principal \(K\)-connection if
\(\omega(TQ)\subset\mathfrak k\). This condition is equivalent
to \(H_q\subset T_qQ\) at every \(q\). If the image
is in \(\mathfrak k\), subtracting the corresponding vertical
\(K\)-vector makes any tangent vector in \(T_qQ\) horizontal.
Since \(T_qQ\) surjects onto \(T_{\pi(q)}M\), it contains all
horizontal lifts there. In the other direction, the
decomposition of \(T_qQ\) into its horizontal vectors and its
vertical \(K\)-directions shows that its image is in
\(\mathfrak k\). The reproduction and equivariance axioms
restrict as well. For instance, an orthogonal reduction of a
real frame bundle arises from a metric, and requiring the
connection to restrict to that reduction is precisely the skew
matrix condition of Chapter~\ref{ch:vector-conn}.

\section{Associated bundles and pullback}

For a smooth linear representation \(\rho\colon G\to
\mathrm{GL}(W)\), define \(E=P\times_G W\) by the relation
\((pg,w)\sim(p,\rho(g)w)\). The chart on \(U_i\) identifies
\([s_i(x),w]\) with \(w\), giving transition \(\rho(g_{ij})\).
A section is equivalently a smooth map \(u\colon P\to W\)
such that \(u(pg)=\rho(g)^{-1}u(p)\). Restrict \(du\) to
horizontal vectors to define a connection on \(E\).
Equivariance of \(H\) ensures this definition is independent of
the representative of a point of \(E\). The Leibniz rule follows
from differentiating \((f\circ\pi)u\).

In local coordinates let \(v_i=u\circ s_i\). Subtracting
\((A_i(X))^\#\) from \(ds_i(X)\) gives the horizontal lift of \(X\)
at \(s_i(x)\). Differentiating \(u\) in that vertical direction
gives \(-\rho_*(A_i(X))v_i\). The local derivative is consequently
\[
                       dv_i+\rho_*(A_i)v_i .
\]
As \(\rho_*\) preserves brackets, the curvature is \(\rho_*(F_i)\).
In the adjoint representation the horizontal equivariant curvature
\(\Omega\) thus gives a form with values in
\(\operatorname{ad}P=P\times_G\mathfrak g\). Here one uses
equivariance to transport the value of \(\Omega\) between
representatives and horizontality to choose tangent lifts.
The curvature endomorphism of the associated adjoint connection
acts by the Lie bracket with the descended value of \(\Omega\).

A smooth map \(f\colon N\to M\) gives a pullback principal bundle
and connection by pulling \(\omega\) back along the projection to
\(P\). Fibers are mapped equivariantly, so the reproduction and
equivariance conditions still hold. Associated bundles commute
with this construction, as both sides identify
\([((y,p),w)]\) with \((y,[p,w])\). The local potential formula
then shows that their induced connections agree.
For the frame bundle of a vector bundle, using its defining
representation recovers the original bundle connection
from the matrices of Chapter~\ref{ch:vector-conn}.

\begin{example}[The unit circle bundle on the projective line]
Give \(P=S^3\subset\mathbb C^2\) the right scalar action of
\(U(1)\), so its quotient is \(\mathbb{CP}^1\). The form
\(\omega=z^\dagger dz\) is imaginary on \(S^3\), since the
derivative of \(z^\dagger z=1\) has zero real part.
It evaluates to \(\xi\) on \(z\xi\), and unitary scalar
multiplication preserves it. It therefore defines a principal
connection. On the chart of lines spanned by \((1,w)\), take
the unit section \((1,w)/\sqrt{1+|w|^2}\). The potential and curvature are
\[
 A=\frac{\bar w\,dw-w\,d\bar w}{2(1+|w|^2)},\qquad
 F=dA=\frac{d\bar w\wedge dw}{(1+|w|^2)^2}.
\]
For example, differentiating the numerator gives
\(2\,d\bar w\wedge dw\), and differentiating the denominator
subtracts the term proportional to \(|w|^2\).
The associated line for the standard representation has fiber
the given line in \(\mathbb C^2\), hence is the tautological
line. In the complex orientation, writing \(w=x+iy\) gives
\[
 \int_{\mathbb{CP}^1}\frac{i}{2\pi}F
      =-\frac{1}{\pi}
          \int_{\mathbb R^2}\frac{dx\,dy}{(1+x^2+y^2)^2}
      =-1.
\]
The omitted point has zero area. In polar coordinates the
real integral is
\[
                  2\pi\int_0^\infty \frac{r}{(1+r^2)^2}\,dr=\pi .
\]
The normalization records both the orientation and the choice
of line in this principal bundle construction.
\end{example}
\chapter{Paths, thin homotopy, and transport}
\label{ch:transport}

A connection can be tested by transporting vectors and frames along
paths. Composition of paths is then a composition of invertible
maps of fibers. Curvature controls the dependence on a homotopy
of paths. The appropriate functor for a general connection uses
homotopies that sweep out no infinitesimal area.

\section{Transport equations}

Write a section along \(\gamma\colon[0,1]\to M\) in a frame
as a column \(v(t)\). It is parallel when
\[
          \dot v(t)=-A_{\gamma(t)}(\dot\gamma(t))v(t).
\]
Existence and uniqueness for linear ordinary differential equations
give a solution on each closed segment contained in a chart.
Along chart changes the relation \(v_i=g_{ij}v_j\) and
equation~\eqref{eq:vector-gauge} give the same equation in the
new frame. Subdivision and uniqueness therefore give a well-defined
transport isomorphism \(T_\gamma\colon E_{\gamma(0)}\to E_{\gamma(1)}\).
On a principal bundle, a lift \(\gamma^H(t)=s_i(\gamma(t))g(t)\)
is horizontal when
\[
             \dot g(t)=-(dR_{g(t)})_1 A_i(\dot\gamma(t)).
\]
This is \eqref{eq:principal-local} evaluated on the lifted tangent.
It has a solution over each compact chart segment. To see the
extension in the group variable, take a short solution starting
at \(1\) and right translate it to the required initial value.
The Lie algebra input is bounded on the segment, so local existence
near \(1\) gives a uniform short time for these restarts.
For a matrix group the right-hand side is simply
\(-A_i(\dot\gamma(t))g(t)\).
The uniqueness of the horizontal lift gives equivariance,
\(T_\gamma(pg)=T_\gamma(p)g\).
In a representation, the horizontal lift induces exactly
the parallel equation with potential \(\rho_*(A_i)\).
This agrees with the frame interpretation in
Tu~\cite[\S 29]{tu} and the principal parallel displacement of
Kobayashi--Nomizu~\cite[Ch.~II]{kobayashi-nomizu}.

We use smooth paths with sitting instants, meaning constant
segments near both endpoints. They can be joined smoothly at an
endpoint. Ordinary changes of speed preserving the endpoint order
leave transport unchanged by the chain rule. If \(\gamma_1\)
ends where \(\gamma_2\) begins, uniqueness on the two pieces gives
\[
               T_{\gamma_2\star\gamma_1}
                   =T_{\gamma_2}\circ T_{\gamma_1}.
\]
The constant path gives the identity. The transport of a reversed
path is the inverse, since reversing the solution of the parallel
equation gives the solution with reversed endpoints.

\section{Thin homotopy and curvature variation}

A thin homotopy of paths with fixed endpoints is a smooth map
\(H(s,t)\) interpolating between them, with constant ends in the
path variable and rank \(dH\leq1\) everywhere. Homotopies can be
reparameterized to sit near their boundaries before concatenating
them. Reparameterizations of a path are thin because their
interpolation factors through the path interval. A path followed
by its reverse contracts by shortening the portion of the path
traversed in both directions. This homotopy likewise factors
through an interval. These facts give associativity, units and
inverses for the thin homotopy classes of sitting paths.
Write \(\mathcal P_1(M)\) for this groupoid, with the points
of \(M\) as objects~\cite[\S 2.1]{schreiber-waldorf-transport}.

\begin{proposition}[Thin invariance]
Parallel transport of a principal connection is invariant under
thin homotopies. For a flat connection it is invariant under
all smooth homotopies relative to the endpoints.
\end{proposition}

\begin{proof}
Lift \(t\mapsto H(s,t)\) horizontally, starting for every \(s\)
at the same point \(p_0\) above the initial endpoint.
Smooth dependence of the differential equation on parameters
gives a smooth family \(p(s,t)\). On the square let
\(a_s=\omega(\partial_s p)\) and \(a_t=\omega(\partial_t p)\).
Then \(a_t=0\), and the structure equation gives
\[
    (p^*\Omega)(\partial_s,\partial_t)
        =\partial_s a_t-\partial_t a_s+[a_s,a_t]
        =-\partial_t a_s.
\]
The curvature is horizontal, so its value on a pair of tangent
vectors is determined by their projections to the base. Under a
thin homotopy those projections are linearly dependent. Thus the
left side vanishes. It also vanishes for an arbitrary homotopy
if the connection is flat. It follows that \(a_s\) is constant
in \(t\). At \(t=0\) it is zero because the lifted initial point
is fixed. At \(t=1\) the field \(\partial_s p\) is vertical
because the terminal base point is fixed, and \(a_s=0\) implies
that this vertical vector is zero. Hence the lifted terminal
point is independent of \(s\).
\end{proof}

It follows that transport assigns a functor
\[
        \mathcal P_1(M)\longrightarrow G\text{-}\mathrm{Tor},
        \qquad x\longmapsto P_x .
\]
The target consists of right \(G\)-torsors and their equivariant
isomorphisms. A connection-preserving bundle map
induces a natural transformation, since it carries horizontal
lifts with specified initial values to the corresponding
horizontal lifts. In particular the following square commutes
for a path \(\gamma\colon x\to y\) and such a bundle map \(\Phi\).
\[
\begin{tikzcd}
 P_x\arrow[r,"T_\gamma"]\arrow[d,"\Phi_x"'] &
 P_y\arrow[d,"\Phi_y"]\\
 P'_x\arrow[r,"T'_\gamma"] & P'_y .
\end{tikzcd}
\]

\section{Small rectangles and global holonomy}

Consider a vector bundle and a coordinate rectangle of side
\(\varepsilon\) in the \(x,y\) plane, beginning at a point taken
as the origin and following \(+x,+y,-x,-y\) in that order.
Take a frame with \(A(0)=0\), as in
Chapter~\ref{ch:vector-conn}. The integral of \(A\) along each
edge is \(O(\varepsilon^2)\). The integral equation for parallel
transport on an edge consequently gives
\(I-\int_{\mathrm{edge}}A+O(\varepsilon^4)\), with the bound
from the second iterated integral and higher terms. Multiplying
the four edge matrices gives
\[
 T_{\partial R_\varepsilon}
     = I-\oint_{\partial R_\varepsilon} A+O(\varepsilon^4)
     = I-\varepsilon^2 F_0(\partial_x,\partial_y)
                       +O(\varepsilon^3).
\]
The last equality uses Stokes entry by entry and \(F_0=(dA)_0\).
The same leading formula holds in any frame at the base point,
by conjugation. For constant \(A=X\,dx+Y\,dy\) one can also
multiply the exact transports in reverse traversal order,
\[
 T_{\partial R_\varepsilon}
   =e^{\varepsilon Y}e^{\varepsilon X}
       e^{-\varepsilon Y}e^{-\varepsilon X}.
\]
Their second-order term is
\(-\varepsilon^2[X,Y]\), in agreement with curvature.
In particular if vector transport descends to the ordinary
fundamental groupoid, every small rectangle has trivial transport.
Dividing by \(\varepsilon^2\) and taking the limit proves
\(F=0\). For a general Lie group, use coordinates on the base centered
at the initial point. The gauge
\(h(x)=\exp(-\sum_\mu x^\mu A_\mu(0))\) makes \(A(0)=0\)
by the same first-derivative argument. Use a group chart at
\(1\) with differential equal to the identity. Each edge solution
starts at \(1\) and stays \(O(\varepsilon^2)\) from it in that
chart, because its right-invariant driving term is
\(O(\varepsilon)\) on an edge of length \(\varepsilon\).
Translating its tangent back to the group chart changes the
driving term by \(O(\varepsilon^3)\), so its displacement is
\(-\int_{\mathrm{edge}}A+O(\varepsilon^4)\).
Group multiplication likewise differs from addition by
\(O(\varepsilon^4)\) for these displacements. This proves the
same curvature limit in the Lie algebra.
Thus flatness is exactly the condition for factoring through
ordinary path homotopy. This formulation uses the usual smooth
relative path model of the fundamental groupoid.

\begin{example}[Holonomy distinguishes connections]
For a unitary line connection \(A=-2\pi i\,a\) in a global
frame, the parallel equation integrates to
\(T_\gamma=\exp(2\pi i\int_\gamma a)\). On \(\mathbb R^2\)
take \(a=x\,dy\). The loop around a rectangle of oriented
area \(b\) has holonomy \(e^{2\pi i b}\) even though it is
contractible. Its phase is determined by the curvature integral.
On \(S^1=\mathbb R/\mathbb Z\), take \(a=\lambda\,d\theta\).
This connection is flat and transport around a loop of winding
\(m\) multiplies by \(e^{2\pi i\lambda m}\). The gauge
\(g=e^{-2\pi i k\theta}\) for integer \(k\) replaces
\(\lambda\) by \(\lambda+k\). Conversely, two such constants
with equal holonomy differ by an integer. Here curvature
vanishes and the global holonomy still detects different
connections up to gauge.
\end{example}

\section{Flat transport and local systems}

A functor from the ordinary fundamental groupoid to right
\(G\)-torsors determines a flat bundle by gluing. Choose a cover
by coordinate balls \(U_i\) with distinguished points \(x_i\).
On \(U_i\) use the product of \(U_i\) with the torsor assigned
to \(x_i\), with zero potential in a chosen model frame.
A path inside \(U_i\) from \(x_i\) to \(x\) identifies that
model torsor with the torsor assigned to \(x\). All such paths
have the same relative homotopy class. On the overlap, compose
one of these identifications with the inverse of the other.
The resulting transition is constant on each connected piece
of the overlap. To verify constancy, join nearby points of the
overlap by a path inside it and append it to both coordinate-ball
paths. The induced map in the target groupoid cancels between
the second identification and the inverse first one.

The transitions satisfy the cocycle law because they all use
the torsor assigned to the point in \(M\). Gluing the local
products uses the same quotient as in
Theorem~\ref{thm:vector-descent}, with multiplication on
the \(G\)-coordinate in place of linear maps.
The local zero potentials agree under the constant transitions,
so the bundle is flat. Its transport recovers the original
functor by subdivision into the coordinate balls. Conversely,
a flat connection has horizontal trivializations by
Chapter~\ref{ch:principal}, and the products above recover
its transition maps. Natural transformations give compatible
maps between these products, locally constant in the chosen
frames, and therefore produce the connection-preserving
isomorphisms. In a connected manifold, a choice of base point
and a point in its torsor describes the resulting groupoid
by representations of the fundamental group up to conjugation.
The description without a chosen base point retains the
conjugating maps as morphisms.

\section{Recovering a connection from transport}

The smooth information in transport is essential for recovery.
In a locally chosen frame, suppose a family of short paths
from \(x\) to \(x+\varepsilon v\) is obtained by smoothly
reparameterizing the straight segment to have sitting instants.
Let \(U(\varepsilon)\) be its arrow in \(G\), written in that
frame at both endpoints. The transport equation yields
\[
                  A_x(v)=-U'(0)\in\mathfrak g .
\]
Indeed the infinitesimal path velocity is proportional to \(v\),
and the integral of the time derivative of the scalar
reparameterization is one. The slope thus recovers the local
one-form. Under a change of frame \(g\), a path arrow from
\(x\) to \(y\) changes to \(g(y)^{-1}U g(x)\).
Its derivative along a short segment recovers precisely the
gauge law for the one-form.

The reconstruction theorem of Schreiber and Waldorf places this
local argument in an equivalence of categories. A transport functor
with \(BG\)-structure admits local trivializations by smooth
\(BG\)-valued functors on a surjective submersion, with smooth
descent data. Under those conditions it comes from a principal
bundle with connection, and its natural morphisms arise from
maps of the corresponding bundles with connection
\cite[Theorem~5.4]{schreiber-waldorf-transport}. The local hypotheses
give the smooth dependence needed for the derivative and the
transition data needed to glue the resulting forms. Thus the
base groupoid and the smooth descent condition together retain
both the local curvature and the global bundle.\par

\bookpart{Chern--Weil theory and transgression}
\chapter{Invariant polynomials and Chern--Weil forms}
\label{ch:chern-weil}

Curvature transforms by conjugation when a frame changes. This transformation law suggests a way to construct ordinary differential forms from it. A polynomial on the Lie algebra invariant under conjugation eliminates dependence on the frame. Closedness and connection independence then follow from two distinct differential identities. Tu develops this passage first for vector bundles and then for principal bundles~\cite[\S\S23,32]{tu}.

\section{Encoding a polynomial by polarization}

Let \(G\) be a Lie group with real Lie algebra \(\mathfrak g\). For a homogeneous polynomial \(p\colon\mathfrak g\to\mathbb R\) of degree \(m\), define
\[
 P(X_1,\ldots,X_m)=\frac{1}{m!}
 \frac{\partial^m}{\partial t_1\cdots\partial t_m}
 p\left(\sum_i t_iX_i\right)\bigg|_{\boldsymbol t=0}.
\]
Then \(P\) is symmetric and multilinear, and \(p(X)=P(X,\ldots,X)\). The formula follows by extracting the coefficient of \(t_1\cdots t_m\), which also proves uniqueness. If \(p\) is invariant under \(\operatorname{Ad}(G)\), simultaneous application of \(\operatorname{Ad}_g\) to all inputs of \(P\) leaves it unchanged. Differentiating at \(g=\exp(sZ)\) gives
\begin{equation}
 \sum_{i=1}^{m}P(X_1,\ldots,[Z,X_i],\ldots,X_m)=0 .
 \label{eq:hdg-inf-invariance}
\end{equation}
For disconnected \(G\), the differentiated equation tests only the identity component. The statement about descent uses invariance under the whole group.

The notation \(P(\beta_1,\ldots,\beta_m)\) also applies to \(\mathfrak g\)-valued differential forms. In a fixed basis, insert their coefficients in the displayed order and wedge the scalar forms. This description is independent of the basis by multilinearity. Permuting two inputs of degrees \(a,b\) gives the sign \((-1)^{ab}\). In particular, degree-two forms can be interchanged inside \(P\). Write \(P(F^m)\) for \(P(F,\ldots,F)\). For example, when \(\mathfrak g\) is a matrix Lie algebra and \(p(X)=\operatorname{tr}(X^2)\), one gets
\[
 P(F^2)=\operatorname{tr}(F\wedge F).
\]
The symmetric polarization is useful even for matrix polynomials. The inputs are multiplied as matrices with form coefficients before taking the trace.

\section{Descent and the Bianchi identity}

Let \(Q\to X\) be a smooth right principal \(G\)-bundle. Give it a connection with local potentials \(A_i\) in sections \(s_i\), where \(s_j=s_i g_{ij}\). Write
\[
 F_i=dA_i+\tfrac12[A_i,A_i].
\]
For matrix-valued forms the bracket is the graded commutator of their wedge products, so this formula reads \(dA_i+A_i\wedge A_i\). If \(D_A=d+[A,-]\) is the covariant exterior derivative, then \(D_AF=0\). In a matrix Lie algebra the cancellation is explicit:
\[
 \begin{split}
 dF&=dA\wedge A-A\wedge dA,\\
 [A,F]&=A\wedge dA-dA\wedge A.
 \end{split}
\]
The intrinsic Lie-algebra calculation replaces associativity of the matrix product by the graded Jacobi identity.

\begin{theorem}[Chern--Weil forms]
\label{thm:hdg-cw}
For every \(\operatorname{Ad}(G)\)-invariant polynomial \(p\) of homogeneous degree \(m\), the expressions \(P(F_i^m)\) patch to a closed \(2m\)-form \(p(F)\) on \(X\). The cohomology class \([p(F)]_{\mathrm{dR}}\) depends only on the principal bundle. It is natural under pullback, and the assignment from invariant polynomials to de Rham cohomology is a graded algebra homomorphism when polynomial degree is doubled.
\end{theorem}

\begin{proof}
Since \(F_j=\operatorname{Ad}_{g_{ij}^{-1}}F_i\), invariance identifies the two expressions on each overlap. Equivalently the curvature on \(Q\) is horizontal and equivariant. Applying \(P\) gives a horizontal invariant form on \(Q\), which is the pullback of the form on \(X\). To see the uniqueness of such a descended form, evaluate on lifts of tangent vectors at a point in a fiber. Horizontality makes the lifts immaterial, and invariance makes the fiber point immaterial. Local sections give smoothness.

The derivative of \(P(F_i^m)\) is \(mP(dF_i,F_i^{m-1})\) because \(F_i\) has even degree. The coefficientwise version of \eqref{eq:hdg-inf-invariance}, applied with the coefficients of \(A_i\) and with its scalar one-form placed first, gives
\[
 mP([A_i,F_i],F_i^{m-1})=0.
\]
Consequently \(dP(F_i^m)=mP(D_{A_i}F_i,F_i^{m-1})=0\).

For independence take connections \(A_0,A_1\). Their difference on \(Q\) is horizontal and equivariant, so their affine combination \(A_t=(1-t)A_0+tA_1\) satisfies the connection identities for every \(t\). Use \(A_t\) as a connection on the pullback of \(Q\) to \([0,1]\times X\), with zero potential in the \(t\)-direction. Its curvature in frames constant in \(t\) is
\[
 \widetilde F=F_t+dt\wedge\dot A_t.
\]
Inserting it into \(P\) and dropping terms with two copies of \(dt\) gives
\[
 P(\widetilde F^m)
 = P(F_t^m)+dt\wedge mP(\dot A_t,F_t^{m-1}).
\]
Everything on the right descends. In particular \(\dot A_t\) transforms homogeneously since the transition functions have no \(t\)-dependence. If a closed form is written \(\alpha_t+dt\wedge\beta_t\), closedness in the coefficient of \(dt\) gives \(\partial_t\alpha_t=d_X\beta_t\). Integrating this identity shows
\begin{equation}
 p(F_1)-p(F_0)
  =d_X\left(m\int_0^1 P(\dot A_t,F_t^{m-1})\,dt\right).
 \label{eq:hdg-cw-path}
\end{equation}

The proof for constant polynomials uses the constant form. Pullback commutes both with curvature and with applying a polynomial. Multiplicativity holds already at the level of forms because the entries being substituted have even degree and commute in the exterior algebra. These properties give the assertion about classes.
\end{proof}

The central gluing step of the proof can be expressed on an overlap by the diagram
\[
 \begin{tikzcd}[column sep=large]
 F_i \arrow[r,"\operatorname{Ad}_{g_{ij}^{-1}}"]
      \arrow[dr,"p"'] & F_j\arrow[d,"p"]\\
 &p(F)|_{U_i\cap U_j}.
 \end{tikzcd}
\]
Here each upper entry is a local section of the curvature-form bundle expressed in its own frame. The lower entry is a genuine scalar form.

\section{Two distinctions visible on a torus}

For \(U(1)\), use a real local potential \(a\) with unitary connection matrix \(A=-2\pi i\,a\). Then \(F=-2\pi i\,da\) and \(p(F)=iF/(2\pi)=da\) for the normalized degree-one polynomial. Take the trivial Hermitian line over \(\mathbb R^2/\mathbb Z^2\) with
\[
 a=\sin(2\pi x)\,dy .
\]
Its normalized curvature is
\[
 p(F)=2\pi\cos(2\pi x)\,dx\wedge dy .
\]
This form varies pointwise and is nonzero on open subsets, while its integral is zero. The trivial line admits the zero connection as well, and \eqref{eq:hdg-cw-path} exhibits the exact difference as \(da\). Thus a fixed characteristic class allows different local curvature forms.

On the same torus, a closed real one-form such as \(a=\theta\,dx\) gives curvature zero. A parallel section along a path solves \(ds-2\pi i\,a\,s=0\), so a loop in the first circle direction transports by multiplication by \(\exp(2\pi i\theta)\). These flat connections share zero curvature and a trivial topological line. Their parallel transport distinguishes them when the parameters differ in \(\mathbb R/\mathbb Z\). Differential characteristic classes will retain that information.

\section{Change of group}

For a Lie-group homomorphism \(\rho\colon G\to H\), extension of structure group constructs \(Q\times_G H\) and its induced connection. Its local potentials and curvature are \(\rho_* A_i\) and \(\rho_*F_i\), since \(\rho_*\) preserves brackets. For an invariant polynomial \(q\) on \(\mathfrak h\), the polynomial \(q\circ\rho_*\) on \(\mathfrak g\) is invariant, and
\[
 q(F_{Q\times_G H})=(q\circ\rho_*)(F_Q).
\]
This identity on forms explains how a representation of a principal bundle leads to the characteristic classes of an associated vector bundle. Replacing a representation by a direct sum replaces its curvature by a block sum. The determinant and trace extract different algebraic expressions from that block sum, as the next chapter develops.


\section{A degree-four matrix polynomial}

For a unitary matrix connection let \(H=iF/(2\pi)\). Eliminating the eigenvalues from the degree-two elementary symmetric polynomial gives
\[
 c_2(\nabla)=\tfrac12\big((\operatorname{tr}H)^2-\operatorname{tr}(H\wedge H)\big)
 =\frac{\operatorname{tr}(F\wedge F)-\operatorname{tr}F\wedge\operatorname{tr}F}{8\pi^2}.
\]
The first equality is a polynomial identity, so it holds over the even scalar forms even where diagonalization would make no sense. On an \(SU(r)\)-bundle the connection and curvature are traceless and only \(\operatorname{tr}(F\wedge F)/(8\pi^2)\) remains. On a compact oriented four-manifold the normalization therefore turns an integral of a quadratic local expression into an integral topological number. Chapter~\ref{ch:characteristic} justifies the normalization through line bundles. Chapter~\ref{ch:chern-simons} compares two such local representatives by an explicit three-form.
\chapter{Chern, Pontryagin, and Euler classes}
\label{ch:characteristic}

The invariant polynomials of Chapter~\ref{ch:chern-weil} yield many closed forms. A normalization selects those corresponding to the named topological classes. The comparison uses a complex line, where the period can be computed, and a flag construction, where the cohomology pullback is injective. Integral classes also contain torsion, which the computation with real forms cannot see~\cite[\S25]{tu}.

\section{Unitary curvature and Chern forms}

Let \(E\) be a complex vector bundle of rank \(r\) with Hermitian metric and unitary connection. In a unitary frame its curvature \(F\) is skew-Hermitian as an endomorphism-valued two-form. Define the total Chern form by
\[
 c(\nabla)=\det\left(I+\frac{i}{2\pi}F\right)
          =1+c_1(\nabla)+\cdots+c_r(\nabla).
\]
The determinant is taken over the commutative algebra of scalar even-degree forms. Its homogeneous term containing \(j\) curvature factors has degree \(2j\). Conjugation invariance allows Chapter~\ref{ch:chern-weil} to be applied. The coefficients are real because the matrix \(iF\) equals its conjugate transpose, and the determinant is unchanged by transpose. All higher Chern forms vanish by rank.

For two bundles, the direct-sum connection has block diagonal curvature. Consequently its Chern form is the product of the two original Chern forms. A metric on a short exact sequence of smooth complex bundles identifies the middle bundle with the direct sum of the ends by taking the orthogonal complement of the subbundle. Together with connection independence this yields the Whitney product identity for the real classes. The corresponding integral Whitney identity is the standard topological one~\cite[\S21]{bott-tu}.

The Chern character uses an additive invariant polynomial in each degree:
\[
 \operatorname{ch}(\nabla)=\operatorname{tr}
 \exp\!\left(\frac{i}{2\pi}F\right).
\]
Only finitely many terms survive on a finite-dimensional base. The constant term is \(r\); the terms of degrees two and four are \(c_1(\nabla)\) and \(\tfrac12(c_1(\nabla)^2-2c_2(\nabla))\). These formulas follow from the Newton identities over the algebra of forms. On \(E\otimes E'\), the induced curvature is \(F\otimes I+I\otimes F'\), with commuting summands. The exponential factors, and the trace of the tensor product is the product of the traces. This gives multiplicativity for the Chern character under tensor products and additivity under direct sums.

\section{The normalization on a projective line}

Let \(L=\mathcal O(-1)\to\mathbb{CP}^1\) be the tautological line with its metric induced from \(\mathbb C^2\). On the affine chart \([1:z]\) the holomorphic frame \((1,z)\) has squared norm \(h=1+|z|^2\). The Chern connection in this frame has potential \(A=\partial\log h\). This formula follows from two conditions: its \((0,1)\) part is ordinary \(\bar\partial\) in a holomorphic frame, and preserving the metric requires \(dh=h(A+\overline A)\). Taking the \((1,0)\) part gives the claimed expression. Although the potential in this frame need not be imaginary, it becomes imaginary in the corresponding unitary frame. Its curvature is independent of the frame:
\[
 A=\frac{\overline z\,dz}{1+|z|^2},
 \qquad
 F=\frac{d\overline z\wedge dz}{(1+|z|^2)^2}.
\]
The normalization integrates to
\[
 \int_{\mathbb{CP}^1}\frac{iF}{2\pi}
   =-\frac1\pi \int_{\mathbb R^2}
        \frac{dx\,dy}{(1+x^2+y^2)^2}
   =-1 .
\]
The integral is computed in polar coordinates, since the omitted point has measure zero for a smooth top form. The bundle's topological first Chern class has the same sign: the dual line \(\mathcal O(1)\) has a section defined by a nonzero linear functional on \(\mathbb C^2\), with one positively oriented transverse zero. Its Euler number and hence its degree are \(1\). The duality of lines reverses the Chern class. One can also recover the sign by using the transition \(z^{-1}\) from the frame \((1,z)\) to the frame \((z^{-1},1)\).

The standard comparison with integral Chern classes reduces to this calculation~\cite[Chapter~IV]{bott-tu}. For a line bundle, naturality applied to the universal line over projective space reduces normalization to the projective-line inclusion, which detects its degree-two generator. For a bundle \(E\), pass to the full flag bundle \(\pi\colon\operatorname{Flag}(E)\to X\). The successive projectivizations have cohomology free over the cohomology of their bases by the projective bundle theorem. The unit is one of the module generators, so each pullback is injective. In particular \(\pi^*\) is injective in real cohomology. The pullback of \(E\) has a flag with line-bundle quotients; choosing a metric splits the filtration as a direct sum of these lines. On the flag bundle, both the real image of the topological total Chern class and the curvature-defined total class are the product of \(1+c_1\) of the line factors. Injectivity proves equality downstairs.

\section{Real bundles and Pontryagin classes}

For a real bundle \(V\), set
\[
 p_j(V)=(-1)^j c_{2j}(V\otimes_{\mathbb R}\mathbb C)
       \in H^{4j}(X;\mathbb Z).
\]
A metric connection on \(V\) complexifies to a unitary connection, so the formula gives a Chern--Weil representative for the real image. In a real orthonormal frame \(F^{\mathsf T}=-F\), and transposing the matrix in \(\det(I+iF/(2\pi))\) shows that this polynomial is even in \(F\). Thus the odd-index Chern forms of the complexification vanish. The integral odd-index Chern classes may still be two-torsion. Indeed conjugation identifies the complexification with its conjugate, and \(c_k(\overline E)=(-1)^k c_k(E)\). It follows that the Pontryagin total class satisfies the product formula after inverting \(2\). Over the integers, products of two odd-index classes in the Whitney formula can contribute two-torsion~\cite[\S15]{milnor-stasheff}.

For an explicit torsion calculation, let \(\lambda\) be the tautological real line on \(\mathbb{RP}^4\) and let \(u=w_1(\lambda)\). The real bundle underlying \(\lambda_{\mathbb C}\) is \(\lambda\oplus\lambda\), whose second Stiefel--Whitney class is \(u^2\). Thus \(a=c_1(\lambda_{\mathbb C})\) reduces to \(u^2\), so \(a^2\) reduces to \(u^4\ne0\). For \(V=\lambda\oplus\lambda\) the integral \(p_1(V)=-a^2\) is nonzero, whereas \(p_1(\lambda)=0\) by rank. The transition functions can be chosen to take values in \(\{\pm1\}\), and the two line summands admit flat connections. Their real characteristic forms vanish. This example exhibits both the two-primary qualification on the Whitney identity and the failure of real forms to detect torsion.

\section{Orientation and the Pfaffian}

For an oriented real bundle of rank \(2m\), take an oriented orthonormal frame. Define the Euler form
\[
 e(\nabla)=\operatorname{Pf}\left(\frac{F}{2\pi}\right).
\]
We use the Pfaffian with \(\operatorname{Pf}\left(\begin{smallmatrix}0&a\\-a&0\end{smallmatrix}\right)=a\). One definition is the coefficient of the ordered top exterior basis vector in the \(m\)-th divided power of \(\sum_{i<j}F_{ij}e_i\wedge e_j/(2\pi)\). Conjugation by an orthogonal matrix multiplies this coefficient by its determinant. The polynomial is therefore invariant under \(SO(2m)\), and reversing orientation changes its sign. On a complex line with its complex orientation, the skew matrix for multiplication by \(i\) is \(\left(\begin{smallmatrix}0&-1\\1&0\end{smallmatrix}\right)\). For \(F=i\,d\alpha\) this gives \(e(\nabla)=-d\alpha/(2\pi)=c_1(\nabla)\). The Pfaffian product on oriented block sums then implies equality of the Euler form and the top Chern form for the realification of a unitary bundle. This can be checked on block diagonal matrices with real scalar entries, which determine the restricted invariant polynomial on \(\mathfrak u(m)\) by unitary diagonalization. Polarization extends the equality to matrices of two-forms.

The comparison of the Pfaffian class with the real image of the integral Euler class is the Euler Chern--Weil theorem. The latter class is obtained by pulling the Thom class back along the zero section~\cite[\S25]{tu}. In odd rank, negation of each fiber reverses orientation and fixes the zero section. It follows that the integral Euler class is killed by \(2\). Its real image is zero.

Take \(E=\mathcal O(1)\oplus\mathcal O(1)\) on \(\mathbb{CP}^2\) and let \(h=c_1(\mathcal O(1))\). The top Chern class is \(h^2\). Since \((\mathcal O(1))_{\mathbb R}\otimes\mathbb C\cong\mathcal O(1)\oplus\mathcal O(-1)\), the first Pontryagin class of \(E_{\mathbb R}\) is \(2h^2\). Two independent linear functionals give hyperplanes intersecting in one positively oriented point, so \(\int_{\mathbb{CP}^2}h^2=1\). The resulting Euler and Pontryagin numbers are \(1\) and \(2\), respectively, for the same oriented real bundle. They detect different invariant polynomials of its curvature.
\chapter[Chern--Simons transgression]{Chern--Simons forms and coherent transgression}
\label{ch:chern-simons}

A form that witnesses the equality of two de Rham classes contains additional information. It records how the two curvature representatives are related, and it changes when the comparison is changed. Chern and Simons used this principle to construct secondary invariants~\cite{chern-simons}. Here we make the comparison depend on a simplex of connections, so comparisons between comparisons have their own differential identities.

\section{An oriented path of connections}

Let \(P\) be an invariant symmetric \(m\)-linear form with \(m\geq1\), and let \(A_0,A_1\) be connections of a fixed principal bundle \(Q\to X\). Set \(B=A_1-A_0\) and \(A_t=A_0+tB\). Define the relative transgression by
\[
 T_1(A_0,A_1)=m\int_0^1P(B,F_t^{m-1})\,dt.
\]
Although written in a frame, the integrand is a global form because \(B\) and \(F_t\) transform by simultaneous conjugation. Its exterior derivative is \(P(F_1^m)-P(F_0^m)\) by \eqref{eq:hdg-cw-path}. A second proof differentiates the curvature formula:
\[
 \dot F_t= D_{A_t}B.
\]
The derivative of \(P(F_t^m)\) is \(mP(D_{A_t}B,F_t^{m-1})\). Invariance converts covariant differentiation to exterior differentiation inside \(P\). Terms differentiating \(F_t\) vanish by Bianchi, so this derivative equals \(m\,d_X P(B,F_t^{m-1})\).

For the quadratic polynomial \(p(X)=\operatorname{tr}(X^2)\) on a matrix Lie algebra, expand
\[
 F_t=F_0+tD_{A_0}B+t^2 B\wedge B.
\]
Integration gives the three-form
\[
 T_1(A_0,A_1)
  =\operatorname{tr}\!\left(2B\wedge F_0
    +B\wedge D_{A_0}B+\tfrac23 B\wedge B\wedge B\right).
\]
On a trivial bundle with the reference connection \(A_0=0\), this reduces to the usual local expression \(\operatorname{tr}(A\wedge dA+\tfrac23 A^{\wedge 3})\). The definition of \(T_1\) also makes sense when neither connection has a global potential, since it uses a reference connection on the same bundle.

\section{The sign from integration over a simplex}

Orient the simplex with ordered vertices \(0,\ldots,q\) by coordinates \(t_1,\ldots,t_q\), with \(t_0=1-\sum_{i=1}^q t_i\). On \(\Delta^q\times X\) use the connection \(\widetilde A=\sum_{i=0}^q t_i A_i\) in the \(X\) directions. It has zero components in the simplex directions in local frames pulled back from \(X\). For a form \(\omega\) on the product, let \(I_q\omega\) mean the integral of the component with simplex degree \(q\), with all simplex differentials written first. Its degree on \(X\) is \(\deg(\omega)-q\). For a closed \(\omega\),
\begin{equation}
 d_X I_q\omega=(-1)^{q+1}I_{\partial\Delta^q}\omega.
 \label{eq:hdg-fiber-stokes}
\end{equation}
Indeed, write \(\omega=\sum_p\omega_p\) by simplex degree. In degree \(q\), the closedness equation is the sum of the simplex derivative of \(\omega_{q-1}\) and \((-1)^q\) times the \(X\)-differential of the coefficients of \(\omega_q\). Stokes on the parameter simplex gives \eqref{eq:hdg-fiber-stokes}.

Apply this to the closed form \(P(\widetilde F^m)\) and define, for \(q\geq0\),
\begin{equation}
 T_q(A_0,\ldots,A_q)
  =(-1)^{q(q-1)/2} I_q P(\widetilde F^m).
 \label{eq:hdg-transgression-simplex}
\end{equation}
At \(q=0\) this is \(P(F_0^m)\). Each oriented face pulls the connection back to the affine family defined by the surviving vertices. Since \((-1)^{q(q-1)/2}(-1)^{q+1}=(-1)^{(q-1)(q-2)/2}\), the sign convention gives
\begin{equation}
 d_X T_q(A_0,\ldots,A_q)
  =\sum_{i=0}^q(-1)^iT_{q-1}(A_0,\ldots,\widehat A_i,\ldots,A_q).
 \label{eq:hdg-transgression-desc}
\end{equation}
The hat omits the indicated connection. The case \(q=2\) is the relation
\[
 T_1(A_0,A_1)+T_1(A_1,A_2)-T_1(A_0,A_2)
 = d_XT_2(A_0,A_1,A_2).
\]
It says that making the comparison in two steps and taking the direct comparison differ by an exact form. For a polynomial of degree \(m\), the form \(T_q\) vanishes for \(q>m\), since each curvature factor carries at most one simplex differential. The relations still hold at those levels because the boundary sum in \eqref{eq:hdg-transgression-desc} then vanishes.

\section{A two-parameter calculation}

For \(q=2\), put \(B_i=A_i-A_0\) for \(i=1,2\). The curvature of the affine family consists of \(F_{\boldsymbol t}\) and the terms \(dt_i\wedge B_i\). Each product of the two mixed curvature terms contributes a minus sign when \(dt_2\) is moved past \(B_1\). The prefactor in \eqref{eq:hdg-transgression-simplex} is also negative, so for \(m\geq2\),
\[
 T_2(A_0,A_1,A_2)=m(m-1)\int_{\Delta^2}
 P(B_1,B_2,F_{\boldsymbol t}^{m-2})\,dt_1\,dt_2 .
\]
If \(m=2\) the integral is over a coefficient independent of the parameter, and \(\operatorname{area}(\Delta^2)=1/2\) in these coordinates. Thus \(T_2=P(B_1,B_2)\). With \(p=c_1^2\) and \(A_i=-2\pi i\,a_i\) for a unitary line, write \(b_i=a_i-a_0\), which is a globally defined real one-form. The identity becomes \(T_2=b_1\wedge b_2\). Expanding the closed curvature expressions directly gives
\[
 T_1(A_0,A_1)+T_1(A_1,A_2)-T_1(A_0,A_2)
   =db_1\wedge b_2-b_1\wedge db_2 .
\]
The two sides agree with \(d(b_1\wedge b_2)\), confirming the orientation choice in a case where the two-form can be nonzero.

\section{Gauge transformations and periods}

Let \(X\) be closed and oriented of dimension \(2m-1\). Suppose \(P(F^m)\) is normalized so that it represents the real image of an integral class on every principal \(G\)-bundle under consideration. A gauge automorphism \(u\colon Q\to Q\) produces \(A_1=u^*A_0\). Choose a path from \(A_0\) to \(A_1\) constant near the endpoints. Gluing the product bundle along \((1,p)\sim(0,u(p))\) gives a principal bundle on \(S^1\times X\). The endpoint conditions make the family of connections descend smoothly. Orient \(S^1\times X\) with the parameter first. Integrating its characteristic form yields
\[
 \int_X T_1(A_0,u^*A_0)
   = \int_{S^1\times X} P(\widetilde F^m) \in \mathbb Z .
\]
Here \(T_1\) can be computed using the path with constant ends. A monotone reparametrization of the affine path gives one such path and leaves its integral unchanged. More generally the transgression for any two smooth paths with the same endpoints differs by an exact form, by Stokes applied to an interpolating square in the affine space of connections. This period identity uses both the closedness of \(X\) and the integral normalization of the polynomial. A choice of an arbitrary multiple of a real polynomial changes the period group.

For \(P=c_1^2\), the comparison of a line connection under a gauge map \(u\colon X\to U(1)\) can also be computed on \(X\) itself. Use the normalized forms
\[
 f=\frac{iF}{2\pi},\qquad
 \beta=\frac{i}{2\pi}u^{-1}du .
\]
The comparison gives \(T_1=2\beta\wedge f\). For \(X=S^1\times S^2\), take \(u(t,x)=e^{-2\pi i t}\) and let the line be pulled back from a degree-one line on \(S^2\). Then \(\beta=dt\) and the period is \(2\). The mapping-torus argument explains its integrality without any choice of a global potential. Later, additive differential cohomology will package curvature and secondary period data in a single object.


\section{A secondary value from a filling}

Let \(X=\partial Z\) with its boundary orientation, and assume that a bundle with connection on \(X\) extends over the compact oriented \(2m\)-manifold \(Z\). For the same integral normalization, define a number in \(\mathbb R/\mathbb Z\) by
\[
 \mathcal S(X,Q,A)=\left[\int_Z P(F^m)\right]\pmod{\mathbb Z}.
\]
Two choices of connections on a fixed extension with equal boundary restriction yield the same integral: their difference is the integral of \(dT_1\), and the restricted \(T_1\) is zero. By this argument, one may arrange that the connection is constant in the normal direction in a collar. Two different fillings can then be glued using the boundary bundle isomorphism, with the second filling oppositely oriented. The difference of integrals is the characteristic number of the glued closed manifold, an integer. This proves independence of the filling whenever at least one filling exists.

This construction determines a change formula. Suppose that both boundary connections \(A_0,A_1\) occur as ends of a cylinder on the same bundle and the first bounds with its connection. Attach the cylinder with its product orientation, with the interval first, to obtain a filling of the second. The two values then differ by \(\int_X T_1(A_0,A_1)\) modulo integers. For the normalized first Chern form on a disk carrying a trivialized line connection \(A=-2\pi i\,a\), the filling rule gives \(\exp(2\pi i\int_D da)=\exp(2\pi i\int_{\partial D}a)\), the actual boundary holonomy by Stokes. A differential character extends this relationship to cycles without requiring a chosen filling.
\chapter[Equivariant curvature]{The Weil algebra and equivariant curvature}
\label{ch:weil}

A principal connection assigns forms to its potential and curvature, subject to the structure equation and Bianchi's equation. The Weil algebra encodes those two equations before a bundle or a base has been chosen. To describe geometry with an extra group action, one combines this algebra with the forms of the space acted upon. The Cartan model results from eliminating the formal connection variables. The comparison is classical~\cite[Chapters~3--5]{g-s-continuous}.

\section{Formal generators and their basic subalgebra}

Let \(K\) be a Lie group with finite-dimensional real Lie algebra \(\mathfrak k\). The Weil algebra as a graded commutative algebra is
\[
 W(\mathfrak k)=
 \Lambda(\mathfrak k^*[1])\otimes
 \operatorname{Sym}(\mathfrak k^*[2]).
\]
Write \(\theta\) and \(u\) for the tautological \(\mathfrak k\)-valued expressions in its generators of degrees \(1\) and \(2\), respectively. Define the differential by
\[
 d\theta=u-\tfrac12[\theta,\theta],\qquad
 du=-[\theta,u].
\]
The structure constants are those of \(\mathfrak k\), with graded signs from the exterior product. Extending \(d\) by the Leibniz rule gives \(d^2=0\) by the graded Jacobi identity. For example,
\[
 d^2u=-[u,u]+\tfrac12[[\theta,\theta],u]-[\theta,[\theta,u]]=0.
\]
The first term vanishes because \(u\) has even degree, and Jacobi identifies the other two terms.

For \(\xi\in\mathfrak k\) the contraction \(\iota_\xi\) is the degree \(-1\) derivation with \(\iota_\xi\theta=\xi,\ \iota_\xi u=0\). Its commutator with \(d\) is the infinitesimal coadjoint action on the generators, expressed as \(L_\xi\theta=-[\xi,\theta]\) and \(L_\xi u=-[\xi,u]\) in the tautological expressions. A form is horizontal if every contraction annihilates it, and basic if it is horizontal and invariant under \(K\). The horizontal subalgebra consists of polynomials in \(u\): the contractions are all derivatives with respect to the exterior generators. Thus
\[
 W(\mathfrak k)_{\mathrm{bas}}
   =\big(\operatorname{Sym}(\mathfrak k^*)\big)^K,\qquad d|_{\mathrm{bas}}=0 .
\]
The differential vanishes there by the infinitesimal invariance identity of Chapter~\ref{ch:chern-weil}. Full invariance under the group accounts for disconnected components.

The cohomology of the entire Weil algebra is \(\mathbb R\) in degree zero. Indeed, replacing the degree-two generators \(u^a\) by \(v^a=d\theta^a\) is an invertible polynomial change of variables. In the new generators \(d\theta^a=v^a\) and \(dv^a=0\). The degree \(-1\) operator \(h=\sum_a\theta^a\partial/\partial v^a\) has \(dh+hd\) equal to the operator that counts all factors of \(\theta\) and \(v\) in a monomial. On the positive-count part, divide \(h\) by the factor count to obtain a contraction. This explains how the basic algebra can carry nontrivial classes although the unqualified algebra is acyclic in positive degrees. Primitives in the whole algebra can fail the basic condition.

Given a right principal \(K\)-bundle with connection, evaluation \(\theta\mapsto A,\ u\mapsto F\) is a morphism of differential graded algebras into forms on the total space. It preserves contractions by the reproduction property for \(A\) and the horizontality of \(F\). It also preserves group actions. Its restriction to basic forms recovers the Chern--Weil homomorphism on representatives discussed in Chapter~\ref{ch:chern-weil}.

\section{Forms on a space with a group action}

Let \(X\) have a smooth right \(K\)-action. Use \(\xi_X(x)=\frac{d}{dt}(x\cdot\exp(t\xi))|_0\) for a fundamental vector field. On \(W(\mathfrak k)\otimes\Omega(X)\) take diagonal contractions and the diagonal group action, using the graded tensor signs. The basic subcomplex is the Weil model. The alternative Cartan model is
\[
 \Omega_K(X)
  =\big(\operatorname{Sym}(\mathfrak k^*)\otimes\Omega(X)\big)^K,
 \qquad
 (d_K\alpha)(\xi)=d\big(\alpha(\xi)\big)-\iota_{\xi_X}\alpha(\xi),
\]
with polynomial generators in degree two. The equivariance convention is
\[
 R_k^*\alpha(\xi)=\alpha(\operatorname{Ad}_k\xi).
\]
Cartan's identity gives \(d_K^2\alpha(\xi)=-L_{\xi_X}\alpha(\xi)=0\), since the one-parameter subgroup generated by \(\xi\) fixes \(\xi\) under the adjoint action.

Here is the comparison at the level of complexes. Choose dual bases \(e_a\) of \(\mathfrak k\) and \(\theta^a\) of degree-one generators. Put \(T=\sum_a\theta^a\iota_{e_a}^X\) on the tensor product, with contractions on the second factor carrying the usual Koszul sign. This degree-zero derivation is nilpotent because it adds one exterior generator at each application. Its exponential is finite. One checks
\[
 [\iota_\xi^{\mathrm{tot}},T]=\iota_\xi^X,\qquad
 [\iota_\xi^X,T]=0 .
\]
The first identity follows by differentiating \(\theta^a\) and the second uses anticommutation of contractions. It follows that \(\exp(-T)\) sends elements independent of \(\theta\) to horizontally constrained elements, and its inverse sends horizontal elements back to ones independent of \(\theta\). The operator \(T\) is invariant by the pairing of a basis with its dual. Conjugating the differential and then setting \(\theta=0\), its only terms that survive are the coefficient differential \(d_X\) and \(-\sum_a u^a\iota_{e_a}^X\). The latter comes from differentiating the first-order term of \(\exp(-T)\), since the degree-two term of \(d\theta\) is \(u\). This proves the Cartan differential formula.

For compact \(K\) the equivariant de Rham theorem identifies the cohomology of this complex with \(H^*(EK\times_K X;\mathbb R)\)~\cite[Chapters~1,4--5]{g-s-continuous}. This comparison invokes the compact-group hypotheses. The algebraic identification of the Weil and Cartan complexes above is a separate calculation. A useful geometric evaluation is obtained from any right principal \(K\)-bundle \(R\to S\) with connection and an equivariant map \(R\to X\). Evaluate its Weil generators on the bundle and its \(X\)-forms by pullback. The resulting forms on \(R\) are basic and descend to \(S\). Thus an equivariant cohomology class yields ordinary cohomology classes for these bundles equipped with a map to \(X\). This pattern will reappear as a characteristic map on a quotient stack.

\section{The moment term of an invariant connection}

Let \(Q\to X\) be a right principal \(G\)-bundle and suppose the right \(K\)-action on \(X\) lifts to \(Q\) commuting with the \(G\)-action. Choose a \(K\)-invariant connection \(A\); when \(K\) is compact one can average any connection using normalized Haar measure. For \(\xi\in\mathfrak k\) put
\[
 \mu(\xi)=\iota_{\xi_Q}A,\qquad F_K(\xi)=F-\mu(\xi).
\]
The term \(\mu\) is an equivariant \(\mathfrak g\)-valued function on \(Q\), linear in the parameter, so both summands of \(F_K\) have total degree two. Our sign follows from Cartan's identity and \(L_{\xi_Q}A=0\):
\[
 \iota_{\xi_Q}F
 =-d\mu(\xi)+[\mu(\xi),A]=-D_A\mu(\xi).
\]
Consequently \((D_A-\iota_{\xi_Q})F_K(\xi)=0\), by the ordinary Bianchi identity. Each \(P(F_K^m)\) is \(G\)-basic, so it descends to an equivariant form on \(X\). Invariance of \(P\) converts \(D_A\) to \(d\) after polarization. Therefore the equivariant form is \(d_K\)-closed. Independence of the invariant connection follows by taking an affine path \(A_t\) and differentiating:
\[
 \partial_t F_{K,t}=(D_{A_t}-\iota_{\xi_Q})\dot A_t.
\]
The same polarization argument as for \eqref{eq:hdg-cw-path} expresses the derivative of \(P(F_{K,t}^m)\) as a Cartan coboundary.

For a concrete calculation take \(X=\ast\) and \(K=G=U(1)\). On \(Q=U(1)\) use \(A=z^{-1}dz\), and lift the \(K\)-action by \(z\cdot k=z k^m\) for an integer \(m\). Write the real Lie-algebra coordinate as \(\xi=2\pi i\,s\). Then \(F=0\) and \(\mu(\xi)=2\pi i\,ms\). The equivariant first Chern polynomial is
\[
 \frac{i}{2\pi}(F-\mu(\xi))=ms.
\]
The degree-two generator of the Cartan model of the point is normalized here by the weight-one action. Taking higher weights multiplies this generator even though the ordinary curvature on the point is zero. The moment term records the isotropy action that ordinary forms on the base cannot record.


\section{Free actions and the sign of the universal generator}

For a free compact \(K\)-action with smooth principal quotient \(X\to X/K\), the projection \(EK\times_KX\to X/K\) has contractible fiber and is a homotopy equivalence. The Cartan map can be evaluated using a connection on the principal quotient and the identity equivariant map \(X\to X\). In the Weil description its evaluation sends \(u\) to the curvature and uses the horizontal part of the coefficient forms. The horizontal projection arises from \(\exp(-T)\): its degree-one effect on a one-form \(\alpha\) is \(\alpha-\sum_a A^a\alpha((e_a)_X)\), and multiplicativity gives the projection on higher forms. Changing the connection gives the same cohomology map by applying the path construction on the quotient times an interval.

For the Hopf circle bundle \(S^3\to\mathbb{CP}^1\), a point over a line is a unit vector in that line. The right circle action is multiplication of the vector by the circle coordinate. Its associated line for the standard representation is \(\mathcal O(-1)\), using \((qg,v)\sim(q,gv)\). If \(A\) is the corresponding circle connection, the variable \(s\) in the preceding weight example evaluates to \(F/(2\pi i)=-iF/(2\pi)\). Its integral over \(\mathbb{CP}^1\) is therefore \(1\) by Chapter~\ref{ch:characteristic}. This also explains the sign for right actions in the quotient construction. A separate \(K\)-action \(q\cdot k=qk^m\) on a circle torsor over a point, pulled back along the principal Hopf bundle, gives the diagonal relation \((r k,qk^m)\sim(r,q)\). The associated character on the underlying Hopf bundle is \(k^{-m}\), so its first Chern number is \(m\), exactly the evaluation of \(ms\).

\bookpart{Smooth stacks and characteristic maps}
\chapter{The smooth site and its higher sheaves}
\label{ch:smooth-site}

A connection can be described in every coordinate chart of its base. The description includes transition functions and their equations. Keeping the transitions is essential. If one retains only the isomorphism class of the object on each chart, the information needed to assemble the object over the whole base can disappear. We now make the difference between these two procedures precise.

\section{Test manifolds and local data}
\label{sec:smooth-tests}

Let \(\mathbf{Mfd}\) denote a category of smooth Hausdorff second-countable finite-dimensional manifolds without boundary, in a fixed universe. We use open covers to define a Grothendieck topology. The later arguments about bundles also use surjective submersions. Such a map has local smooth sections by the submersion theorem, so it is a cover for the topology generated by open covers. When integrating a family, a boundary hypothesis will be stated explicitly.

For a manifold \(X\), put \(yX(U)=C^\infty(U,X)\). This is a sheaf of sets because smoothness is local on the source. If \(\mathcal F\) is a presheaf, the action of a map \(v:V\to U\) goes in the direction \(\mathcal F(U)\to\mathcal F(V)\). For differential forms it is the ordinary pullback \(v^*\). The sheaf \(\Omega^q\) assigns to \(U\) its real \(q\)-forms, and \(\Omega_{\mathrm{cl}}^q\) assigns the kernel of \(d:\Omega^q(U)\to\Omega^{q+1}(U)\). A family of forms on \(U_i\) glues uniquely when its two restrictions to each \(U_i\cap U_j\) agree. Checking the exterior derivative locally shows that the glued form is closed when the given forms are closed.

On a cover \(\{U_i\to U\}\), write \(U_{i_0\cdots i_p}=U_{i_0}\cap\cdots\cap U_{i_p}\). The Čech nerve has at level \(p\) the union of the \(U_{i_0\cdots i_p}\), with a separate copy for every tuple. Its face maps omit one index and its degeneracies repeat an index. If necessary we use formal coproducts of representable sheaves for this construction. At the first two levels this is the diagram
\[
  \coprod_{i,j}yU_{ij}\;\rightrightarrows\;
  \coprod_i yU_i\;\longrightarrow\;yU .
\]
For a sheaf of sets the limit test amounts to the equalizer at levels zero and one. The higher levels record identities already forced by equality of restrictions.

\section{Descent with isomorphisms}
\label{sec:smooth-groupoid-descent}

A presheaf of groupoids has objects and their isomorphisms over each \(U\). We allow restriction functors associative up to specified coherent natural isomorphisms. This is how pullback of bundles behaves without choosing a strict model. The descent groupoid over a cover has as objects \(a_i\in\mathcal F(U_i)\) and isomorphisms
\(\phi_{ij}:a_j|_{U_{ij}}\to a_i|_{U_{ij}}\). We suppress the canonical pullback isomorphisms from the notation. The equations are
\[
 \phi_{ii}=1,\qquad \phi_{ij}\phi_{jk}=\phi_{ik}.
\]
A morphism is a family \(f_i:a_i\to b_i\) with
\(f_i\phi_{ij}=\psi_{ij}f_j\). Composition and inversion are componentwise. The restriction functor from \(\mathcal F(U)\) to this groupoid is fully faithful exactly when isomorphisms between global objects glue uniquely. It is essentially surjective exactly when compatible local objects can be glued, up to isomorphism. A stack satisfies both tests.

\begin{proposition}[Vector bundles satisfy descent]
Let \(\mathcal V_r(U)\) be the groupoid of real rank \(r\) smooth vector bundles over \(U\) and bundle isomorphisms over its identity. Then \(\mathcal V_r\) is a stack.
\end{proposition}

\begin{proof}
Given descent data, take the disjoint union of the local total spaces \(E_i\). Over a base point in \(U_{ij}\), identify \(v_j\) with \(\phi_{ij}(v_j)\). The equations make this an equivalence relation. Every equivalence class over a point of \(U_i\) has exactly one representative in \(E_i\). Give the quotient the topology for which the quotient charts from \(E_i\) over \(U_i\) are open, and use the vector bundle charts of \(E_i\) as smooth trivializations. The identifications on overlaps are smooth and linear, so they supply the change of coordinates. Points over different base points have disjoint neighborhoods by Hausdorffness of \(U\). Points over the same base point lie in one of the open bundle charts and are separated there. With a countable refinement of the base cover and countable trivializations this gives the usual second-countable total space. Applying a compatible family of local isomorphisms to each representative gives the unique global isomorphism. This verifies both parts of descent.
\end{proof}

This proof does not identify nonisomorphic gluing choices. For example, cover a circle by two connected arcs whose intersection has two components. Take trivial real lines on the two arcs. Gluing by the sign \(+1\) on one component and \(-1\) on the other produces the Möbius line. A change of local frames consists of two continuous nonzero functions, each with a constant sign on its arc. Their sign ratio is therefore the same on both components. It cannot convert the chosen transition to the identity on both. Consequently a presheaf keeping just isomorphism classes is not separated: the Möbius line and the trivial line give different classes on the circle and identical classes on all chart domains. The different transition signs are precisely the information missing from the local isomorphism classes.

\section{From groupoids to spaces}
\label{sec:smooth-higher-descent}

The nerve of a groupoid is a space with no homotopy above degree one. Its points are objects, paths represent isomorphisms, and the composition law determines the higher simplices. The groupoid of descent data above models a homotopy limit of the Čech diagram of nerves. One way to see this is to read the data of a coherent cone over the diagram. Level zero specifies the \(a_i\). Level one supplies paths between their restrictions. A triangle filling in level two is the displayed composition equation, since the mapping spaces between fixed objects of a groupoid nerve are discrete. Higher coherence contains no new independent choices once the equations hold.

For sheaves of arbitrary spaces we use the corresponding homotopy-limit condition
\[
  \mathcal F(U)\simeq
  \operatorname*{holim}_{[p]\in\mathbf\Delta}
                \prod_{i_0,\ldots,i_p}\mathcal F(U_{i_0\cdots i_p}).
\]
In this case an agreement may itself have nontrivial higher paths. A weakly associative system of transition maps can therefore carry additional information in the fillers and their compatibilities. We use \(\mathcal H\) for this \(\infty\)-category of smooth sheaves of spaces. It is an \(\infty\)-topos by the general theory of sheaves of spaces on a site \cite[Chapter~6]{lurie-htt}. In these chapters descent refers to cover descent unless a hypercover is specified. Degree-one descriptions only require groupoid descent. For the bounded differential complexes used later, the constructions can be checked on refinements by good covers of manifolds.

Representables remain sheaves, and the Yoneda equivalence has the form
\[
  \operatorname{Map}_{\mathcal H}(yU,\mathcal F)\simeq \mathcal F(U).
\]
In particular, a natural operation on families can be written as a morphism of sheaves of spaces. Pullback along a map of test spaces is built into the morphism.

\begin{example}[A homotopy fiber retains an identification]
Let \(F:\mathcal X\to\mathcal Y\) be a functor between groupoids and \(y\) an object of \(\mathcal Y\). A model of the homotopy fiber has objects \((x,\alpha:F(x)\cong y)\) and arrows \(u:x\to x'\) satisfying \(\alpha'F(u)=\alpha\). The square of nerves is a homotopy pullback. To check the assertion, replace the target point by the groupoid of arrows with target \(y\), with morphisms the commuting triangles. This groupoid is contractible, and its source projection to \(\mathcal Y\) lifts isomorphisms. The ordinary pullback along that projection is precisely the indicated groupoid. An automorphism of \(x\) survives in the fiber if \(F(u)=1\). We will use this test to fix the underlying bundle while allowing its connection to vary.
\end{example}

\section{Local presentations of a moduli object}
\label{sec:smooth-atlases}

A presentation by an atlas starts with a map \(yX\to\mathcal X\) admitting local lifts of points in the sense of sheaves. Its relation object is the homotopy fiber product \(yX\times^h_{\mathcal X}yX\). A point of this relation has two lifts and an identification between their images. Iterating the construction gives the Čech nerve of the atlas. When the map is an effective epimorphism, the geometric realization of that nerve is \(\mathcal X\). The local lifting condition is the ordinary criterion for effective epimorphisms of sheaves of spaces in terms of locally surjective maps on components. In the groupoid case the assertion follows directly by gluing the chosen lifts and the isomorphisms on their overlaps.

The atlas by itself supplies a collection of local representatives. Curvature and characteristic forms computed on those representatives must be compatible with the full relation object in order to define natural operations. This is one reason to distinguish basic forms on a principal bundle, differential forms on a chosen chart, and a cohomology class on the stack. The next chapter works out the relation object for torsor classification.
\chapter{The classifying stack of smooth torsors}
\label{ch:classifying}

The familiar classifying space \(BG\) provides information about the isomorphism class of a topological \(G\)-bundle. A smooth classifying stack records bundles, their isomorphisms, and smooth pullback along test maps. This refinement lets us express a construction such as extension of structure group by an actual map of moduli objects.

\section{The stack \texorpdfstring{\(\mathbf BG\)}{BG}}
\label{sec:bg}

Let \(G\) be a Lie group, regarded as the sheaf \(U\mapsto C^\infty(U,G)\). The one-object groupoid with automorphism sheaf \(G\) gives a presheaf of groupoids \(\mathcal P_G\). Its stackification is denoted \(\mathbf BG\). Stackification permits local objects with descent isomorphisms to define global objects. Thus, for a cover \(\{U_i\}\), an object is represented by smooth functions \(g_{ij}:U_{ij}\to G\) satisfying
\[
 g_{ii}=1,\qquad g_{ij}g_{jk}=g_{ik}.
\]
A choice of right-principal local sections \(s_i\) relates them by \(s_j=s_i g_{ij}\). With this convention the overlap gluing identifies \((x,h)_j\) with \((x,g_{ij}(x)h)_i\).

\begin{theorem}[Classification on a test manifold]
\label{thm:bg-classifies}
For every \(U\), \(\mathbf BG(U)\) is equivalent to the groupoid of smooth right principal \(G\)-bundles over \(U\), and these equivalences commute with pullback up to the canonical isomorphisms.
\end{theorem}

\begin{proof}
A principal bundle is locally a product \(U_i\times G\). Choose local sections to express each change of frame by \(g_{ij}\). Multiplying changes of frame establishes the cocycle identity. Conversely the indicated gluing of the local products produces a smooth locally trivial right \(G\)-space with base \(U\). The equivalence relation commutes with the right action and a local product identifies each fiber with \(G\) acting on the right. Given an equivariant isomorphism between such products over \(U_i\), it is left multiplication by a smooth function on \(U_i\). The compatibility condition on the overlaps is exactly the condition that these local maps respect the identifications defining the quotient. They therefore glue to the given global isomorphisms, and descriptions from different covers agree on a common refinement. Pullback replaces every local product and every transition function by their pullbacks. This identifies the stackification with the indicated groupoid on each test.
\end{proof}

The terminal smooth sheaf \(*\) maps to \(\mathbf BG\) by picking the trivial torsor. Its relation object
\[
                 *\times^h_{\mathbf BG}*\simeq G
\]
is the sheaf of its automorphisms. On a manifold \(U\), an isomorphism between the two trivial right bundles has the form
\((x,h)\mapsto (x,k(x)h)\), for a unique \(k:U\to G\). This computes the fiber product directly, including its functorial action on tests. Iteration gives the nerve with \(G^n\) at level \(n\), with faces given by multiplication and deletion and degeneracies by identities. In particular, the smooth group is recovered by looping the stack at the trivial torsor.

\section{Fibers over a given torsor}
\label{sec:bg-fibers}

A map \(yU\to\mathbf BG\) classifying \(P\) has a homotopy pullback square
\[
\begin{tikzcd}[column sep=3.2em,row sep=2.6em]
yP \arrow[r] \arrow[d,"\pi"] &
* \arrow[d] \\
yU \arrow[r,"{[P]}"] & \mathbf BG .
\end{tikzcd}
\]
For a test map \(v:V\to U\), a point in the upper left consists of an isomorphism from the trivial bundle over \(V\) to \(v^*P\). Evaluation at the identity picks a section of \(v^*P\), and that section extends equivariantly to the isomorphism. Such sections are the lifts \(V\to P\). There are no further automorphisms of a given identification in this fiber. Thus the square reconstructs the original bundle, including its projection, without a choice of coordinates. Right multiplication of the evaluation point corresponds to precomposing the isomorphism by the appropriate automorphism of the trivial right bundle.

For a homomorphism \(\rho:G\to K\), extension of structure group sends \(P\) to \(P\times_G K\), the quotient of \(P\times K\) by \((p g,k)\sim(p,\rho(g)k)\). It gives a map \(\mathbf BG\to\mathbf BK\). The transition functions become \(\rho(g_{ij})\), which already proves compatibility with any base pullback and with iterated extension. A linear representation \(\rho:G\to GL(V)\) works in the same way with \(V\) in place of \(K\), giving the associated vector bundle \(P\times_G V\). Its section over \(U\) is an equivariant map \(P\to V\) satisfying \(t(p g)=\rho(g)^{-1}t(p)\). Evaluation in the local sections \(s_i\) gives \(t(s_j)=\rho(g_{ij})^{-1}t(s_i)\), which is exactly the change of columns in the associated frames.

\begin{proposition}[Reductions as sections]
\label{prop:reductions}
Let \(H\) be a closed Lie subgroup of \(G\), and let \(P\) be a smooth right \(G\)-bundle over \(U\). The groupoid of \(H\)-bundles together with an identification of their extended \(G\)-bundle with \(P\) is equivalent to the set of smooth sections of \(P/H\to U\), viewed as a discrete groupoid.
\end{proposition}

\begin{proof}
An identification \(Q\times_H G\cong P\) sends the canonical copy of \(Q\) to an \(H\)-subbundle of \(P\). Taking its cosets gives a section of \(P/H\). Conversely a section selects one right \(H\)-coset in each fiber. Pulling back \(P\to P/H\) along the section gives an \(H\)-bundle. Here \(G\to G/H\) is a locally trivial smooth \(H\)-bundle, so the pullback gives exactly the indicated smooth subbundle. Multiplication extends it to \(P\). Constructions in both directions are inverse up to unique isomorphism, since a map between reductions is fixed by its required extension to \(P\). Thus no additional automorphisms remain when the identification is specified.
\end{proof}

For \(H=O(r)\subset GL_r(\mathbb R)\), this is the familiar relation between a positive definite inner product and its orthonormal frame bundle. In a real vector bundle \(E\), a frame \(p\) transports the Euclidean inner product onto the fiber. Replacing \(p\) by \(p h\), with \(h\in O(r)\), leaves that inner product unchanged. Conversely frames defining the same inner product differ by an orthogonal matrix. Therefore \(P/H\) is the bundle of positive definite forms and the section in the proposition is a metric. A metric is a reduction of structure group. A connection is different data and is treated by a separate moduli stack in Chapter~\ref{ch:bg-connection}.

\section{Stacks and underlying homotopy types}
\label{sec:bg-coarse}

A classifying stack is functorial on every smooth test. Its set of connected components on a test gives isomorphism classes of torsors, but it also retains their gauge groups. For a bundle over a point all objects of \(\mathbf BG(*)\) are isomorphic, while its automorphisms are the elements of \(G\). Over a manifold the group of gauge automorphisms of \(P\) is the group of sections of the conjugation bundle \(P\times_G G\). Indeed an automorphism can be written \(p\mapsto p\,u(p)\), with \(u(p g)=g^{-1}u(p)g\). This formula also exhibits the group law by pointwise multiplication in the conjugation bundle.

One can apply a realization or shape construction that forgets the smooth structure of the sheaf and obtains an ordinary homotopy type. For the Čech presentation above this recovers the classifying homotopy type of the topological Lie group, with the usual model obtained by realizing the nerve of the topological group. It answers a different question from evaluation on a fixed smooth test. A smooth-family map into \(\mathbf BG\) specifies a bundle over the product of the parameter manifold with the base, together with its descent data when no global trivialization is available. The same distinction will be needed for connections, differential cohomology, and higher bundles.
\chapter{Connections as a smooth moduli stack}
\label{ch:bg-connection}

To a bundle with connection we can assign its transport, its curvature, or a characteristic form. Each assignment is compatible with pullback, but they retain different amounts of the original data. A stack makes the domain of these assignments explicit. Its automorphisms are the gauge transformations that preserve the given connection.

\section{Gluing connective data}
\label{sec:bgconn-trivial}

In a local right-principal section, the connective data for a Lie group \(G\) are a form \(A\in\Omega^1(U;\mathfrak g)\). Define
\[
    A^g=\operatorname{Ad}_{g^{-1}}A+g^*\theta_L ,
\]
where \(\theta_L\) is the left Maurer--Cartan form on \(G\). For a matrix group \(g^*\theta_L=g^{-1}dg\). The product rule for \(gh\) gives \((A^g)^h=A^{gh}\). Thus there is an action groupoid of these forms and smooth gauge changes on each \(U\). Its stackification is written \(\mathbf BG_\nabla\).

An object of \(\mathbf BG_\nabla(U)\) has local data \((g_{ij},A_i)\) satisfying
\[
    g_{ij}g_{jk}=g_{ik},\qquad A_j=A_i^{g_{ij}} .
\]
Replacing frames by \(s_i k_i\) sends \(A_i\) to \(A_i^{k_i}\) and sends \(g_{ij}\) to \(k_i^{-1}g_{ij}k_j\). When the local descriptions come from different covers one first restricts to a common refinement. This presentation of the connection stack is used in higher Chern--Weil theory \cite{fiorenza-schreiber-stasheff,schreiber-differential}.

\begin{theorem}[Classification with connection]
For a test manifold \(U\), \(\mathbf BG_\nabla(U)\) is equivalent to the groupoid of smooth principal \(G\)-bundles with connection and connection-preserving isomorphisms.
\end{theorem}

\begin{proof}
A global connection \(\omega\) on a right bundle \(P\) gives \(A_i=s_i^*\omega\). Right equivariance for a constant group element and reproduction of fundamental vector fields for the derivative of a varying one give \(A_j=A_i^{g_{ij}}\). Conversely, glue the local trivial bundles using \(g_{ij}\), as in Theorem~\ref{thm:bg-classifies}. On the local product with group coordinate \(h\) set
\[
              \omega_i=\operatorname{Ad}_{h^{-1}}A_i+\theta_L(h).
\]
The form reproduces vertical fundamental vectors and has the required behavior under right translation. The chart change from \(j\) to \(i\) replaces \(h\) by \(g_{ij}(x)h\). Pulling \(\omega_i\) back along it gives \(\operatorname{Ad}_{h^{-1}}A_i^{g_{ij}}+\theta_L(h)\), which is \(\omega_j\). Hence the forms glue to a connection. Local gauge isomorphisms between two presentations preserve the local connection forms by the corresponding formula for \(A^g\), and they glue by the compatibility of the underlying bundle maps. All operations commute with pullback along maps of test manifolds.
\end{proof}

The pointwise curvature of the local connection is \(F_i=dA_i+\frac12[A_i,A_i]\), using the graded bracket on forms. The change of frame gives \(F_j=\operatorname{Ad}_{g_{ij}^{-1}}F_i\) by the Maurer--Cartan equation. Consequently the \(F_i\) glue to a two-form with values in \(\operatorname{ad}P=P\times_G\mathfrak g\). One cannot in general use a single \(\mathfrak g\)-valued form on the whole base without choosing a gauge. An invariant polynomial on \(\mathfrak g\) eliminates this dependence and produces a scalar form.

\section{Forgetting the connection}
\label{sec:bgconn-forget}

The map \(u:\mathbf BG_\nabla\to\mathbf BG\) forgets the connective data. Take its homotopy fiber on \(U\) over a bundle \(P\). A point consists of \((Q,\omega)\) and an isomorphism \(Q\cong P\). Transport of \(\omega\) across the isomorphism gives a connection on \(P\). Every connection arises this way. An isomorphism over a prescribed identification of the underlying bundle is forced to be the identity after transport. Thus the fiber on \(U\) is equivalent to the discrete groupoid of connections on \(P\). When nonempty it is an affine space under \(\Omega^1(U;\operatorname{ad}P)\): the difference of two principal connections is horizontal and equivariant, and adding an adjoint-valued one-form preserves the defining conditions. In the present manifold setting, local connections patch by a partition of unity, so the affine space is nonempty.

Discreteness here refers to automorphisms in a groupoid over a fixed test. On mathematical families it still has smooth variation. A map \(y(S\times U)\to\mathbf BG_\nabla\) specifies a connection on the whole product. In a chosen trivialization it may include one-form components in the parameter directions of \(S\). A family of connections only along fibers of \(S\times U\to S\) uses relative one-forms and a relative exterior derivative. A relative connection by itself gives no chosen components along \(S\). The difference can be seen on \(U=*\): a connection along the point fibers is empty data, whereas an ordinary connection on \(S\) has nontrivial curvature in general.

Extension of structure group induces a square
\[
\begin{tikzcd}[column sep=3.3em,row sep=2.5em]
\mathbf BG_\nabla \arrow[r,"\rho_\nabla"] \arrow[d,"u"] &
\mathbf BK_\nabla \arrow[d,"u"]\\
\mathbf BG \arrow[r,"\rho_*"] & \mathbf BK
\end{tikzcd}
\]
for a Lie group homomorphism \(\rho:G\to K\). In local data the upper map uses \(\rho(g_{ij})\) and the Lie algebra map \(d\rho(A_i)\). The naturality of \(\theta_L\) under \(\rho\) verifies the gauge law, and \(d\rho(F_i)\) gives the resulting curvature. For a representation this recovers the connection on the associated vector bundle.

\section{Infinitesimal gauge data}
\label{sec:moduli-tangent}

Choose a model \(P\) for the underlying bundle and a connection \(A\). In the first-order problem below we retain gauge transformations of this model that reduce to the identity. In the marked fiber of Section~\ref{sec:bgconn-forget}, the identification of the underlying bundle removes these transformations and the variations form \(\Omega^1(U;\operatorname{ad}P)\) alone. A change is \(A+\epsilon a\), where \(\epsilon^2=0\) and \(a\in\Omega^1(U;\operatorname{ad}P)\). An infinitesimal change of frame with parameter \(\epsilon\xi\), for \(\xi\in\Omega^0(U;\operatorname{ad}P)\), adds \(d_A\xi\) to \(a\), with
\[
            d_A\xi=d\xi+[A,\xi].
\]
This follows by expanding the gauge formula to first order. It is independent of the chosen matrix realization because the derivative of the adjoint action is the Lie bracket. The first-order gauge action groupoid is therefore encoded by
\[
          \Omega^0(U;\operatorname{ad}P)\xrightarrow{d_A}
          \Omega^1(U;\operatorname{ad}P).
\]
Place the two terms in degrees \(-1,0\). Infinitesimal automorphisms lie in the kernel, and isomorphism classes of changes form the cokernel.

The derivative of curvature is \(d_Aa\). It follows by expanding
\(F(A+\epsilon a)=F(A)+\epsilon(da+[A,a])\). A direct calculation with the graded Jacobi identity gives \(d_A^2\eta=[F(A),\eta]\). At a flat connection, imposing flatness to first order requires \(d_Aa=0\), and the two arrows extend to the ordinary covariant de Rham complex. At a curved connection the linearized curvature of a gauge change is \([F(A),\xi]\), which is the infinitesimal conjugation law. The flat deformation complex thus has a flatness hypothesis as part of its definition.

\begin{example}[A flat circle connection on the two-torus]
On \(U=\mathbb R^2/\mathbb Z^2\) consider \(A=-2\pi i(r\,dx+s\,dy)\) on the trivial circle bundle. Its curvature vanishes. The two fundamental loops have holonomies \(\exp(2\pi i r)\) and \(\exp(2\pi i s)\). A gauge change \(g=\exp(-2\pi i(mx+ny))\) for \(m,n\in\mathbb Z\) adds the integer periods \((m,n)\). Conversely the logarithmic derivative of any smooth circle-valued gauge has integral periods in this normalization. Hence two members of this family are gauge equivalent exactly when their two periods differ by integers. Constant circle gauges preserve every member. This example exhibits curvature, holonomy and automorphisms as three separate pieces of data of the connection stack.
\end{example}
\chapter{Characteristic maps and their differential lifts}
\label{ch:characteristic-maps}

Classical Chern--Weil theory produces a closed form and identifies its cohomology class. The two constructions, the form itself and its cohomology class, have different domains of naturality. There is a map on bundles with connection that produces the form. Integrality concerns the corresponding bundle after forgetting the connection. The space of differential refinements is a way to keep both together with their comparison.

\section{The characteristic form as a natural operation}
\label{sec:curvature-map}

Let \(P_m\) be an invariant symmetric \(m\)-linear form on \(\mathfrak g\). Applying it to the curvatures of a connection gives
\[
                 \omega_{P_m}=P_m(F,\ldots,F)\in\Omega^{2m}(U).
\]
Invariance implies that the local expressions agree under changes of frame. Their exterior derivative is the sum of the terms obtained by replacing one input by \(d_AF\). To see why covariant differentiation can replace ordinary differentiation in this expression, differentiate the invariance identity along a one-parameter subgroup. The resulting sum with \([\xi,F]\) in each position vanishes. In local coordinates this cancels the terms \([A,F]\) in the graded derivative. The Bianchi identity then proves \(d\omega_{P_m}=0\). This repeats the defining calculation in a form suitable for descent.

The assignment commutes with smooth pullback, since exterior differentiation and substitution into \(P_m\) commute with pullback. A connection-preserving isomorphism fixes the resulting form, so this is a morphism
\[
                 \mathbf BG_\nabla\longrightarrow\Omega_{\mathrm{cl}}^{2m}
\]
of smooth sheaves of spaces. Its values are actual closed forms. If two connections on the same bundle are joined by their affine interpolation, the transgression calculation of Chapter~\ref{ch:chern-simons} gives an exact difference between their characteristic forms. Equality of classes here uses the interpolation, whereas equality of forms would require further hypotheses.

\section{Representing the real comparison}
\label{sec:represent-real}

For \(n\geq 1\), write \(K(C,n)\) for the \(n\)-fold delooping of the locally constant sheaf of an abelian group \(C\). On a test manifold its space of sections represents cohomology with those coefficients, with \(\pi_0\) given by \(H^n(U;C)\). One model uses the sheafified cochain complex of \(C\) placed in degree \(-n\), followed by the Dold--Kan construction. Equivalently, cover \(U\) by a good cover, apply the cochain construction on its nerve, and keep the homotopies between cocycles. This computes the same space because the good cover is acyclic for the locally constant coefficients.

The Poincaré lemma resolves the locally constant reals by differential forms. In fact
\[
 \Omega^0 \xrightarrow{d}\Omega^1\xrightarrow{d}\cdots
      \xrightarrow{d}\Omega^{n-1}\xrightarrow{d}\Omega_{\mathrm{cl}}^n
\]
has only degree-zero local cohomology, the locally constant functions, when displayed starting in cohomological degree zero. Placing its rightmost term in degree zero instead realizes \(K(\mathbb R,n)\), after Dold--Kan and sheafification. Inclusion of the rightmost term gives the morphism \(\mathrm{dR}:\Omega_{\mathrm{cl}}^n\to K(\mathbb R,n)\). On connected components over \(U\) it sends a closed form to its de Rham class. The comparison uses sheaves because locally exact potentials need not be globally exact.

The map \(\mathrm{real}:K(\mathbb Z,n)\to K(\mathbb R,n)\) comes from changing coefficients. Form the pullback
\[
\begin{tikzcd}[column sep=3.3em,row sep=2.65em]
\widehat{\mathbf H}^{n} \arrow[r,"I"]\arrow[d,"R"] &
K(\mathbb Z,n)\arrow[d,"\mathrm{real}"]\\
\Omega_{\mathrm{cl}}^n\arrow[r,"\mathrm{dR}"]&
K(\mathbb R,n).
\end{tikzcd}
\]
All pullbacks in this square are homotopy pullbacks. Sections of \(\widehat{\mathbf H}^{n}\) therefore include an integral cocycle, a closed form, and a path comparing their real images. The path can be represented by a cochain with precisely the prescribed differential. Replacing a path by a homotopic one, or altering the chosen cocycle representatives with their compatibilities, gives the usual equivalence in this space. Its set of components is the degree-\(n\) differential cohomology group modeled in Chapter~\ref{ch:diff-cohomology}. This is the discrete-coefficient specialization of the differential refinement constructions in \cite{hopkins-singer,bunke-nikolaus-voelkl}.

\begin{proposition}[A lift supplies the real comparison]
Let \(X\) be a smooth sheaf of spaces and let \(c:X\to K(\mathbb Z,n)\) and \(\omega:X\to\Omega_{\mathrm{cl}}^n\) be given. The space of lifts to \(\widehat{\mathbf H}^{n}\) with these components fixed is the space of paths from \(\mathrm{real}\circ c\) to \(\mathrm{dR}\circ\omega\) in \(\operatorname{Map}(X,K(\mathbb R,n))\).
\end{proposition}

\begin{proof}
The universal property of the homotopy pullback identifies maps into it with the homotopy pullback of the three mapping spaces. Taking the fiber over the pair \((c,\omega)\) gives a model with exactly the indicated path as its remaining datum.
\end{proof}

For a normalized Chern polynomial, the integral component is the ordinary characteristic map of the underlying bundle and the form component is its Chern--Weil form. A differential characteristic map records a \emph{natural} comparison in the above sense. Existence of a refinement on each individual test does not itself produce the coherent map on all tests. The classical differential refinements have concrete cocycle constructions \cite{cheeger-simons,fiorenza-schreiber-stasheff}. The circle case gives an elementary illustration.

\section{The differential first Chern class in coordinates}
\label{sec:line-diff-cocycle}

Take a unitary frame of a complex line over each \(U_i\). Use \(A_i=-2\pi i\,a_i\) and, on a good cover, write \(g_{ij}=\exp(-2\pi i\,\lambda_{ij})\). The connection law gives \(a_j-a_i=d\lambda_{ij}\), and \(\delta\lambda\) is integer-valued. The ordinary exponential sequence with exponential \(t\mapsto\exp(2\pi i t)\) computes the integral class from the logarithms \(-\lambda_{ij}\). With those signs put
\[
      n=-\delta\lambda,\qquad \omega=da_i .
\]
The \(da_i\) agree on overlaps. For total Čech degree \(p\) and form degree \(q\), use the total differential \(D=\delta+(-1)^p d\). Then
\[
                D(a+\lambda)=\omega-n .
\]
Thus the cochain \(a+\lambda\) gives the real comparison between the two displayed representatives. The curvature is exactly \(iF/(2\pi)\), since \(F=dA_i=-2\pi i\,da_i\).

For completeness, a lift of the comparison on a good cover is represented by \(n\) in integral Čech degree two, a closed global \(\omega\), and total real degree-one data \(a+\lambda\) with the displayed differential. Reading its bidegrees gives \(da_i=\omega,\ \delta a=d\lambda,\ \delta\lambda=-n\). Consequently the exponentials reconstruct the cocycle of a circle bundle and the \(a_i\) reconstruct its connection. Changing a local frame by \(\exp(-2\pi i\mu_i)\) replaces \(a_i\) by \(a_i+d\mu_i\) and \(\lambda\) by \(\lambda+\delta\mu+k\), for an integral overlap cochain \(k\). The comparison cochain changes by \(D\mu+k\), and \(n\) by \(-\delta k\), giving the equivalent triple. Changing \(\mu\) by an integral local function \(\ell\) and \(k\) by \( -\delta\ell\) gives the same gauge transformation. These changes of representatives recover the Čech equivalences, retaining the circle-valued gauge automorphisms. Locally refinements of covers merely restrict the data. This yields the identification of the degree-two differential cocycle stack with \(\mathbf BU(1)_\nabla\), sending a line and connection to its first differential Chern class.

In particular a flat line on a circle may have nontrivial holonomy although \(\omega\) and \(n\) can both be taken zero. It is the comparison, modulo integral changes, that remembers the flat class. For a nontrivial bundle the same comparison also records how the curvature periods agree with the integral class. Higher differential cohomology replaces these overlap logarithms by longer strings of cochains, as the next part develops.

\bookpart{Differential cohomology}
\chapter{Differential cohomology as a pullback}
\label{ch:diff-cohomology}

A characteristic form has a real cohomology class. An integral characteristic class can contain torsion, and a connection can have nontrivial holonomy even when both that form and that class vanish. Differential cohomology records all three phenomena. We develop it in ordinary integral degrees, using smooth singular chains to specify exactly how differential forms and holonomy fit together.

\section{Characters on cycles}

Write \(C_k(M;\mathbb Z)\) for smooth singular chains, \(Z_k(M;\mathbb Z)=\ker\partial\), and \(C^k(M;A)=\operatorname{Hom}(C_k(M;\mathbb Z),A)\). Smooth means smooth up to the faces of each simplex. The cochain differential is \((\delta a)(c)=a(\partial c)\). Integration gives a cochain map \(J:\Omega^\bullet(M)\to C^\bullet(M;\mathbb R)\). The smooth singular and ordinary singular cohomologies agree for manifolds. Write \(\Omega^q_{\mathbb Z}(M)\) for closed \(q\)-forms whose integrals on smooth integral \(q\)-cycles lie in \(\mathbb Z\).

\begin{definition}
A degree \(q\geq 1\) differential character on \(M\) is a homomorphism \(h:Z_{q-1}(M;\mathbb Z)\to\mathbb R/\mathbb Z\) for which there is a \(q\)-form \(\omega\) satisfying
\begin{equation}
 h(\partial c)=\int_c\omega\pmod{\mathbb Z}
 \qquad(c\in C_q(M;\mathbb Z)).
 \label{eq:diff-stokes}
\end{equation}
The group under pointwise addition is denoted \(\widehat H^q(M;\mathbb Z)\).
\end{definition}

Pullback along \(f:N\to M\) sends \(h\) to \(h\circ f_*\) on cycles and sends \(\omega\) to \(f^*\omega\). Both conclusions follow from the corresponding statements for smooth singular chains and forms. This is the contravariant functoriality of differential characters~\cite{cheeger-simons}.

\begin{lemma}
The form \(\omega\) in \eqref{eq:diff-stokes} is unique and belongs to \(\Omega^q_{\mathbb Z}(M)\).
\end{lemma}
\begin{proof}
The difference of two such forms integrates to an integer over every smooth \(q\)-simplex. Take a shrinking affine simplex in a chart, with edge vectors \(t v_1,\ldots,t v_q\). The difference of the integrals tends to zero, so it is zero for small \(t\). The limit after division by \(t^q/q!\) is evaluation of the difference of the forms on \(v_1,\ldots,v_q\). This proves uniqueness. Replacing \(c\) in \eqref{eq:diff-stokes} by a cycle proves integral periods. If \(b\) is a smooth \((q+1)\)-simplex, the same equation for \(c=\partial b\) gives \(\int_b d\omega\in\mathbb Z\) by Stokes. The shrinking-simplex argument now proves \(d\omega=0\).
\end{proof}

Write \(R(h)=\omega\) for the curvature. When \(q=1\), a character is precisely a smooth map \(f:M\to\mathbb R/\mathbb Z\), regarded as a homomorphism on the free group of zero-chains. Indeed the boundary of a path gives the change in \(f\) as the integral of the curvature. On a small coordinate ball this expresses \(f\) as a constant plus a smooth real primitive of the curvature, reduced modulo \(\mathbb Z\). Conversely, the normalized differential of a circle-valued function satisfies the boundary rule on paths.

\section{Differential cocycles and integral classes}

A real cochain \(\widetilde h\in C^{q-1}(M;\mathbb R)\) lifting \(h\) on cycles exists. The cycles and boundaries in each degree are free abelian as subgroups of free abelian groups. The surjection \(C_{q-1}\to B_{q-2}\) therefore splits, and one may lift the values on a basis of \(Z_{q-1}\) and extend by zero on a complement. In degree zero the complement can be taken to be zero. The equation for curvature implies that \(c=J\omega-\delta\widetilde h\) takes integral values. Since \(d\omega=0\), it satisfies \(\delta c=0\).

\begin{proposition}
A character is represented by a triple \((c,\widetilde h,\omega)\) with \(c\in C^q(M;\mathbb Z)\), \(\delta c=0\), \(\omega\in\Omega^q(M)\), \(d\omega=0\), and
\[
 \delta\widetilde h=J\omega-c .
\]
Two triples represent the same character precisely when they differ by
\begin{equation}
 (\delta b,-b-\delta u,0),
 \quad b\in C^{q-1}(M;\mathbb Z),\quad
 u\in C^{q-2}(M;\mathbb R).
 \label{eq:diff-gauge}
\end{equation}
In particular \(I(h)=[c]\in H^q(M;\mathbb Z)\) is well defined and its realification is \([\omega]_{\mathrm{dR}}\).
\end{proposition}
\begin{proof}
A triple defines a character by evaluation of \(\widetilde h\) on cycles modulo integers. The defining cochain equation gives \eqref{eq:diff-stokes}. Terms of the form \eqref{eq:diff-gauge} evaluate to zero on cycles and leave the form unchanged. Conversely, two lifts of the same character differ on the cycle group by an integral homomorphism. Subtract a cochain extending that integral homomorphism using the splitting above. The remaining real difference vanishes on cycles, hence factors through \(B_{q-2}\). Extend its real functional from \(B_{q-2}\) to \(C_{q-2}\), using divisibility of \(\mathbb R\). Thus the lift difference is an integral cochain plus a real coboundary, with the change in \(c\) then forced by the equation. Realifying the equation and applying de Rham's theorem gives the last assertion.
\end{proof}

Let \(a(\eta)\) be the character \(z\mapsto\int_z\eta\bmod\mathbb Z\) for a \((q-1)\)-form \(\eta\). Its representative is \((0,J\eta,d\eta)\). A flat class \(\xi\in H^{q-1}(M;\mathbb R/\mathbb Z)\) defines a character \(j(\xi)\) by evaluation on cycles. Under the convention \(\delta\widetilde h=J\omega-c\), the integral class of \(j(\xi)\) is the negative of the Bockstein calculated by lifting \(\xi\) to a real cochain and taking its coboundary. The sign follows directly by taking \(\omega=0\).

\begin{theorem}
There are natural short exact sequences
\begin{align*}
 0&\longrightarrow H^{q-1}(M;\mathbb R/\mathbb Z)
   \xrightarrow{j}\widehat H^q(M;\mathbb Z)
   \xrightarrow{R}\Omega_{\mathbb Z}^q(M)\longrightarrow 0,\\
 0&\longrightarrow \Omega^{q-1}(M)/\Omega^{q-1}_{\mathbb Z}(M)
   \xrightarrow{a}\widehat H^q(M;\mathbb Z)
   \xrightarrow{I} H^q(M;\mathbb Z)\longrightarrow 0.
\end{align*}
Here \(\Omega^0_{\mathbb Z}\) consists of locally constant integer-valued functions.
\end{theorem}
\begin{proof}
A flat character factors through \(H_{q-1}(M;\mathbb Z)\), and every such homomorphism is a unique cohomology class by the universal coefficient theorem with injective coefficients \(\mathbb R/\mathbb Z\). For an integral-period form \(\omega\), the integer homomorphism on \(H_q(M;\mathbb Z)\) given by periods lifts to integral cohomology by the universal coefficient theorem. Choose a cocycle \(c\) representing that lift. The real cochain \(J\omega-c\) has zero real class and is a coboundary. This produces a character with curvature \(\omega\).

The condition \(a(\eta)=0\) says that the integrals of \(\eta\) on all \((q-1)\)-cycles are integers. Integrating over boundaries and applying the shrinking-simplex argument gives \(d\eta=0\). Thus its kernel is the indicated subgroup. If \(I(h)=0\), the cochain representative can be chosen with \(c=0\). Then the curvature is exact by de Rham's theorem, so write \(\omega=d\eta\). The real cochain \(\widetilde h-J\eta\) is closed and can be represented up to coboundary by \(J\beta\) for a closed \((q-1)\)-form \(\beta\). Evaluating on cycles gives \(h=a(\eta+\beta)\). Finally, for any integral cocycle \(c\) one can choose a closed \(\omega\) representing its real class, then solve \(\delta\widetilde h=J\omega-c\). This proves surjectivity of \(I\).
\end{proof}

For example, on \(S^1=\mathbb R/\mathbb Z\) the degree-two group is \(H^1(S^1;\mathbb R/\mathbb Z)\cong\mathbb R/\mathbb Z\). All its classes have \(R=I=0\). The form \(a_0=\theta\,dx\) on the trivial line defines the character \(a(a_0)\), whose value on the positively oriented circle is \(\theta\bmod\mathbb Z\). This is the phase of the flat connection with matrix \(-2\pi i a_0\).

\section{The space of cocycles}

Fix \(q\) and use the form complex truncated by setting degrees less than \(q\) to zero. The data in a triple are cochains in the homotopy pullback
\begin{equation}
\begin{tikzcd}[column sep=2.4em,row sep=2.4em]
 D(q)(M)\arrow[r]\arrow[d] &
 \Omega^{\ge q}(M)\arrow[d,"J"]\\
 C^\bullet(M;\mathbb Z)\arrow[r] &
 C^\bullet(M;\mathbb R).
\end{tikzcd}
 \label{eq:diff-homotopy-pullback}
\end{equation}
The lower map is realification. Concretely the pullback complex has degree-\(n\) terms \((c,\widetilde h,\omega)\) with respective degrees \(n,n-1,n\) and differential
\[
 (c,\widetilde h,\omega)\longmapsto
 (\delta c,J\omega-c-\delta\widetilde h,d\omega).
\]
The middle component makes the comparison of the two real images part of the data. Its second differential vanishes because \(J\) commutes with the differentials. At degree \(q\), cocycles and coboundaries are precisely those above. Dropping the comparison component before passing to cohomology would retain only the pair \((I,R)\), which already loses the circle example. This mapping-cone model is the ordinary integral case of differential cocycles~\cite{hopkins-singer}.

The space attached to \eqref{eq:diff-homotopy-pullback} is obtained by shifting degree \(q\) to degree zero and taking the associated infinite loop space. Its connected components are \(\widehat H^q(M;\mathbb Z)\). For \(i>0\) its homotopy groups are \(H^{q-i}(D(q)(M))\). In degrees less than \(q\) the form term does not enter the cohomology calculation, leaving the homotopy fiber of \(\mathbb Z\to\mathbb R\) on cochains. Its cohomology in degree \(r\) is \(H^{r-1}(M;\mathbb R/\mathbb Z)\). Thus degree two has connection-preserving gauge automorphisms \(H^0(M;\mathbb R/\mathbb Z)\) at each object, and degree three has \(H^1(M;\mathbb R/\mathbb Z)\) and \(H^0(M;\mathbb R/\mathbb Z)\) in the first two homotopy degrees. The groups of components alone do not specify these automorphisms.

For a manifold of parameters \(S\), evaluating the cocycle space on \(S\times M\) gives families with forms in all directions of the product, including the parameter directions. Restrictions at points of \(S\) give individual characters. A collection of such pointwise restrictions can miss curvature components involving tangent vectors to \(S\). The presheaf of spaces defined by \eqref{eq:diff-homotopy-pullback} satisfies descent. Smooth singular cochains satisfy homotopy descent for covers by subdivision, and each sheaf of forms is fine on manifolds. These give descent for the constituents and hence for their homotopy pullback. This use of sheaves of complexes separates the family object from evaluation in a single degree~\cite{bunke-nikolaus-voelkl}.
\chapter{Products and differential characteristic classes}
\label{ch:diff-products}

The Chern--Weil construction gives a closed form from a connection. To refine an integral characteristic class, that form must be supplied with its integral class and a comparison between them in real cochains. The comparison participates in products and remembers the secondary change under variation of the connection.

\section{Multiplication at the cochain level}

Use \(J\) and the cocycle model from Chapter~\ref{ch:diff-cohomology}. Fix the usual singular cup product, for instance the Alexander--Whitney product, and a natural real cochain homotopy \(B\) between integrating the wedge product and taking the cup product of the integrations. For forms \(\omega,\eta\) with \(\omega\) of degree \(p\), the sign convention is
\begin{align}
 \delta B(\omega,\eta)+B(d\omega,\eta)
      +(-1)^p B(\omega,d\eta)
   &=J(\omega\wedge\eta)-J\omega\smile J\eta .
 \label{eq:de-rham-product-homotopy}
\end{align}
Such a homotopy is a form of the multiplicative de Rham comparison. The construction and its coherent choices are treated in the differential cochain setting by Hopkins and Singer~\cite[\S 3.2]{hopkins-singer}.

Take cocycles \(x=(c,h,\omega)\) of degree \(p\) and \(y=(c',h',\eta)\) of degree \(q\). Define their product by the triple
\begin{equation}
 \left(c\smile c',\ 
 h\smile J\eta+(-1)^p c\smile h'+B(\omega,\eta),\
 \omega\wedge\eta\right).
 \label{eq:diff-cup-representative}
\end{equation}

\begin{proposition}
Formula \eqref{eq:diff-cup-representative} defines a product on differential cohomology classes. Its integral class and curvature are the products of the corresponding integral classes and curvatures. For a form \(\alpha\) of degree \(p-1\),
\begin{equation}
 a(\alpha)\smile y=a\bigl(\alpha\wedge R(y)\bigr).
 \label{eq:form-product}
\end{equation}
\end{proposition}
\begin{proof}
The coboundary of the middle component of \eqref{eq:diff-cup-representative} is
\begin{align*}
 &(J\omega-c)\smile J\eta
   +c\smile(J\eta-c')
   +J(\omega\wedge\eta)-J\omega\smile J\eta\\
 &\hspace{2em}=J(\omega\wedge\eta)-c\smile c' .
\end{align*}
This is the cocycle equation. The first and last components give the claimed product formulas for \(I\) and \(R\).

For invariance under changing \(x\), replace \((c,h)\) by \((c+\delta b,h-b-\delta u)\). The change in the first component is \(\delta(b\smile c')\). If \(\Delta H\) is the change in the middle component, the coboundary rule for the cup product gives
\[
 \Delta H+b\smile c'
   =(-1)^p\delta(b\smile h')-\delta(u\smile J\eta).
\]
Thus the change is an equivalence of the form \eqref{eq:diff-gauge}. Changing \(y\) by \((\delta b',-b'-\delta u',0)\) instead changes the first component by \(\delta((-1)^p c\smile b')\) and the middle component by \(-(-1)^p c\smile b'-\delta(c\smile u')\). This too is a cocycle equivalence.

For \eqref{eq:form-product}, substitute \(c=0\), \(h=J\alpha\), and \(\omega=d\alpha\). Equation \eqref{eq:de-rham-product-homotopy} says that the resulting middle component and \(J(\alpha\wedge\eta)\) differ by a coboundary. The first component is zero in both triples and the last component is \(d(\alpha\wedge\eta)\).
\end{proof}

The coherent multiplicative comparison makes the operation associative and graded commutative on classes~\cite[\S 3.2]{hopkins-singer}. The correction \(B\) has a concrete purpose. The two cochain products in \eqref{eq:de-rham-product-homotopy} agree in real cohomology, and the term \(B\) supplies the comparison before passing to cohomology. Consequently multiplication preserves both the integral class and the curvature while controlling the character on cycles. Omitting this correction generally breaks the cocycle equation.

\section{The cylinder and changes of connection}

\begin{lemma}[Homotopy formula]
\label{lem:diff-cylinder}
For \(x\in\widehat H^q([0,1]\times M;\mathbb Z)\) and the inclusions \(i_0,i_1\),
\[
 i_1^*x-i_0^*x
 =a\left(\int_{[0,1]} R(x)\right).
\]
Here the interval direction is first in the integral. The smooth cochain and character constructions admit manifolds with boundary in this calculation.
\end{lemma}
\begin{proof}
For a cycle \(z\), triangulate the prism \([0,1]\times z\) so that its boundary is \((i_1)_*z-(i_0)_*z\). This is the chain cross product with the interval as the first factor. The boundary formula for the cross product gives the two endpoint terms, since \(\partial z=0\). Apply \eqref{eq:diff-stokes} and integrate first over the interval. The equality on every cycle proves the formula.
\end{proof}

For two connections on one bundle, a smooth path of connections gives a connection on the pullback bundle over the cylinder. The curvature of a degree-\(2j\) differential characteristic class there is the Chern--Weil form of the cylinder connection. Its interval integral is the transgression of that form. The homotopy formula says that the secondary change of the differential characteristic class is exactly the character \(a\) of that transgression. A constant curvature form on each slice does not determine the interval components of the cylinder curvature.

\section{Characteristic refinements for unitary bundles}

Let \(G\) be a Lie group with finitely many components, and let \(P\) be a principal \(G\)-bundle with connection \(\nabla\). Suppose an invariant polynomial \(p\) of degree \(j\) and a class \(\xi\in H^{2j}(BG;\mathbb Z)\) have the same real universal Chern--Weil class. The Cheeger--Simons characteristic refinement for this compatible pair is natural under pullback and satisfies
\[
 R\bigl(\widehat{\xi}(P,\nabla)\bigr)=p(F_\nabla),
 \qquad I\bigl(\widehat{\xi}(P,\nabla)\bigr)=\xi(P).
\]
Existence in this setting is the characteristic-class construction of Cheeger and Simons~\cite[Theorem 2.2]{cheeger-simons}. We use its specialization to \(G=U(r)\), with the polynomial obtained as the degree-\(j\) coefficient of the determinant
\[
 \det\left(1+t\,\frac{iF_\nabla}{2\pi}\right).
\]
It refines the Chern class \(c_j(E)\) of the associated Hermitian vector bundle. We denote the differential class by \(\widehat c_j(E,\nabla)\). Compatibility between the determinant form and the real integral Chern class is the classical input~\cite{tu}.

\begin{theorem}[Uniqueness for the unitary construction]
\label{thm:diff-chern-unique}
On smooth manifolds, there is at most one natural rule on Hermitian rank-\(r\) bundles with unitary connection whose integral class is \(c_j\) and whose curvature is the corresponding determinant coefficient.
\end{theorem}
\begin{proof}
Subtract two such rules. Their difference has zero curvature on every manifold, including the cylinder. Interpolate two unitary connections by their affine path on the cylinder. The homotopy formula applied to the difference shows that it is independent of the connection. Naturality then reduces it to a rule on Hermitian bundles.

First consider a bundle admitting a finite trivializing cover. Using a smooth partition of unity subordinate to that cover, it embeds isometrically into a finite-dimensional trivial Hermitian bundle. For example, choose smooth functions \(\rho_i\) supported in local trivializations with \(\sum_i\rho_i^2=1\), and send a vector to the tuple of its coordinates weighted by \(\rho_i\). Such functions are obtained by normalizing the squares of a finite partition of unity. The image gives a smooth classifying map into a finite complex Grassmannian, and the bundle is pulled back from its tautological bundle. Since the Grassmannian has only even-dimensional cells, its \(H^{2j-1}(-;\mathbb R/\mathbb Z)\) vanishes. The curvature exact sequence then forces the difference for its tautological bundle to be zero. Naturality forces the same for every bundle admitting a finite trivializing cover.

For an arbitrary base, the image of each smooth singular cycle is compact. Choose an open neighborhood of that image formed from finitely many trivializing opens. Naturality and the finite-cover argument show that the difference evaluates to zero on that cycle. A character with zero value on every cycle is zero.
\end{proof}

Write \(\widehat c=1+\sum_{j\geq1}\widehat c_j\), extending degree zero by \(H^0(M;\mathbb Z)\). The proof also gives the differential Whitney sum formula by applying the same argument on products of Grassmannians. Both candidate total classes have the same curvature because the block-diagonal curvature of a direct sum gives a product of determinant forms. Their integral classes agree by the ordinary Whitney sum theorem. Each positive-degree coefficient of their difference is flat and becomes zero on a product of the finite Grassmannians used for the two bundles. This proves
\[
 \widehat c(E\oplus E',\nabla\oplus\nabla')
 =\widehat c(E,\nabla)\smile\widehat c(E',\nabla').
\]

\section{First Chern class as holonomy}

For a unitary line, take the real local one-form \(a_i\) such that the frame connection is \(A_i=-2\pi i a_i\). Parallel transport in a frame on an interval multiplies by \(\exp(2\pi i\int a_i)\). Along a loop, include the changes of frame where the path crosses charts. The abelian holonomy defines a character \(h\) on smooth one-cycles. One can choose frames at the vertices of the one-chain and multiply the transport scalars of its oriented edges with their integer powers. Changing a vertex frame contributes opposite powers on incoming and outgoing edges, so the result is unchanged. On a smooth two-simplex the pulled back bundle is trivial. Stokes in a frame on that simplex gives the boundary rule with curvature \(da_i=iF/(2\pi)\).

The induced integral class is \(c_1\). In local frames \(s_j=s_i g_{ij}\), write \(g_{ij}=\exp(-2\pi i\lambda_{ij})\). Then \(a_j-a_i=d\lambda_{ij}\) and the integral representative for \(c_1\) is \(-\delta\lambda\) using the exponential \(t\mapsto\exp(2\pi it)\). With \(D=\delta+(-1)^p d\) in \v{C}ech degree \(p\), one has \(D(a+\lambda)=da+\delta\lambda\), exhibiting the real comparison with curvature. The higher-holonomy interpretation of \(\widehat c_1\) thus recovers the actual phase of parallel transport. For a flat line its curvature is zero, but that phase can still detect either a nonzero torsion integral class or a continuous holonomy parameter.
\chapter{Integration, holonomy, and transgression}
\label{ch:fiber-integration}

Integrating along an oriented fiber takes a differential cocycle on a total space to a cocycle on the space of parameters. The integration acts before curvature is taken. This lets the output retain the holonomy on the fibers even when integrating the curvature alone would lose it.

\section{Coherent transfers of smooth chains}

Let \(\pi:E\to M\) be a smooth proper submersion with closed fibers of dimension \(d\), equipped with an orientation of the vertical tangent bundle. One may also take an oriented smooth fiber bundle with a closed fiber as the starting point. For a smooth simplex \(\sigma:\Delta^k\to M\), put \(E_\sigma=\Delta^k\times_M E\), a compact smooth manifold with corners. Orient it with the simplex direction first and the vertical direction second. Proper local triviality of the oriented family and the contractibility of the simplex give \(E_\sigma\) the homotopy type of its fiber. The same conclusion follows by pulling back a smooth fiber bundle and taking a trivialization over the simplex. Homology here may be calculated by smooth singular chains, including for the manifolds with corners that occur in this construction.

Let \(E_{\sigma_i}\) be the pullback over face \(i\) and \(j_i:E_{\sigma_i}\to E_\sigma\) its inclusion. We choose, for each simplex, a smooth relative fundamental chain \(s_\sigma\) of degree \(k+d\), with prescribed boundary
\begin{equation}
 \partial s_\sigma=\sum_{i=0}^k(-1)^i(j_i)_*s_{\sigma_i}
 \quad(k>0).
 \label{eq:transfer-boundary}
\end{equation}
For \(k=0\), choose an oriented smooth fundamental cycle of the fiber. Suppose lower degrees have been chosen. The right side of \eqref{eq:transfer-boundary} is a cycle in \(\partial E_\sigma\) because the contributions from codimension-two faces cancel. It represents the boundary fundamental class. Indeed on the interior of each facet it has the corresponding local orientation, and those interiors meet every component of the boundary. The boundary of any relative fundamental representative also represents this class. Their difference bounds a chain in \(\partial E_\sigma\). Adding such a chain to the representative gives the desired boundary. This argument requires orientations of the entire family, since the face choices have to agree under every identification of fibers.

Compose \(s_\sigma\) with the projection \(E_\sigma\to E\) and extend linearly. The resulting degree-\(d\) transfer \(T\) on smooth singular chains satisfies \(\partial T=T\partial\) in the base-first convention. For a \((k+d)\)-form \(\omega\), relative fundamental chains compute its integral on \(E_\sigma\) after pullback. This holds for every top form on \(E_\sigma\). A boundary change integrates to zero by Stokes, and a chain supported in the boundary has zero integral by dimension. Fiberwise Fubini therefore gives
\begin{equation}
 \int_{T(\sigma)}\omega
   =\int_{E_\sigma}\omega|_{E_\sigma}
   =\int_{\sigma}\pi_!\omega ,
 \label{eq:transfer-fubini}
\end{equation}
where the first term uses the projection and the middle term uses its pullback. Here \(\pi_!\) on forms is defined by evaluating first on horizontal lifts of base directions, then on the vertical directions. Changing a lift by a vertical vector gives a vanishing term by the fiber dimension, so this definition needs no horizontal connection. It gives, in particular, \(\pi_!(\pi^*\alpha\wedge\beta)=\alpha\wedge\pi_!\beta\).

\begin{lemma}
Two choices \(T,T'\) in this construction are chain homotopic by a homotopy \(K\) for which \(K(\sigma)\) factors through \(E_\sigma\).
\end{lemma}
\begin{proof}
Induct on \(k\), keeping the homotopy values as chains in the pullbacks before projecting to \(E\). In degree zero the difference of the two fundamental cycles bounds in the fiber. Suppose chains \(k_{\sigma_i}\) giving the homotopy have been chosen on the faces. In \(E_\sigma\), the chain
\[
 s'_\sigma-s_\sigma-\sum_i(-1)^i(j_i)_*k_{\sigma_i}
\]
is a cycle because of \eqref{eq:transfer-boundary} and the identities on the faces. For \(k>0\) its degree is \(k+d>d\), so it bounds in \(E_\sigma\) by the fiber homotopy type. Let \(k_\sigma\) be a bounding chain there. Projecting to \(E\) gives \(T'-T=\partial K+K\partial\). In each degree \(K(\sigma)\) thus factors through a smooth space of dimension one less than the degree of that chain, a property stronger than the existence of an abstract chain homotopy.
\end{proof}

\section{Pushforward of differential characters}

Suppose \(q>d\) and \(h\in\widehat H^q(E;\mathbb Z)\). For a smooth integral \((q-d-1)\)-cycle \(z\) of \(M\), define
\begin{equation}
 \bigl(\widehat\pi_!h\bigr)(z)=h(T(z)).
 \label{eq:diff-pushforward}
\end{equation}
By the chain equation \(T(z)\) is a cycle. The boundary of a \((q-d)\)-chain \(c\) is sent to the boundary of \(T(c)\). The character boundary rule and \eqref{eq:transfer-fubini} give
\[
 \bigl(\widehat\pi_!h\bigr)(\partial c)
  =\int_{T(c)}R(h)
  =\int_c\pi_!R(h)\pmod{\mathbb Z}.
\]
Thus the pushforward is a character. For different choices of transfer its change on \(z\) is \(h(\partial K(z))\). The integral of \(R(h)\) on \(K(z)\) vanishes because each summand factors through a pullback of dimension \(q-1\). The character is consequently independent of the fundamental chains used to define it.

In an oriented pullback square
\[
 \begin{tikzcd}[column sep=3em,row sep=2.5em]
 E'\arrow[r,"\widetilde g"]\arrow[d,"\pi'"'] & E\arrow[d,"\pi"]\\
 M'\arrow[r,"g"'] & M
 \end{tikzcd}
\]
the bundles over a simplex of \(M'\) are canonically the bundles over the composite simplex of \(M\). Using corresponding fundamental chains and applying independence of choices gives
\[
 g^*\widehat\pi_!(h)=\widehat\pi'_!(\widetilde g^*h).
\]
The curvature and integral class formulas are
\[
 R(\widehat\pi_!h)=\pi_!R(h),
 \qquad I(\widehat\pi_!h)=\pi_!I(h).
\]
For the second formula, use a triple \((c,\widetilde h,\omega)\) on \(E\) and precompose the two cochains with \(T\). Equation \eqref{eq:transfer-fubini} identifies the resulting triple with curvature \(\pi_!\omega\). The integral cochain operation is the ordinary oriented transfer dual to the relative fundamental chains of the bundle. This is a smooth singular description of the integral Gysin map. Fiber integration for characters has several equivalent constructions~\cite[\S 7]{baer-becker}.

A useful special case is \(h=a(\eta)\). By \eqref{eq:transfer-fubini} in degree \(q-1\), \(\widehat\pi_!a(\eta)=a(\pi_!\eta)\). For a product \(M\times F\to M\), the transfer can be chosen as the chain cross product with a smooth fundamental cycle of \(F\), with the base cycle written first. A relative fundamental chain of a fiber with boundary would instead contribute an additional term to \(\partial T(z)\). The definition with absolute cycles above uses closed fibers. Relative characters describe integration when boundary data, such as a trivialization, are specified~\cite[\S\S 7.2 and 8.3]{baer-becker}.

\section{Loop families}

Let \(B\) be a finite-dimensional smooth parameter manifold and let \(\Gamma:B\times S^1\to M\) be a smooth family of loops. For every \(h\in\widehat H^q(M;\mathbb Z)\) with \(q\geq2\), set
\[
 \tau_\Gamma(h)=\widehat\pi_!\Gamma^*h
 \in\widehat H^{q-1}(B;\mathbb Z),
 \qquad \pi:B\times S^1\longrightarrow B.
\]
For \(q=2\), this character is the smooth \(S^1\)-valued holonomy function on the family. Evaluating it at \(b\) evaluates \(h\) on the oriented loop \(\Gamma(b,-)\). The differential of the function is described in real angular units by the fiber integral of \(\Gamma^*R(h)\), including its component with one base direction and one circle direction. The formula explains how holonomies vary in a family even when each individual loop sees only the phase.

For \(q=3\), the output is a degree-two class, so it is represented by a line with connection on \(B\). Evaluate on a loop \(b:S^1\to B\). Equation \eqref{eq:diff-pushforward} identifies the resulting phase with \(h\) evaluated on the mapped torus
\[
 S^1\times S^1\longrightarrow M,\qquad (s,t)\longmapsto\Gamma(b(s),t).
\]
The orientation of the torus is \(s\) first, \(t\) second. This characterization lets one use loop-space transgression without choosing a particular calculus on the whole infinite-dimensional free loop space. Each smooth family gives a well-defined pullback datum, and the construction is natural in the parameter manifold.

\begin{example}
On \(M=\mathbb R^2/\mathbb Z^2\) take \(h=a(\theta\,dx\wedge dy)\) of degree three with \(\theta\in\mathbb R\). Let \(B=S^1\) and \(\Gamma(s,t)=(s,t)\). The resulting degree-two class is \(a(\theta\,ds)\). Its curvature is zero, and its phase on the positive generator is \(\theta\bmod\mathbb Z\). Transgressing the curvature alone would give zero for every \(\theta\), whereas transgressing the character distinguishes the phases.
\end{example}
\chapter{Bundle gerbes and surface phases}
\label{ch:gerbes}

A unitary line with connection has one-form potentials on charts and circle-valued gluing functions on double intersections. A bundle gerbe with connection and curving has two-form potentials on charts, one-form connection data on double intersections, and circle-valued coherence on triple intersections. Its invariant is a degree-three character, so the boundary rule compares its phase on a surface with the integral of a three-form over a filling.

\section{The Deligne complex and its signs}

At the level of sheaves, the real cochain sheaf in the pullback model of Chapter~\ref{ch:diff-cohomology} can be computed by the de Rham complex, using the Poincar\'e lemma. Write \(\underline{\mathbb Z}\) for the locally constant integer sheaf. The homotopy pullback then identifies with the homotopy fiber of \(\underline{\mathbb Z}\to\Omega^{<q}\), since \(\Omega^{<q}\) is the quotient by the subcomplex \(\Omega^{\ge q}\). A signed presentation of that fiber is the Deligne complex
\[
 \underline{\mathbb Z}\ \longrightarrow\
 \Omega^0\ \xrightarrow{d}\ \cdots\
 \xrightarrow{d}\ \Omega^{q-1},
\]
with \(\underline{\mathbb Z}\) in degree zero and the first map the inclusion. The sign convention can be read directly from the middle component \(-c-\delta\widetilde h\) of the fiber differential. In internal degree \(n>0\), multiply that component's cochain variable by \((-1)^n\) to obtain the displayed positive differentials. In degree \(q\) the curvature map on the top form is therefore \((-1)^q d\). Projection onto \(\underline{\mathbb Z}\) gives the integral class. The hypercohomology in degree \(q\) is \(\widehat H^q(M;\mathbb Z)\), and the shifted complex retains the higher equivalences between cocycles~\cite{hopkins-singer,bunke-nikolaus-voelkl}.

Choose a good open cover, so that every nonempty finite intersection is contractible and the \v{C}ech complex computes the required hypercohomology. Use \(\delta\) for the alternating \v{C}ech differential. In \v{C}ech degree \(p\) the total differential is \(D=\delta+(-1)^p d_{\rm Del}\), where \(d_{\rm Del}\) includes the first map of the Deligne complex. In total degree three write the entries in order of increasing internal degree as
\[
 (n,\lambda,-A,-B).
\]
The entries have \v{C}ech degrees \(3,2,1,0\), respectively, with \(n\) integer-valued, \(\lambda\) real-valued, \(A\) a one-form cochain and \(B\) a two-form cochain. The cocycle conditions become
\begin{equation}
 \delta n=0,\qquad
 \delta\lambda=n,\qquad
 \delta A=d\lambda,\qquad
 \delta B=dA .
 \label{eq:gerbe-descent}
\end{equation}
The curvings \(B_i\) give a global three-form \(H=dB_i\), because the difference on a double intersection is the differential of \(A_{ij}\). This is the curvature of the class. For comparison, \(D(B+A-\lambda)=H-n\) in the full real \v{C}ech--de Rham complex with \(D=\delta+(-1)^p d\). Thus the real image of \([n]\) is \([H]\). The integer cochain remains part of the cocycle even when it represents torsion.

For an integer two-cochain \(m\), a real one-cochain \(\mu\), and a zero-cochain \(\Lambda\) of one-forms, adding a total coboundary with components \((m,\mu,-\Lambda)\) in degree two gives
\begin{align*}
 B'&=B+d\Lambda,& A'&=A+\delta\Lambda+d\mu,\\
 \lambda'&=\lambda+\delta\mu+m,& n'&=n+\delta m .
\end{align*}
These signs follow by applying \(D\) to each component. They preserve \eqref{eq:gerbe-descent} and \(H\). Higher gauge relations between the changes are obtained from the preceding degrees of the same total complex. This explains why the gerbe cocycle belongs to a space with nontrivial higher homotopy groups.

In degree two the analogous entries are \((n,\lambda,a)\), and the equations are \(\delta a=d\lambda\) and \(n=-\delta\lambda\). Exponentiating with \(g_{ij}=\exp(-2\pi i\lambda_{ij})\) gives precisely the frame changes for a line with \(A_i=-2\pi i a_i\). Thus the Deligne model identifies the integral component of its holonomy character with the integral Chern cocycle, including its possible torsion. This also verifies the normalization in Chapter~\ref{ch:diff-products}.

\section{Geometric realization by bundle gerbes}

Take a surjective submersion \(p:Y\to M\) possessing local sections. Denote its repeated fiber products by \(Y^{[k]}\). A unitary bundle gerbe has a Hermitian line \(L\to Y^{[2]}\), with multiplication over \(Y^{[3]}\)
\[
 L_{y_1,y_2}\otimes L_{y_2,y_3}\longrightarrow L_{y_1,y_3}.
\]
The map is a unitary isomorphism and the two composites on \(Y^{[4]}\) agree. This is the line-bundle realization of the gerbe descent condition~\cite{murray}. A connection is compatible when multiplication preserves the tensor product connection. Its normalized curvature \(F^L_{\mathbb R}=iF_L/(2\pi)\) therefore has zero alternating sum of pullbacks to \(Y^{[3]}\). A curving is a two-form \(B\) on \(Y\) with
\[
 F^L_{\mathbb R}=\operatorname{pr}_2^*B-\operatorname{pr}_1^*B .
\]

A curving exists for a compatible line connection on a paracompact base. Choose local sections \(s_i\) and a subordinate smooth partition of unity \(\rho_i\) on \(M\). On \(p^{-1}(U_i)\) let \(B_i'\) be the pullback of \(F^L_{\mathbb R}\) by \(y\mapsto(s_i(p(y)),y)\). The cocycle identity for \(F^L_{\mathbb R}\) gives \(B_i'(y_2)-B_i'(y_1)=F^L_{\mathbb R}(y_1,y_2)\) as pullback maps on forms. Hence \(B=\sum_i p^*\rho_i B_i'\), with terms extended by zero outside their supports, is a curving. Taking \(d\) gives equality of the two pullbacks of \(dB\). A form with this equality descends along \(p\), since pullback to any two local sections gives the same local form on the overlap. Thus \(dB=p^*H\).

Choose local sections of \(Y\) on a good cover and unit frames of the induced lines on double intersections. Write the connection form in each line frame as \(-2\pi i A_{ij}\). With the frame multiplication scalar \(\exp(-2\pi i\lambda_{ijk})\), compatibility of multiplication with the connection gives \(\delta A=d\lambda\). Associativity gives \(\delta\lambda\in\mathbb Z\). The curving identity gives \(\delta B=dA\) for \(B_i=s_i^*B\). In these conventions the integral Dixmier--Douady cocycle is the Bockstein of the inverse frame multiplication scalars, which is \(\delta\lambda\) for the positive exponential. This assigns the sign so that the integral class realifies to the normalized three-curvature.

Conversely, on a good cover, \eqref{eq:gerbe-descent} constructs a bundle gerbe with \(Y=\coprod_i U_i\). Put a trivial line with connection \(-2\pi i A_{ij}\) on each double intersection. Use the exponential of \(-2\pi i\lambda\) for multiplication. The equations are exactly the compatibility, associativity and curving conditions. Coboundaries describe changes of local representatives and tensoring by the corresponding trivialization data. Up to refinement and stable equivalence with connection, this gives the gerbe classes described by degree-three differential cohomology. For the underlying topological gerbe the equivalence relation is stable isomorphism, and the class is the signed Dixmier--Douady class above~\cite{murray-stevenson}. Surface phases are evaluations of the associated differential character on smooth two-cycles.

\section{A three-sphere and its equator}

Let \(H\) be a closed three-form on an oriented \(S^3\) with \(\int_{S^3} H=N\in\mathbb Z\). Cover \(S^3\) by \(U_+\) and \(U_-\), the complements of the south and north pole. Choose two-forms \(B_+\) and \(B_-\) there with \(dB_\pm=H\), since each open is contractible. The difference \(K=B_--B_+\) is closed on the overlap. Orient the equator as the boundary of the southern closed hemisphere. Stokes on the two hemispheres gives
\[
 \int_{S^2} K=\int_{S^3} H=N.
\]
The overlap has the homotopy type of that \(S^2\), so \(K\) has integral periods. Choose a line \(L\) on the overlap with \(c_1\) pairing to \(N\) on the equator. Its normalized curvature can be chosen to equal \(K\), because any initial normalized curvature represents the same real class and can be adjusted by an exact form. Use \(L\) on the ordered overlap \((+,-)\), its dual on \((-,+)\), and the trivial lines on the equal-index overlaps. Evaluation and tensor product multiplication give the associative gerbe multiplication. The two curvings yield exactly \(H\), and the integral class of the gerbe has value \(N\) on the fundamental class of \(S^3\).

Every integral two-cycle \(z\) on \(S^3\) bounds a three-chain \(C\), so the gerbe phase is
\[
 \exp\!\left(2\pi i\int_C H\right).
\]
A different filling leaves the phase unchanged because periods are integral. If \(H\) is the round volume form normalized to have integral \(N\), its integral over either hemisphere is \(N/2\). The equatorial phase is \((-1)^N\). This calculation uses both the orientation of the fiberwise boundary and the period convention. A global curving on \(S^3\) could not give this nonzero integral of \(H\) for \(N\ne0\), since the integral of its exterior derivative would vanish by Stokes.

\bookpart{Higher groups and surface transport}
\chapter{Strict two-groups and crossed modules}
\label{ch:crossed-modules}

An ordinary gauge transformation changes a frame. In higher gauge theory
there are also transformations between frame changes. A useful local
model is a groupoid whose objects and arrows have group structures and
whose structure maps respect multiplication. It has a particularly
explicit description by two groups, a boundary homomorphism, and an
action. The distinction between an arrow and its boundary will
become the distinction between a surface transport and the path
transports on its edges.

\section{The groupoid of higher arrows}

\begin{definition}[Crossed module]
A crossed module of Lie groups consists of a smooth homomorphism
\(\partial:H\to G\) and a smooth left action
\(g\triangleright h\) of \(G\) on \(H\) by group automorphisms, such that
\begin{align}
 \partial(g\triangleright h)&=g\,\partial(h)\,g^{-1},\label{eq:crossed-equiv}\\
 \partial(h)\triangleright k&=hkh^{-1}.\label{eq:crossed-peiffer}
\end{align}
The second identity is the Peiffer identity.
\end{definition}

The associated strict Lie two-group has objects \(g\in G\). An arrow
\((h,g)\) starts at \(g\) and ends at \(\partial(h)g\). Composition in
the groupoid is multiplication of the \(H\)-labels in chronological
order from right to left:
\[
 (k,\partial(h)g)\circ(h,g)=(kh,g).
\]
The identity of \(g\) is \((1,g)\) and the inverse for groupoid
composition is \((h^{-1},\partial(h)g)\). There is a second
multiplication, the tensor product, with
\[
 (h,g)\otimes(k,\ell)
   =\bigl(h(g\triangleright k),g\ell\bigr).
\]
The object group multiplies by \(g\otimes\ell=g\ell\). Equation
\eqref{eq:crossed-equiv} verifies the target of the tensor product:
\[
 \partial\bigl(h(g\triangleright k)\bigr)g\ell
   =(\partial(h)g)(\partial(k)\ell).
\]
The group law on arrows is the semidirect product \(H\rtimes G\).
Its inverse uses action by \(g^{-1}\), whereas the inverse of an
arrow in the groupoid uses the target \(\partial(h)g\).

Tensor and groupoid composition must satisfy interchange. If the
upper arrows in two composable pairs carry labels \(h'\) and \(k'\)
and the lower arrows carry \(h\) and \(k\), respectively, multiplying
each pair first gives the label
\[
 h'h\bigl(g\triangleright(k'k)\bigr).
\]
Tensoring at each level first gives the label
\[
 h'\bigl((\partial(h)g)\triangleright k'\bigr)
   h(g\triangleright k).
\]
Since \((\partial(h)g)\triangleright k'
=h(g\triangleright k')h^{-1}\), the two labels agree. This
calculation explains why equivariance and the Peiffer identity need
separate hypotheses.

A strict monoidal groupoid with such an arrow group can also be read
as a one-object strict two-category. Its one-morphisms are \(G\), its
two-morphisms are the arrows just described, and horizontal
composition is the tensor product. In this reading, when one path
runs after another, the corresponding one-morphism is multiplied on
the left. This agrees with successive parallel transport.

\section{Recovering a crossed module from a group object}

\begin{proposition}
The crossed module construction and passage to the kernel of the
source map are inverse up to natural isomorphism. They identify
crossed modules of Lie groups with internal groupoids in Lie groups.
\end{proposition}

\begin{proof}
Let \(\mathcal G_1\rightrightarrows\mathcal G_0\) be an internal
groupoid with source \(s\), target \(t\), and unit \(u\). Put
\(G=\mathcal G_0\), \(H=\ker s\), and \(\partial=t|_H\).
The split surjection \(s\) is a submersion and its kernel is a
Lie subgroup. Conjugation by \(u(g)\) defines the action on \(H\)
and applying \(t\) verifies \eqref{eq:crossed-equiv}.

To prove the Peiffer identity, use that groupoid composition is a
group homomorphism on composable pairs. If \(a\in\ker s\) and
\(b\in\ker t\), then the composable pairs \((a,1)\) and \((1,b)\)
commute in the group of composable pairs. Their composites are \(a\)
and \(b\), so \(a\) commutes with \(b\) in \(\mathcal G_1\).
For \(h\in H\), the element \(h^{-1}u(\partial(h))\) belongs to
\(\ker t\). Its commutation with \(k\in H\) is exactly
\[
 u(\partial(h))ku(\partial(h))^{-1}=hkh^{-1}.
\]

The map \((h,g)\mapsto hu(g)\) identifies \(H\rtimes G\) with
\(\mathcal G_1\), with inverse
\(a\mapsto \bigl(au(s(a))^{-1},s(a)\bigr)\). This identification
respects source and target. For a composable pair with lower arrow
\(hu(g)\), the upper arrow has the form \(k u(\partial(h)g)\).
Factor the pair as
\[
 (k,1)\cdot \bigl(u(\partial(h)g),hu(g)\bigr)
\]
in \(\mathcal G_1\times_G\mathcal G_1\). Applying the composition
homomorphism gives \(khu(g)\). This is the composition formula of the
crossed module groupoid. The constructions on homomorphisms follow
by restricting to \(H\) and applying the same identification.
\end{proof}

The result addresses a strict model. A weak two-group has coherent
associators, and questions about smooth presentations must keep
track of how it is presented. In particular a choice of crossed
module as a smooth model is extra structure. Baez and Lauda develop
the corresponding algebraic and internal notions and the role of
coherence~\cite{baez-lauda}.

\section{The nerve and the two homotopy groups}

The groupoid nerve has an explicit \(n\)-simplex. It consists of
\(n\) consecutive \(H\)-labels and an initial object:
\[
 (h_1,\ldots,h_n;g)
 \ \longleftrightarrow\
 g\xrightarrow{h_1}\partial(h_1)g
   \xrightarrow{h_2}\partial(h_2h_1)g
   \longrightarrow\cdots .
\]
In degree two its middle face replaces \(h_1,h_2\) by \(h_2h_1\).
The first face adjusts the initial object by \(\partial(h_1)\).
Degeneracies insert identity arrows. Since all structure maps of the
groupoid respect group multiplication, the nerve is a simplicial
group. This gives a nerve description of the same strict model,
including the operation responsible for the triangular faces.

For the next interpretation take the underlying \emph{discrete} groups.
The natural components of the groupoid are cosets of
\(\partial(H)\), which is normal by \eqref{eq:crossed-equiv}. At each
object the automorphism group is \(\ker\partial\). A groupoid can be
replaced, up to equivalence of categories, by one representative of
each component together with all of its automorphisms. Indeed
inclusion of that subcategory is fully faithful and essentially
surjective. The groupoid's nerve therefore has components of the
form \(B(\ker\partial)\), with contractible universal cover. Delooping
the simplicial group gives a connected two-type with
\[
 \pi_1\cong G/\partial(H),\qquad
 \pi_2\cong\ker\partial.
\]
Here the delooping can be defined by the simplicial classifying
construction, whose loop space is the realization of the simplicial
group. Passing instead to smooth sheaves or keeping the topology of
the Lie groups changes that interpretation. For example, the
classifying space of the topological circle has its first nonzero homotopy
group in degree two, while the classifying space of the discrete
circle has fundamental group the circle regarded as an abstract
group.

If \(a\in\ker\partial\), then \eqref{eq:crossed-peiffer} gives
\(aka^{-1}=k\) for every \(k\in H\). Thus \(\ker\partial\) is abelian,
and the action of \(\partial(H)\) on it is trivial. This supplies the
\(G/\partial(H)\)-action on the second homotopy group. The action is
visible in the tensor formula as conjugation of a two-cell by the
one-cell \(g\).

\section{Models and their geometric interpretations}

\begin{example}[Ordinary groups and abelian two-groups]
For \(H=\{1\}\) the associated groupoid is discrete on \(G\), and its
delooping gives \(BG\) in the discrete interpretation. If \(G=\{1\}\)
and \(H\) is abelian, the groupoid has a single object and all its
arrows are elements of \(H\). The resulting two-category models
a twice-delooped abelian group. When \(H=U(1)\) is kept as a smooth
group, the corresponding smooth bundle theory is the gerbe theory
of Chapter~\ref{ch:gerbes}.
\end{example}

\begin{example}[An inner action]
Let \(H\) be a Lie group, and take \(G\) to be a Lie group of smooth
automorphisms of \(H\) containing its inner automorphisms, with smooth
evaluation and inner-automorphism map. The map
\(h\mapsto \operatorname{Ad}_h\) with the evaluation action gives a
crossed module since
\(\alpha\operatorname{Ad}_h\alpha^{-1}=\operatorname{Ad}_{\alpha(h)}\).
Its kernel is the center \(Z(H)\), and its quotient of objects is
the chosen group of automorphisms modulo inner automorphisms.
A choice of finite-dimensional Lie group \(G\) is part of this
example. For instance \(H=U(n)\) and
\(G=PU(n)\) with conjugation work. This model keeps central
two-automorphisms even when every object belongs to a single
component.
\end{example}

\begin{example}[An abelian kernel with a nontrivial action]
Let \(V\) be a finite-dimensional real representation of \(G\), with
\(H=(V,+)\) and trivial boundary. The Peiffer identity holds because
\(V\) is abelian. There are no arrows between distinct \(g\)'s, but
tensoring with \(g\) acts on the automorphism arrows by the given
representation. This provides a model in which the two-form
connection takes values in an abelian group and still depends on
transport by a nonabelian one-form connection.
\end{example}
\chapter{Two-connections and their curvatures}
\label{ch:2connections}

A strict two-group assigns a boundary in \(G\) to each arrow
label in \(H\). On a small surface, the path transport around the
boundary is controlled by the curvature of an ordinary
connection. A two-form suitable for surface transport must have
that curvature as its infinitesimal boundary. This yields a
compatibility condition between the one- and two-form data.

\section{Forms for a differential crossed module}

Differentiate the crossed module of Chapter~\ref{ch:crossed-modules}.
Write \(\mathfrak h,\mathfrak g\) for the Lie algebras,
\(\partial_*\) for the differential of the boundary, and
\(X\triangleright b\) for the induced action of \(\mathfrak g\) on
\(\mathfrak h\) by derivations. The differentiated identities are
\[
 \partial_*(X\triangleright b)=[X,\partial_*b],
 \qquad
 (\partial_*a)\triangleright b=[a,b].
\]
They extend to forms by combining wedge of scalar forms with
the indicated bracket or action on coefficients. In particular,
brackets of forms are graded antisymmetric. When both inputs have
degree one, the factor \(\tfrac12\) in \(\tfrac12[A,A]\)
accounts for the two cross terms.

On a trivialization over \(U\), a pair of potentials consists of
\(A\in\Omega^1(U;\mathfrak g)\) and
\(B\in\Omega^2(U;\mathfrak h)\). Write \(D_A=d+A\triangleright(-)\)
on \(\mathfrak h\)-valued forms and \(D_A=d+[A,-]\) on
\(\mathfrak g\)-valued forms. Our conventions are
\begin{align}
 F_A&=dA+\tfrac12[A,A],\label{eq:two-f}\\
 K&=F_A-\partial_* B,\label{eq:two-fake}\\
 \mathcal H&=D_AB.\label{eq:two-h}
\end{align}
The form \(K\) is the fake curvature and \(\mathcal H\) is the
three-form curvature. We reserve the term \emph{fake-flat} for
\(K=0\). In the path two-functor model this condition is part of
the local correspondence; a pair of forms can also be considered
without imposing it~\cite{baez-schreiber,schreiber-waldorf-surface}.

\begin{proposition}[Higher Bianchi identities]
For a pair of potentials with conventions \eqref{eq:two-f}--\eqref{eq:two-h},
\[
 D_A K=-\partial_*\mathcal H,\qquad
 D_A\mathcal H=K\triangleright B.
\]
If \(K=0\), then \(\mathcal H\) takes values in \(\ker\partial_*\)
and is \(D_A\)-closed there. The operator \(D_A\) is flat on
\(\ker\partial_*\).
\end{proposition}

\begin{proof}
Equivariance of \(\partial_*\) gives
\(D_A\partial_*B=\partial_*D_AB\). The ordinary Bianchi identity
\(D_AF_A=0\), applied to \eqref{eq:two-fake}, gives the first formula.
On an \(\mathfrak h\)-valued form \(u\), expansion gives
\[
 D_A^2u
 =dA\triangleright u+A\triangleright(A\triangleright u)
 =F_A\triangleright u.
\]
The two terms involving \(du\) cancel, and the representation
identity makes \(A\triangleright(A\triangleright u)\) equal to
\(\tfrac12[A,A]\triangleright u\). In particular,
\[
 D_A\mathcal H
   =K\triangleright B+(\partial_*B)\triangleright B
   =K\triangleright B+[B,B].
\]
Here \([B,B]=0\) because \(B\) has even degree and the coefficients
are over \(\mathbb R\). When \(K=0\), the first identity gives
\(\partial_*\mathcal H=0\). The differentiated Peiffer identity
makes \(\ker\partial_*\) central in \(\mathfrak h\), and
equivariance makes it stable under the \(G\)-action. On that
subspace \(D_A^2u=(\partial_*B)\triangleright u=[B,u]=0\).
\end{proof}

Even under fake flatness \(D_A\)-closedness is a statement about
the flat local system on the kernel, rather than ordinary
closedness of each coefficient in an arbitrary frame. In a
parallel local frame for that system it becomes ordinary
closedness. The distinction is meaningful when the
\(G/\partial(H)\)-action on \(\ker\partial_*\) is nontrivial.

\section{Two elementary changes of potentials}

The ordinary frame change by \(g:U\to G\) acts by
\[
 A^g=\operatorname{Ad}(g^{-1})A+g^{-1}dg,
 \qquad B^g=g^{-1}\triangleright B.
\]
Differentiating the representation sends \(g^{-1}dg\) to the
logarithmic derivative of the matrix acting on \(\mathfrak h\).
This cancels the derivative of \(g^{-1}\) in
\(D_{A^g}(g^{-1}\triangleright B)\). It follows that \(K\) transforms
by \(\operatorname{Ad}(g^{-1})\) and \(\mathcal H\) by
\(g^{-1}\triangleright(-)\).

There is also a one-form change. For
\(a\in\Omega^1(U;\mathfrak h)\), put
\[
 A'=A+\partial_*a,\qquad
 B'=B+D_Aa+\tfrac12[a,a].
\]
Expanding the ordinary curvature gives
\[
 F_{A'}=F_A+\partial_*\bigl(D_Aa+\tfrac12[a,a]\bigr),
 \qquad K'=K.
\]
By the Peiffer identity \(D_{A'}=D_A+[a,-]\).
Use the graded derivation rule
\(D_A[a,a]=-2[a,D_Aa]\) and the graded Jacobi identity
\([a,[a,a]]=0\). Expanding the new three-form curvature leaves
\[
 \mathcal H'=\mathcal H+F_A\triangleright a+[a,B]
             =\mathcal H+K\triangleright a.
\]
In the last equality the terms \([B,a]\) and \([a,B]\) cancel.
Thus the one-form change preserves \(\mathcal H\) in the
fake-flat case. These formulas are local. Gluing a two-bundle
requires transition data and coherence on overlaps in addition
to these changes of representatives.

\section{The flat quotient connection}

Suppose the image \(N=\partial(H)\) is a closed Lie subgroup of
\(G\) with Lie algebra \(\partial_*\mathfrak h\). Equivariance
makes it normal. Write \(Q=G/N\), and use \(q\) for the quotient
map and \(q_*\) for its differential. In a fake-flat model
\[
 F_{q_*A}=q_*F_A=q_*\partial_*B=0.
\]
The equality uses the fact that \(q_*\) preserves brackets.
The change \(A\mapsto A+\partial_*a\) has no effect on \(q_*A\).

This calculation also gives a descent statement with explicit
hypotheses. Suppose local fake-flat pairs have changes of frame
\(g_{ij}\) on double overlaps, and \(h_{ijk}\) on triple overlaps
with \(g_{ij}g_{jk}=\partial(h_{ijk})g_{ik}\). Suppose the local
one-forms satisfy
\[
 A_j=A_i^{g_{ij}}+\partial_*a_{ij}
\]
for one-forms \(a_{ij}\) valued in \(\mathfrak h\).
Then \(q(g_{ij})q(g_{jk})=q(g_{ik})\), so the quotient transition
functions glue an ordinary principal \(Q\)-bundle.
The projected forms satisfy the ordinary gauge relation
with these transition functions. They define a flat connection
on that bundle. This is the part of the higher connection
visible after identifying two path transports joined by a
two-arrow. A smooth Lie group quotient was an assumption in
this argument. At the Lie-algebra level the projection to
\(\mathfrak g/\partial_*\mathfrak h\) and its local flatness
still make sense.

\section{Local models and transport compatibility}

\begin{example}[Abelian curving]
Let \(G\) be trivial and \(H=U(1)\). In real coordinates on the
Lie algebra, the two-form potential \(b\) has three-form
curvature \(db\). A one-form change is \(b\mapsto b+d\lambda\).
On \(\mathbb R^3\), take \(b=x\,dy\wedge dz\). Then
\(db=dx\wedge dy\wedge dz\), so a fake-flat potential can have
nonzero curvature. The leading transport on a surface comes
from integrating \(b\), and its change along a three-dimensional
deformation comes from integrating \(db\).
\end{example}

\begin{example}[A twisted abelian kernel]
For the crossed module \(V\to G\) with zero boundary from
Chapter~\ref{ch:crossed-modules}, fake flatness forces \(F_A=0\).
The form \(B\) is \(V\)-valued and its curvature is \(D_AB\).
Take \(G=GL(V)\) and \(U=(\mathbb R/\mathbb Z)\times S^2\).
Let \(T\in\operatorname{End}(V)\),
\(A=T\,d\theta\), and \(B=v\,\omega\), with \(\theta\) the
circle coordinate, \(v\in V\), and \(\omega\) an area form on
\(S^2\) pulled back to \(U\). Then \(F_A=0\) and
\[
 \mathcal H=Tv\,d\theta\wedge\omega.
\]
The local system can have monodromy \(\exp(-T)\) around the
circle. The three-form uses the action even though its
coefficients lie in an abelian Lie algebra.
\end{example}

\begin{example}[Identity boundary]
For \(\partial=\operatorname{id}_G\) with conjugation action,
fake flatness sets \(B=F_A\), hence
\(\mathcal H=D_AF_A=0\). The underlying groupoid has a unique
arrow between any pair of objects. It records a comparison
between transports but has no nonidentity automorphism at a
fixed transport.
\end{example}

For the relevance of \(K\) to surface integration, consider a
small coordinate rectangle oriented by \(dx\wedge dy\). Its
bottom and right edges give a path \(\gamma_0\) from the lower
left to the upper right. The left and top edges give a parallel
path \(\gamma_1\). Use \(U_\gamma\) for parallel transport with
equation \(\dot U=-A(\dot\gamma)U\). Comparing the two endpoint
operators gives \(U_{\gamma_1}U_{\gamma_0}^{-1}\). This product
corresponds to the clockwise boundary based at the common
endpoint. On a square of side length \(\epsilon\), the standard
curvature expansion, in a smooth frame, reads
\[
 U_{\gamma_1}U_{\gamma_0}^{-1}
   =1+\epsilon^2 F_A(\partial_x,\partial_y)+O(\epsilon^3)
\]
in coordinates at the identity. One can see the coefficient
by choosing a frame with \(A=0\) at the lower left. The line
integral around the rectangle gives the exterior derivative
term to order two, and the bracket term at that point is
zero in the chosen frame. Covariance gives the formula in
the original frame. A surface arrow \((h,U_{\gamma_0})\)
has target \(U_{\gamma_1}\) exactly when
\(\partial(h)=U_{\gamma_1}U_{\gamma_0}^{-1}\). The coefficient
of area in \(h\) is \(B(\partial_x,\partial_y)\). Equating
coefficients gives \(K=0\). The next chapter describes the
global composition law for such arrows.
\chapter{Thin homotopy and surface transport}
\label{ch:surface-transport}

A bundle with connection sends a path to an isomorphism between
fibers. A higher connection sends a deformation between two paths
to a comparison between their isomorphisms. The parameterization of
the deformation should have no effect on this comparison. The
rank conditions defining thin homotopy express the precise
independence needed here. A homotopy whose image sweeps out a
three-dimensional volume can change the comparison, and the
three-form curvature measures that effect.

\section{The path two-groupoid}

Take paths with sitting instants, so a path is constant on collars
of its endpoints. One-morphisms from \(x\) to \(y\) are paths
modulo endpoint-preserving smooth homotopies of differential rank
at most one. A \emph{bigon} from \(\gamma_0\) to \(\gamma_1\) is a
smooth map \(\Sigma(t,s)\) on the square. The parameter \(t\) runs
along a path and \(s\) varies it:
\[
 \Sigma(t,0)=\gamma_0(t),\qquad
 \Sigma(t,1)=\gamma_1(t),\qquad
 \Sigma(0,s)=x,\qquad \Sigma(1,s)=y.
\]
Require collars on which the map is independent of the coordinate
normal to each side. We orient its integration domain by
\(dt\wedge ds\), so its boundary is \(\gamma_0-\gamma_1\) after
discarding the constant edges. Two bigons are equivalent if they
are joined by a collared smooth homotopy of rank at most two,
with fixed end objects and with rank at most one on the two side
faces that vary the boundary paths. The two-morphisms are these
classes of bigons.

Vertical composition stacks squares in \(s\). If boundary paths
have only the same thin class, a thin bigon joins the chosen
representatives before stacking. Horizontal composition uses
concatenation in \(t\). The collars make concatenation smooth.
Different choices of rescaling on the concatenated intervals give
homotopic parameterizations through the same surface. Such a
homotopy has rank at most two because it factors through that
surface, and its restrictions to the moving path faces have rank
at most one. In the full quotient these compositions give the
strict path two-groupoid \(\Pi_2^{\mathrm{thin}}(M)\). In particular
the independence of the connecting thin bigon and the interchange
law are part of the path two-groupoid construction
in~\cite[\S 2.1]{schreiber-waldorf-surface}. Thin equivalence of two
bigons is more restrictive than ordinary homotopy, even when
the boundary paths agree pointwise.

\section{Smooth two-functors and local potentials}

Let \(\mathcal G\) be the finite-dimensional strict Lie two-group
defined by a smooth crossed module. Its one-object two-category
is denoted by \(B\mathcal G\) in this chapter. A two-functor
\(T:\Pi_2^{\mathrm{thin}}(U)\to B\mathcal G\) sends the paths to
elements of \(G\). It associates to a bigon \(\Sigma\) with
source \(\gamma_0\) an arrow \((h_\Sigma,T(\gamma_0))\) as in
Chapter~\ref{ch:crossed-modules}. In particular
\[
 \partial(h_\Sigma)T(\gamma_0)=T(\gamma_1).
\]
For successive vertical bigons with labels \(h,h'\), the composite
has label \(h'h\). For horizontal composition, if \(\Sigma_1\)
runs before \(\Sigma_2\) and their sources have transports \(g_1,g_2\),
the label is
\[
 h_2(g_2\triangleright h_1).
\]
The action accounts for transporting the first surface label to
the frame used by the second. Diagrammatically, the comparison
belongs to the hom-groupoid of the target two-category:
\[
\begin{tikzcd}[column sep=7em]
 {*} \arrow[r,bend left=40,"{T(\gamma_0)}",""{name=upper,inner sep=0pt}]
     \arrow[r,bend right=40,"{T(\gamma_1)}",""{name=lower,inner sep=0pt}]
   & {*}
 \arrow[Rightarrow,from=upper,to=lower,shorten <=8pt,shorten >=8pt,"h_\Sigma"]
\end{tikzcd}
\]
Smoothness here means smoothness on smooth families in the path and
bigon quotients with their quotient diffeologies. Use the smooth
structure of \(G\) and \(H\rtimes G\) on the values. Thus a
set-theoretic two-functor need not have local differential-form data.

With a fixed trivialization on \(U\), Schreiber and Waldorf prove
that these smooth two-functors correspond to pairs
\((A,B)\in\Omega^1(U;\mathfrak g)\times\Omega^2(U;\mathfrak h)\)
satisfying \(F_A=\partial_*B\), in the conventions for orientation
and infinitesimal extraction used there
\cite[Theorem~2.21]{schreiber-waldorf-surface}. One extracts the
one-form from short paths and the two-form from shrinking
rectangles. The converse uses ordinary differential equations for
path transport and a second integration for the bigon label.
The boundary calculation of Chapter~\ref{ch:2connections}
is the infinitesimal necessity of the equation in our conventions.
The theorem supplies sufficiency and reconstructs the two-functor.
Global transport for a nontrivial higher bundle uses local
trivializations and their higher transformation data. The global
formulation with transport two-functors and descent is developed
in~\cite{schreiber-waldorf-gerbe}.

\section{Abelian transport from a two-form}

The abelian case has a direct construction. Take \(G=\{1\}\)
and \(H=U(1)\). Let \(b\in\Omega^2(M)\) be a global curving for
a trivial gerbe. All path transports are the sole \(G\)-value,
and the surface assignment is
\[
 h_b(\Sigma)=\exp\!\left(2\pi i\int_{[0,1]^2}\Sigma^*b\right).
\]
For either form of concatenation, the integral over the new square
is the sum of the integrals over the two old squares after
orientation-preserving changes of variable. Thus the assignment
respects both compositions. The identity bigon factors through a
path, so its integral is zero. Smoothness on parameter families
follows by differentiating an integral over a fixed compact domain.

For thin invariance take a bigon homotopy \(V\) on the cube.
At every point \(\operatorname{rank}(dV)\le 2\), so the pullback
of \(db\) is zero. Two side faces are constant and the other two
have rank at most one, so their pullbacks of \(b\) are zero.
Using the cube orientation with the deformation parameter
first and Stokes' theorem gives
\[
 \int \Sigma'^*b-\int\Sigma^*b=\int V^*db=0.
\]
This proves that the surface formula descends to the thin quotient
even if \(db\) is nonzero. For an unrestricted homotopy with
fixed boundary paths the same calculation gives
\[
 \frac{h_b(\Sigma')}{h_b(\Sigma)}
   =\exp\!\left(2\pi i\int V^*db\right).
\]
In particular the local curvature is recoverable from infinitesimal
three-dimensional changes of holonomy.

If \(b'=b+d\lambda\) for a real one-form \(\lambda\), then
\[
 \frac{h_{b'}(\Sigma)}{h_b(\Sigma)}
 =\exp\!\left(2\pi i\bigl(
     \textstyle\int_{\gamma_0}\lambda-
     \textstyle\int_{\gamma_1}\lambda\bigr)\right).
\]
The ratio is a boundary comparison for an open bigon and is one for
a closed surface. On smooth singular two-cycles \(z\), the rule
\(z\mapsto\int_z b\bmod\mathbb Z\) is a degree-three differential
character. For any smooth singular three-chain \(c\), its value on
\(\partial c\) is \(\int_c db\bmod\mathbb Z\). The underlying integral
class is zero for this globally defined potential. A gerbe with
nonzero integral class uses locally defined curvings with
transition terms, as in Chapter~\ref{ch:gerbes}.

For example, on \(\mathbb R^3\) take
\(b=\tfrac12 x\,dy\wedge dz\) and let \(c\) be the unit cube with
its standard orientation. The curvature integral is \(1/2\).
Equivalently the face at \(x=1\) contributes \(1/2\) with outward
orientation while the other faces contribute zero. Its surface
holonomy is \(-1\) even though \(\partial c\) is a boundary.
This is consistent with the differential character law and
with the invariance of holonomy under thin deformations.

\section{Transgression and orientation}

For a global \(b\), a smooth family of loops carries the
transgressed one-form
\[
 (\tau b)_\gamma(V)
  =\int_{S^1} b_{\gamma(t)}(\dot\gamma(t),V(t))\,dt .
\]
Here the variation vector \(V\) is along the loop. The formula
defines a form on families without putting a choice of
infinite-dimensional manifold structure on the free loop space.
A path of loops makes a cylinder. With the loop tangent ordered
first, the integral of \(\tau b\) over the path equals the integral
of \(b\) on this cylinder. Consequently the trivial line with
connection form \(-2\pi i\,\tau b\) has precisely the cylinder
transport coming from \(h_b\).

The sign in its curvature can be checked on any finite-dimensional
parameter space \(S\). Write the pulled-back two-form on
\(S^1\times S\) as \(dt\wedge\alpha(t)+\beta(t)\),
with \(\alpha,\beta\) forms on \(S\). The coefficient of \(dt\)
in its exterior derivative is \(\partial_t\beta-d_S\alpha\).
Integrating in \(t\), the derivative of a periodic function
integrates to zero. Thus \(d\,\tau b=-\tau(db)\) for this
fiber-first convention. One can instead choose a convention that
reverses both transgression and cylinder orientation. The formula
on a trivial gerbe explains why transgression on a global gerbe
produces a line bundle with connection on loop families and why
the resulting curvature is a transgressed three-form.
\chapter{Weil algebras and higher characteristic maps}
\label{ch:higher-chern-weil}

For an ordinary Lie group a characteristic form is obtained by
evaluating an invariant polynomial on curvature. Chapter~\ref{ch:weil}
expressed that evaluation in the Weil algebra. A higher analogue
starts from a graded algebra whose differential encodes brackets
of several arities. The characteristic form, the Chern--Simons
element, and the infinitesimal cocycle then appear as three
parts of one calculation. Passing from that calculation to a
differential characteristic class requires global integration
and descent.

\section{The algebra of connections and curvatures}

Work over \(\mathbb R\) with cohomological degrees. For the local
models here, a finite-dimensional \(L_\infty\)-algebra is presented
by its Chevalley--Eilenberg algebra
\[
 \operatorname{CE}=\bigl(\operatorname{Sym}(V),q\bigr)
\]
where \(V\) is a finite graded space in positive degrees,
\(\operatorname{Sym}\) is graded-commutative, and \(q\) is an
augmentation-preserving derivation of degree one with \(q^2=0\).
The homogeneous components of \(q\) in each word length are dual
to the brackets. For an ordinary Lie algebra the generators have
degree one and \(q\) is quadratic. A degree-two generator can
record a higher abelian direction. We only use free graded
algebras with polynomial differential, so no completion is needed.

Write \(sV\) for a copy of \(V\) raised by one degree. As a graded
algebra, the Weil algebra is
\[
 W=\operatorname{Sym}(V\oplus sV).
\]
Extend \(s(v)=sv\), \(s(sv)=0\) as a derivation of degree one.
The differential is determined by
\[
 d_W v=qv+sv,\qquad d_W(sv)=-s(qv).
\]
On the unshifted algebra, \(d_W\) is \(q+s\). The identity
\(d_Ws+sd_W=0\) holds on the generators and hence everywhere.
On \(v\), \(d_W^2v=s(qv)-s(qv)=0\). On \(sv\) it is
\(s\,d_Wqv=s^2qv=0\), so this is indeed a differential graded
algebra. The map \(j^*:W\to\operatorname{CE}\) kills \(sV\) and
is a chain map. It restricts the algebra of connections to
the corresponding flat infinitesimal data.

\begin{proposition}[Acyclicity of the Weil algebra]
The cohomology of \(W\) is \(\mathbb R\) in degree zero and
zero in positive degrees.
\end{proposition}

\begin{proof}
Use \(v\) and \(t(v)=sv+qv\) as algebra generators; the change
is invertible because \(qv\) belongs to the unshifted algebra.
In these coordinates \(d_W v=t(v)\) and \(d_W t(v)=0\).
Define the degree-minus-one derivation \(h\) by
\(h(v)=0\) and \(h(t(v))=v\). The derivation
\(d_Wh+hd_W\) acts as the identity on generators, so on
monomials of word length \(r\) it acts as multiplication by
\(r\). The differential preserves this word length. On
positive word length define the contraction by \(h/r\).
It contracts the augmentation ideal, leaving the constants.
\end{proof}

A map of differential graded algebras
\(\Phi:W\to\Omega^\bullet(U)\) evaluates \(v\) as a potential
of degree \(|v|\), and \(sv\) as its raw curvature of degree
\(|v|+1\). It is specified by the potentials: the shifted
values must be \(d\Phi(v)-\Phi(qv)\), and the equation for
\(d_W(sv)\) then follows from \(d_W^2=0\). For a Lie algebra,
write \(A\) for the collection of degree-one generators and
\(F\) for their shifts. The universal identities are exactly
\[
 d_W A=F-\tfrac12[A,A],\qquad
 d_WF=-[A,F].
\]
They recover the structure equation and Bianchi identity after
evaluation. A map that annihilates the shifted generators
factors through \(\operatorname{CE}\).

\section{An invariant form and its two primitives}

The subalgebra generated by \(sV\) is the algebra of horizontal
polynomials. Call a horizontal polynomial \(P\) invariant if
\(d_W P=0\). For ordinary reductive Lie algebras this gives the
usual invariant polynomials on their curvature variables
\cite[Remark~4.1.17]{fiorenza-schreiber-stasheff}. This elementary
definition suffices for the examples below.

When \(P\) has positive degree \(n+1\), choose
\(C\in W^n\) with \(d_W C=P\) using acyclicity. Since \(j^*P=0\),
\(\mu=j^*C\) is a cocycle of degree \(n\). Two choices of \(C\)
differ by a closed element. If \(n>0\), acyclicity gives an
exact difference, so the class of \(\mu\) is independent of
that choice. This is the transgression relation in
\cite[Definition~4.1.20]{fiorenza-schreiber-stasheff}.
On \(U\), evaluation gives
\[
 d\Phi(C)=\Phi(P).
\]
Under flat evaluation it gives \(\Phi(C)=\Phi(\mu)\), using
the factorization through the Chevalley--Eilenberg algebra.
Thus \(\Phi(C)\) is a curvature primitive on a chart and
\(\mu\) records its flat restriction.

The diagram separates algebraic restriction from form evaluation:
\[
 \begin{tikzcd}[column sep=3.8em,row sep=2.5em]
 W \arrow[r,"\Phi"] \arrow[d,"j^*"'] &
    \Omega^\bullet(U)\\
 \operatorname{CE} \arrow[ur,dashed,"\Phi_{\mathrm{flat}}"'] &
 \end{tikzcd}
\]
The dashed factorization is defined when \(\Phi(sV)=0\).
The algebra gives a local real form for \(P\) even without
that factorization. Gluing this form and refining it to an
integral differential class requires the global structure
group, compatible descent data, and an integral normalization.
For example, if \(\Phi(P)\) is the curvature of a differential
character, its integrals on integral cycles must be integers.
The local Weil calculation by itself has no such period
condition.

\begin{example}[The degree-three abelian case]
Take \(\operatorname{CE}=\mathbb R[b]\) with \(|b|=2\) and \(qb=0\).
The Weil algebra adds \(h=sb\) with \(|h|=3\) and
\(d_Wb=h,\ d_Wh=0\). Thus \(P=h,\ C=b,\ \mu=b\).
A map of Weil algebras into forms assigns \(B=\Phi(b)\)
and \(\mathcal H=\Phi(h)=dB\). Taking the smooth circle
as the integrated coefficient group gives the curving
of a gerbe on a trivializing open set. Integral Čech descent
data can change its underlying degree-three class while
leaving the same equations on each open set. The
holonomy and curvature law for a single global \(B\)
were computed in Chapter~\ref{ch:surface-transport}.
\end{example}

\section{The string transgression}

Let \(\mathfrak g\) be an ordinary Lie algebra with an invariant
symmetric bilinear form \(Q\). Its degree-four characteristic
polynomial is \(P=Q(F,F)\). The connection interpolation
\(A_t=tA\) in the ordinary Weil algebra has curvature
\[
 F_t=tF+\tfrac12(t^2-t)[A,A].
\]
Differentiating the polynomial with respect to \(t\) gives
\(2d_WQ(A,F_t)\) by invariance of \(Q\) and the Bianchi
identity for \(A_t\). Integrating from \(0\) to \(1\) gives
\[
 C=Q(A,F)-\tfrac16Q(A,[A,A]),\qquad d_W C=P .
\]
In these signs, \(\mu=-\tfrac16Q(A,[A,A])\) is the
Chevalley--Eilenberg restriction. Reversing the sign of the
new abelian generator gives the alternate convention for
the corresponding central extension.

Adjoin a degree-two generator \(b\) to the Chevalley--Eilenberg
algebra with \(qb=-\mu\). This is a differential because
\(q\mu=0\). In its Weil algebra, write \(h=sb\).
Then \(d_Wb=h-\mu\), \(d_Wh=s\mu\), and
\(d_W\mu=s\mu\). The modified element
\[
 \mathcal K=h+C-\mu
\]
satisfies \(j^*\mathcal K=0\) and \(d_W\mathcal K=P\).
After evaluation this reads
\[
 H=dB+\Phi(C),\qquad dH=\Phi(P).
\]
The new three-form \(H\) uses a Chern--Simons term. This
Bianchi equation is different from the fake-flat strict
crossed module equation in Chapter~\ref{ch:2connections}.
It arises from an extension whose Chevalley--Eilenberg
differential has the cubic term \(\mu\).

For the spin group in stable rank, take \(Q\) normalized
so \(P\) represents the real image of the integral class
\(\lambda=\tfrac12 p_1\). The topological string lift is a lift
through the homotopy fiber of
\[
 B\operatorname{Spin}\xrightarrow{\ \lambda\ } K(\mathbb Z,4).
\]
By the universal property of a homotopy fiber, a lift of
a fixed spin classifying map is a chosen nullhomotopy of
its composite with \(\lambda\). Consequently existence
is equivalent to vanishing of the integral class on the
base. Vanishing of the de Rham image alone leaves a
possible torsion obstruction. At the differential level,
if the underlying integral class vanishes, the differential
character is represented by a three-form \(H\) modulo closed
forms of integral periods, and its curvature is \(dH\).
A geometric differential string structure keeps track of a
choice of trivialization and its coherences. Treating it
as a nullhomotopy to the zero \emph{differential} character
would impose zero curvature and would miss the three-form
in this Bianchi identity.

Fiorenza, Schreiber, and Stasheff construct the global differential
refinement of the fractional Pontryagin map at the level of
smooth cocycle spaces and describe the corresponding
differential trivializations
\cite[\S\S 5.3 and 6]{fiorenza-schreiber-stasheff}. Their construction
uses integration of the infinitesimal model and compatible
periods. The local equations above explain which forms such
a refinement must recover. The global characteristic map
also retains the integral class and the higher automorphisms
that the form equations leave undetermined.

\bookpart{Obstructions, equivariance, and derived geometry}
\chapter{The string obstruction and its differential lifts}
\label{ch:string}

A spin connection has a curvature polynomial of degree four. The same
polynomial represents a distinguished integral class, whose vanishing
is a lifting problem. A primitive of the curvature form can exist even
when the integral obstruction survives. Conversely an integral lift
can exist while a particular differential refinement is nonzero.
Keeping track of the spaces in which those conditions live separates
the choices.

\section{The topological fiber}

Let \(E\to M\) be an oriented Euclidean real vector bundle of rank
\(n\geq 5\) over a smooth manifold, with a specified spin lift \(P\)
of its oriented orthonormal frame bundle. A spin lift supplies a
classifying map \(c_P:M\to B\operatorname{Spin}(n)\), where the
domain denotes the underlying homotopy type. The universal integral
class \(\lambda\in H^4(B\operatorname{Spin}(n);\mathbb Z)\)
satisfies \(2\lambda=p_1\) pulled back along
\(B\operatorname{Spin}(n)\to B\operatorname{SO}(n)\). We use this
universal class to define the half-Pontryagin class of \(P\).
An equation \(2x=p_1(E)\) in the cohomology of a given \(M\)
would not define it uniquely when \(H^4(M;\mathbb Z)\) has
two-torsion. The integral assertion is part of the spin
classification \cite[Section~5]{stolz-teichner}.

Represent \(\lambda\) by a map to \(K(\mathbb Z,4)\). Define the
topological classifying space for string lifts by the homotopy
pullback
\[
\begin{tikzcd}[row sep=2.7em,column sep=3.7em]
 B\operatorname{String}(n)\arrow[r]\arrow[d] &
 {*} \arrow[d,"0"]\\
 B\operatorname{Spin}(n)\arrow[r,"\lambda"] & K(\mathbb Z,4).
\end{tikzcd}
\]
The right-hand arrow chooses the zero class. A lift is taken
over \(c_P\), so the choice includes the comparison with that
specified spin bundle.

\begin{proposition}[Lifts of a fixed spin bundle]
\label{prop:string-lifts}
The space of lifts of \(c_P\) is nonempty precisely when
\(\lambda(P)=0\). When it is nonempty, its components form a
torsor for \(H^3(M;\mathbb Z)\). After choosing a lift as
basepoint, its homotopy groups in degrees \(1\) through \(3\)
are \(H^{3-j}(M;\mathbb Z)\) in degree \(j\), and the higher
groups vanish.
\end{proposition}

\begin{proof}
Mapping out of \(M\) preserves limits in the category of spaces.
The lift space is therefore the space of paths from
\(\lambda c_P\) to the constant map inside
\(\operatorname{Map}(M,K(\mathbb Z,4))\). The description
of connected components by cohomology gives the obstruction.
When a path exists, composing with it identifies every other
path with a loop at the constant map, up to coherent homotopy.
Looping commutes with taking this mapping space, giving
\(\operatorname{Map}(M,K(\mathbb Z,3))\). Its components are
\(H^3(M;\mathbb Z)\), and the action by path composition is
simply transitive on components. For \(j\geq 1\), a based
\(j\)-sphere of maps to \(K(\mathbb Z,3)\) has adjoint a map
from \(S^j\wedge M_+\) to \(K(\mathbb Z,3)\), where \(M_+\)
adds a disjoint basepoint. The suspension isomorphism gives
the formula.
\end{proof}

The proof distinguishes classification from existence. For instance,
the trivial spin bundle over \(S^3\) has infinitely many lift classes,
indexed by \(\mathbb Z\) after choosing one of them. Its spin
curvature may be identically zero throughout this family. The same
proof yields automorphisms and higher automorphisms, since a choice
of lift class identifies a component without eliminating its
higher homotopy.

\section{Normalization and refinement}

Take the curvature \(F=dA+A\wedge A\) of a spin connection in the
real defining representation of \(\mathfrak{so}(n)\). In our
Chern normalization, the real form for \(\lambda\) is
\[
 \lambda(F)=-\frac{1}{16\pi^2}\operatorname{tr}(F\wedge F).
\]
Indeed the degree-four part of
\(\det(1+iF/(2\pi))\) for the complexification of \(E\) is
\(\operatorname{tr}(F\wedge F)/(8\pi^2)\), since
\(\operatorname{tr}F=0\). The relation
\(p_1(E)=-c_2(E\otimes\mathbb C)\) supplies the minus sign.
Trace invariance and the Bianchi identity make this form closed
and independent of the frame. The universal integral normalization
specifies which refinement is meant.

The differential characteristic class \(\widehat\lambda(P,A)\)
has integral class \(\lambda(P)\) and curvature \(\lambda(F)\).
It can be realized by the Chern--Simons bundle \(2\)-gerbe of
the spin bundle, equipped with its induced connection
\cite[Theorems~1.1.3 and~1.3.3]{waldorf-string}. A geometric string
structure in this model consists of a trivialization of that
bundle \(2\)-gerbe and compatible connective data. Such data
produce a \(3\)-form \(H\) with \(dH=\lambda(F)\). Tensoring
the trivialization by a bundle gerbe with connection adds that
gerbe's curvature to \(H\). Waldorf's existence theorem gives
compatible connections on every topological trivialization
\cite[Theorem~1.3.4 and Section~3.2]{waldorf-string}.

The exact sequence of differential cohomology makes the
role of \(H\) explicit. Write \(\Omega^3_{\mathbb Z}(M)\)
for closed forms with integral periods. We have
\[
 \Omega^3(M)/\Omega^3_{\mathbb Z}(M)
 \xrightarrow{\ a\ } \widehat H^4(M;\mathbb Z)
 \xrightarrow{\ I\ } H^4(M;\mathbb Z)\longrightarrow 0 .
\]
Here \(a(H)\) evaluates on a \(3\)-cycle by integration modulo
\(\mathbb Z\). If \(\lambda(P)=0\), the refinement has the form
\(a(H)\). Its curvature is \(dH\), which may be nonzero.
Changing \(H\) by a closed form of integral periods keeps
\(a(H)\) fixed. The exact sequence describes the image of
the differential class as a form modulo periods. A specified
trivialization includes gluing data beyond that image.
In particular the requirement
\(\widehat\lambda(P,A)=0\) is stronger than existence of a
geometric string structure, which allows a topologically
trivial \(2\)-gerbe carrying a nonzero global curving.

The difference of two such compatible trivializations, for
the same spin connection, is a bundle gerbe with connection.
One way to see this is to compare their local descent
cochains. Their coboundaries represent the same obstruction
cocycle, so the quotient of their data satisfies the
unobstructed gerbe descent equations. Conversely multiplication
by such a cocycle changes the trivialization and preserves
its relation to the obstruction cocycle. On isomorphism
classes this is the torsor action of \(\widehat H^3(M;\mathbb Z)\).
Taking underlying integral classes recovers the \(H^3\)-action
in Proposition~\ref{prop:string-lifts}, and taking curvature
recovers addition to \(H\). In the \(2\)-gerbe description
these operations are tensor product and inverse of gerbes
with connection \cite[Corollary~1.3.2 and Section~3.2]{waldorf-string}.

\section{Nonzero integral and differential obstructions}

The quaternionic Hopf bundle \(Q\) on \(S^4=\mathbb H P^1\)
has complex rank two and structure group
\(\operatorname{Sp}(1)=\operatorname{SU}(2)\).
Its unit sphere bundle is \(S^7\to S^4\), so the Gysin
sequence identifies its Euler class with a generator of
\(H^4(S^4;\mathbb Z)\), up to orientation. This Euler class
is \(c_2(Q)\) for the complex orientation. The real
five-plane bundle \(Q_{\mathbb R}\oplus\mathbb R\) has a
spin lift. The group \(\operatorname{SU}(2)\) is simply
connected, so its orthogonal representation on
\(\mathbb H\oplus\mathbb R\) lifts to \(\operatorname{Spin}(5)\).
For a complex two-plane bundle,
\(p_1(Q_{\mathbb R})=c_1(Q)^2-2c_2(Q)=-2c_2(Q)\),
as follows from the formal Chern roots.
Here \(H^4(S^4;\mathbb Z)\) is torsion-free, so the spin
\(\lambda\) is \(-c_2(Q)\). Proposition~\ref{prop:string-lifts}
then obstructs a string structure.

There is also a flat model with a torsion obstruction.
Let \(Y=S^5/\mu_m\), where the \(m\)-th roots of unity act by
common scalar multiplication on the unit sphere in \(\mathbb C^3\)
and \(m\geq 2\). The map \(p:Y\to\mathbb C P^2\) is the
unit circle bundle of \(L^{\otimes m}\), with \(L\) the
tautological line. Indeed the fiberwise tensor-power map
on the unit circle bundle of \(L\) is surjective with
kernel \(\mu_m\). Put \(x=c_1(L)\).
The Gysin sequence for \(p\) contains
\[
 H^2(\mathbb C P^2;\mathbb Z)\xrightarrow{\ \smile mx\ }
 H^4(\mathbb C P^2;\mathbb Z)\xrightarrow{\ p^*\ }
 H^4(Y;\mathbb Z)\longrightarrow 0 .
\]
The final zero is \(H^3(\mathbb C P^2;\mathbb Z)\), and
the first two groups are generated by \(x\) and \(x^2\).
It follows that \(u=p^*(x^2)\) has order exactly \(m\).
The line \(p^*L\) admits a flat connection because its
pullback to \(S^5\) has the unit frame \(q\mapsto q\),
and the deck group changes this frame by constant phases.
The rank-two bundle \(Q=p^*L\oplus(p^*L)^{-1}\) carries
the direct sum of the flat connection and its dual.
Its determinant reduction has structure group
\(\operatorname{SU}(2)\) and \(c_2(Q)=-u\).
For the associated spin five-plane bundle the identity
\(\lambda=-c_2\) follows by naturality from \(B\operatorname{SU}(2)\),
whose integral \(H^4\) is free, by the complexification computation
above. Thus \(\lambda=u\) while \(F=0\). This obstruction comes
entirely from the integral gluing data.

For the other direction take the trivial spin bundle on
\(T^3=(\mathbb R/\mathbb Z)^3\) and the connection
\(A=tX\alpha\) in a global frame. Here \(X\) is a nonzero
real skew-symmetric matrix from \(\mathfrak{spin}(n)\),
\(t\in\mathbb R\), and
\[
 \alpha=\cos(2\pi z)\,dx+\sin(2\pi z)\,dy.
\]
The transgression from the trivial connection to \(A\)
is \(-t^2\operatorname{tr}(X^2)\alpha\wedge d\alpha/(16\pi^2)\).
This follows by inserting \(sA\) into the curvature polynomial
and integrating its derivative for \(0\leq s\leq 1\).
The cubic term vanishes because \(A\) has just one form factor.
Direct differentiation gives
\(\alpha\wedge d\alpha=-2\pi\,dx\wedge dy\wedge dz\),
so the transgression integral is a nonzero multiple of \(t^2\).
The differential characteristic class changes by \(a\) of
this transgression. For small nonzero \(t\) the resulting
degree-four differential character has nonzero value on \(T^3\),
while its integral class and its \(4\)-form curvature both
vanish. It records the comparison of two spin connections
on the same topologically trivial bundle. A geometric
trivialization can retain that comparison as a curving.
\chapter{Connections on quotient stacks}
\label{ch:quotients}

A group action carries information at each point about its stabilizer.
Taking an orbit space discards those groups. A quotient stack keeps
them as automorphisms, and characteristic classes can detect
representations of those automorphisms even when the base orbit
space is a point.

\section{The action groupoid and its test objects}

Let a Lie group \(K\) act smoothly from the left on \(X\). The
action groupoid has objects \(x\in X\) and arrows \(x\to kx\)
for \(k\in K\). Given arrows labelled \(k\) and \(l\) in that
order, their composite is labelled \(lk\). Its nerve in
degree \(q\) is \(K^q\times X\). A sequence of composable
arrows is determined by its initial point and its \(q\)
group elements, and the inner face maps multiply adjacent
elements in their composition order. The outer faces forget
the first or last vertex. In particular the first face
also moves the initial point.

The quotient stack \([X/K]\) is described on a test manifold
\(T\) by pairs \((P,u)\) with \(P\to T\) a right principal
\(K\)-bundle and \(u:P\to X\) a smooth map satisfying
\(u(pk)=k^{-1}u(p)\). Arrows are principal bundle isomorphisms
commuting with the maps to \(X\). A frame \(p_i:T_i\to P\)
gives a map \(x_i:T_i\to X\). On \(T_{ij}\) write
\(p_j=p_i k_{ij}\), so \(x_j=k_{ij}^{-1}x_i\).
On triples \(k_{ij}k_{jk}=k_{ik}\). Conversely this local
data glue the trivial principal bundles by their transition
functions, and equivariance of the local maps makes them
glue to \(u\). The same argument on isomorphisms identifies
the arrows. This exhibits the passage from the action
groupoid to its stack as descent for locally trivial bundles.

Taking \(K\) trivial recovers \(X\). Taking \(X=*\) gives
\(BK\), because the map \(u\) is then forced. For a free
proper action with smooth quotient and principal bundle
\(X\to X/K\), this stack is represented by \(X/K\). Indeed
the composition \(P\to X\to X/K\) descends to \(T\), and the
resulting map to \(X/K\) makes \(P\) isomorphic to its pullback
of \(X\), with the right action on \(X\) obtained by inversion.
This follows fiberwise since an equivariant map between
torsors is an isomorphism, and local sections give smoothness
of the inverse. In general a test map \(*\to[X/K]\) determined
by \(x\in X\) has automorphisms identified with its stabilizer.
The projection to the orbit set sends each such automorphism
to the identity.

\section{Bundles and the condition on a connection}

Let \(G\) be another Lie group. A principal \(G\)-bundle on
\([X/K]\) amounts to a \(K\)-equivariant principal \(G\)-bundle
\(Q\to X\). Pullback to the atlas \(X\to[X/K]\) gives \(Q\).
The chosen descent isomorphism on \(K\times X\) acts in the
fiber over \(x\) by \(Q_x\to Q_{kx}\). The cocycle condition on
\(K^2\times X\) says this action respects multiplication.
Conversely these isomorphisms give the bundle over any test
\((P,u)\), namely \((u^*Q)/K\to T\). The right action
on \(u^*Q\) sends \((p,q)\) to \((pk,k^{-1}q)\).
Its quotient is a principal \(G\)-bundle, as can be checked
after a local trivialization of \(P\). This also proves
the assertion for morphisms.

A connection form \(\omega\) on \(Q\) descends with equality
on the atlas nerve exactly when
\[
 b^*\omega=\operatorname{pr}_2^*\omega,\qquad
 b:K\times Q\longrightarrow Q,\quad b(k,q)=kq.
\]
Evaluating on the \(Q\)-directions at each fixed \(k\)
requires \(K\)-invariance. Evaluating on vectors tangent to \(K\)
at the identity requires \(\omega(\xi_Q)=0\), where
\(\xi_Q(q)=\frac{d}{dt}|_0\exp(t\xi)q\).
Those conditions imply the equality on all tangent directions
by their behavior under translations. Thus a \(K\)-invariant
connection can fail this descent condition. An invariant
connection still has a role in computing equivariant
characteristic forms. The groupoid formulation and the
comparison with the usual equivariant Chern--Weil map are
developed in \cite{lgtx}.

For a \(K\)-invariant connection write \(\mu_\xi=\omega(\xi_Q)\),
an equivariant \(\mathfrak g\)-valued function. If \(F\) is the
ordinary curvature and \(D_\omega=d+[\omega,-]\), then
\[
 \iota_{\xi_Q}F=-D_\omega\mu_\xi .
\]
In fact Cartan's formula gives
\(\iota_{\xi_Q}d\omega=-d\mu_\xi\), and contraction of
\(\tfrac12[\omega,\omega]\) is \([\mu_\xi,\omega]\).
Consequently \(F-\mu_\xi\) satisfies
\((D_\omega-\iota_{\xi_Q})(F-\mu_\xi)=0\),
using the Bianchi identity and the fact that contraction of a
function vanishes. For an invariant polynomial, evaluating on
this equivariant curvature gives a closed expression in the
Cartan complex with differential \(d-\iota_{\xi_X}\).
The calculation comes from the derivation rule for the
invariant polynomial, which kills the adjoint-bracket terms.
The \(\xi\)-polynomial degree is counted as twice its ordinary
degree. For compact \(K\) the Cartan comparison computes the
real cohomology of the Borel construction, or equivalently
the simplicial de Rham model of the action groupoid.
This use of an equivariant complex differs from the complex
of invariant horizontal forms on \(X\).

An example is \(X=*\), \(K=U(1)\), and \(Q=U(1)\) with
left action \(kq=k^rq\) for some integer \(r\). The connection
\(q^{-1}dq\) is invariant, but \(\mu_\xi=ir\xi\) for the
generator parameterization \(\exp(i t\xi)\). The ordinary
curvature is zero. The equivariant Chern polynomial in degree
two gives \(r\xi/(2\pi)\). Thus the real first Chern class
over \(BU(1)\) can be nonzero, whereas the positive-degree
ordinary forms on the atlas point vanish.

\section{A flat torsion line on a stack}

Let \(\Gamma=\mathbb Z/m\) with \(m\geq 2\). A complex line on
\(B\Gamma\) is supplied by the unitary character
\(\chi(j)=\exp(-2\pi i j/m)\). A trivial connection on the
line over the point is preserved by each constant unitary
isomorphism. Since \(\Gamma\) is discrete, the horizontal
condition has no infinitesimal group directions to impose,
so this is a flat line with connection on the stack.
Its holonomy on a generator of the stabilizer is nontrivial.

Compute its integral class using
\[
 0\longrightarrow\mathbb Z\longrightarrow\mathbb R
 \xrightarrow{\ t\mapsto e^{2\pi i t}\ } U(1)
 \longrightarrow 0 .
\]
For homogeneous real group cochains in positive degree,
averaging over the first group entry gives a contracting
homotopy: insert a first entry \(h\), sum over \(h\), and
divide by \(m\). The alternating deletion differential and
this homotopy have anticommutator the identity in positive
degree because the term deleting \(h\) is the original
cochain and the other terms cancel in pairs.
Invariance is preserved by reindexing the average.
Thus \(H^q(B\Gamma;\mathbb R)=0\) for \(q>0\).
The connecting homomorphism gives an isomorphism
\[
 \operatorname{Hom}(\Gamma,U(1))
 =H^1(B\Gamma;U(1))\xrightarrow{\ \sim\ }H^2(B\Gamma;\mathbb Z),
\]
where the coefficients on the left are constant coefficients
on the realized groupoid. Lifting the character to the real
cochain \(j\mapsto -j/m\) for \(0\le j<m\) gives the integer
cocycle \((j,k)\mapsto-\lfloor(j+k)/m\rfloor\).
This is the Chern cocycle in the chosen exponential
normalization. The character has order \(m\), so its first
Chern class generates a cyclic group of order \(m\).
The curvature and the coarse space are both unable to show
this class. The stack retains the discrete descent arrows,
and the differential class retains their phases.

The same mechanism occurs on a quotient with a nontrivial
atlas. Let \(\Gamma=\mathbb Z/m\) rotate \(\mathbb C P^1\)
by \(j\cdot[a:b]=[a:e^{2\pi i j/m}b]\).
The corresponding unitary diagonal matrices act on the
ambient \(\mathbb C^2\) and carry each tautological line to
the tautological line at its image. The projection connection
of Chapter~\ref{ch:synthesis} is equivariant because these
matrices preserve both the trivial differential and
orthogonal projection. Since the group is finite, this is
a descent connection on \([\mathbb C P^1/\Gamma]\).
At \([1:0]\) the action on the fiber is trivial, and at
\([0:1]\) the action is the inverse of \(\chi\) above.
Each fixed point defines a map
\(B\Gamma\to[\mathbb C P^1/\Gamma]\).
The restrictions of the equivariant first Chern class
along those two maps are respectively zero and a generator
of \(H^2(B\Gamma;\mathbb Z)\). Its restriction to the atlas
is the first Chern class of the tautological line on the
sphere, which has Chern number \(-1\).
These are three different tests of the same characteristic
class. If the tautological equivariant bundle were pulled
back from the orbit space, each stabilizer would act
trivially on its fiber, because the pullback from a space
acts on fibers by the identity along arrows mapping to
identities. The action at the second fixed point disproves
that descent to the coarse orbit space.
\chapter{Derived intersections and Atiyah classes}
\label{ch:derived}

The first infinitesimal neighborhood of a diagonal gives a local
description of connections. A derived intersection of that diagonal
with itself retains the relations among its equations. This chapter
works in smooth algebraic geometry over a characteristic-zero field
\(k\), using the algebraic differential \(d\). Smooth real connections
and algebraic connections have different existence properties. The
comparison rests on the common Leibniz identity.

\section{Intersecting schemes through a derived tensor product}

For affine schemes mapping to \(\operatorname{Spec} R\), let their
coordinate algebras be \(B\) and \(C\). The ordinary pullback has
coordinate algebra \(B\otimes_R C\). In derived algebraic geometry one
uses \(B\otimes_R^{\mathbf L}C\), the homotopy pushout of the
corresponding commutative algebras. We use cohomological degrees, so
\[
 H^{-j}(B\otimes_R^{\mathbf L}C)
       =\operatorname{Tor}^R_j(B,C),\qquad j\geq 0.
\]
The homotopy pushout can be modeled by commutative differential graded
algebras in nonpositive degrees in characteristic zero
\cite[Section~2]{toen}. The formula is understood locally on schemes
when the rings vary from chart to chart.

Consider two embeddings of the line \(\{y=0\}\) in
\(\mathbb A^2_k\), so \(R=k[x,y]\) and \(B=C=R/(y)\).
The algebra obtained by adjoining one odd generator \(e\) of
degree \(-1\) to \(R\) with \(de=y\) resolves \(B\).
Indeed its underlying complex is \(R\xrightarrow{y}R\),
with zero kernel and cokernel \(B\). As a free extension in
graded commutative algebras this is a Koszul model for the derived
base change. Tensoring with \(C\) gives a differential of zero,
and the resulting algebra is \(k[x]\oplus k[x]e\), with \(e^2=0\).
The ordinary intersection is just the line, while \(H^{-1}\)
records a normal equation that occurs on both sides.

For a different second embedding take the parabola \(y=x^2\).
On \(R/(y-x^2)\cong k[x]\), the same differential is multiplication
by \(x^2\), hence it is injective. The degree-zero algebra is
\(k[x]/(x^2)\), with no higher Tor. Here failure of transversality
appears as a nonreduced ordinary intersection. The two examples
distinguish nilpotents in degree zero from negative-degree homology.

\section{The diagonal and its self-intersection}

Let \(X\) be smooth, separated and of finite type over \(k\). Its
diagonal \(\Delta:X\to X\times_k X\) is a regular closed immersion.
Writing \(\mathcal I\) for its ideal, the conormal
\(\mathcal I/\mathcal I^2\) is \(\Omega^1_{X/k}\).
The identification takes the class of \(p_2^*f-p_1^*f\) to \(df\).
Indeed modulo \(\mathcal I^2\) the difference obeys Leibniz,
and on an étale coordinate neighborhood the coordinate
differences generate the conormal freely. The same statement
also follows from the universal derivation of a commutative algebra
by applying its defining property to the multiplication map.

\begin{proposition}[Tor sheaves along the diagonal]
\label{prop:diag-tor}
The homology sheaves on the derived self-intersection
\(X\times^{h}_{X\times_k X}X\) satisfy
\[
 \mathcal H^{-j}(\mathcal O_{X\times^{h}_{X\times_k X}X})
       \cong \bigwedge\nolimits^j\Omega^1_{X/k}.
\]
These identifications respect the graded homology algebra.
\end{proposition}

\begin{proof}
Take local generators \(f_1,\ldots,f_d\) of the diagonal ideal
forming a regular sequence. The Koszul algebra adjoins
odd generators \(e_i\) with \(de_i=f_i\).
It resolves the diagonal, as shown by induction on the length
of the sequence. Passing from \(d-1\) to \(d\) takes the
cone of multiplication by \(f_d\) on the preceding Koszul
complex, with the shift that puts the new exterior generator
in degree \(-1\). The preceding homology is the quotient by
the first \(d-1\) generators. The regular-sequence hypothesis
makes the multiplication injective on that quotient, so the
cone has only the requisite degree-zero homology. This
completes the induction. Pulling back along the diagonal
kills the \(f_i\), leaving the exterior algebra with zero
differential. Under a change of generators, the induced map
in degree \(-1\) is the change-of-basis map of
\(\mathcal I/\mathcal I^2\). Multiplication extends the
identification to the exterior powers and makes it
independent of the local resolution.
\end{proof}

The loop-space expression here is categorical. Cotensoring by
\(S^1\simeq *\coprod^h_{S^0}*\) turns the homotopy pushout
into the pullback of the two diagonal maps. A degree-zero
scheme of points would miss the derived directions of this
pullback. The global Hochschild--Kostant--Rosenberg comparison
for the smooth diagonal gives the corresponding splitting
as a derived object. For other embeddings a local Koszul model
by itself only determines the homology sheaves. The extension
condition for a global splitting is treated in
\cite[Theorem~1]{arinkin-caldararu}.

\section{First jets and the obstruction to a connection}

For a vector bundle \(E\) on \(X\), let \(Z\) be the first
infinitesimal neighborhood of the diagonal, with
\(\mathcal O_Z=\mathcal O_{X\times X}/\mathcal I^2\).
The two projections of \(Z\) to \(X\) are still distinct.
Define \(J^1E=p_{1*}p_2^*E\), regarded as an
\(\mathcal O_X\)-module by the first projection.
There is a sequence
\[
 0\longrightarrow\Omega^1_{X/k}\otimes E
   \longrightarrow J^1E\longrightarrow E\longrightarrow 0.
\]
After trivializing \(E\), this is the exact sequence coming
from reduction \(\mathcal O_Z\to\mathcal O_\Delta\);
the conormal module multiplies by coefficients restricted
to the diagonal. Thus the sequence is exact independently
of the chosen frame. Its extension class
\[
 \operatorname{At}(E)\in
   \operatorname{Ext}^1_{\mathcal O_X}(E,\Omega^1_{X/k}\otimes E)
\]
uses the first jet convention. Atiyah's obstruction
construction treats the holomorphic counterpart
\cite[Sections~2--4]{atiyah}.

The jet operator \(j\) pulls a section along \(p_2\).
On \(\mathcal O_Z\), multiplication by \(p_2^*f\)
differs from multiplication by \(p_1^*f\) by \(df\) in the
conormal ideal. Thus \(j(fs)=fj(s)+df\otimes s\).
If \(\sigma:E\to J^1E\) is an \(\mathcal O_X\)-linear
splitting, \(j-\sigma\) takes values in the kernel and
satisfies the connection Leibniz rule. Conversely a
connection \(\nabla\) gives the splitting \(j-\nabla\).
The extension class vanishes exactly when a connection of
this kind exists. Its vanishing alone makes no integrability
assertion. Flatness imposes the separate curvature equation.

Here is a coordinate account of the class. Take local frames
with \(e_j=e_i g_{ij}\) and in frame \(i\) take the connection
\(\nabla^{(i)}\) for which \(e_i\) is parallel. Write
\(\sigma_i=j-\nabla^{(i)}\). On a double overlap
\[
 \sigma_j-\sigma_i=\nabla^{(i)}-\nabla^{(j)}.
\]
In the \(j\)-frame this difference is \(g_{ij}^{-1}dg_{ij}\).
The overlap differences form a Čech cocycle of
\(\Omega^1\otimes\operatorname{End}E\), and changing the
splittings adds a coboundary. This is the same inhomogeneous
term that enters the connection descent law.

\begin{proposition}[Tensor products and determinant]
With the jet convention above, the class of \(E\otimes G\)
is the sum of \(\operatorname{At}(E)\otimes 1_G\) and
\(1_E\otimes\operatorname{At}(G)\), with the tensor factors
reordered so that both take values in
\(\Omega^1\otimes E\otimes G\).
The class of \(\det E\) is the trace of \(\operatorname{At}(E)\).
\end{proposition}

\begin{proof}
On a common trivializing cover, the tensor connection
built from the chosen local connections satisfies the
Leibniz rule on \(s\otimes t\) by differentiating each factor
in turn. Taking the difference between the two local
tensor connections on an overlap leaves the sum of the
differences in each input, since all terms are
\(\mathcal O_X\)-linear after subtraction. This proves
the tensor formula on the extension cocycles. For the
determinant use the local connection induced on the top
exterior power. In a wedge of a frame, replacing its \(i\)-th
vector by a multiple of any other frame element gives zero,
so only the diagonal entries of the difference matrix
survive. Their sum gives the trace formula.
\end{proof}

On \(\mathbb P^1_k\), the tautological bundle has frames
\((1,z)\) and \((w,1)\) with \(w=1/z\), giving transition
\(g=z^{-1}\). Its jet cocycle is \(-dz/z\).
In the two standard charts,
\[
 H^1(\mathbb P^1,\Omega^1)
 =\frac{k[z,z^{-1}]\,dz}
 {k[z]\,dz+z^{-2}k[z^{-1}]\,dz}.
\]
This quotient follows by rewriting regular forms in the
\(w\)-chart using \(dw=-z^{-2}dz\).
The class of \(dz/z\) survives, so the tautological bundle
has no algebraic connection. Over \(\mathbb C\) the same
underlying smooth line has a Hermitian connection, computed
explicitly in Chapter~\ref{ch:synthesis}. The change of
coefficient sheaf accounts for the different answers.
\chapter{From a Hopf connection to surface holonomy}
\label{ch:synthesis}

The Hopf fibration gives a sequence of explicit geometric models.
Its line bundle over the two-sphere has a curvature form with
nonzero period. On the total space this form admits a global
primitive. The primitive and the curvature together give an
integral \(3\)-form, leading to surface holonomy on the total
space and ordinary line holonomy on its loop space.

\section{The tautological connection and its periods}

Let \(L\to\mathbb C P^1\) be the tautological Hermitian line.
It is a subbundle of \(\mathbb C P^1\times\mathbb C^2\).
Projecting the trivial differential orthogonally to \(L\)
gives a metric connection. For a section \(s\), define
\(\nabla s=\pi_L ds\). The Leibniz identity follows from
\(\pi_Ls=s\), and metric compatibility follows because
the perpendicular component of \(ds\) pairs to zero
with any section of \(L\).

In the affine \(z\)-chart use the unit frame
\[
 s_0(z)=\frac{(1,z)}{\sqrt{1+|z|^2}}.
\]
The connection matrix is \(A_0=s_0^*ds_0\), namely
\[
 A_0=\frac{\bar z\,dz-z\,d\bar z}{2(1+|z|^2)}
     =i\,\frac{r^2}{1+r^2}\,d\theta ,
 \qquad z=re^{i\theta}.
\]
The polar expression holds away from the origin. The
Cartesian formula is smooth there. On the chart \(w=1/z\)
take \(s_1=(w,1)/\sqrt{1+|w|^2}\).
On the overlap \(s_1=s_0 e^{-i\theta}\), so the two matrices
obey \(A_1=A_0-i\,d\theta\). This is the transformation
law from Chapter~\ref{ch:principal}.

Set \(\omega=iF/(2\pi)\). On the \(z\)-chart,
\[
 \omega=-\frac{1}{\pi}\frac{r}{(1+r^2)^2}\,dr\wedge d\theta,
 \qquad
 \int_{\mathbb C P^1}\omega=-1.
\]
For the integral one first integrates over \(r<R\) and sends
\(R\) to infinity. The missing point has measure zero for the
smooth form in the other chart. With the complex orientation
this gives the first Chern number of \(L\)
\cite{tu,bott-tu}. Its sign also follows by using Stokes
on the two chart disks, since \(A_0-A_1=i\,d\theta\)
on their common boundary oriented from the \(z\)-disk.

On a positively oriented \(z\)-circle of radius \(R\), the
parallel transport equation for the coefficient in the frame
is \(\dot v=-A_0(\dot\gamma)v\), so
\[
 \operatorname{Hol}_L(\gamma)=
 \exp\!\left(-2\pi i\frac{R^2}{1+R^2}\right).
\]
The integral of \(\omega\) over the disk it bounds is
\(-R^2/(1+R^2)\). Choosing the complementary filling with
the same boundary changes that integral by an integer.
The exponent therefore comes from a differential character
with curvature \(\omega\), not just from a choice of
potential on the chart.

\section{An integral 3-form on the total space}

The unit circle bundle \(p:S^3\to\mathbb C P^1\) of \(L\)
has global real \(1\)-form
\(\alpha=q^*dq/(2\pi i)\) for a point \(q\in S^3\)
viewed as a unit complex column vector.
Its value on a positively oriented fiber integrates to one.
In the frame \(q=s_0(z)e^{i\phi}\) the form is
\(d\phi/(2\pi)+A_0/(2\pi i)\). Consequently
\[
 d\alpha=-p^*\omega,\qquad
 \eta=\alpha\wedge p^*(-\omega),\qquad
 \int_{S^3}\eta=1.
\]
The orientation here uses the positive circle and complex
base orientations. The integral follows by fiber integration:
on each trivialization only the \(d\phi\) term contributes,
and the local base integrals glue by a partition of unity.
Also \(d\eta=p^*((-\omega)\wedge(-\omega))=0\).
Thus \(\eta\) represents the integral generator of \(H^3(S^3;\mathbb Z)\)
in real cohomology.

One can construct the associated degree-three differential
character directly. A smooth singular \(2\)-cycle \(z\) in
\(S^3\) bounds a smooth singular \(3\)-chain \(D\).
Put \(h(z)=\int_D\eta\bmod\mathbb Z\).
The difference between two choices of \(D\) is a \(3\)-cycle,
and integrality of \(\eta\) gives independence. Linearity
follows by adding fillings. Taking \(z=\partial D\) gives
exactly the character boundary identity with curvature \(\eta\).
The integral class of \(h\) maps to the de Rham class
of \(\eta\); here the degree-three cohomology is free,
so this class is the generator. This character can also
be represented by a bundle gerbe with connection.

To see the overlap geometry, cover \(S^3\) by the complements
of the two poles, with local \(2\)-forms \(B_N,B_S\)
satisfying \(dB_N=dB_S=\eta\). Orient the equator as the
boundary of the northern hemisphere. Stokes gives
\(\int_{S^2}(B_N-B_S)=1\).
On the overlap this difference is a closed \(2\)-form of
integral periods because the equator generates its \(H_2\).
It can be realized as a line curvature there by the
differential cohomology curvature sequence. For the convention
in which the overlap curvature is \(B_S-B_N\), take the
dual line. Use the reverse overlap line as its dual and the
identity lines on repeated charts. Evaluation of the
line against its dual supplies the product on alternating
overlaps. Associativity reduces to the tensor identities
for a line and its dual, which can be checked in a unit frame.
This is the connective descent presentation whose curvature is
\(\eta\). The difference of local \(2\)-forms carries the
same integer that the original Hopf curvature detected.

\section{Evaluating families of loops}

Let \(LS^3\) consist of smooth loops, with a smooth family
defined by its adjoint smooth map to \(S^3\).
For a smooth singular \(1\)-cycle \(c\) of families of loops,
write \(\operatorname{ev}:LS^3\times S^1\to S^3\) for evaluation.
Then the transgressed holonomy of \(c\) is
\[
 (\tau h)(c)=h\big(\operatorname{ev}_*(c\times[S^1])\big).
\]
A smooth singular fundamental cycle of \(S^1\) is used in
this expression. Replacing it by a homologous cycle changes
the argument by a boundary evaluated through the
one-dimensional circle factor. The lifted \(3\)-chain has
rank at most two under evaluation, so its curvature
integral vanishes. This proves independence of that choice.

For a smooth \(2\)-chain \(C\) of loop families, the boundary
of \(C\times[S^1]\) is \(\partial C\times[S^1]\).
The character property gives
\[
 (\tau h)(\partial C)=
 \int_C\tau\eta\bmod\mathbb Z,\qquad
 (\tau\eta)_\gamma(V,W)
    =\int_0^1\eta_{\gamma(t)}(\dot\gamma(t),V(t),W(t))\,dt .
\]
In the orientation of \(C\times S^1\), moving the loop
tangent from the final slot to the first makes two
interchanges, explaining the sign. The gerbe-to-loop-space
transgression theorem realizes this degree-two character as
a line bundle with connection. Waldorf's convention uses
inverse surface holonomy \cite[Proposition~4.3.2]{waldorf-transgression}.
The dual of his transgressed bundle realizes the positive
evaluation convention above. Its fusion structure gives
compatible products along triples of paths with common
endpoints. The conditions on this structure for the inverse
regression theorem are given in \cite[Theorem~A]{waldorf-transgression}.

More explicitly, for paths \(\gamma_i\) from \(x\) to \(y\)
that are constant near their endpoints, let \(\ell_{ij}\)
traverse \(\gamma_i\) and then \(\gamma_j\) backwards.
The line fibers \(T_{\ell_{ij}}\) of the transgression admit
fusion maps
\[
 T_{\ell_{12}}\otimes T_{\ell_{23}}\longrightarrow T_{\ell_{13}} .
\]
For four paths the two composites of fusion maps agree.
The maps express the cancellation of the common path
segment in the surface description. A bare line bundle
over the loop space does not supply the fusion maps or their
compatibility with transport. The cited theorem uses
fusion bundles with connections satisfying the stated
superficiality and compatibility conditions, and takes
the homotopy category on the gerbe side.

The transgressed curvature has a nonzero integral period.
For this, choose an embedded oriented \(3\)-ball in
\(S^2\times S^1\) avoiding both \(S^2\times\{t_0\}\) and
\(\{s_0\}\times S^1\). Collapse its complement to a
point in \(S^3\), using a smooth representative
\(f:S^2\times S^1\to S^3\) constant near the boundary
of the ball. Such a representative is obtained in
polar coordinates by mapping the radius to a polar
angle from zero to \(\pi\), with a smooth constant
collar at \(\pi\) and a linear radial behavior near
zero. Choosing the orientation of the chart gives
degree \(+1\). The adjoint family
\(S^2\to LS^3\) takes values in based loops since
\(f(s,t_0)\) is constant. Its curvature period is
\[
 \int_{S^2}(\operatorname{adj}f)^*(\tau\eta)
   =\int_{S^2\times S^1}f^*\eta=1.
\]
Thus this higher holonomy produces a line bundle with
nontrivial first Chern class on loop space.

The examples identify the roles of the data at each step.
A local matrix measures parallel transport, an integral class
controls gluing, and a differential character compares fillings
of cycles. Changes of presentation preserve these comparisons.
Increasing the dimension of the cycle changes the descent
data that carry the holonomy, and transgression converts one
such comparison into another by evaluation on a family of loops.

\appendix
\bookpart*{Appendices}
\chapter{Notation and comparison conventions}
\label{app:notation}

\section{Bundles, frames, and curvature}

For a principal bundle \(P\to X\) the group acts on the right. Local sections are written \(s_i\), with \(s_j=s_i g_{ij}\) on overlaps. Matrix group connection forms transform as
\[
 A_j=g_{ij}^{-1}A_i g_{ij}+g_{ij}^{-1}dg_{ij},
 \qquad F_j=g_{ij}^{-1}F_i g_{ij}.
\]
A column representing a section of the associated vector bundle changes by \(v_j=g_{ij}^{-1}v_i\) for the defining representation. The local covariant derivative on the column is \(dv_i+A_i v_i\). When the Lie algebra-valued form is expressed in matrices, \(F=dA+A\wedge A\). In bracket notation the last term is \(\tfrac12[A,A]\). On an ordinary tangent bundle we take
\[
 R(X,Y)Z=\nabla_X\nabla_Y Z-\nabla_Y\nabla_X Z-\nabla_{[X,Y]}Z.
\]
Sectional curvature uses \(\langle R(X,Y)Y,X\rangle\) divided by
\(\langle X,X\rangle\langle Y,Y\rangle-\langle X,Y\rangle^2\).

For a unitary line \(A=-2\pi i\,a\), the real first Chern form is
\(iF/(2\pi)=da\), and in a local frame parallel transport along a path
is multiplication by \(\exp(2\pi i\int a)\). The curvature on a
general principal bundle is an adjoint-valued form on the base.
Invariant polynomials give scalar differential forms. For a unitary
vector bundle the total Chern form is \(\det(I+iF/(2\pi))\).
Changing the orientation of a real even-rank bundle changes the sign
of its Euler class.

\section{Differentials, local cochains, and transport}

The Čech differential \(\delta\) is the alternating sum of the restrictions obtained by omitting an index. In a Čech complex with \(p\) the Čech degree and \(q\) the form degree, the total differential is \(D=\delta+(-1)^p d\) on bidegree \((p,q)\). In degree two, if \(g_{ij}=\exp(-2\pi i\lambda_{ij})\), then \(a_j-a_i=d\lambda_{ij}\), and the integral first Chern cocycle for the positive exponential sequence is \(-\delta\lambda\). The comparison \(D(a+\lambda)=da+\delta\lambda\) verifies the sign in real cohomology. The gerbe chapter specifies its multiplication and cocycle conventions separately.

Manifolds used as smooth tests are Hausdorff, second countable, and finite dimensional. Smooth singular simplices, paths with sitting instants, and fiber integration are considered with the boundary, orientation, and support conventions in the relevant chapters. In the path transport formulas, composing two paths means traversal of the first path followed by the second, with the corresponding linear maps composed in the usual function order.

\section{Moduli and differential coefficients}

\(\mathbf BG\) is the smooth stack of right \(G\)-torsors, and
\(\mathbf BG_\nabla\) includes their connections. A smooth test \(X\)
is identified with its representable sheaf when this causes no
ambiguity. A superscript \({\mathrm h}\) on a pullback or a limit
emphasizes the use of paths and their coherences. The pullback in
Chapter~\ref{ch:characteristic-maps} uses locally constant
coefficient sheaves \(K(\mathbb Z,n)\) and \(K(\mathbb R,n)\).
In differential cohomology \(I\) denotes the underlying integral class
and \(R\) denotes curvature. Where \(R\) acts on a vector via a
tangent connection it denotes the Riemannian curvature tensor.
The section and page links in Appendix~\ref{app:index} point to the
main treatments.

\paragraph{Authorship.}
The manuscript has an anonymous byline. A language model assisted
with drafting, mathematical review, typesetting, and revision under
the submitter's direction.
\chapter{Index of concepts and constructions}
\label{app:index}

Select a page number to open the corresponding treatment.

\begingroup\small
\renewcommand{\arraystretch}{1.18}
\setlength{\LTpre}{8pt}
\setlength{\LTpost}{0pt}
\begin{longtable}{@{}>{\raggedright\arraybackslash}p{.30\textwidth}@{\hspace{.025\textwidth}}>{\raggedright\arraybackslash}p{.49\textwidth}@{\hspace{.025\textwidth}}>{\raggedleft\arraybackslash}p{.16\textwidth}@{}}
\textbf{Concept}&\textbf{Construction}&\textbf{Page}\\
\hline
\endfirsthead
\textbf{Concept}&\textbf{Construction}&\textbf{Page}\\
\hline
\endhead
\textbf{Adjoint bundle}&Lie algebra bundle carrying curvature and covariant differentiation.&\bookindexpage{ch:principal}, \bookindexpage{ch:bg-connection}\\
\textbf{Associated bundle}&Turns a torsor and a representation into a vector bundle.&\bookindexpage{ch:principal}, \bookindexpage{ch:classifying}\\
\textbf{Atiyah class}&The extension class of the first jet sequence.&\bookindexpage{ch:derived}\\
\textbf{Basic form}&Horizontal invariant form that descends to the base.&\bookindexpage{ch:chern-weil}, \bookindexpage{ch:weil}\\
\textbf{Bianchi identity}&Covariant differential of curvature and its higher analogue.&\bookindexpage{ch:vector-conn}, \bookindexpage{ch:2connections}\\
\textbf{Cartan model}&Equivariant forms and the contraction term in their differential.&\bookindexpage{ch:weil}\\
\textbf{Characteristic form}&Closed form obtained by evaluating an invariant polynomial on curvature.&\bookindexpage{ch:chern-weil}, \bookindexpage{ch:characteristic-maps}\\
\textbf{\v{C}ech descent}&Local objects, transition maps, and higher comparisons.&\bookindexpage{ch:bundles}, \bookindexpage{ch:smooth-site}\\
\textbf{Chern classes}&Classes represented over the reals by the normalized determinant form.&\bookindexpage{ch:characteristic}, \bookindexpage{ch:diff-products}\\
\textbf{Chern--Simons form}&Transgresses the difference of characteristic forms.&\bookindexpage{ch:chern-simons}\\
\textbf{Christoffel symbols}&Local coefficients of the affine connection.&\bookindexpage{ch:levi}\\
\textbf{Classifying stack}&\(\mathbf BG\) for torsors, \(\mathbf BG_\nabla\) for torsors with connection.&\bookindexpage{ch:classifying}, \bookindexpage{ch:bg-connection}\\
\textbf{Crossed module}&Two groups, an action, and Peiffer identities presenting a strict \(2\)-group.&\bookindexpage{ch:crossed-modules}\\
\textbf{Curvature}&Obstruction detected by the square of the covariant differential and small loops.&\bookindexpage{ch:vector-conn}, \bookindexpage{ch:transport}\\
\textbf{Curving}&Local two-form involved in a gerbe connection.&\bookindexpage{ch:gerbes}\\
\textbf{Derived intersection}&Homotopy pullback computed locally by a derived tensor product.&\bookindexpage{ch:derived}\\
\textbf{Differential character}&Evaluates smooth cycles with a Stokes law for its curvature.&\bookindexpage{ch:diff-cohomology}\\
\textbf{Differential refinement}&An integral class, a form, and a comparison of their real images.&\bookindexpage{ch:characteristic-maps}, \bookindexpage{ch:diff-products}\\
\textbf{Dixmier--Douady class}&Underlying degree-three characteristic class of a circle gerbe.&\bookindexpage{ch:gerbes}\\
\textbf{Euler class}&Top class of an oriented real bundle, with a Pfaffian form in even rank.&\bookindexpage{ch:characteristic}\\
\textbf{Fake curvature}&Difference between ordinary curvature and the boundary of the two-form in a strict two-connection.&\bookindexpage{ch:2connections}\\
\textbf{Fiber integration}&Pushforward reducing cohomological degree by fiber dimension.&\bookindexpage{ch:fiber-integration}\\
\textbf{Gauss and Codazzi equations}&Compare ambient and intrinsic derivatives of a submanifold.&\bookindexpage{ch:intrinsic}\\
\textbf{Gauss--Bonnet}&Identifies the total Gaussian curvature with \(2\pi\) times the Euler characteristic.&\bookindexpage{ch:geodesics}\\
\textbf{Gauge transformation}&Changes a frame or acts as an isomorphism in the connection stack.&\bookindexpage{ch:vector-conn}, \bookindexpage{ch:bg-connection}\\
\textbf{Gerbe holonomy}&Surface evaluation encoded by local two-forms and overlap data.&\bookindexpage{ch:gerbes}, \bookindexpage{ch:surface-transport}\\
\textbf{Holonomy}&Transport around a closed path or higher cycle.&\bookindexpage{ch:transport}, \bookindexpage{ch:fiber-integration}\\
\textbf{Homotopy pullback}&A pair of maps with their specified compatibility path.&\bookindexpage{ch:smooth-site}, \bookindexpage{ch:characteristic-maps}\\
\textbf{Jacobi field}&First variation of a geodesic through geodesics.&\bookindexpage{ch:geodesics}\\
\textbf{Levi--Civita connection}&The metric-compatible tangent connection with zero torsion.&\bookindexpage{ch:levi}\\
\textbf{Pontryagin classes}&Even Chern classes of the complexification with the conventional signs.&\bookindexpage{ch:characteristic}, \bookindexpage{ch:string}\\
\textbf{Principal connection}&An equivariant horizontal splitting, equivalently a normalized Lie algebra-valued one-form.&\bookindexpage{ch:principal}\\
\textbf{Quotient stack}&Remembers inertia and descent along an action groupoid.&\bookindexpage{ch:quotients}\\
\textbf{Reduction}&An \(H\)-subbundle of a \(G\)-bundle, or a section of its \(G/H\)-bundle.&\bookindexpage{ch:classifying}\\
\textbf{Sectional curvature}&Scalar curvature associated to an unoriented tangent two-plane.&\bookindexpage{ch:intrinsic}\\
\textbf{Smooth site}&Manifolds as tests, with gluing over open covers.&\bookindexpage{ch:smooth-site}\\
\textbf{String structure}&A lift across the homotopy fiber of \(\tfrac12p_1\).&\bookindexpage{ch:string}\\
\textbf{Surface transport}&Two-dimensional transport on bigons with their thin relations.&\bookindexpage{ch:surface-transport}\\
\textbf{Thin homotopy}&A rank-constrained homotopy respected by transport.&\bookindexpage{ch:transport}, \bookindexpage{ch:surface-transport}\\
\textbf{Transgression}&Integrates along a parameter family or along the evaluation circle.&\bookindexpage{ch:chern-simons}, \bookindexpage{ch:fiber-integration}\\
\textbf{Weil algebra}&Universal algebra for connection and curvature generators.&\bookindexpage{ch:weil}, \bookindexpage{ch:higher-chern-weil}\\
\end{longtable}
\endgroup
\backmatter
\begin{thebibliography}{99}
\addtolength{\itemsep}{-1pt}
\addcontentsline{toc}{chapter}{Bibliography}

\bibitem{tu}
Loring W. Tu.
\emph{Differential Geometry. Connections, Curvature, and Characteristic Classes}.
Graduate Texts in Mathematics 275, Springer, 2017.
\url{https://link.springer.com/book/10.1007/978-3-319-55084-8}.

\bibitem{lee-riemannian}
John M. Lee. \emph{Introduction to Riemannian Manifolds}.
2nd ed., Graduate Texts in Mathematics 176, Springer, 2018.
\url{https://link.springer.com/book/10.1007/978-3-319-91755-9}.

\bibitem{kobayashi-nomizu}
Shoshichi Kobayashi and Katsumi Nomizu. \emph{Foundations of Differential Geometry}, Vol. I.
Wiley, 1963.
\url{https://www.wiley.com/en-us/foundations-of-differential-geometry-volume-1-p-9780471157335}.

\bibitem{schreiber-waldorf-transport}
Urs Schreiber and Konrad Waldorf. \emph{Parallel Transport and Functors}.
Journal of Homotopy and Related Structures 4(1), 187--244, 2009.
\url{https://arxiv.org/abs/0705.0452}.

\bibitem{bott-tu}
Raoul Bott and Loring W. Tu. \emph{Differential Forms in Algebraic Topology}.
Graduate Texts in Mathematics 82, Springer, 1982.
\url{https://link.springer.com/book/10.1007/978-1-4757-3951-0}.

\bibitem{milnor-stasheff}
John W. Milnor and James D. Stasheff. \emph{Characteristic Classes}.
Annals of Mathematics Studies 76, Princeton University Press, 1974.
\url{https://press.princeton.edu/books/paperback/9780691081229/characteristic-classes}.

\bibitem{chern-simons}
Shiing-Shen Chern and James Simons. \emph{Characteristic Forms and Geometric Invariants}.
Annals of Mathematics 99(1), 48--69, 1974.
\url{https://annals.math.princeton.edu/1974/99-1/p03}.

\bibitem{g-s-continuous}
Victor W. Guillemin and Shlomo Sternberg. \emph{Supersymmetry and Equivariant de Rham Theory}.
Springer, 1999.
\url{https://link.springer.com/book/10.1007/978-3-662-03992-2}.

\bibitem{lurie-htt}
Jacob Lurie. \emph{Higher Topos Theory}.
Annals of Mathematics Studies 170, Princeton University Press, 2009.
\url{https://www.math.ias.edu/~lurie/papers/HTT.pdf}.

\bibitem{schreiber-differential}
Urs Schreiber. \emph{Differential cohomology in a cohesive infinity-topos}.
2013.
\url{https://arxiv.org/abs/1310.7930}.

\bibitem{fiorenza-schreiber-stasheff}
Domenico Fiorenza, Urs Schreiber, and Jim Stasheff.
\emph{\v{C}ech cocycles for differential characteristic classes: An infinity-Lie theoretic construction}.
Advances in Theoretical and Mathematical Physics 16(1), 149--250, 2012.
\url{https://arxiv.org/abs/1011.4735}.

\bibitem{cheeger-simons}
Jeff Cheeger and James Simons. \emph{Differential characters and geometric invariants}.
In Geometry and Topology, Lecture Notes in Mathematics 1167, 50--80, Springer, 1985.
\url{https://math.mit.edu/juvitop/pastseminars/notes_2019_Fall/cheeger-simons.pdf}.

\bibitem{hopkins-singer}
Michael J. Hopkins and Isadore M. Singer.
\emph{Quadratic functions in geometry, topology, and M-theory}.
Journal of Differential Geometry 70(3), 329--452, 2005.
\url{https://arxiv.org/abs/math/0211216}.

\bibitem{bunke-nikolaus-voelkl}
Ulrich Bunke, Thomas Nikolaus, and Michael V\"olkl.
\emph{Differential cohomology theories as sheaves of spectra}.
Journal of Homotopy and Related Structures 11, 1--66, 2016.
\url{https://arxiv.org/abs/1311.3188}.

\bibitem{baer-becker}
Christian B\"ar and Christian Becker.
\emph{Differential Characters and Geometric Chains}.
\url{https://arxiv.org/abs/1303.6457}.

\bibitem{murray}
Michael K. Murray. \emph{Bundle gerbes}.
Journal of the London Mathematical Society 54(2), 403--416, 1996.
\url{https://arxiv.org/abs/dg-ga/9407015}.

\bibitem{baez-lauda}
John C. Baez and Aaron D. Lauda.
\emph{Higher-dimensional algebra V: 2-groups}.
Theory and Applications of Categories 12, 423--491, 2004.
\url{https://arxiv.org/abs/math/0307200}.

\bibitem{baez-schreiber}
John C. Baez and Urs Schreiber.
\emph{Higher gauge theory: 2-connections on 2-bundles}.
2004.
\url{https://arxiv.org/abs/hep-th/0412325}.

\bibitem{schreiber-waldorf-surface}
Urs Schreiber and Konrad Waldorf.
\emph{Smooth functors vs. differential forms}.
Homology, Homotopy and Applications 13(1), 143--203, 2011.
\url{https://arxiv.org/abs/0802.0663}.

\bibitem{schreiber-waldorf-gerbe}
Urs Schreiber and Konrad Waldorf.
\emph{Connections on non-abelian gerbes and their holonomy}.
Theory and Applications of Categories 28, 476--540, 2013.
\url{https://arxiv.org/abs/0808.1923}.

\bibitem{stolz-teichner}
Stephan Stolz and Peter Teichner. \emph{What is an elliptic object?}
In Topology, Geometry and Quantum Field Theory, London Mathematical Society Lecture Note Series 308, 247--343, Cambridge University Press, 2004.
\url{https://doi.org/10.1017/CBO9780511526398.013}.

\bibitem{waldorf-string}
Konrad Waldorf. \emph{String connections and Chern--Simons theory}.
Transactions of the American Mathematical Society 365, 4393--4432, 2013.
\url{https://arxiv.org/abs/0906.0117}.

\bibitem{lgtx}
Camille Laurent-Gengoux, Jean-Louis Tu, and Ping Xu.
\emph{Chern--Weil map for principal bundles over groupoids}.
Mathematische Zeitschrift 255, 451--491, 2007.
\url{https://arxiv.org/abs/math/0401420}.

\bibitem{toen}
Bertrand To\"en. \emph{Derived algebraic geometry}.
EMS Surveys in Mathematical Sciences 1, 153--240, 2014.
\url{https://arxiv.org/abs/1401.1044}.

\bibitem{atiyah}
Michael F. Atiyah. \emph{Complex analytic connections in fibre bundles}.
Transactions of the American Mathematical Society 85, 181--207, 1957.
\url{https://doi.org/10.1090/S0002-9947-1957-0086359-5}.

\bibitem{arinkin-caldararu}
Dima Arinkin and Andrei C\u{a}ld\u{a}raru.
\emph{When is the self-intersection of a subvariety a fibration?}
Advances in Mathematics 231, 815--842, 2012.
\url{https://arxiv.org/abs/1007.1671}.

\bibitem{waldorf-transgression}
Konrad Waldorf.
\emph{Transgression to loop spaces and its inverse, II: Gerbes and fusion bundles with connection}.
Asian Journal of Mathematics 20, 59--116, 2016.
\url{https://arxiv.org/abs/1004.0031}.

\bibitem{murray-stevenson}
Michael K. Murray and Daniel Stevenson.
\emph{Bundle gerbes: Stable isomorphism and local theory}.
Journal of the London Mathematical Society 62, 925--937, 2000.
\url{https://arxiv.org/abs/math/9908135}.

\end{thebibliography}
\end{document}

End of manuscript · Back to top ↑