Independent scholarship · Open accessThe monthly digest

Full manuscript · browser edition

Differential Topology in Higher Categories

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

Contents · 126 sections

Differential Topology
in Higher Categories

From regular values and cobordism
to derived smooth families

Anonymous

November 2026

Smooth manifolds and their probes

A smooth manifold may be specified by an atlas, reconstructed by gluing its charts, or characterized by the smooth families of points that it admits. The three descriptions support different operations. Charts give access to calculus. Gluing exposes compatibility conditions. Smooth families provide a test for morphisms that continues to make sense even when a construction has no manifold representative. We begin with the ordinary smooth objects of Hirsch’s first chapter [15, Ch. 1], and pass from these to their functors of points.

Coordinates and gluing

A smooth nn-manifold is a Hausdorff, second-countable topological space with coordinate neighborhoods in Rn\mathbb R^n and smooth transition diffeomorphisms. The local dimension is constant within a connected component. For a manifold with boundary, charts instead take values in relatively open subsets of the half-space [0,∞)×Rn−1[0,\infty)\times\mathbb R^{n-1}, with the usual zero-dimensional convention. Smoothness in these coordinates means local extension to smooth maps on open subsets of the ambient Euclidean space. The boundary is the locus whose charts have first coordinate zero. It is intrinsic, as each transition and its inverse preserve the cone of inward velocities. At a boundary point this cone is a half-space, whereas at an interior point every direction occurs. The derivatives of smooth extensions agree on overlaps by continuity from the interior. Coordinate transitions therefore make the same smoothness test in every chart [18].

A map f:M→Nf:M\to N is smooth if its coordinate expressions are smooth on the open sets where they are defined. Composition is smooth by the coordinate chain rule. The resulting category has a terminal point, and maps out of the point recover the underlying sets. A map of manifolds with boundary need not preserve the boundary. For example, a map from a closed interval that is constant at an interior point is smooth. When an argument requires a map of pairs (M,∂M)→(N,∂N)(M,\partial M)\to(N,\partial N), the boundary condition is part of the structure.

One can construct manifolds from coordinates given in advance. Let UiU_i be chart domains, let Uij⊂UiU_{ij}\subset U_i be open subsets, and let ϕji:Uij→Uji\phi_{ji}:U_{ij}\to U_{ji} be smooth diffeomorphisms. Assume the overlap domains and these maps obey the identity, inverse, and triple-overlap conditions. The quotient of the disjoint union ∐iUi\coprod_i U_i by the resulting identifications has each UiU_i as an open chart. To see this, the saturate of an open subset of UiU_i is a union of open sets obtained by the transition diffeomorphisms. The overlap identities ensure that the quotient map is injective on each UiU_i. When the quotient is Hausdorff and second countable, these charts give a smooth manifold, with the manifold dimension and boundary behavior specified on each chart.

Hausdorffness is a condition on this gluing. Take two copies of the real line and identify them on R∖{0}\mathbb R\setminus\{0\}. The resulting space has two origins. A neighborhood of either contains the image of some punctured interval, so two neighborhoods of the different origins cannot be disjoint. The coordinate changes are smooth, but the result is outside the category of Hausdorff manifolds. A setting for smooth families can still retain the gluing diagram and its compatibility data.

Smooth families and full faithfulness

Fix a small site U\mathsf{U} of coordinate domains, with open covers as its coverings. For boundary arguments we allow relatively open subsets of quadrants [0,∞)k×Rn−k[0,\infty)^k\times\mathbb R^{n-k} with the same convention of local smooth extension. Morphisms are all smooth maps between the domains. Including quadrants allows finite products of tests. The manifolds in a particular argument need only use the boundary or boundaryless charts required there. Let E\mathcal E denote the ∞\infty-category of sheaves of spaces on this site.

Associate to a manifold MM the discrete-valued functor M‾(U)=C∞(U,M),U∈U.\underline M(U)=C^\infty(U,M),\qquad U\in\mathsf{U}. A map U→MU\to M can be recovered from compatible restrictions to an open cover of UU. Its assembled function is continuous because continuity is local, and it is smooth because every point has a smooth local expression. This proves the sheaf condition on M‾\underline M. If a test UU has coordinates (u1,…,ud)(u_1,\dots,u_d), a section of M‾(U)\underline M(U) is a smooth dd-parameter family, possibly with parameters on the faces.

Proposition 1.1 (Reconstruction from test families). The association M↦M‾M\mapsto\underline M is fully faithful. A natural transformation M‾→N‾\underline M\to\underline N is composition with a unique smooth map M→NM\to N. As maps into N‾\underline N in E\mathcal E, these form a discrete mapping space.

Proof. Let ai:Ui→Ma_i:U_i\to M be open chart maps from coordinate domains whose images cover MM. A natural transformation τ\tau assigns the smooth map τUi(ai):Ui→N\tau_{U_i}(a_i):U_i\to N to each of these families. On an overlap, use an open subdomain of UiU_i and its coordinate change to UjU_j. Naturality identifies the two restrictions. The maps thus glue to a smooth f:M→Nf:M\to N. Given any u:U→Mu:U\to M, restrict its domain to open pieces on which it factors through one of the aia_i. Naturality on each piece shows that τU(u)\tau_U(u) restricts to f∘uf\circ u. The sheaf condition then gives τU(u)=f∘u\tau_U(u)=f\circ u. This also proves uniqueness. The representable smooth objects under consideration are 00-truncated sheaves. By definition all mapping spaces into a 00-truncated object are 00-truncated, giving the last assertion. ◻

The value of M‾\underline M at the point remembers only its underlying set. The extra naturality requirement is substantial. For instance, x↦x3x\mapsto x^3 is bijective on points, but the identity plot u↦uu\mapsto u of its target has no smooth lift on a neighborhood of 00. The inverse is therefore absent even as a local map of smooth sheaves. Test families detect a singularity of a set-theoretic bijection.

Atlases and interval homotopies

An atlas takes a universal form in the category of sheaves. Set A=∐iU‾iA=\coprod_i\underline U_i in E\mathcal E and let A→M‾A\to\underline M be induced by the chart maps. Every test-section of M‾\underline M locally lifts to AA, by pulling back the coordinate cover. In this sense the atlas is an effective epimorphism, meaning that M‾\underline M is reconstructed as the colimit of its Čech nerve. The first stages are image The iterated fiber products of AA in this nerve record all multiple overlaps. The reconstruction can be checked by testing against a sheaf FF of spaces. A map M‾→F\underline M\to F is given on plots into MM by sections of FF. A compatible system on the nerve restricts, for each plot, to compatible sections on its pulled-back open cover, with the higher homotopies supplied on repeated intersections. Sheaf descent glues these into a section over the plot. This gluing is compatible with changes of plot and has a contractible space of choices once the descent data are fixed. Equivalently, mapping out of M‾\underline M gives the limit of the mapping spaces out of the nerve.

Two different levels of smooth comparison are relevant. Proposition 1.1 compares maps in the unlocalized sheaf category. A smooth homotopy, on the other hand, is a map H:M×[0,1]→NH:M\times[0,1]\to N restricting to specified maps at the two endpoints. It is a family parametrized by the smooth interval. For example, the maps x↦0x\mapsto0 and x↦xx\mapsto x from R\mathbb R to R\mathbb R are joined by (x,t)↦tx(x,t)\mapsto tx. They are still distinct points of the discrete mapping space of the corresponding unlocalized sheaves. If one subsequently localizes at the projections X×[0,1]→XX\times[0,1]\to X, the two endpoint inclusions become equivalent, since both are sections of an inverted arrow. Smooth homotopies then induce equal maps in the homotopy category of the localization. This explains which extra operation makes an interval family into a homotopical comparison. The shape of a smooth object and its sheaf of plots serve different purposes.

Finally, objects representing plot families can exist without being manifolds of the chosen kind. For instance, the product [0,∞)‾×[0,∞)‾\underline{[0,\infty)}\times\underline{[0,\infty)} is represented on the chosen site by the quadrant. Its corner has an inward cone with zero-dimensional lineality space, whereas an inward half-plane has one-dimensional lineality and R2\mathbb R^2 has two-dimensional lineality. Here lineality means the largest vector subspace contained in the cone. An isomorphism of smooth coordinate models would induce an invertible derivative carrying one cone onto the other. Thus at the corner this product has no local model by an ordinary two-dimensional manifold with boundary. The sheaf product, with maps given by pairs of smooth maps, supplies the expected universal property. Function objects and singular pullbacks will use this enlargement in a similar way [2].

Tangent functors and smooth maps

Differentiation associates a linear map to a smooth map at each point. Its functoriality comes from the fact that each linear map describes the effect on first-order probes. Pointwise linear algebra detects local immersion, submersion, and transversality conditions. Global properties such as embedding also require control of the topology of the image.

First-order probes

At a point xx of a smooth manifold MM, let CM,x∞C^\infty_{M,x} be the real algebra of germs of smooth functions at xx. A derivation at xx is an R\mathbb R-linear map v:CM,x∞→Rv:C^\infty_{M,x}\to\mathbb R satisfying v(ab)=a(x)v(b)+v(a)b(x).v(ab)=a(x)v(b)+v(a)b(x). Equivalently, it is a unital R\mathbb R-algebra map from CM,x∞C^\infty_{M,x} to R[ϵ]/(ϵ2)\mathbb R[\epsilon]/(\epsilon^2) reducing to evaluation at xx. The map is a↦a(x)+ϵv(a)a\mapsto a(x)+\epsilon v(a). Multiplication of such expressions gives the derivation rule exactly because ϵ2=0\epsilon^2=0. This is a description by infinitesimal algebra. The ordinary site of Chapter [ch:manifolds] uses smooth coordinate domains as probes. To use the dual numbers as an actual probe object one enlarges the site or otherwise specifies a theory of infinitesimals.

In a chart centered at xx, write the coordinate germs as t1,…,tnt_1,\ldots,t_n. The identity a(t)−a(0)=∑iti∫01∂ia(st) dsa(t)-a(0) =\sum_i t_i\int_0^1 \partial_i a(st)\,ds holds in a sufficiently small coordinate neighborhood, using a smooth extension if xx is on the boundary. A derivation thus has v(a)=∑iv(ti)∂ia(0)v(a)=\sum_i v(t_i)\partial_i a(0). Conversely every choice of the numbers v(ti)v(t_i) gives a derivation by this formula. This identifies the tangent space TxMT_xM with Rn\mathbb R^n in coordinates. At an interior point the same coordinates identify tangent vectors with first derivatives of smooth curves. On the boundary the full tangent space has dimension nn. Velocities of curves defined only for nonnegative times form its inward half-space, and two-sided curves through the point have velocities tangent to the boundary if their images lie in the manifold.

For smooth f:M→Nf:M\to N, define Dxf(v)D_xf(v) by applying vv after pulling back the germ from NN to MM. Composition of pullbacks gives Dx(gf)=Df(x)g∘DxfD_x(gf)=D_{f(x)}g\circ D_xf. For the identity the derivative is the identity. Under coordinate changes the derivation coordinates transform by the Jacobian matrices, so gluing U×RnU\times\mathbb R^n gives the smooth tangent bundle TMTM and a smooth map Tf:TM→TNTf:TM\to TN. In bundle notation its defining square is image This construction preserves products, allowing corner charts if necessary. A product chart separates the two lists of variables, giving the canonical identification T(M×N)≅TM×TNT(M\times N)\cong TM\times TN. For example, the tangent map of multiplication m:R2→Rm:\mathbb R^2\to\mathbb R satisfies D(x,y)m(v,w)=yv+xwD_{(x,y)}m(v,w)=yv+xw. This formula follows as well by multiplying x+ϵvx+\epsilon v and y+ϵwy+\epsilon w in the dual numbers.

Rank and normal forms

An immersion is a smooth map with injective derivative at every point. A submersion has surjective derivative. We use the constant-rank theorem for boundaryless coordinate neighborhoods in the following form [15, Ch. 1].

Proposition 1.1 (Local normal form). Suppose ff is a smooth map between boundaryless manifolds, of constant rank rr near xx. There are source and target coordinates in which f(u,v)=(u,0)f(u,v)=(u,0), with u∈Rru\in\mathbb R^r.

Proof. Work near the origin and choose coordinates so the first rr components have an invertible minor in the first rr variables. Use (f1,…,fr,tr+1,…,tm)(f_1,\ldots,f_r,t_{r+1},\ldots,t_m) as a change of source coordinates, by the inverse function theorem. The map is now (u,v)↦(u,h(u,v))(u,v)\mapsto(u,h(u,v)). Since the derivative has rank exactly rr, its columns in the vv variables, whose first rr entries vanish, must themselves vanish. Restricting to a coordinate box shows that h(u,v)=h(u,0)h(u,v)=h(u,0) by integration along the vv-segments. Subtracting h(u,0)h(u,0) from the last target coordinates gives the stated form. Rank zero is the same argument with an empty first minor. ◻

Full rank is open, so the hypothesis applies near every point of an immersion or a submersion in the interior. For an immersion the form is a coordinate inclusion. For a submersion it is a coordinate projection. Boundary versions require hypotheses on the boundary behavior. For instance t↦t2t\mapsto t^2 on the half-line has a critical boundary point even though it is a submersion on the interior. Requiring a boundary-preserving map between half-spaces does not by itself ensure a smooth inverse up to the boundary.

An embedding is an immersion that is a homeomorphism onto its image with the subspace topology. The irrational-slope map R→R2/Z2\mathbb R\to\mathbb R^2/\mathbb Z^2 given by t↦(t,αt) mod Z2t\mapsto (t,\alpha t)\bmod\mathbb Z^2 illustrates the second condition. For irrational α\alpha it is an injective immersion. There are unbounded integers njn_j with αnj\alpha n_j tending to zero modulo Z\mathbb Z. Indeed the pigeonhole principle applied to the fractional parts of 0,α,…,Nα0,\alpha,\ldots,N\alpha gives a nonzero integer of size at most NN whose phase is within 1/N1/N of zero. A fixed finite set of nonzero integers cannot do this as NN grows. The images of njn_j tend to the image of 00, so the inverse on the image is discontinuous.

A proper injective immersion between the manifolds here is an embedding. To check that it is closed, take yy in the closure of the image of a closed set. In the first-countable target there is a sequence yj→yy_j\to y in that image. The sequence together with its limit is compact, so the preimages of its terms lie in a compact set by properness. A convergent subsequence of the preimages has limit in the closed set and maps to yy. Thus the immersion is closed, and injectivity makes the inverse on its image continuous. For an embedded S⊂MS\subset M, the quotient bundle TM∣S/TSTM|_S/TS is its normal bundle. A metric identifies it with the orthogonal complement, whereas the quotient comes directly from the embedding.

At the boundary a useful class is given by neat embeddings. In this chapter this means an embedding f:M→Nf:M\to N with f−1(∂N)=∂Mf^{-1}(\partial N)=\partial M and with derivative transverse to ∂N\partial N at each point of ∂M\partial M. The boundary normal form is a coordinate inclusion preserving the first coordinate. In fact, let tt and ss denote source and target boundary coordinates, with t,s≥0t,s\ge0. In a neighborhood of a boundary point the first component of ff is tb(t,u)t b(t,u) for a smooth bb by integrating its derivative in tt, since the component vanishes at t=0t=0. Transversality and nonnegativity give b>0b>0 after shrinking. The other target coordinates can be linearly recombined along ∂N\partial N so that the first dim⁡M\dim M components, including ss, have an invertible derivative. They become source coordinates by the inverse function theorem on smooth extensions. The sign of tb(t,u)t b(t,u) ensures this coordinate change respects the half-spaces. Subtracting the graph of the remaining components gives the coordinate inclusion by a change that fixes ss. This proves the claimed local form and exhibits the boundary hypothesis needed for it.

Transverse pullbacks

Let f:M→Zf:M\to Z and g:N→Zg:N\to Z be smooth maps of boundaryless manifolds. At a coincidence (m,n)(m,n) with f(m)=g(n)=zf(m)=g(n)=z, their derivatives define the linear map (u,v)⟼Dmf(u)−Dng(v)from TmM⊕TnN to TzZ.(u,v)\longmapsto D_mf(u)-D_ng(v)\quad \text{from }T_mM\oplus T_nN\text{ to }T_zZ . When this is surjective at every coincidence, call the pair transverse. In a target chart the coincidence condition is the zero set of the difference of the coordinate expressions. Its derivative on the zero set is precisely the displayed map, in coordinates. The local submersion normal form gives the solution set a smooth structure and its dimension dim⁡M+dim⁡N−dim⁡Z\dim M+\dim N-\dim Z when nonempty. The change of target chart does not change this embedded submanifold structure, since it is the subspace preimage of the diagonal with the induced charts. A smooth map into M×NM\times N meeting this set factors through it smoothly by its coordinate inclusion. Hence this is the smooth pullback M×ZNM\times_Z N. Its tangent exact sequence at (m,n)(m,n) is 0⟶T(m,n)(M×ZN)⟶TmM⊕TnN→ Df−Dg TzZ⟶0.0\longrightarrow T_{(m,n)}(M\times_ZN) \longrightarrow T_mM\oplus T_nN \xrightarrow{\,Df-Dg\,}T_zZ\longrightarrow0 . The initial tangent term is the fiber product of the tangent spaces over TzZT_zZ. This explains the precise sense in which differentiation preserves this pullback.

For f(s,t)=s2+t2f(s,t)=s^2+t^2 and gg the inclusion of 11 in R\mathbb R, the pullback is the unit circle, with tangent line su+tv=0s u+t v=0 at (s,t)(s,t). Replacing 11 by 00 leaves the singleton {(0,0)}\{(0,0)\}, while the linear equation given by the derivative has the whole plane as its kernel. Both pullbacks exist in manifolds in this particular example. At the singular value the ordinary tangent to the solution set loses the first-order equation. Testing with an infinitesimal point sends (s,t)(s,t) to (aϵ,bϵ)(a\epsilon,b\epsilon) and makes this difference explicit, since s2+t2=0s^2+t^2=0 in the dual numbers for all a,ba,b. Later intersection constructions will distinguish the agreement of smooth families from the infinitesimal information of a singular equation.

Function spaces and smooth families

Convergence of smooth maps can be imposed on one compact set at a time or simultaneously over the entire source. A smooth family of maps has another kind of structure: it is evaluated jointly in its parameter and source variables. Distinguishing these structures is especially important for noncompact sources. We follow Hirsch’s distinction between the weak and strong topologies [15, Ch. 2], then describe the object that represents jointly smooth families.

Compact control and global control

For a fixed nonnegative integer rr, write Jr(M,N)J^r(M,N) for the bundle of order-rr jet data. A jet at a point consists of the value and the coordinate derivatives through order rr, interpreted in all charts by the chain rule. Chapter [ch:jets] proves the basic properties of this bundle. The weak CrC^r topology is the compact-open topology on jet prolongations. In particular { g:jrg(K)⊂O },K⊂M compact,O⊂Jr(M,N) open,\{\,g:j^r g(K)\subset O\,\},\qquad K\subset M\text{ compact},\quad O\subset J^r(M,N)\text{ open}, is a subbasic set. A neighborhood of ff in this topology requires such a condition on finitely many compact sets. After finitely many charts are chosen around the compact jet data, these neighborhoods amount to uniform coordinate control through order rr on compact subsets.

For the strong CrC^r topology we allow a locally finite family (Ki)(K_i) of compact subsets of MM, with an open jet neighborhood OiO_i for each. The basic condition is jrg(Ki)⊂Oij^r g(K_i)\subset O_i for every ii. This is the same topology as that given by tubes about jet graphs, namely sets of maps whose entire jet graph lies in one open set of Jr(M,N)J^r(M,N). To see one direction, let π\pi be the source map of the jet bundle. The complement of ⋃i(π−1(Ki)∖Oi)\bigcup_i\bigl(\pi^{-1}(K_i)\setminus O_i\bigr) is open, since the union is locally finite and each term is closed. Requiring a graph to lie in that complement imposes exactly the compact-set conditions. For the other direction, use a locally finite compact cover of MM and impose the tube condition on each compact member. Such a cover is available from a compact exhaustion or from a locally finite refinement by precompact coordinate domains. Throughout this discussion the order rr is fixed. A smooth map can be tested at each finite order without committing to a single topology that controls indefinitely many orders at infinity.

Consider a smooth bump function b:R→[0,1]b:\mathbb R\to[0,1] supported in (−1,1)(-1,1) with b(0)=1b(0)=1. The functions gj(x)=b(x−j)g_j(x)=b(x-j) converge weakly to zero through every finite order, since their supports eventually miss any compact set. They miss the strong neighborhood of zero given by the condition ∣g(x)∣<1/2|g(x)|<1/2 at every xx. On a compact source every locally finite family of nonempty compact sets is finite, so the two topologies agree. The bump sequence shows why the compactness hypothesis cannot simply be dropped.

Baire category arguments commonly use spaces in which an exhaustion allows a completeness proof. For instance, on Cr(Rm,Rn)C^r(\mathbb R^m,\mathbb R^n) take the seminorm pjp_j of uniform control of all partial derivatives through order rr on the closed ball of radius jj. A compatible complete metric for the weak topology is d(f,g)=∑j≥12−jmin⁡{1,pj(f−g)}.d(f,g)=\sum_{j\geq1}2^{-j}\min\{1,p_j(f-g)\}. For a Cauchy sequence the functions and each derivative converge uniformly on each ball. The fundamental theorem of calculus along coordinate segments shows that the limits of the derivatives are the corresponding derivatives of the function limit. Induction on the order gives a global CrC^r limit. This proves completeness and the Baire property in this setting. On smooth Euclidean-valued maps one can use the same construction with pjp_j controlling orders up to jj, proving completeness in the weak smooth topology by the same argument. Hirsch proves more generally that the weak topology on Cr(M,N)C^r(M,N) admits a complete metric, and every weakly closed subset is a Baire space for the induced strong topology, using closure of holonomic jets and completeness on the target of jet maps [15, Ch. 2, §4]. In later genericity arguments one has to justify the topology and the Baire property of the actual class of maps being perturbed.

Approximation with fixed data

Here is a useful approximation statement in a form that separates the local analytic argument from the geometric target.

Proposition 1.1 (Relative smoothing). Let M,NM,N be ordinary smooth manifolds without boundary and let f:M→Nf:M\to N be CrC^r for finite rr. If ff is smooth on a neighborhood of a closed set A⊂MA\subset M, it can be approximated in any strong CrC^r neighborhood by a smooth gg agreeing with ff on a possibly smaller neighborhood of AA. The approximation can be joined to ff by a homotopy fixed near AA.

Proof. For an RL\mathbb R^L-valued map FF take a locally finite smooth partition of unity by functions ψi\psi_i with compact support in coordinate domains. In each coordinate chart mollify ψiF\psi_i F with sufficiently small support radius. The mollified functions F~i\widetilde F_i are smooth and supported in slightly larger compact subsets of the same coordinate domains, chosen locally finite from the start. Since convolution commutes with derivatives through order rr, their differences from ψiF\psi_iF approach zero in CrC^r in the coordinate chart. Extend by zero and sum.

To meet a given strong neighborhood, use locally finite compact controls on the sum. Each enlarged support meets only finitely many of these control sets. On each meeting, derivatives in the control charts are estimated by the chart derivatives using the chain rule on compact subsets. Choose the smoothing radius for the iith summand so its contribution to each relevant constraint is at most 2−i−12^{-i-1} times the allowed error there. The sum then lies in the prescribed neighborhood. If FF is smooth near AA, first choose a smooth cutoff χ\chi with χ=1\chi=1 near AA and with support inside that smooth locus. Replace the sum F~\widetilde F by χF+(1−χ)F~\chi F+(1-\chi)\widetilde F. Multiplication by 1−χ1-\chi has a bounded derivative estimate on each compact constraint. Choosing smaller local errors gives the same approximation with the exact agreement required.

For a manifold target use a proper smooth embedding i:N↪RLi:N\hookrightarrow\mathbb R^L and a tubular retraction ρ:V→i(N)\rho:V\to i(N) from an open neighborhood [15, Chs. 2, 4]. Apply the Euclidean argument to F=i∘fF=i\circ f. Before smoothing, tighten the compact controls so that the straight segments between FF and its approximation remain in VV. This is possible on each compact set since FF maps it into a compact subset of the open set VV. The chain rule and compact control also allow the approximation to be chosen small enough so that ρ∘F~\rho\circ\widetilde F meets the original jet conditions. Composing the straight-line homotopy with i−1ρi^{-1}\rho gives the assertion, including the fixed neighborhood. ◻

Real-analytic approximation has another source. Local partitions of unity cannot serve as analytic cutoffs, since an analytic function on a connected domain that vanishes on a nonempty open set vanishes identically. A compact interval illustrates a different mechanism. If ff is CrC^r on [a,b][a,b], approximate f(r)f^{(r)} uniformly by a polynomial qq. For r≥1r\geq1 integrate qq successively rr times, choosing the constants to match ff and its lower derivatives at aa. The resulting polynomial approximates the jjth derivative for 0≤j<r0\le j<r with error bounded by ∥q−f(r)∥∞ (b−a)r−j(r−j)!.\|q-f^{(r)}\|_\infty\,\frac{(b-a)^{r-j}}{(r-j)!}. This follows by integrating the derivative error at each step. For r=0r=0 it is the original polynomial approximation. On general real-analytic manifolds the global approximation theorem uses complex analytic methods. Hirsch states the density of analytic maps in the strong topology for targets and sources equipped with analytic structures [15, Ch. 2, §5]. Density does not impose exact agreement with arbitrary prescribed smooth data.

The object representing jointly smooth families

Let E\mathcal E be the category of sheaves of spaces from Chapter [ch:manifolds]. It has an exponential [M‾,N‾][\underline M,\underline N], characterized by a natural equivalence Map⁡(S,[M‾,N‾])≃Map⁡(S×M‾,N‾).\operatorname{Map}(S,[\underline M,\underline N]) \simeq \operatorname{Map}(S\times\underline M,\underline N). If S=U‾S=\underline U is a coordinate probe, the right-hand side is the discrete space of jointly smooth maps U×M→NU\times M\to N. To see this, the product is represented by the usual product manifold, allowing corners for boundary probes, and test-family reconstruction applies to it. The exponential therefore treats a family and its evaluation under a single universal property. Unlike a topology alone, it specifies which parameter tests have smooth evaluation [2].

A jointly smooth family F:U×M→NF:U\times M\to N is continuous into the weak CrC^r topology. For a point u0u_0 and compact set KK, joint continuity of the partial derivatives in the MM variables and compactness of KK yield one parameter neighborhood for each required compact control. This proof does not supply infinitely many controls at once. For F(s,x)=sxF(s,x)=sx on R×R\mathbb R\times\mathbb R, the parameter family is not continuous at s=0s=0 into the strong topology: the strong neighborhood ∣g(x)∣<1|g(x)|<1 on the whole source contains F(0,−)F(0,-) and no other member. Thus the smooth-family structure does not implicitly choose the strong topology.

When MM is compact, familiar infinite-dimensional charts explain the first-order geometry of this function object. Choose a Riemannian metric on NN. For gg sufficiently close to ff in C0C^0, the inverse to the exponential map near the zero section is defined for all pairs (f(x),g(x))(f(x),g(x)), by compactness of f(M)f(M). The resulting section of f∗TNf^*TN is a local coordinate for gg, and its inverse sends a small section vv to x↦exp⁡f(x)v(x)x\mapsto\exp_{f(x)}v(x). For a smoothly evaluated curve through ff, differentiating in its curve parameter gives a smooth section of f∗TNf^*TN. Every such section occurs as a velocity, since compactness permits one common small time interval for this exponential construction. This gives the first-order tangent model Γ(M,f∗TN)\Gamma(M,f^*TN). A full manifold structure on the space of maps requires an infinite-dimensional calculus. For the compact-source convenient model and its exponential law see [17].

Jets and differential constraints

A map and its first derivative determine whether it is an immersion or a submersion at a point. A second derivative can distinguish a nondegenerate critical point from a more singular one. A jet is the portion of a map needed for such finite-order questions. Its construction turns differential conditions into geometric conditions on a bundle, which can then be tested on local families.

Finite Taylor data

Use boundaryless manifolds for the constructions here. Two germs of smooth maps M→NM\to N at xx have the same rr-jet if their values are the same and their coordinate expressions have equal partial derivatives through order rr at xx. In charts centered at xx and at that common value yy, this is equality of Taylor polynomials through order rr. The chain rule for compositions of Taylor polynomials, discarding terms of degree >r>r, shows that the equivalence relation is independent of the charts. The same calculation shows that changing charts gives a smooth change of jet coordinates. Such a change uses derivatives of the coordinate transitions evaluated at xx and yy, and polynomial expressions in the given coefficients. Thus the classes form a smooth bundle Jr(M,N)J^r(M,N) over M×NM\times N. At order zero the bundle is M×NM\times N. The first-jet bundle has fiber Hom⁡(TxM,TyN)\operatorname{Hom}(T_xM,T_yN) at (x,y)(x,y).

For r≥1r\geq1 the truncation Jr(M,N)→Jr−1(M,N)J^r(M,N)\to J^{r-1}(M,N) is an affine bundle. If two jets lie over the same lower jet at (x,y)(x,y), the difference of the rrth-order coefficients defines an element of Sym⁡r(Tx∗M)⊗TyN.\operatorname{Sym}^r(T_x^*M)\otimes T_yN . Indeed, in a target coordinate change the linear term of the change acts on the difference by its derivative at yy. Any nonlinear term that would contain a difference of order rr has another factor of positive order and contributes only in order >r>r. Source coordinate change acts on the difference by the rrth symmetric power of the derivative of the source change. Terms depending only on lower derivatives cancel. The tensor difference is therefore intrinsic. In coordinates one can add any homogeneous degree-rr term to a fixed Taylor polynomial, and that polynomial is the germ of a smooth map to the target chart after restricting the source. This gives the free and transitive action on each fiber that defines the affine structure. The resulting tower and its applications to singularities are developed in [11].

A map f:M→Nf:M\to N has a prolongation jrf:M→Jr(M,N)j^r f:M\to J^r(M,N), given by its jet at each point. Over M×NM\times N it lies over the graph of ff: image If F:P×M→NF:P\times M\to N is smooth, differentiating only in the MM variables gives a smooth family (p,x)↦jxrFp(p,x)\mapsto j^r_xF_p. It commutes with every smooth parameter change P′→PP'\to P. This is the relative prolongation used when the parameters sample different maps from the same source.

Composition, formal data, and rank

Write jxrgj^r_x g for the jet of a germ with specified source point xx. For a composable pair of germs, truncated composition gives jf(x)rg∘jxrf=jxr(g∘f).j^r_{f(x)}g\circ j^r_x f=j^r_x(g\circ f). This is well-defined for arbitrary representatives of the two classes, by the Taylor argument above. Associativity follows from composition of germs. Consequently postcomposition by a fixed smooth map gives a map of jet bundles, also changing the target value. Precomposition by a fixed h:M′→Mh:M'\to M first pulls back jet data by its source map hh and then uses the jet of hh at each point. The two operations commute. Between points of a fixed manifold, the composable jets of germs of local diffeomorphisms, for r≥1r\geq1, form a groupoid. Inverse jets come from smooth local inverses. In coordinates, taking the inverse of an invertible linear term and solving successively in each higher degree gives smooth inversion on this groupoid.

A smooth section of Jr(M,N)→MJ^r(M,N)\to M is called formal differential data. It is holonomic when it equals jrfj^r f. For M=N=RM=N=\mathbb R, coordinates (x,y,a)(x,y,a) on J1J^1 designate source, value, and derivative. A section (x,y(x),a(x))(x,y(x),a(x)) is holonomic exactly when dy−a dxdy-a\,dx pulls back to zero. The choice y(x)=0y(x)=0 and a(x)=1a(x)=1 fails this condition. Thus a bundle of possible derivatives does not by itself supply a space of solutions by ordinary maps. A global example is the first-jet section over S1S^1 with value identically zero and derivative the angular 11-form dθd\theta. It specifies an injective linear map on each tangent line. There can be no smooth immersion S1→RS^1\to\mathbb R, since a global smooth real function on the circle has a critical point. In particular this everywhere regular formal data cannot be deformed through formal immersions to a holonomic one between these fixed manifolds.

The rank strata of first jets provide useful smooth submanifolds. For a linear map between spaces of dimensions m,nm,n, the locus of rank ℓ\ell has codimension (m−ℓ)(n−ℓ)(m-\ell)(n-\ell) in the space of matrices. To prove this, choose a nonvanishing ℓ×ℓ\ell\times\ell minor AA and decompose a nearby matrix into blocks (ABCD).\begin{pmatrix}A&B\\ C&D\end{pmatrix}. The rank is ℓ\ell exactly when D−CA−1B=0D-CA^{-1}B=0. In these coordinates (A,B,C,D−CA−1B)(A,B,C,D-CA^{-1}B) is a smooth coordinate change, so the displayed condition imposes (m−ℓ)(n−ℓ)(m-\ell)(n-\ell) independent equations. Every rank-ℓ\ell matrix belongs to one of these charts. At rank zero the condition is that all entries vanish. This proof applies in local bundle frames, giving the strata in J1(M,N)J^1(M,N).

For example, the family fa(x)=x3−axf_a(x)=x^3-a x has derivative 3x2−a3x^2-a. The total parameter-source family meets the zero-derivative stratum transversely, since differentiating that equation in aa gives −1-1. Its critical-point locus is therefore the smooth parabola a=3x2a=3x^2. The projection of this locus to parameters has a critical point at the origin. For a>0a>0 the fixed map has two critical points, for a<0a<0 it has none, and for a=0a=0 the two branches meet at a degenerate critical point. Regularity of a total family and regularity of each parameter fiber are separate questions even in this elementary example.

The construction admits a description by smooth families. Let Fr(U)\mathcal F_r(U) consist of the smooth maps U×M→Jr(M,N)U\times M\to J^r(M,N) over the projection to MM. These sets are a sheaf in UU, and relative prolongation is a natural map [M‾,N‾]→Fr[\underline M,\underline N]\to\mathcal F_r. Taking the value coordinate gives a retraction, since every relative jet has the original map as its zeroth component. The subfunctor of relative holonomic data consists precisely of the image. Indeed, if formal data are holonomic over an open cover of a parameter-source product, their local realizing maps must be the value coordinate there. These maps consequently glue, and their prolongations agree on overlaps. Holonomicity is a local compatibility of derivatives with values, rather than a restriction on the topology of the parameter alone.

Jets at distinct source points

For a positive integer qq let M(q)M^{(q)} be the ordered configuration space of qq distinct points of MM. Over it the order-rr multijet bundle has fiber at (x1,…,xq)(x_1,\ldots,x_q) equal to the product of the jet fibers at the specified source points, with values free in NN. The multijet jqrfj^r_q f is evaluation of jrfj^r f in all the slots. Its domain excludes repeated source points. The purpose of this exclusion is that local perturbations can be varied independently in the slots.

Here is the local independence argument. Fix a map ff and distinct points xix_i. Take disjoint coordinate neighborhoods of the xix_i with ff mapping each into a target chart. In each source coordinate centered at xix_i, choose a bump function equal to one near the center and supported compactly within the neighborhood. Add a polynomial of degree at most rr times this bump to the target coordinates of ff. For small polynomial coefficients all perturbed points still lie in the target chart, since the change is made on a compact support. Continue by ff outside the supports. At xix_i the derivatives of the change are exactly the derivatives of the chosen polynomial, and at all other chosen points the jet is unchanged. Therefore the derivative of multijet evaluation in the polynomial coefficients surjects onto the product of the vertical tangent spaces of the jet bundle at the configuration. This finite-dimensional family is the input to multijet transversality arguments [11, Ch. II].

At order zero and q=2q=2, the pair of values (f(x1),f(x2))(f(x_1),f(x_2)) can be tested against the diagonal of N×NN\times N. If this test is transverse, the double-point locus away from the source diagonal is a submanifold of dimension 2dim⁡M−dim⁡N2\dim M-\dim N when nonempty, by the transverse pullback calculation of Chapter [ch:tangent]. The rank condition is a one-point test. Accidental equality of values is a two-point test. The distinction explains why injectivity questions require multijets even when the immersion condition is already settled.

Critical values and smooth pullbacks

A level equation f(x)=yf(x)=y asks for a pullback by the point y:∗→Ny:*\to N. To identify this pullback with a smooth manifold of the expected dimension, the crucial hypothesis is a rank condition on the derivative along the fiber. Sard’s theorem makes that condition available for almost every choice of the point. The point of view in this chapter is to keep the rank test and the genericity argument distinct, so that each can subsequently be applied to families.

The fiber of a regular value

Let M,NM,N be finite-dimensional second-countable smooth manifolds without boundary, with dim⁡M=m\dim M=m and dim⁡N=n\dim N=n. A point x∈Mx\in M is critical for a smooth map ff if dfx:TxM→Tf(x)Ndf_x:T_xM\to T_{f(x)}N fails to be surjective. The critical values form the image of the critical points. A regular value is a point outside this image, so a point outside f(M)f(M) is a regular value with empty fiber.

Theorem 1.1 (Regular preimages). If yy is a regular value, then f−1(y)f^{-1}(y) is an embedded submanifold, of dimension m−nm-n when nonempty. At a point x∈f−1(y)x\in f^{-1}(y) its tangent space is ker⁡dfx\ker df_x. More generally, for an embedded submanifold W⊆NW\subseteq N of codimension cc, suppose dfx(TxM)+Tf(x)W=Tf(x)Ndf_x(T_xM)+T_{f(x)}W=T_{f(x)}N whenever f(x)∈Wf(x)\in W. Then f−1(W)f^{-1}(W) is an embedded submanifold of codimension cc when nonempty, with tangent space dfx−1(Tf(x)W)df_x^{-1}(T_{f(x)}W).

Proof. For the general statement choose a local submersion q:U→Rcq:U\to\mathbb R^c defining WW near f(x)f(x), so that W∩U=q−1(0)W\cap U=q^{-1}(0). The displayed hypothesis says exactly that d(qf)xd(qf)_x is surjective. In source coordinates whose first cc columns give an invertible minor, augment qfqf by the remaining source coordinates. The derivative of the augmented map is invertible, so by the inverse function theorem its components are local coordinates. In these coordinates the preimage is the zero set of the first cc coordinates. Differentiating the equations identifies its tangent space with ker⁡d(qf)x=dfx−1(Tf(x)W)\ker d(qf)_x=df_x^{-1}(T_{f(x)}W). The case W={y}W=\{y\} has c=nc=n. The local descriptions use the subspace topology and thus assemble to the asserted embedded structures. ◻

The square in the theorem is a smooth pullback. image Indeed, a smooth test map into MM whose image lands in f−1(W)f^{-1}(W) factors smoothly through it. This follows from the adapted coordinates in the proof. Testing against all smooth manifolds gives the same pullback square of representable sheaves. Such sheaves are 00-truncated, so their pullback in sheaves of spaces is also computed by the ordinary fiber-product condition. Representability of that sheaf pullback requires a separate argument. For f(x)=x2f(x)=x^2 and y=0y=0, the fiber sheaf is represented by a point, but df0=0df_0=0 and its kernel is one-dimensional. A derived enhancement of a smooth equation can record this failure of the differential rank test.

Sard’s theorem in the smooth case

A subset of a manifold is null if its intersection with each coordinate chart has Lebesgue outer measure zero in that chart. This definition is invariant under smooth change of coordinates because the change is locally Lipschitz on a countable cover by compact coordinate boxes. A smooth positive density induces the same local class of null sets.

Theorem 1.2 (Sard). For a smooth map between finite-dimensional second-countable manifolds without boundary, the critical values are null. In particular the regular values are dense.

Proof. The statement is local in source and target. Use countable charts with the image of each source chart in a target chart, and work with f:U→Rnf:U\to\mathbb R^n for U⊆RmU\subseteq\mathbb R^m open. The case n=0n=0 has no critical points. For n≥1n\geq1, induct on mm, with the case m=0m=0 giving at most a single image point on each chart. Write CC for the critical set and CjC_j for the subset on which every partial derivative of every component of ff, of orders 11 through jj, vanishes. Thus C1⊆CC_1\subseteq C.

Positive derivative rank. Near a point of C∖C1C\setminus C_1 there is a nonzero first partial of some component. After permuting coordinates, use that component in place of the first source coordinate. The inverse function theorem makes this a local coordinate change and the new expression for ff has the form f(t,u)=(t,F(t,u)).f(t,u)=(t,F(t,u)). If n=1n=1, there are no critical points in this chart. Otherwise the derivative has rank 1+rank⁡duF(t,u)1+\operatorname{rank}d_u F(t,u). At each tt, the corresponding section of the image of the critical set lies among the critical values of u↦F(t,u)u\mapsto F(t,u). It is null by induction. To apply Fubini, exhaust the chart by countably many compact subsets and intersect the critical set with each one. Since criticality is a closed condition, the images of these intersections are compact, and their slices are measurable. Fubini gives measure zero on each compact image. Countably many such charts cover C∖C1C\setminus C_1.

A first nonvanishing higher derivative. Near any x∈Cj∖Cj+1x\in C_j\setminus C_{j+1}, for j≥1j\geq1, some scalar partial derivative gg of order jj has nonzero differential. It vanishes on CjC_j, so the portion of CjC_j in this neighborhood lies in the smooth hypersurface g−1(0)g^{-1}(0) through xx. At every point of CjC_j there, df=0df=0, and therefore the restriction of ff to the hypersurface is critical at each of these points. Its source dimension is m−1m-1, so the induction hypothesis gives a null image of this part of CjC_j. A countable collection of hypersurface charts covers Cj∖Cj+1C_j\setminus C_{j+1}.

High-order vanishing. Choose jj with (j+1)n>m(j+1)n>m. On a closed coordinate box K⊂UK\subset U, fix a bound for all derivatives of order j+1j+1. Subdivide the box into O(ϵ−m)O(\epsilon^{-m}) pieces of diameter at most a constant times ϵ\epsilon. For each piece meeting CjC_j, expand at one such point. Taylor’s formula on the segment inside KK bounds the oscillation from that point by O(ϵj+1)O(\epsilon^{j+1}). Thus f(Cj∩K)f(C_j\cap K) is covered by O(ϵ−m)O(\epsilon^{-m}) balls of radius O(ϵj+1)O(\epsilon^{j+1}). Their total nn-volume is O(ϵ(j+1)n−m)O(\epsilon^{(j+1)n-m}), which tends to zero. A countable box cover handles all of CjC_j. The three cases exhaust CC. ◻

This smooth proof follows the induction underlying [24, Section 3]. Hirsch also treats finite differentiability. For CrC^r maps, the sufficient threshold is r>max⁡{m−n,0}r>\max\{m-n,0\} [15, Chapter 3]. In the smooth argument above, choosing a high order of vanishing helps obtain the last covering estimate. The sharper finite-differentiability threshold uses a different argument. Sard’s conclusion concerns the image of the critical set. The set itself can contain open regions, as for any map constant on an open set.

Boundary strata and openness

Suppose MM has boundary and NN has none. If a value is regular for both f∣int⁡Mf|_{\operatorname{int}M} and f∣∂Mf|_{\partial M}, the fiber is a manifold with boundary, with boundary equal to the fiber of the second restriction. At a boundary point write a source chart as (u,t)(u,t) with t≥0t\geq0. The surjectivity on the boundary stratum lets us choose the invertible minor in the uu-directions. The coordinate change in the regular preimage proof then leaves tt unchanged. Apply Sard’s theorem to the two strata and take a union of null sets to see that almost every value has this property. Testing only the ambient tangent space at a boundary point is insufficient. For the inclusion of [0,∞)[0,\infty) in R\mathbb R, the ambient derivative at 00 is surjective, but the boundary restriction is critical there.

For a proper ff without boundary, the critical set is closed and its image is closed, so the regular values are open as well as dense. Properness cannot be discarded in this statement. For a connected example, let f:R→Rf:\mathbb R\to\mathbb R be f(x)=e−x(2+sin⁡(2x))f(x)=e^{-x}(2+\sin(2x)). It is strictly positive, so 00 is a regular value with empty fiber. At x=jπx=j\pi the derivative f′(x)=e−x(2cos⁡(2x)−2−sin⁡(2x))f'(x)=e^{-x}\bigl(2\cos(2x)-2-\sin(2x)\bigr) vanishes, and the resulting critical values 2e−jπ2e^{-j\pi} approach zero as j→∞j\to\infty. This illustrates why openness statements for families typically include compactness or control of the behavior at infinity.

Example 1.3 (Completing the tangent exact sequence). For the function f(x1,…,xm)=∑ixi2f(x_1,\ldots,x_m)=\sum_i x_i^2, every positive real number aa is regular and its fiber is the sphere of radius a\sqrt{a}. At a point xx of the sphere, 0⟶Txf−1(a)⟶Rm→ v↦2⟨x,v⟩ R⟶00\longrightarrow T_x f^{-1}(a) \longrightarrow \mathbb R^m \xrightarrow{\ v\mapsto 2\langle x,v\rangle\ } \mathbb R\longrightarrow 0 is exact. Dividing by ∥x∥2\|x\|^2 constructs a splitting in the radial direction, which varies smoothly for a>0a>0. At a=0a=0 the ordinary fiber contracts to the origin and the derivative has nontrivial cokernel. The comparison exhibits the information in the linearized equation that its set of solutions cannot retain.

Transversality in parameter families

The preceding chapter used regular values to vary a point in the target. A family of maps allows another kind of variation. The parameter can repair a normal derivative that fails to be onto in the input variables. Parametric transversality makes precise when this collective freedom gives individual members with well-behaved pullbacks.

The projection of an incidence manifold

Let P,M,NP,M,N be finite-dimensional second-countable smooth manifolds without boundary, and let W⊂NW\subset N be an embedded submanifold. Say f⋔Wf\pitchfork W when ff satisfies the normal surjectivity condition in Theorem [thm:regular-preimage]. Write Fp(x)=F(p,x)F_p(x)=F(p,x) for the members of a smooth family F:P×M→NF:P\times M\to N.

Theorem 1.1 (Parametric transversality). If F⋔WF\pitchfork W, the incidence space Z=F−1(W)Z=F^{-1}(W) is an embedded submanifold. For π:Z→P\pi:Z\to P the restricted projection, pp is a regular value of π\pi exactly when Fp⋔WF_p\pitchfork W. The set of such parameters is conull, dense, and residual. If π\pi is proper, it is also open.

Proof. Theorem [thm:regular-preimage] constructs ZZ. At (p,x)∈Z(p,x)\in Z, write Q=TF(p,x)N/TF(p,x)WQ=T_{F(p,x)}N/T_{F(p,x)}W. The derivative followed by projection to QQ gives a pair of linear maps A:TpP→QA:T_pP\to Q, B:TxM→QB:T_xM\to Q, with im⁡A+im⁡B=Q\operatorname{im}A+\operatorname{im}B=Q. The tangent space of ZZ consists of (u,v)(u,v) such that Au+Bv=0Au+Bv=0. The map dπd\pi sends this pair to uu, so it is onto if and only if im⁡A⊆im⁡B\operatorname{im}A\subseteq\operatorname{im}B. Joint surjectivity makes this equivalent to surjectivity of BB. That is the transverse condition for FpF_p at xx. The criterion is vacuous when π−1(p)\pi^{-1}(p) is empty.

By Sard’s theorem, the bad parameters form a null set. The critical set of π\pi is closed in a second-countable, locally compact manifold ZZ, hence is a countable union of compact subsets. Each compact subset has compact, null image in PP, which has empty interior. The bad set is therefore a countable union of closed nowhere dense sets. If π\pi is proper its whole critical image is closed. ◻

The diagram compares the fibers with the total incidence space. image The left square gives a smooth manifold pullback at every regular parameter. If MM is compact and WW is closed in NN, the incidence space is a closed subset of P×MP\times M, which makes its projection proper. Its restriction to the open regular locus is a proper submersion and is smoothly locally trivial. Over a coordinate ball choose a metric and lift radial paths from the center horizontally. Properness keeps their lifts over a closed smaller ball in a compact inverse image. The lifts can therefore be integrated all the way along each path, and can be reversed. Smooth dependence on the endpoints and initial values gives the local trivialization. In particular, after any smooth change of parameter into this regular locus, the incidence manifold pulls back as an ordinary smooth manifold. This is a representability result for that family pullback, proved by the normal derivative criterion together with local transport.

Example 1.2 (A constrained family on the circle). Consider F((a,b),t)=sin⁡t+acos⁡t+bF((a,b),t)=\sin t+a\cos t+b on R2×S1\mathbb R^2\times S^1, with W={0}W=\{0\}. The derivative with respect to bb is 11, so the total incidence manifold is smooth. The trigonometric summand sin⁡t+acos⁡t\sin t+a\cos t has amplitude R=1+a2R=\sqrt{1+a^2}. Its derivative vanishes exactly at its two extrema. The parameters with a nonregular zero are therefore the two curves b=±Rb=\pm R. In the region ∣b∣<R|b|<R each fiber has two points. In the region ∣b∣>R|b|>R it is empty. On a discriminant curve the vanished input derivative has a nonzero second derivative. Although the full family is transverse, freezing bb and aa at such a point gives a nontransverse member. A different family F(a,x)=ax2F(a,x)=ax^2 fails even joint transversality along x=0x=0 because both partial derivatives vanish there.

Genericity in a space of maps

Parametric transversality can also be used to construct dense sets of transverse maps. Here the family is invented locally for a given map, so the argument needs no global finite dimensional chart on the space of maps.

Proposition 1.3 (Transversality on a compact set). Let W⊂NW\subset N be closed and embedded, and let K⊂MK\subset M be compact. Smooth maps that are transverse to WW at all points of KK form an open dense subset of C∞(M,N)C^\infty(M,N) in the weak Whitney topology. With a compact exhaustion of MM, global transversality is residual in this topology, and dense by its Baire property.

Proof. For openness, cover KK by finitely many relatively compact source charts on which the original map either avoids WW or is tested against a target chart adapted to WW. Avoidance on a compact subset persists because WW is closed. Where the map meets WW, the relevant minor of the normal derivative is nonzero and persists after shrinking the source chart and restricting to small C1C^1 changes. Together finitely many such conditions work on KK.

For density, start with a smooth f:M→Nf:M\to N and a compact neighborhood K′K' of KK. Pick finitely many smooth sections viv_i of f∗TNf^*TN, supported in a compact neighborhood of K′K', whose values span at every point of K′K'. To construct them, work in local bundle frames, multiply frame vectors by cutoff functions that are nonzero on smaller neighborhoods, and use a finite cover. For small a=(a1,…,as)a=(a_1,\ldots,a_s) define F(a,x)=exp⁡f(x)(∑i=1saivi(x)).F(a,x)=\exp_{f(x)} \left(\sum_{i=1}^s a_i v_i(x)\right). Use a Riemannian metric on NN. Compact support lets all these exponentials be defined uniformly for small aa. At a=0a=0 the parameter derivative is surjective at every point of K′K', and openness and compactness make it surjective on K′K' for all small parameters. Consequently the family is transverse on the open manifold int⁡K′\operatorname{int}K', and Theorem 1.1 gives arbitrarily small parameters whose maps are transverse on KK. The maps converge smoothly on compacta to ff as a→0a\to0, so every weak neighborhood of ff contains one of them. Intersect the open dense tests for the members of a countable compact exhaustion. The weak smooth mapping space is Baire [15, Chapter 2], proving the last assertion. ◻

The same local mechanism applies to jet conditions from Chapter [ch:jets]. In coordinate neighborhoods of a point xx, perturb a map by vector-valued polynomials of degree at most rr multiplied by a cutoff equal to one near xx. Varying their coefficients spans all derivatives through order rr at any point of this smaller neighborhood: translation of a polynomial is an invertible change of basis of the finite-dimensional polynomial space. The source coordinate of the jet is supplied by the input variable. The evaluation of prolonged jets is thus a submersion there. Applying parametric transversality to a smooth submanifold in the jet bundle gives the local density step of the jet transversality theorem [15, Chapter 3]. In particular, maps into a jet bundle arising from actual smooth functions can be varied through such maps. Arbitrary perturbations of a section need not remain a prolongation.

Boundary and relative constraints

When MM has boundary, test transversality of the full family on both P×int⁡MP\times\operatorname{int}M and P×∂MP\times\partial M, with P,N,WP,N,W as above. Their incidence spaces are the interior and boundary strata of ZZ. At a boundary point one can choose a nonsingular normal minor in the boundary variables, leaving the collar parameter unchanged. Apply Sard to the two projections into PP. Outside their union of critical images, both stratumwise restrictions of FpF_p are transverse, and the fiber is a manifold with boundary as in Chapter [ch:sard]. When the parameter manifold itself has boundary, the same argument can be applied separately to the parameter strata, using the appropriate total family on each product stratum.

A relative problem can require the map to stay fixed on a closed subset A⊂MA\subset M. If it is already transverse on a neighborhood of AA, compact-set perturbations can be supported away from AA. For the proof of Proposition 1.3, choose an open neighborhood VV of AA with closure inside the region where ff is transverse. Then K∖VK\setminus V is compact and disjoint from AA. Build the spanning sections using supports in M∖AM\setminus A and apply the parametric argument to an open neighborhood of K∖VK\setminus V. The existing transversality on K∩V‾K\cap\overline V persists for sufficiently small perturbations, since this set is compact. At a point constrained to remain fixed on an open region, any preexisting failure of transversality in that region persists as well. The support of the allowed perturbation is therefore part of the geometric hypothesis.

Bundles, frames, and classifying objects

The tangent exact sequence of a transverse preimage is a sequence of vector bundles rather than a list of separate linear maps. Bundles make linear operations vary with the base, and frame changes record how a local description varies. They provide a useful first example of the distinction between classifying objects up to existence of an isomorphism and retaining the groupoid of their presentations and symmetries.

Linear gluing and pullback

Let XX be a smooth manifold. A smooth rank-rr real vector bundle E→XE\to X has smooth local trivializations that are linear on fibers. Choose smooth frames sis_i over open sets Ui⊂XU_i\subset X, viewing si(x)s_i(x) as a linear isomorphism Rr→Ex\mathbb R^r\to E_x. On an overlap write sj=sigijs_j=s_i g_{ij}, where gij:Ui∩Uj→GLr(R)g_{ij}:U_i\cap U_j\to \mathrm{GL}_r(\mathbb R) is smooth. On triple overlaps these matrices satisfy gijgjk=gikg_{ij}g_{jk}=g_{ik}. Conversely, such a cocycle, with gii=1g_{ii}=1, glues the products Ui×RrU_i\times\mathbb R^r by (x,v)j∼(x,gij(x)v)i(x,v)_j\sim(x,g_{ij}(x)v)_i. The cocycle makes the relations consistent on all overlaps, and the inclusion of each product gives a chart of the glued total space. Changing frames by si′=sihis'_i=s_i h_i replaces gijg_{ij} by hi−1gijhjh_i^{-1}g_{ij}h_j. A section is a family of smooth functions viv_i satisfying vi=gijvjv_i=g_{ij}v_j.

A map a:Y→Xa:Y\to X pulls the bundle back by pulling back its frames and transitions. The diagram describes the pullback on total spaces. image If bundles E,FE,F have transitions gij,hijg_{ij},h_{ij} on a common cover, the direct sum and tensor product have transitions gij⊕hijg_{ij}\oplus h_{ij} and gij⊗hijg_{ij}\otimes h_{ij}. The bundle of linear maps E→FE\to F has transition A↦hijAgij−1A\mapsto h_{ij} A g_{ij}^{-1}. Thus a global linear map of bundles is a section of a bundle defined by this action on matrices. This local transformation law matters whenever a model takes a feature in one bundle to a feature in another, because all coordinate matrices must satisfy the overlap conditions.

A smooth positive definite fiber metric exists on any bundle over a paracompact smooth manifold. Choose local metrics and average them with a smooth partition of unity. Each nonzero vector has positive length in every metric with nonzero coefficient, so the average is again positive. Orthonormalizing local frames for this metric gives O(r)\mathrm O(r)-valued transitions. Orthogonal complements then split smooth subbundles. For example, a short exact sequence of smooth bundles splits by choosing a metric in the middle term, but the splitting depends on that choice.

Finite Grassmannians and stabilization

Let Grr(RN)\mathrm{Gr}_r(\mathbb R^N) be the Grassmannian of rr-planes with tautological bundle γr,N\gamma_{r,N}. Its fiber over a plane is the plane itself.

Theorem 1.1 (A classifying map for a compact smooth base). Let XX be a compact smooth manifold. A smooth rank-rr bundle over XX is isomorphic to c∗γr,Nc^*\gamma_{r,N} for some smooth map c:X→Grr(RN)c:X\to\mathrm{Gr}_r(\mathbb R^N) and some finite NN. Two such maps give isomorphic bundles if and only if they are smoothly homotopic after inclusion in a common larger Grassmannian.

Proof. Take a finite cover by bundle frames and smooth cutoff functions ϕi\phi_i supported in the corresponding frame domains, with at least one ϕi(x)>0\phi_i(x)>0 at each xx. For v∈Exv\in E_x define a linear embedding JxJ_x by concatenating the vectors ϕi(x)si(x)−1v\phi_i(x)s_i(x)^{-1}v, extending by zero outside their domains. One nonzero coefficient suffices for injectivity. These embeddings are smooth over XX, and their image planes form a smooth map cc to the Grassmannian. Locally the image is spanned by a smooth full-rank matrix C(x)C(x). Its orthogonal projection is C(x)(C(x)TC(x))−1C(x)TC(x)(C(x)^{\mathsf T}C(x))^{-1}C(x)^{\mathsf T}, which also shows that the planes vary smoothly. The maps JxJ_x give a bundle isomorphism onto the tautological pullback.

If two maps classify isomorphic bundles, identify their pullbacks with the same EE and regard the resulting fiberwise injections J0,J1J_0,J_1 as injections into the same RN\mathbb R^N by padding with zeros. In the direct sum of two copies of RN\mathbb R^N, the family v⟼(cos⁡t J0v, sin⁡t J1v),0≤t≤π/2,v\longmapsto (\cos t\,J_0v,\ \sin t\,J_1v), \qquad 0\leq t\leq\pi/2 , stays injective, since at least one coefficient is nonzero. Its endpoints are J0J_0 in the first block and J1J_1 in the second. A rotation of the two blocks moves the latter to J1J_1 in the first. Reparametrizing the two paths to be stationary near their concatenation gives a smooth homotopy of the stabilized classifying maps.

Conversely, a smooth homotopy pulls the tautological bundle back to a bundle on X×[0,1]X\times[0,1]. A connection gives parallel transport along each interval {x}×[0,1]\{x\}\times[0,1]. One may take the connection induced by orthogonal projection of the trivial connection in the ambient trivial bundle. The resulting transport is linear, invertible by reverse transport, and smooth in xx by smooth dependence of linear ordinary differential equations. It identifies the endpoint bundles. Stabilizing a plane by inclusion of its ambient vector space pulls the tautological bundle back to the original one, so this is also an isomorphism before stabilization. ◻

Hirsch gives finite-dimensional bounds from the dimension of the base [15, Chapter 4]. The construction here uses a finite cover and lets the ambient dimension grow. The topological theorem for paracompact bases classifies numerable bundles by homotopy classes of continuous maps to an infinite Grassmannian, with the relevant topology on that Grassmannian [14, Chapter 1]. For a compact smooth base, a continuous map into a finite Grassmannian can be smoothed up to homotopy, and continuous homotopies between smooth maps into one can be smoothed relative to their ends after making them stationary near the ends. The smooth formulation retains a smooth bundle and smooth changes of frames throughout its proof.

Classifying the whole groupoid

Consider the site of smooth test manifolds with covers by open sets, using sheaves of spaces for the ambient category. Let Gr\mathcal G_r be the group sheaf Gr(U)=C∞(U,GLr(R))\mathcal G_r(U)=C^\infty(U,\mathrm{GL}_r(\mathbb R)). The geometric realization of the simplicial bar construction of Gr\mathcal G_r, taken in sheaves, is its classifying stack BGrB\mathcal G_r. Its groupoid of points on a test manifold XX is the groupoid of smooth rank-rr bundles on XX and smooth bundle isomorphisms.

To see the classification, the frames of EE form a right Gr\mathcal G_r-torsor. They are locally nonempty, and any two frames differ by a unique smooth pointwise matrix. Conversely a torsor PP gives the associated bundle with fibers Px×GLr(R)RrP_x\times_{\mathrm{GL}_r(\mathbb R)}\mathbb R^r. Local sections of PP reduce this construction to the gluing equations above. These operations are inverse on objects and on isomorphisms. The bar construction presents the groupoid with one object and automorphism sheaf Gr\mathcal G_r. Descent adds the torsors obtained by gluing this object, with the changes of local frames giving their isomorphisms. This computes the stack interpretation.

For example, the trivial line over a point has automorphism group R×\mathbb R^\times as an ordinary group in this groupoid. Its single isomorphism class says nothing about these automorphisms. The topological classifying space uses GLr(R)\mathrm{GL}_r(\mathbb R) with its Lie group topology and classifies bundles through homotopy classes of maps. It answers a corresponding isomorphism-class question in a category with continuous homotopies. A map to the smooth stack, in contrast, records actual smooth changes of frame over each test manifold. The classifying map of Theorem 1.1 concerns only the isomorphism class, and its equivalence relation is stabilized homotopy.

Orientation as reduction of structure group

For r≥1r\geq1, an orientation on a real rank-rr bundle is a coherent choice of orientation on its fibers. The functions sgn⁡det⁡gij\operatorname{sgn}\det g_{ij} define a double cover with fiber the two possible orientations. A section amounts to choosing local signs that agree under these transitions, equivalently to reducing the frame group to the positive determinant subgroup. With a metric this is a reduction from O(r)\mathrm O(r) to SO(r)\mathrm{SO}(r). On stacks the sign of determinant acts as a map to the classifying stack of a two-element group, and an orientation is a trivialization of the resulting torsor. This interpretation also shows why the orientation pulls back with the bundle.

Example 1.2 (Twisted coefficients on a circle). Over S1S^1 identify (2π,v)∼(0,−v)(2\pi,v)\sim(0,-v) in [0,2π]×R[0,2\pi]\times\mathbb R to obtain the Möbius line LL. A section is represented by an antiperiodic real function on the real line. A nowhere-zero section would have constant sign by continuity, contradicting antiperiodicity. Thus LL has no orientation or global frame. The tensor-square line has transition (−1)2=1(-1)^2=1 and is trivial. A local feature that is linear in a vector of LL inherits the sign change. A quadratic scalar in the same vector can be invariant and global even when the frame for the original feature does not exist.

Tubular neighborhoods and collars

An embedded submanifold determines normal directions by a quotient of tangent bundles. Turning the quotient into coordinates in the ambient manifold requires a choice. Tubular neighborhoods provide this realization, and their isotopy properties explain how subsequent constructions can use these coordinates and still be independent up to homotopy. Boundary collars play the same role for an inward half-line.

Normal coordinates near an embedding

Let i:A↪Mi:A\hookrightarrow M be an embedded submanifold, with A,MA,M both without boundary, and set νi=i∗TM/TA\nu_i=i^*TM/TA. A normalized tubular embedding is an embedding e:U→Me:U\to M from an open neighborhood UU of the zero section in νi\nu_i, with open image, such that e(0x)=i(x)e(0_x)=i(x) and the induced derivative in the normal quotient at 0x0_x is the identity of (νi)x(\nu_i)_x. The quotient property of νi\nu_i makes the normalization meaningful without a metric. The source projection determines a retraction on the image. image

Theorem 1.1 (Existence near a compact submanifold). If AA is compact, a normalized tubular embedding exists on an open disk bundle of some positive radius.

Proof. Choose a Riemannian metric on MM, identifying νi\nu_i with the orthogonal complement of TATA in i∗TMi^*TM. Restrict the Riemannian exponential to this complement. At 0x0_x, the derivative on the tangent of the zero section is dixdi_x, and the derivative along a fiber is the inclusion of the normal complement into TxMT_xM. Their sum is an isomorphism. It follows that the restriction is a local diffeomorphism on some neighborhood of the zero section. Because AA is compact, a single small positive radius works for definition of the exponential and invertibility of its derivative on the resulting disk bundle.

If injectivity failed on every smaller disk bundle, there would be pairs uj,vju_j,v_j of distinct vectors of norms tending to zero with equal images. By compactness their footpoints have convergent subsequences, with limits xx and yy. Continuity and injectivity of ii force x=yx=y. Both sequences then approach 0x0_x, contradicting injectivity in a coordinate neighborhood where the map is a diffeomorphism. An injective local diffeomorphism has open image and smooth inverse on it. The derivative calculation provides the normalization. ◻

For a closed noncompact embedded submanifold one can use a positive radius depending on the footpoint [15, Chapter 4]. Uniform radius can fail at infinity. For example, embed a countable discrete set in R2\mathbb R^2 as the points (j,0)(j,0) and (j,1/j)(j,1/j) for j≥1j\geq1. This is a closed zero-dimensional submanifold. Under the Euclidean exponential its normal disks are ambient disks, and no uniform positive radius makes all of them disjoint. Choosing radii less than one third of the distance to the nearest other point remedies the problem.

Isotopy of normalized germs

Proposition 1.2 (Independence after shrinking). Let AA be compact, with ambient manifold without boundary. Any two normalized tubular embeddings e0,e1e_0,e_1 of the same embedding ii, defined on possibly different neighborhoods, are joined by a smooth family of normalized tubular embeddings on a common smaller neighborhood. The family is constant as a germ wherever the two original germs agree.

Proof. After shrinking, form the transition g=e0−1e1g=e_0^{-1}e_1 near the zero section of νi\nu_i. It fixes the zero section, and its normal derivative is the identity. For s>0s>0 let gs=δ1/s∘g∘δsg_s=\delta_{1/s}\circ g\circ\delta_s, where δs\delta_s scales the fibers by ss. In a local trivialization gg has components (b(x,v),w(x,v))(b(x,v),w(x,v)) and w(x,0)=0w(x,0)=0. The base component of gsg_s tends to xx, and its fiber component has the formula w(x,sv)s=∫01Dvw(x,τsv) v dτ.\frac{w(x,sv)}{s} =\int_0^1 D_vw(x,\tau s v)\,v\,d\tau . The equality supplies a smooth extension at s=0s=0, with fiber component vv by normalization. It also shows that sufficiently small vectors, uniformly over the compact zero section and s∈[0,1]s\in[0,1], have images in the domain of e0e_0. For positive ss the transition is an embedding on this smaller domain, by its definition as a composition of embeddings. At zero it is the identity. Its derivative on the quotient remains the identity, and e0∘gse_0\circ g_s gives the required family. On a region where the transition is the identity as a germ, every rescaled transition is the identity as a germ there. ◻

The formula uses compatible transition maps of the normal bundle, so the extension shown in each trivialization glues globally. This conclusion concerns tubular germs. An isotopy on a prescribed radius without shrinkage requires control away from the zero section. There is also a precise base change statement for the quotient normal bundle. If f:X→Mf:X\to M is transverse to AA and B=f−1(A)B=f^{-1}(A), then dfdf followed by projection to νi\nu_i is onto along BB with kernel TBTB. Consequently TX∣B/TB ≅ (f∣B)∗νi.TX|_B/TB\ \cong\ (f|_B)^*\nu_i . This identification is natural for repeated transverse base change. An exponential map uses additional geometric data. Comparing two chosen exponentials requires a corresponding comparison of these choices.

Collars and neat intersections

A collar of a manifold with boundary MM is an embedding c:∂M×[0,ϵ)→Mc:\partial M\times[0,\epsilon)\to M onto a neighborhood of the boundary with c(x,0)=xc(x,0)=x. For compact boundary such a collar is obtained by choosing a smooth vector field near it pointing strictly inward there. Use inward coordinate vectors in local boundary charts, and glue with a partition of unity. Positivity of the derivative of an inward boundary coordinate is invariant under changing the defining coordinate, and is preserved by positive sums of vector fields. Thus the glued field remains inward. Its positive-time flow on the boundary is defined for a uniform short interval by compactness. At time zero its derivative has image the sum of the boundary tangent space and the transverse inward direction. The local inverse function theorem in half-space charts gives a collar near each point. The same compactness argument as for tubular neighborhoods gives a common interval on which this flow is injective and has smooth inverse onto its image.

For a compact manifold MM with boundary, suppose the compact submanifold AA is neat. This means ∂A=A∩∂M\partial A=A\cap\partial M and near each of those points there are half-space charts with coordinates (u,v,t)(u,v,t), t≥0t\geq0, in which AA is defined by v=0v=0. One can construct a collar of MM whose restriction to ∂A\partial A is a collar of AA. In the neat charts the vector ∂t\partial_t is tangent to AA along AA. Choose all local inward fields meeting ∂A\partial A in this form, and choose fields on the remaining patches disjoint from AA near those patches of the boundary. A partition of unity preserves tangency to AA and inwardness. Uniqueness of solutions of the flow equation then gives the compatible collar near ∂A\partial A.

Use product metrics on a smaller common collar and double MM and AA across their boundaries. The double is smooth in the signed collar coordinate, and the doubles form an embedded pair without boundary. Apply the normal exponential construction to this pair. On the product collar, the normal vectors to AA have zero component in the collar direction. Their short normal geodesics have constant collar coordinate. Away from a smaller collar the feet of the vectors are a compact set separated from the boundary. After shrinking further, their geodesics remain in the original half of the double. The resulting restriction is a tubular neighborhood for the original neat pair, compatible with its boundary product structure [15, Chapter 4]. The product condition avoids treating the inward half-line as an unrestricted normal direction.

Example 1.3 (Collapse to a Thom space). Let AA be compact inside a compact manifold MM without boundary, and choose a closed radius-rr disk bundle Dr(νi)D_r(\nu_i) inside a tubular domain. The Thom space here is Dr(νi)/Sr(νi)D_r(\nu_i)/S_r(\nu_i), with the sphere bundle collapsed to a base point. For positive normal rank, the tubular embedding defines a map M→Dr(νi)/Sr(νi)M\to D_r(\nu_i)/S_r(\nu_i): on the embedded disk bundle use the inverse embedding followed by the quotient, and on the complement of its interior use the base point. The definitions agree on the embedded sphere bundle, giving continuity by gluing over two closed subsets. Shrinking the radius identifies Thom spaces by fiberwise rescaling and gives homotopic collapse maps by varying the radius continuously. Likewise, a family of tubular embeddings on a common disk bundle gives a homotopy of collapse maps. The inverse embedding on that family is continuous jointly with the parameter by compactness of the total closed disk bundle over the compact parameter interval. Thus the collapse construction can use a tubular choice and still define a homotopy class independent of it.

Degree and oriented fibers

A regular fiber of a map between manifolds of the same dimension is a finite set when the source is compact. Orientation gives one sign to each point of this set. The degree is the sum of these signs, and its invariance comes from two compatible constructions. The first uses the map on fundamental classes. The second compares regular fibers by an oriented one-dimensional preimage. Both constructions are useful in families. Here a closed manifold is a compact manifold without boundary.

Local orientations and the signed count

Let MM and NN be closed oriented nn-manifolds with NN connected. For a smooth map f:M→Nf:M\to N, a value yy is regular if DxfD_xf is surjective at every point of f−1(y)f^{-1}(y). At such a point the tangent map is an isomorphism. Define ϵx(f)\epsilon_x(f) to be 11 or −1-1 according as it preserves or reverses orientation. The inverse function theorem gives a discrete fiber. Its closedness in a compact space makes it finite, including the possibility of an empty fiber.

The orientation at xx is equivalently a choice of generator of Hn(M,M∖{x};Z)H_n(M,M\setminus\{x\};\mathbb Z). This follows by excision to a coordinate ball and the relative homology of that ball punctured at its center. A global fundamental class [M][M] restricts to this generator at each point. If MM is disconnected, [M][M] sums the fundamental classes of the oriented components. In particular, if ff is surjective on some components and misses a given value on others, every component can still be included in the same formula.

Theorem 1.1 (Regular-fiber degree). The homomorphism on top homology satisfies f∗[M]=d[N],d=∑x∈f−1(y)ϵx(f)f_*[M]=d[N], \qquad d=\sum_{x\in f^{-1}(y)}\epsilon_x(f) at every regular value yy. Thus the signed count is independent of the regular value and depends only on the homotopy class of ff.

Proof. Since Hn(N;Z)H_n(N;\mathbb Z) is generated freely by [N][N], the first equation uniquely defines dd. Put Q=f−1(y)Q=f^{-1}(y). Apply homology to the maps from absolute to relative pairs. image All groups in the square use integer coefficients. Disjoint coordinate balls about the finitely many points of QQ and excision identify the lower left group with the direct sum of the local orientation groups. The image of [M][M] is the tuple of their chosen generators. If QQ is empty, the relative group is zero. On each summand the bottom arrow sends the local generator to ϵx(f)\epsilon_x(f) times the local orientation at yy. Indeed, in small charts the map is a diffeomorphism, and its action on local homology is its orientation sign. The right vertical arrow sends [N][N] to the local orientation at yy. The square proves the signed count. Homotopy invariance follows from that of the induced homology map. ◻

This argument also explains why the signs must be computed relative to both orientations. Reversing the orientation of MM changes [M][M] by a minus sign, and the differential signs all change accordingly. Reversing both orientations leaves the degree unchanged. If f:M→Nf:M\to N and g:N→Lg:N\to L have connected closed oriented targets of the same dimension, functoriality gives deg⁡(gf)=deg⁡(g)deg⁡(f)\deg(gf)=\deg(g)\deg(f). Disjoint union adds degrees for maps to a common target. A map of nonzero degree is surjective, since an omitted value would be regular with empty fiber. The formula can apply even when the critical points occupy part of the source. Only the chosen fiber needs an isomorphism of tangent spaces at each of its points.

Homotopy and the boundary of a fiber

Suppose f0,f1:M→Nf_0,f_1:M\to N have a common regular value yy and are joined by a smooth homotopy FF. Reparametrize near the ends so that FF agrees with the end maps on their collars. Relative transversality permits replacing the interior of FF by a nearby homotopy transverse to yy without altering these collars [15, Chs. 3 and 5]. Its fiber over yy is a compact one-manifold WW with boundary. Orient [0,1]×M[0,1]\times M using the time coordinate followed by the orientation on MM. The preimage orientation on WW is given by 0⟶TW⟶T([0,1]×M)∣W→ DF TyN⟶00\longrightarrow TW\longrightarrow T([0,1]\times M)|_W \xrightarrow{\ DF\ } T_yN\longrightarrow 0 using the kernel first. On each end collar a positive tangent vector of WW points in the increasing-time direction exactly when the corresponding local sign of fif_i is positive. The boundary convention therefore gives ∂W=f1−1(y)⨿−f0−1(y)\partial W=f_1^{-1}(y)\amalg -f_0^{-1}(y) as oriented zero-manifolds. Each compact oriented interval has one positive and one negative endpoint, and circle components contribute none. This proves equality of the signed cardinalities with a geometric comparison. Points of opposite signs can therefore be the two boundary points of one oriented component of a transverse interpolating fiber.

The same reasoning for a homotopy of targets can be expressed as transversality to a graph. For a smooth path γ:[0,1]→N\gamma:[0,1]\to N, one forms the preimage of its graph in [0,1]×N[0,1]\times N under (t,x)↦(t,f(x))(t,x)\mapsto(t,f(x)). If this is transverse, oriented by the normal orientation pulled back from TNTN, it has the corresponding endpoint fibers. The cylinder argument and the graph argument are instances of the same normal exact sequence. The top-homology proof covers paths through arbitrary singular values without choosing such a transverse comparison.

Families of regular values

Let PP be a smooth parameter manifold and MM compact. Suppose F:P×M→NF:P\times M\to N is smooth and a smooth choice y:P→Ny:P\to N is regular for each Fp=F(p,−)F_p=F(p,-), with dim⁡M=dim⁡N\dim M=\dim N. Consider the diagram image The diagram is a transverse pullback. In target coordinates the equation is F(p,x)=y(p)F(p,x)=y(p), and its derivative in xx is invertible at every solution. The implicit function theorem makes π\pi a local diffeomorphism. It is proper because ZZ is closed in P×MP\times M and MM is compact. On each component with nonempty fibers it is a finite covering. To see the local triviality, choose disjoint inverse branches about the points of a fiber. Properness allows shrinking the base neighborhood until no points remain outside these branches. Otherwise points outside the branches over a convergent sequence would have a limit in the original fiber.

For oriented M,NM,N, the signs are locally constant on ZZ. Transport around a loop can permute points, while preserving their signs. Thus the signed fiber is a locally constant family of finite sets with a map to {−1,1}\{-1,1\}. Forgetting transport and taking signed cardinality yields a locally constant integer. For P=M=N=S1P=M=N=S^1, let F(p,z)=zkF(p,z)=z^k and y(p)=py(p)=p with k>0k>0. The regular fiber has kk positive points. A single turn in PP cyclically permutes them. Its degree is kk, and the permutation is additional data of the family.

Parity and examples

Over F2\mathbb F_2, each closed nn-manifold has a canonical fundamental class even if it is nonorientable. The same excision argument gives the parity formula f∗[M]2=(#f−1(y) mod 2)[N]2f_*[M]_2=\bigl(\# f^{-1}(y)\bmod 2\bigr)[N]_2 for closed manifolds of dimension nn and connected NN. The mod-two degree is homotopy invariant. An unoriented compact one-manifold has an even number of boundary points, so the geometric comparison works with parity as well. If integral orientations exist, coefficient reduction sends the integer degree to the mod-two degree. An unbranched two-sheeted covering, for instance the antipodal covering S2→RP2S^2\to\mathbb{RP}^2, has mod-two degree zero, although the pullback of a point contains two geometrically distinct points.

For a different example let 0<a<π0<a<\pi and f(eit)=exp⁡(iasin⁡t)f(e^{it})=\exp(ia\sin t). The value 11 has two preimages, at t=0t=0 and t=πt=\pi, since asin⁡ta\sin t lies in (−π,π)(-\pi,\pi). Their local angular derivatives are aa and −a-a. The count is zero, in agreement with the homotopy exp⁡(isasin⁡t)\exp(isa\sin t) from the constant map to ff. By contrast the angle map z↦zkz\mapsto z^k has degree kk for every integer kk. For k<0k<0 all ∣k∣|k| regular preimages have negative local derivative. For k=0k=0 the map misses a value.

There is also a relative version for compact oriented manifolds with boundary. If f:(M,∂M)→(N,∂N)f:(M,\partial M)\to(N,\partial N) is smooth and NN is connected, use the map on Hn(−,∂−)H_n(-,\partial-) and its relative fundamental classes. A regular value in the interior of NN has no preimage in the boundary of MM. Excision gives the same local formula by mapping the relative fundamental classes into the local pairs. This version makes the behavior on the boundary part of the hypotheses. Without it the absolute top homology used for closed manifolds does not supply the desired comparison.

Intersections and the Euler class

The sign in degree theory compares two vector spaces of equal dimension. For intersections the relevant target is a normal space. A transverse preimage is oriented using its tangent and normal exact sequence. The Thom class expresses the same comparison in cohomology, and it gives a formula when the intersection has larger dimension than transversality predicts.

Oriented transverse pullbacks

Work with smooth manifolds without boundary in this chapter. Let A↪WA\hookrightarrow W be a closed embedded submanifold of codimension dd with oriented normal bundle νA/W\nu_{A/W}. Let f:B→Wf:B\to W be transverse to AA and put Z=f−1(A)Z=f^{-1}(A). The normal quotient of DfD f is surjective on ZZ and the transverse preimage theorem yields the diagram image It is a pullback of smooth manifolds. Smoothness of a factorization through ZZ follows from the defining local submersion coordinates, so the square has the same universal property on all smooth test manifolds. The differential identifies the normal bundle of ZZ in BB with the pullback of νA/W\nu_{A/W}, by the exact sequence 0⟶TZ⟶TB∣Z⟶(f∣Z)∗νA/W⟶0.0\longrightarrow TZ\longrightarrow TB|_Z \longrightarrow (f|_Z)^*\nu_{A/W}\longrightarrow 0 . If BB is oriented, we orient TZTZ so that an oriented basis of TZTZ, followed by lifts of a positive normal basis, gives the orientation on BB. When dim⁡B=d\dim B=d, each point of ZZ is oriented by the determinant sign of the normal differential. For compact BB the intersection number is the sum of these signs. Taking AA to be a point in an oriented target recovers the regular-value formula of Chapter [ch:degree].

If ftf_t is a homotopy between transverse maps with compact source without boundary, a transverse replacement relative to end collars has oriented preimage with boundary the two end intersections, taken with opposite signs. The preimage is compact by closedness of AA. This proves invariance of the intersection number and gives an oriented bordism in higher dimension. The tangent sequence provides the sign convention throughout. Specifying normal orientations also shows why the order of intersection can matter. Interchanging two independent oriented normal blocks of ranks aa and bb changes their combined orientation by (−1)ab(-1)^{ab}.

Thom classes and zero sets

For an oriented rank-dd real vector bundle E→BE\to B with zero section zz, the Thom class is the class UE∈Hd(E,E∖z(B);Z)U_E\in H^d(E,E\setminus z(B);\mathbb Z) which restricts on each fiber to the chosen orientation generator. Fiber restriction and the Thom isomorphism give its uniqueness. One construction glues the local generator on each oriented product chart, using the trivializations and the orientation compatibility in the Thom isomorphism [26, §9]. Its pullback along zz, after forgetting the relative condition, is the Euler class e(E)∈Hd(B;Z)e(E)\in H^d(B;\mathbb Z).

Let ss be a section transverse to zz and let Z=s−1(0)Z=s^{-1}(0). The class s∗UEs^*U_E belongs to Hd(B,B∖Z;Z)H^d(B,B\setminus Z;\mathbb Z). Its absolute image is e(E)e(E), by the homotopy t↦tst\mapsto ts of sections after forgetting the complements. In a local bundle chart the derivative of the coordinate vector of ss identifies νZ/B\nu_{Z/B} with E∣ZE|_Z. This identification preserves the induced normal orientation. The Thom class s∗UEs^*U_E is consequently the normal Thom class of ZZ. For compact oriented BB with dim⁡B=d\dim B=d, evaluation on its fundamental class gives ⟨e(E),[B]⟩=∑x∈Zsgn⁡det⁡(Dxverts).\langle e(E),[B]\rangle =\sum_{x\in Z}\operatorname{sgn}\det(D_x^{\mathrm{vert}}s). Indeed, excise to disjoint oriented coordinate balls and use the local fiber generators as in the degree proof. This also proves independence of the signed zero count from the transverse section. If ss is nowhere zero, the pullback of the relative class is zero, so e(E)e(E) vanishes. The Euler class detects a necessary condition for a nonvanishing section.

More generally, an oriented closed embedding j:Z↪Bj:Z\hookrightarrow B of codimension kk has a cohomological pushforward j!:Hr(Z)→Hr+k(B)j_!:H^r(Z)\to H^{r+k}(B) constructed by placing the orientation Thom class first in the cup product with the pulled-back cohomology class. Compute this in a tubular neighborhood, excise to the relative pair in BB, and forget the complement. We write PD⁡W[A]\operatorname{PD}_W[A] for the class obtained by applying this operation to 11 for A↪WA\hookrightarrow W. These definitions give f∗PD⁡W[A]=j!1f^*\operatorname{PD}_W[A]=j_!1 for the transverse square. On a normal fiber the pullback is an orientation isomorphism, which identifies the corresponding Thom classes.

The excess bundle in a clean intersection

Let i:B↪Wi:B\hookrightarrow W and A↪WA\hookrightarrow W be closed embedded submanifolds with a clean intersection. This means that Z=A∩BZ=A\cap B is a submanifold and TzZ=TzA∩TzBT_zZ=T_zA\cap T_z B for every z∈Zz\in Z. For simplicity assume the normal ranks are constant. Clean intersection makes the normal derivative injective, giving 0⟶νZ/B⟶νA/W∣Z⟶Eexc⟶0.0\longrightarrow \nu_{Z/B} \longrightarrow \nu_{A/W}|_Z \longrightarrow E_{\mathrm{exc}} \longrightarrow 0 . The excess bundle EexcE_{\mathrm{exc}} measures the normal directions of AA that are absent from the image of TzBT_z B. Suppose that νA/W\nu_{A/W} and νZ/B\nu_{Z/B} are oriented and orient EexcE_{\mathrm{exc}} by the displayed order of the sequence.

Theorem 1.1 (Clean intersection formula). Under these hypotheses, i∗PD⁡W[A]=j!e(Eexc)in Hd(B;Z),j:Z↪B.i^*\operatorname{PD}_W[A]=j_!e(E_{\mathrm{exc}}) \quad\text{in }H^d(B;\mathbb Z),\qquad j:Z\hookrightarrow B .

Proof. We compute the pullback near ZZ in tubular neighborhoods. A neighborhood of ZZ in BB is represented by normal vectors ξ\xi in νZ/B\nu_{Z/B}. Over the corresponding neighborhood of AA in WW the inclusion is a map into νA/W\nu_{A/W}. Write it as ξ↦(α(ξ),v(ξ))\xi\mapsto (\alpha(\xi),v(\xi)), where the first entry lies in AA and the second in the normal fiber over α(ξ)\alpha(\xi). Here v(0z)=0v(0_z)=0 and α(0z)=z\alpha(0_z)=z. Shrink the domain fiberwise if necessary. For t>0t>0 the rescaled map has value (α(tξ),t−1v(tξ)).\bigl(\alpha(t\xi),t^{-1}v(t\xi)\bigr). As t→0t\to 0, it extends smoothly to the normal derivative over ZZ, as is seen by integrating the first derivative in a bundle chart along the line from 0z0_z to ξ\xi. For t>0t>0 it is nonzero off the zero section since i−1(A)=Zi^{-1}(A)=Z. At t=0t=0 it is nonzero there since the normal derivative is injective. Thus the relative Thom pullback is computed by the linear injection of normal bundles.

Choose a splitting so that the target normal bundle over ZZ becomes νZ/B⊕Eexc\nu_{Z/B}\oplus E_{\mathrm{exc}}. The Thom class of the sum is the product of the two Thom classes in the displayed order. This follows by restricting to each fiber, where the product is the orientation generator of the oriented direct sum. Pulling back by the inclusion as the first summand leaves its Thom class and pulls back the second class by its zero section. The latter pullback is e(Eexc)e(E_{\mathrm{exc}}). Extending the resulting relative class to BB and forgetting the complement gives precisely j!e(Eexc)j_!e(E_{\mathrm{exc}}). ◻

An ordinary clean pullback displays the manifold ZZ, but its tangent sequence also contains the cokernel bundle. Transversality makes that cokernel vanish and recovers the rank-zero Thom formula. A tangent-complex description keeps it in a second degree, as developed later for derived intersections. The formula above identifies a classical contribution of that bundle without replacing the clean intersection by a finite set.

Self-intersection and tangent zeros

Take ii to be the inclusion of AA itself. The intersection inside B=AB=A is all of AA, and the excess bundle is νA/W\nu_{A/W}. Thus restricting PD⁡W[A]\operatorname{PD}_W[A] to AA gives the Euler class of the normal bundle. In CP2\mathbb{CP}^2, let AA be the projective line z0=0z_0=0. The line z1=0z_1=0 is isotopic to AA, through the rotations of the first two homogeneous coordinates. The two lines meet in one point [0:0:1][0:0:1]. In the affine chart at that point they are complementary complex coordinate lines. Complex linear identifications preserve the underlying real orientation, so the intersection sign is positive. The normal Euler number of AA is therefore 11. The self-intersection can be computed by a transverse motion even though the original pullback is a two-dimensional manifold.

When E=TBE=TB for a closed oriented manifold, a vector field with nondegenerate zeros computes the tangent Euler number by the same formula. Take a Morse function and a metric, using its gradient as the section. At a critical point of index λ\lambda, the derivative of the gradient is the inverse metric composed with the Hessian, so its determinant sign is (−1)λ(-1)^\lambda. Each corresponding Morse handle contributes that same sign to the Euler characteristic by Chapter [ch:morse]. Thus ⟨e(TB),[B]⟩=χ(B)\langle e(TB),[B]\rangle=\chi(B), the Poincaré–Hopf equality in the transverse setting [15, Ch. 5]. For example, a height function of S2S^2 has one minimum and one maximum. Both indices are even, and the two contributions give 22.

Morse functions and handle filtrations

The sublevel sets of a smooth real-valued function form an increasing family of spaces. A regular band contributes a product, while an isolated nondegenerate critical point adds a handle. This turns the real-valued function into a description of the topology of its source. We use the regular and critical level arguments of differential topology [15, Ch. 6] to describe the filtration and its homotopy pushouts explicitly.

The Hessian and the local form

Let f:M→Rf:M\to\mathbb R be smooth and let pp satisfy dfp=0df_p=0. Second partial derivatives in a chart define a symmetric bilinear form on TpMT_pM. Under a change of coordinates the Hessian is conjugated by the Jacobian, since the term involving the second derivatives of the coordinate change is multiplied by dfp=0df_p=0. A critical point is nondegenerate if this form is nonsingular, and its index is the dimension of a maximal negative definite subspace. A Morse function has only nondegenerate critical points. They are isolated by the inverse function theorem applied to the coordinate components of dfdf. For a compact source the critical set is closed and therefore finite.

Theorem 1.1 (Morse local form). If pp has index λ\lambda in an nn-manifold, there are local coordinates (u,v)(u,v) centered at pp, of dimensions λ\lambda and n−λn-\lambda, such that f(u,v)=f(p)−∥u∥2+∥v∥2.f(u,v)=f(p)-\|u\|^2+\|v\|^2 .

Proof. In coordinates xx near p=0p=0, Taylor’s integral formula writes the difference as xTA(x)xx^{\mathsf T}A(x)x with A(x)=∫01(1−t) D2f(tx) dt.A(x)=\int_0^1(1-t)\,D^2 f(tx)\,dt . This is a smooth symmetric matrix and A(0)A(0) is nonsingular. A fixed linear change of basis makes A(0)A(0) diagonal with entries −1-1 and 11. In a neighborhood of 00 all successive pivots in the symmetric elimination of A(x)A(x) are nonzero. Eliminating one row and column at a time is the completion of a square, using only division by the current nonzero pivot. The signs of the pivots are locally constant. Taking square roots of their absolute values gives a smooth factorization A(x)=B(x)TDB(x)A(x)=B(x)^{\mathsf T} D B(x), where DD is the same diagonal sign matrix. The map x↦B(x)xx\mapsto B(x)x has invertible derivative B(0)B(0) at 00. Its components are the required local coordinates by the inverse function theorem. ◻

The index is stable under small perturbations of the Hessian through nonsingular forms. If fsf_s is a smooth family and fs0f_{s_0} has a nondegenerate critical point pp, the equation dfs(x)=0df_s(x)=0 near (s0,p)(s_0,p) has a unique smooth local solution p(s)p(s) by the implicit function theorem. The index is constant on this branch. This describes how critical points move through regular parameter values. When the Hessian becomes singular, the equation no longer has to be locally the graph of a smooth section.

Regular bands

Write Ma={x:f(x)≤a}M^a=\{x:f(x)\le a\}. Suppose a closed band f−1([a,b])f^{-1}([a,b]) is compact and contains no critical point. Fix a Riemannian metric and use the vector field V=−∇f/∥∇f∥2V=-\nabla f/\|\nabla f\|^2 on a neighborhood of the band. Along its flow ff decreases at unit rate. A flow line starting at level ss therefore reaches level aa in time s−as-a. Existence for this duration follows by compactness, since the flow stays in the band until it reaches the designated end. Flowing backward gives the diffeomorphism of the band with f−1(a)×[a,b]f^{-1}(a)\times[a,b]. This proves that both boundaries are regular copies of the same level. Flowing each point of Mb∖MaM^b\setminus M^a for time t(f(x)−a)t(f(x)-a), for 0≤t≤10\le t\le1, and fixing MaM^a, gives a deformation retraction Mb→MaM^b\to M^a. The formula is continuous at level aa because the travel time tends to zero.

This argument requires compactness of the band, or another condition giving the same flow existence. A nonproper function on a noncompact manifold can change level behavior at infinity with no critical point. For example, f:R→Rf:\mathbb R\to\mathbb R given by f(x)=exf(x)=e^x has empty levels at nonpositive values and single-point levels at all positive values. Its derivative is nowhere zero, and a band straddling zero is noncompact.

Crossing one critical level

Assume MM is closed and pp is the only critical point in a band around c=f(p)c=f(p). Set a=c−εa=c-\varepsilon, b=c+εb=c+\varepsilon with ε>0\varepsilon>0 small. Write DjD^j for the closed unit disk in Rj\mathbb R^j, including D0={0}D^0=\{0\} and ∂D0=∅\partial D^0=\varnothing. The topological handle of index λ\lambda is Dλ×Dn−λD^\lambda\times D^{n-\lambda} attached along ∂Dλ×Dn−λ\partial D^\lambda\times D^{n-\lambda}.

Theorem 1.2 (Sublevel attachment). The upper sublevel MbM^b has the homotopy type, relative to MaM^a, of MaM^a with one index-λ\lambda handle attached. Consequently there is a homotopy pushout of spaces image

Proof. Work in the chart of Theorem 1.1. Choose a metric equal to the Euclidean metric there near pp. Choose r>0r>0 with r2<εr^2<\varepsilon and both small enough that H={(u,v):∥v∥≤r,∥u∥≤ε+∥v∥2}H=\{(u,v):\|v\|\le r,\quad \|u\|\le\sqrt{\varepsilon+\|v\|^2}\} is contained in this metric neighborhood. Dividing uu by ε+∥v∥2\sqrt{\varepsilon+\|v\|^2} and rescaling vv gives the disk product. The inequality f≥af\ge a on HH is an equality precisely on the attaching face. The inequality f≤c+r2<bf\le c+r^2<b puts the handle inside the upper sublevel. Hence Ma∩HM^a\cap H is the attaching face.

Follow negative gradient flow outside Ma∪HM^a\cup H. There are no critical points on the compact closure of the region still to be added, since pp is interior to HH relative to MbM^b. The decrease rate of ff there is bounded below in absolute value by a positive constant. The flow reaches Ma∪HM^a\cup H in finite time. At its level-aa boundary the flow crosses strictly toward the lower sublevel. At the other face of HH, ∥v∥=r\|v\|=r, the local gradient equation is v′=−2vv'=-2v, so the flow crosses strictly inward there as well. These inequalities give continuous first-entry time, using the first of the two transverse crossings where the faces meet. Defining time zero inside the union yields a deformation retraction of MbM^b onto Ma∪HM^a\cup H by running a fixed fraction of this time. If a disk factor has dimension zero, its empty boundary requires no crossing argument.

The attaching face and its inclusions admit collars in their respective spaces when the face is nonempty. The resulting pushout computes the homotopy pushout. Projection from the disk product to DλD^\lambda is an equivalence of pairs with the corresponding boundary. Using the center v=0v=0 for the attaching map therefore gives the displayed square. ◻

The smooth handle attachment statement uses a smoothing of the corners of such a model. It describes the boundary surgery as well as the homotopy type [25, §3]. For the homotopy filtration, each interval between critical values can be replaced by a single arrow. If the critical values coincide, a small bump constant near each critical point separates them. The transition regions of these bumps lie in a compact set where ∥df∥\|df\| has a positive lower bound, so sufficiently small bumps create no new critical points. The filtration then starts at the empty set and successively attaches a cell of each critical index. The relative homology of each step is free of rank one in the critical index and zero in the other degrees. Additivity in the long exact sequences of the pairs gives χ(M)=∑p∈Crit⁡(f)(−1)ind⁡(p).\chi(M)=\sum_{p\in\operatorname{Crit}(f)}(-1)^{\operatorname{ind}(p)} .

On the torus use f(s,t)=cos⁡s+2cos⁡tf(s,t)=\cos s+2\cos t. The four critical points occur where s,ts,t are 00 or π\pi. The Hessian is diagonal with entries −cos⁡s-\cos s and −2cos⁡t-2\cos t. The critical values in order are −3,−1,1,3-3,-1,1,3, with indices 0,1,1,20,1,1,2. The filtration first creates a component, then adds two one-dimensional cells, and finally fills a two-dimensional cell. It gives χ(T2)=1−2+1=0\chi(T^2)=1-2+1=0.

Indices in a family do not determine handle transport. For θ∈R/(2πZ)\theta\in\mathbb R/(2\pi\mathbb Z), let R(ϕ)R(\phi) be the rotation of R2\mathbb R^2 by ϕ\phi and define the quadratic form xTR(θ/2)(−1001)R(θ/2)Tx.x^{\mathsf T}R(\theta/2) \begin{pmatrix}-1&0\\0&1\end{pmatrix} R(\theta/2)^{\mathsf T}x . It is well defined on the circle because adding 2π2\pi changes both rotation matrices by a sign. The origin is a nondegenerate critical point of index one in each fiber. Its negative eigenspace is spanned locally by R(θ/2)(1,0)R(\theta/2)(1,0), which reverses sign on a full turn. The negative spaces form a Möbius line, so the local orientations of the index-one disks reverse under transport. The critical point and its index are constant throughout, while the relative homology generator of the local handle has sign monodromy.

Trajectory spaces and Morse homology

A handle filtration determines the ranks of the Morse chain groups. To recover the boundary maps one needs information about the attaching maps. Negative gradient trajectories provide that information through compact spaces whose boundary strata consist of successive trajectories. Counting zero-dimensional spaces produces a chain complex. Retaining all the spaces instead produces a category enriched in spaces.

Flow lines and the boundary operator

Let MM be closed, f:M→Rf:M\to\mathbb R Morse, and gg a metric. The negative gradient is complete by compactness. For a critical point pp, let Wu(p)W^u(p) and Ws(p)W^s(p) be the unstable and stable manifolds. They consist of points whose flow converges to pp at time −∞-\infty and +∞+\infty, respectively. The stable manifold theorem gives their dimensions as ∣p∣=ind⁡(p)|p|=\operatorname{ind}(p) and dim⁡M−∣p∣\dim M-|p|. These dimensions come from the positive and negative eigenspaces of the linearized negative gradient. Assume the pair is Morse–Smale, meaning all Wu(p)W^u(p) and Ws(q)W^s(q) intersect transversely. For p≠qp\ne q, the intersection, when nonempty, consists of nonstationary trajectories and has dimension ∣p∣−∣q∣|p|-|q|. Quotienting by time translation gives the moduli space M(p,q)\mathcal M(p,q). When the space is nonempty, f(p)>f(q)f(p)>f(q). Every trajectory crosses a fixed regular level between the two values exactly once, so one can realize the quotient as a smooth level slice. Its dimension is ∣p∣−∣q∣−1|p|-|q|-1.

The Morse compactification theorem adds broken trajectories to these spaces. The stratum with a sequence of breaks through p1,…,pkp_1,\ldots,p_k is M(p,p1)×M(p1,p2)×⋯×M(pk,q).\mathcal M(p,p_1)\times\mathcal M(p_1,p_2) \times\cdots\times\mathcal M(p_k,q). In particular, under a metric Euclidean in Morse coordinates near the critical points, the spaces have smooth corner models with these strata in codimension kk [32, Thms. 2.3 and 2.6]. To see the source of the boundary, cut trajectories at regular levels between the critical values. The compact slices control where segments can travel. If the time between two slices tends to infinity, a limiting segment can approach an intervening critical point. The compactness theorem shows that only finitely many such breaks occur and gives the resulting limiting tuple. The gluing theorem supplies a nonnegative collar parameter for each break, with zero recording the broken tuple. The proofs below use both compactness and this description of the boundary. A set of broken trajectories alone does not establish the existence of the collars.

Let Cj(f)C_j(f) be the F2\mathbb F_2-vector space on critical points of index jj. For a difference of indices equal to one, each nonconstant segment drops the index by at least one, so the compactification has no breaks. Thus M(p,q)\mathcal M(p,q) is compact and zero-dimensional, hence finite. Set ∂p=∑∣q∣=∣p∣−1#2M(p,q) q,\partial p= \sum_{|q|=|p|-1}\#_2\mathcal M(p,q)\,q , where #2\#_2 denotes parity. For ∣p∣−∣r∣=2|p|-|r|=2, M‾(p,r)\overline{\mathcal M}(p,r) is a compact one-manifold whose boundary is the disjoint union of M(p,q)×M(q,r)\mathcal M(p,q)\times\mathcal M(q,r) over ∣q∣=∣p∣−1|q|=|p|-1. The coefficient of rr in ∂2p\partial^2p is the parity of that boundary. It is zero since each interval has two endpoints. Over the integers, compatible orientations of unstable manifolds give the trajectory signs, and the oriented-boundary formula yields the signed version [28, Chs. 3–4]. The Morse homology theorem identifies the resulting homology with singular homology. Its comparison uses the generator of a handle pair from Chapter [ch:morse] and identifies boundary coefficients by flowing an attaching sphere to a transverse normal slice at the next lower index [28, Ch. 4].

Continuation and homotopies of continuations

Let (fs,gs)(f_s,g_s), s∈Rs\in\mathbb R, interpolate between two Morse–Smale pairs and be constant for ∣s∣|s| sufficiently large. A continuation trajectory satisfies x˙(s)=−∇gsfs(x(s)),lim⁡s→−∞x(s)=p,lim⁡s→+∞x(s)=q.\dot x(s)=-\nabla_{g_s}f_s(x(s)),\qquad \lim_{s\to-\infty}x(s)=p,\quad \lim_{s\to+\infty}x(s)=q . The fixed transition interval distinguishes a time coordinate. In particular one does not divide out an R\mathbb R-translation action for this equation. For regular interpolation data, the dimension is ∣p∣−∣q∣|p|-|q|. Indeed, transfer Wu(p)W^u(p) by the time-dependent flow through the transition interval and intersect it transversely with the terminal Ws(q)W^s(q). Counting the zero-dimensional continuation spaces gives a degree-zero map Φ:C∗(f−)→C∗(f+)\Phi:C_*(f_-)\to C_*(f_+). Compactified one-dimensional spaces have breaking at the two stationary ends. The two types of boundary points give the two terms in ∂+Φ+Φ∂−=0\partial_+\Phi+\Phi\partial_-=0 over F2\mathbb F_2.

A regular one-parameter comparison of interpolations has the additional parameter direction. For an index difference of −1-1, counting it gives a degree-one operator KK. In index difference zero, its compactified one-manifold has boundary from the two parameter ends and from breaking at either stationary end. The boundary parity therefore gives Φ1+Φ0=∂+K+K∂−.\Phi_1+\Phi_0=\partial_+K+K\partial_- . Gluing along a long intermediate interval compares composite continuation with a single interpolation. Reversing the interpolation and comparing the concatenation to the constant interpolation gives inverse maps on homology [28, Ch. 4]. These statements take regularity and compactified gluing for the continuation equations as inputs. The identities explain which boundaries are needed for invariance and why a comparable chain map is specified only up to chain homotopy.

The enriched flow category

The objects of the flow category Cf\mathcal C_f are critical points. Its nonidentity morphism spaces are M‾(p,q)\overline{\mathcal M}(p,q). Concatenating broken tuples is continuous and associative. In particular a codimension-two corner for a three-stage trajectory gives the same ordered tuple under either grouping. The following diagram records the two face inclusions. We abbreviate a compactified space by CpqC_{pq}. image The nerve is a simplicial space, with level nn given by composable strings of nn morphisms. Faces compose or forget arrows and degeneracies insert identities. Applying the singular complex to each mapping space gives a simplicial category, whose homotopy-coherent nerve is an ∞\infty-category. A path within one trajectory space now gives a higher comparison of arrows. The classifying-space result of Cohen, Jones, and Segal recovers the homotopy type of MM from the realization of the topological nerve in this Morse–Smale case [8, Thms. 2.6 and 3.1]. Composition here uses recorded break data. The continuation calculation above explains the one-dimensional coherence needed at the chain level. An analogous construction of continuation on all the enriched data would require compatible multiparameter gluing.

Two explicit flow categories

On the circle take f(eit)=cos⁡tf(e^{it})=\cos t and the standard metric. The flow equation is t˙=sin⁡t\dot t=\sin t. The maximum at 00 and the minimum at π\pi are connected by two unparametrized trajectories, one in each open semicircle. There are no intermediate critical points at which to break. The mod-two differential vanishes because the trajectory count is two. The realization of the flow category has two vertices and two edges joining them, which gives a circle directly.

On the round sphere take the height function. The maximum has index two, the minimum has index zero, and the nonstationary trajectories are the meridians, parametrized by their intersections with the equator. There are no broken trajectories. The morphism space from maximum to minimum is therefore S1S^1. The nerve realization is obtained from S1×[0,1]S^1\times[0,1] by collapsing its two end circles separately to the vertices. This is the suspension S2S^2. If we retain only the path components of the morphism spaces, the ordinary category has one nonidentity arrow and its realization is an interval. Thus taking components of the mapping spaces can lose the topology of the manifold, even when all the objects and their indices are retained.

The product torus gives an example with actual breaking. Use the flat metric and f(s,t)=cos⁡s+2cos⁡tf(s,t)=\cos s+2\cos t. The negative gradient preserves the two coordinate equations, and the stable and unstable manifolds are products of those on the two circles. Transversality holds because it holds in each factor. For index difference one, exactly one coordinate moves, and the two semicircles in that factor give parity zero. Thus the mod-two complex has one generator in each of degrees zero and two, and two in degree one, with zero differential.

For a trajectory from the maximum to the minimum both coordinates move. Choosing one of the two semicircles in each factor leaves a time shift for each of the two one-dimensional motions. Modulo common translation the difference of shifts is a real coordinate. Each of the four choices therefore gives a real line in the unbroken moduli space. In the two limits of that coordinate, one coordinate transition occurs before the other, giving the two orders of a broken trajectory. The four closed intervals in the compactification exhibit eight boundary points. The two possible intermediate critical points account for four each.

Geometric bordism and the Thom construction

A closed manifold is a possible boundary, and a bordism asks whether two closed manifolds can occur as the ends of a single manifold. The question keeps track of tangent or normal structures when these are part of the geometry. Its stable answer is a homotopy group. The passage from a manifold to that group uses the same two constructions that govern intersections: a tubular neighborhood and a transverse inverse image.

Bordisms and their elementary groups

A closed dd-manifold is compact and without boundary. An unoriented bordism from M0M_0 to M1M_1 is a compact (d+1)(d+1)-manifold WW with designated ends ∂W≃M0⨿M1\partial W\simeq M_0\amalg M_1. In the oriented case the boundary orientation identifies ∂W\partial W with (−M0)⨿M1(-M_0)\amalg M_1. The two conventions use an outward normal first in the orientation of a boundary. One can choose collars, as in Chapter [ch:tubular], to glue successive bordisms smoothly. The cylinder proves reflexivity, and reversing the roles of the ends proves symmetry, using −W-W in the oriented setting. Disjoint union gives an abelian group of equivalence classes. The cylinder also gives [M]+[−M]=0[M]+[-M]=0 in the oriented group, and 2[M]=02[M]=0 in the unoriented group. Denote the respective groups by ΩdSO\Omega_d^{SO} and ΩdO\Omega_d^O.

Example 1.1 (Dimension zero). An oriented closed zero-manifold is a finite list of positive and negative points, with multiplicities. Each oriented interval has one endpoint of each sign, and a circle has no endpoints. The signed count consequently gives an invariant in Z\mathbb Z. A finite collection with count zero bounds by pairing the positive and negative points and taking one oriented interval for each pair. Thus Ω0SO≃Z\Omega_0^{SO}\simeq\mathbb Z. Dropping the orientations retains the parity of the number of points, and the same interval construction proves Ω0O≃Z/2\Omega_0^O\simeq\mathbb Z/2.

Products endow the direct sum of the degree groups with graded multiplication. A bordism WW for M0,M1M_0,M_1 produces the bordism W×NW\times N when NN is closed. In the oriented case the factor order contributes the sign (−1)de(-1)^{de} on swapping oriented manifolds of dimensions dd and ee. A related construction equips a manifold with a continuous map to a space XX, with both this map and its structure required to extend across a bordism. This is a geometric way for a theory of manifolds to run over variable spaces. We first consider the coefficient groups where XX is a point.

Example 1.2 (An unoriented surface that does not bound). For a compact smooth 33-manifold WW, the double of WW is a closed 33-manifold. Its Euler characteristic is zero, by Poincaré duality over F2\mathbb F_2. Additivity of Euler characteristic gives χ(∂W)=2χ(W)\chi(\partial W)=2\chi(W), so no closed surface of odd Euler characteristic can be its whole boundary. In particular RP2\mathbb{RP}^2 represents a nonzero element of Ω2O\Omega_2^O. This computation uses duality only as a vanishing statement on the double. It applies regardless of orientability.

Collapsing a normal neighborhood

Embed MdM^d in Rd+k\mathbb R^{d+k} and let ν\nu be its normal bundle of rank kk. For the Thom space write Th⁡(ν)=D(ν)/S(ν)\operatorname{Th}(\nu)=D(\nu)/S(\nu), using a metric on ν\nu. A tubular identification of D(ν)D(\nu) with a closed neighborhood of MM in the one-point compactification Sd+kS^{d+k} gives a pointed collapse. The classifying map of ν\nu induces a bundle map to the universal rank-kk bundle γk\gamma_k on BO(k)BO(k). With compatible metrics it induces the second arrow of image An oriented MM induces a normal orientation using the ambient orientation and the order TM⊕νTM\oplus\nu. In that case use the universal oriented bundle over BSO(k)BSO(k).

The collapse sends the complement of the open normal disk bundle to the base point. For a smaller radius the annulus between the two boundaries can be absorbed by a radial change of coordinates. An isotopy of tubular neighborhoods gives a homotopy of collapse maps by applying this construction over the interval. Increasing the ambient dimension by one replaces the normal bundle by ν⊕R\nu\oplus\mathbb R. Under this operation the collapse suspends: image The right arrow comes from adjoining a trivial line to each normal plane. Choices of embeddings disappear after stabilization. Indeed, for embeddings i0,i1i_0,i_1 placed in independent coordinate summands, the family x↦((1−t)i0(x),ti1(x))x\mapsto ((1-t)i_0(x),t i_1(x)) stays embedded for 0≤t≤10\le t\le 1. Initial embeddings may therefore be compared using a smooth cylinder in a common larger ambient space, with orientation-preserving rotations of ambient coordinates if an ordering of the normal directions has been specified.

Transversality gives the inverse

Theorem 1.3 (Stable Pontryagin–Thom correspondence). Collapse induces isomorphisms ΩdO≃πdMO,ΩdSO≃πdMSO.\Omega_d^O\simeq\pi_d MO, \qquad \Omega_d^{SO}\simeq\pi_d MSO . The kkth spaces in these Thom spectra are the Thom spaces of the universal rank-kk bundles over BO(k)BO(k) and BSO(k)BSO(k). The stable group on the right is the colimit of the groups in degrees d+kd+k under suspension and the Thom structure maps.

Proof. We give the normal-bundle argument [15, Ch. 7][9, Lectures 2–5]. Use finite Grassmannians for the bundle classifying maps, enlarging their ambient dimension as needed. Representatives with compact source, and their homotopies, can be taken at finite stages by cellular approximation. This permits the usual smooth transversality argument near the zero section. The disk bundle away from the sphere identification is a smooth total space, with the zero section a closed submanifold of codimension kk.

For a pointed map a:Sd+k→Th⁡(γk)a:S^{d+k}\to\operatorname{Th}(\gamma_k) represented at such a stage, approximate it smoothly near the inverse image of the zero section and make it transverse there. One may leave the map fixed near the base point of the source. To see why local smoothing is enough, first choose a smaller disk bundle about the zero section. The complement of its inverse image maps away from zero and transversality needs to be arranged only on a neighborhood of the smaller inverse image. An arbitrarily small perturbation there, kept equal to the original outside the larger inverse image, introduces no zeros in the intervening closed annulus.

The transverse inverse image MM is a closed dd-manifold in the finite chart Rd+k\mathbb R^{d+k} of the source. The differential identifies its normal bundle with the pullback of γk\gamma_k along a∣Ma|_M. This is just the isomorphism from the normal space to the quotient by the tangent space of the zero section. In the oriented version this gives its orientation by the stated convention. The same argument applied to a homotopy, made transverse relative to its endpoints, yields a bordism in Sd+k×[0,1]S^{d+k}\times[0,1]. It has the required collars after the homotopy is made constant in the parameter near its ends. Suspension appends a trivial direction to the normal bundle of the inverse image.

Conversely, collapse along an embedded manifold is transverse near the zero section and recovers that manifold with its normal data. For a transverse aa, shrink to a tubular neighborhood of its inverse image by radially pushing the complement of a smaller bundle neighborhood to the base point. The resulting map is homotopic to normal collapse: on a smaller tube one compares the normal-coordinate map with its derivative along MM. In local bundle coordinates this uses the homotopy given by the rescaled normal component t−1F(x,tv)t^{-1}F(x,tv), which extends to the derivative at t=0t=0. Parallel transport in the target bundle identifies the nearby fibers with the fibers over a∣Ma|_M for this calculation. The derivative is a normal isomorphism. Compactness gives a common tube on which the comparison is defined, and a radial collapse sends its exterior to the base point. This gives inverse operations on stable classes.

An abstract bordism has a neat embedding into a large Rd+k×[0,1]\mathbb R^{d+k}\times[0,1] with product form near its ends. This can be obtained by prescribing the collar coordinate as the interval coordinate near the boundary and combining it with a relative Euclidean embedding in the remaining coordinates. After stabilization the end embeddings can be adjusted by the embedding isotopies above. Collapsing its normal neighborhood then gives the required homotopy between end collapses. Finally, disjoint union corresponds to addition by placing the two tubes in disjoint balls and pinching the complement to the base point. ◻

What stabilization remembers

A normal framing is a specified ordered trivialization of the normal bundle, stabilized by adding the standard last coordinate. It gives a collapse to SkS^k, since the classifying map of the trivialized bundle factors through the sphere model of the framed theory. The same preimage operation recovers a framing from a transverse inverse image of a regular point of SkS^k. Stable framed bordism therefore has groups πdS=lim→⁡kπd+k(Sk)\pi_d^S=\varinjlim_k\pi_{d+k}(S^k). Forgetting a framing and retaining its induced orientation gives the map from the sphere spectrum to MSOMSO on coefficient groups. Distinct framings on the same underlying manifold can give distinct classes. In dimension zero, the relative orientation of an ordered normal frame and the ambient frame contributes the sign of the point, recovering the stable degree.

The theorem classifies manifolds after passing to bordism classes. A space of embeddings and collars also retains paths, homotopies of paths, and automorphisms of individual geometric representatives. The collapse construction runs in families as well as on points. The next chapter retains composable geometric bordisms before taking a classifying space and distinguishes the Thom spectrum in that setting from the normal Thom spectrum used here.

Bordism as a higher category

Collapsing a manifold records its bordism class. Bordisms also form a system of composable arrows, and many of their geometric constructions compare arrows rather than objects. At the next level, a family of diffeomorphisms compares such comparisons. A categorical model must specify which differences are retained at each level.

Gluing and its space of choices

Fix dd. At a real level aa, an object is a closed (d−1)(d-1)-manifold embedded in {a}×RN\{a\}\times\mathbb R^N. An arrow to an object at level b>ab>a is a compact dd-submanifold W⊂[a,b]×RNW\subset[a,b]\times\mathbb R^N whose only boundary consists of the designated objects and whose embedding is a product in a neighborhood of each end. At an interface level bb, two consecutive arrows agree on the object and their product neighborhoods glue. At the level of these embedded representatives the composition is union, so equality of the two composites of three consecutive arrows follows from associativity of union. This defines the positive-length part of an embedded cobordism category CdC_d. One includes identities when forming its category or uses the corresponding model with nonidentity arrows and a semisimplicial classifying space.

Without fixed ambient levels one can glue two abstract bordisms by identifying boundary collars. The overlap chart on the union uses a signed coordinate on the two inward half collars. Its smooth structure is independent up to a diffeomorphism of collars, and repeated gluing keeps the same signed coordinate along each seam. Retaining the space of such gluings is useful for comparison with embedded families.

For a fixed abstract bordism WW with its collared ends, the embedding model has a space of embeddings of WW modulo its diffeomorphisms. A diffeomorphism acts freely on parametrized embeddings by precomposition. On a closed manifold, stable embeddings in increasing Euclidean dimensions form a weakly contractible space: a compact family at a finite stage can be deformed to one fixed embedding using an independent block, as in Chapter [ch:bordism]. For a sphere of parameters this deformation fills the sphere by contracting it. In these models one obtains classifying spaces of diffeomorphism groups by taking the homotopy quotient of the embedding space under reparametrization. The use of homotopy quotients retains families with nontrivial monodromy.

Example 1.1 (Parity and the classifying space). Taking the nerve and geometric realization of CdC_d gives a space BCdBC_d. An arrow becomes a path in this space and can be traversed in either direction. On connected components this identifies objects exactly up to unoriented bordism. To verify the claim in one direction, any abstract bordism can be embedded in a sufficiently large Euclidean slab with collared ends. For the other direction, paths in the space of object embeddings preserve their diffeomorphism type, and one-simplices preserve the bordism class. A finite zigzag with backward arrows gives a bordism after reversing the corresponding slabs. In particular π0BC1≃Z/2\pi_0 BC_1\simeq\mathbb Z/2 via parity of points. This calculation alone leaves the higher homotopy groups of BC1BC_1 undetermined.

The spectrum of an embedded cobordism category

The Pontryagin–Thom spectrum of the preceding chapter stabilizes the normal rank of individual closed manifolds. For the classifying space of a category with fixed dd-dimensional arrows one instead fixes the tangent rank and varies the complementary normal directions. Over Gr⁡d(Rd+n)\operatorname{Gr}_d(\mathbb R^{d+n}) let γd,n\gamma_{d,n} be the tautological dd-plane bundle and γd,n⊥\gamma_{d,n}^{\perp} its rank-nn orthogonal complement. The spectrum MTO(d)MTO(d) is the Thom spectrum of the virtual bundle −γd-\gamma_d on BO(d)BO(d). Its (d+n)(d+n)th space can be modeled by Th⁡(γd,n⊥)\operatorname{Th}(\gamma_{d,n}^{\perp}). Adding an ambient coordinate adds a trivial line to the complements along the inclusion of Grassmannians and gives the structure maps.

Theorem 1.2 (Galatius–Madsen–Tillmann–Weiss). The classifying space of the stable embedded unoriented dd-cobordism category is weakly equivalent to Ω∞−1MTO(d)\Omega^{\infty-1} MTO(d). For the category with tangent structures classified by a fibration θ:X→BO(d)\theta:X\to BO(d), the corresponding spectrum is the Thom spectrum of −θ∗γd-\theta^*\gamma_d.

Here the shifted infinite loop space is represented by the stabilization of the spaces Ωd+n−1Th⁡(γd,n⊥)\Omega^{d+n-1}\operatorname{Th}(\gamma_{d,n}^{\perp}). The theorem of [10], whose proof uses parametrized geometry, starts with collapse near embedded manifolds over their level coordinate. The tangent dd-plane specifies the Grassmannian coordinate of the collapse. The remaining directions supply a point in the complementary normal disk. Compatibility with a cut recovers categorical composition at the level of the classifying space. The theorem’s converse compares families of such local data to geometric bordisms, a stronger statement than the inverse-image computation on individual maps in Theorem [thm:pontryagin-thom]. The difference between the spectra is visible in their bundle ranks. In MOMO the universal normal bundle’s rank increases in the stabilization. In MTO(d)MTO(d) the positive tangent plane has fixed rank dd, and its negative virtual class is stabilized.

An oriented version uses BSO(d)→BO(d)BSO(d)\to BO(d). The structure on a (d−1)(d-1)-dimensional object is a lift of its tangent bundle with a distinguished added interval direction. This recovers the boundary orientation convention when gluing two oriented arrows. In dimension zero the positive-length category has only the empty manifold as an object type, and its arrows are finite point configurations in open slabs. The emptiness of the boundary in this case is consistent with the shift: MTO(0)MTO(0) is the sphere spectrum and its shifted infinite loop space is connected.

Corners, duals, and field theories

A fully extended bordism category starts at dimension zero. Its jj-morphisms for 0≤j≤d0\le j\le d are jj-dimensional bordisms with the relevant corners and collars. Geometrically a corner has several independent boundary coordinates, so that composition is gluing along a selected face. As in the ordinary bordism category, one retains diffeomorphisms and their homotopies above the geometric dimensions. With disjoint union this gives a symmetric monoidal (∞,d)(\infty,d)-category. In the dd-framed version Borddfr\mathrm{Bord}^{\mathrm{fr}}_d, a jj-manifold is equipped with a compatible trivialization of its tangent bundle with d−jd-j additional ordered directions. Along collars these added directions identify the boundary framing with the restriction of the framing in the next dimension.

A dual for an object AA consists of an object A∨A^\vee and coevaluation and evaluation arrows. They have composites that are equivalent to the respective identities. One of the two triangle identities has diagram image In dimension one the intervals with both ends outgoing or both ends incoming supply these maps for the two framings of a point. The triangle identities come from gluing intervals and rectifying the result as a collared interval. In higher dimensions the evaluation and coevaluation morphisms have adjoints and the units and counits have coherent adjointability data through the permitted levels. This is the geometric structure behind full dualizability. Full dualizability is defined in the target symmetric monoidal (∞,d)(\infty,d)-category using these successive levels of adjointability.

The framed cobordism hypothesis identifies the space of symmetric monoidal functors out of Borddfr\mathrm{Bord}^{\mathrm{fr}}_d with the maximal ∞\infty-groupoid of fully dualizable objects of the target, by evaluation at the framed point [21, §2.4]. The functor space here uses equivalences of functors. Other tangential structures require compatible symmetry data on a fully dualizable object. They do not result from taking the same target space and merely relabeling the point. This universal property refers to functors from the extended higher category. The classifying-space theorem refers to a groupoid completion of the embedded category with adjacent dimensions, so it answers a different question about the geometry of bordisms.

Example 1.3 (A trace from gluing). In dimension one take the target to be finite-dimensional vector spaces over a field kk with the tensor product. If the positively framed point is sent to VV, its dual point is sent to V∗V^*. The evaluation and coevaluation are the pairing and the tensor ∑iei⊗ei∗\sum_i e_i\otimes e_i^*. The tensor is independent of basis because it corresponds to the identity in End⁡(V)\operatorname{End}(V). Gluing the outgoing ends of the coevaluation to the incoming ends of the evaluation, with the symmetry exchanging the middle factors, makes a circle. Its image is multiplication by ∑iei∗(ei)=dim⁡kV\sum_i e_i^*(e_i)=\dim_k V viewed in kk. More generally, inserting an endomorphism on the strand gives its usual trace in the target. The circle evaluation is therefore a consequence of coherent gluing and the choice of dual.

Example 1.4 (A surface cut into pairs of pants). In dimension two the ordinary oriented bordism category has circles and their disjoint unions as objects. In a symmetric monoidal functor to finite-dimensional vector spaces, put AA for the vector space assigned to one circle. The twice-incoming pair of pants supplies a multiplication m:A⊗A→Am:A\otimes A\to A, and the outgoing disk is its unit. There are two ways to glue two pairs of pants to get three incoming circles and one outgoing circle. The resulting labeled surfaces are diffeomorphic relative to their boundary collars, so their images give associativity. A diffeomorphism interchanging the two incoming circles gives commutativity. Reversing the pants and the disk gives a comultiplication and a counit ϵ\epsilon.

A cylinder drawn as a cup can be cut into an outgoing disk and a one-incoming pair of pants. Its reverse gives the pairing (u,v)↦ϵ(m(u,v))(u,v)\mapsto\epsilon(m(u,v)). The cup and cap satisfy the cylinder identities, so this pairing is nondegenerate. The two ways to cut a surface from two circles to two circles into oppositely directed pairs of pants give the Frobenius equality between comultiplication after multiplication and multiplication on one side of a comultiplication. These statements are consequences of labeled gluing and the elementary classification of the relevant surfaces (Chapter [ch:surfaces]). They concern theories defined on circles and surfaces. A fully extended theory also has values on points and compatible additional gluing.

A geometric family also has a different source of higher data: paths in an embedding space act through ambient diffeomorphisms, with a stabilizer of the embedding. The next chapter constructs these lifts, then describes the distinction between a path of images and a chosen lift through the ambient transformations.

Isotopies and ambient transformations

An isotopy is a family of embeddings. An ambient isotopy is a family of diffeomorphisms of the target starting at its identity. The extension theorem turns the first type of family into the second by extending its time-dependent velocity. The resulting factorization has a whole space of choices, which can be understood as a stabilizer in a space of transformations.

Extending velocities

Theorem 1.1 (Isotopy extension with smooth parameters). Let AA be a closed smooth manifold, and let NN be a smooth manifold without boundary. Let BB be a compact smooth parameter manifold. Suppose e:B×[0,1]×A→Ne:B\times[0,1]\times A\to N is smooth and every eb,te_{b,t} is an embedding. There exists a smooth family of compactly supported ambient isotopies Φb,t\Phi_{b,t} of NN with Φb,0=1N,Φb,t∘eb,0=eb,t.\Phi_{b,0}=1_N,\qquad \Phi_{b,t}\circ e_{b,0}=e_{b,t}. The support can be taken in one compact subset of NN independent of b,tb,t. If eb,te_{b,t} is constant in tt for designated parameters bb, then Φb,t\Phi_{b,t} can be chosen equal to the identity for those parameters.

Proof. The velocity along eb,te_{b,t} is vb,t(eb,t(a))=∂te(b,t,a)v_{b,t}(e_{b,t}(a))=\partial_t e(b,t,a). This is well-defined because eb,te_{b,t} is injective. Choose a metric on NN. The normal spaces to the embedded slices form a bundle over B×[0,1]×AB\times[0,1]\times A. The normal exponential map with b,tb,t unchanged is a tubular coordinate map for a uniform sufficiently small radius. In more detail, it has invertible derivative along the zero section in the manifold with parameters. Local inverse maps give neighborhoods near that section. If there were intersections between distinct normal disks at arbitrarily small radii, compactness would give a limiting pair of points over the same b,tb,t. The injectivity of eb,te_{b,t} forces the pair to coincide, and a local inverse then rules out the intersections for the subsequence. This gives the required uniform radius.

For a point in the tube at parameters b,tb,t, transport the slice velocity from the foot point along the normal geodesic. Multiply by a cutoff in the normal radius that is one near the zero section and zero near the boundary of the tube. Extend by zero on the complement. The resulting vector field Vb,tV_{b,t} on NN depends smoothly on the parameters and has uniform compact support. The extension uses the velocity at the same b,tb,t and sends a zero velocity on a whole slice to the zero field. The time-dependent flow exists for 0≤t≤10\le t\le1 by compact support, and smooth dependence on initial conditions and parameters gives Φb,t\Phi_{b,t}. Solving the equation backward shows each time map is a diffeomorphism. The two curves t↦eb,t(a)t\mapsto e_{b,t}(a) and t↦Φb,t(eb,0(a))t\mapsto\Phi_{b,t}(e_{b,0}(a)) have the same initial point and satisfy the same differential equation. Uniqueness proves the result. ◻

The theorem gives a local extension principle as well. An open subset of NN containing the track of the isotopy contains the tubes of sufficiently small radius. The vector field may therefore be supported there. When the images miss a closed obstacle, this produces an ambient isotopy fixed close to the obstacle. Compactness is used to obtain a common tube and support. Without it a pointwise prescription of velocities may produce a field that escapes to infinity in finite time. For sources with boundary or a target with boundary, prescribed tangencies and compatible collar data require a relative version. The classical extension and gluing results for those settings are developed in [15, Ch. 8].

Restriction and the stabilizer

For this section take NN compact without boundary. Use the C∞C^\infty topologies on spaces of smooth maps and their subspaces. Write i:A↪Ni:A\hookrightarrow N for one embedding, D0(N)\mathcal D_0(N) for the identity path component of the diffeomorphism group, and Ei\mathcal E_i for the path component of ii in the embedding space. The restriction map q:D0(N)⟶Ei,q(F)=F∘iq:\mathcal D_0(N)\longrightarrow\mathcal E_i,\qquad q(F)=F\circ i has fiber Ki={F∈D0(N)∣F∘i=i}K_i=\{F\in\mathcal D_0(N)\mid F\circ i=i\} at ii. In particular KiK_i fixes the specified parametrized embedding pointwise and may have many path components.

Proposition 1.2 (The restriction bundle). The map qq is a locally trivial principal KiK_i-bundle. It yields a homotopy fiber sequence Ki⟶D0(N)⟶Ei.K_i\longrightarrow\mathcal D_0(N) \longrightarrow\mathcal E_i .

Proof. Here is the local-section argument behind the classical restriction theorem [27]. When a smooth embedding jj is sufficiently C1C^1-close to ii, the vector vj(a)=exp⁡i(a)−1(j(a))v_j(a)=\exp_{i(a)}^{-1}(j(a)) is defined and varies continuously with jj in the C∞C^\infty topology. Extend this section along ii to a vector field VjV_j on NN by normal projection, transport, and a fixed cutoff. For jj sufficiently close, the formula Fj,t(x)=exp⁡x(tVj(x))F_{j,t}(x)=\exp_x(tV_j(x)) defines a local diffeomorphism for each tt, by C1C^1 closeness to the identity. Compactness makes it a covering map on each connected component. Being homotopic to the identity, it induces an isomorphism on the fundamental group of each component, so the covering has one sheet. Thus it is a diffeomorphism, and j↦Fj,1j\mapsto F_{j,1} is a continuous section near ii with values in D0(N)\mathcal D_0(N).

The same construction based at every embedding shows that all D0(N)\mathcal D_0(N)-orbits in the full embedding space are open. An orbit is path connected, and each orbit’s complement is open because it is a union of other open orbits. A path cannot meet both an orbit and its complement. Consequently the orbit of ii is precisely Ei\mathcal E_i. Translation of the local section by elements of D0(N)\mathcal D_0(N) gives sections throughout that orbit. If ss is a section over UU, the map (j,k)↦s(j)k(j,k)\mapsto s(j)k identifies U×KiU\times K_i with q−1(U)q^{-1}(U), since any two diffeomorphisms sending ii to jj differ by a unique element of KiK_i on the right. The inverse is the restriction together with F↦s(q(F))−1FF\mapsto s(q(F))^{-1}F. Local triviality gives the stated homotopy fiber sequence. ◻

For example, an ambient lift of an embedding loop, started at 1N1_N, ends at an element of KiK_i. Its component is independent of the chosen lift. This follows because two lifts Ft,Ft′F_t,F'_t starting at the identity satisfy Ft−1Ft′∈KiF_t^{-1}F'_t\in K_i, giving a path from the identity that relates their endpoints. With conventional base-point choices this is the connecting map from π1(Ei,i)\pi_1(\mathcal E_i,i) to π0(Ki)\pi_0(K_i). It detects when a loop cannot be lifted to a loop of ambient diffeomorphisms.

Example 1.3 (Point pushing on a surface). Take AA to be a point and NN the closed oriented surface of genus two. Then Ei=N\mathcal E_i=N and qq evaluates a diffeomorphism at the designated point x0x_0. Every loop FtF_t in D0(N)\mathcal D_0(N) based at the identity evaluates to a loop representing a central element of π1(N,x0)\pi_1(N,x_0). Indeed, for any based loop γs\gamma_s the map (t,s)↦Ft(γs)(t,s)\mapsto F_t(\gamma_s) defines a map from the torus to NN. The two generators of the torus group map to the evaluated loop and γ\gamma, so their classes commute.

There exists a loop α\alpha in NN whose class is noncentral. In the genus-two polygon presentation with generators a,b,c,da,b,c,d and relation [a,b][c,d]=1[a,b][c,d]=1, sending b,db,d to 11 realizes a surjection onto the free group on a,ca,c. The image of aa does not commute with cc, so one may take α\alpha to represent aa. Its ambient lift, supplied by the restriction bundle, cannot end in the identity component of KiK_i. If it did, a path in the stabilizer to 1N1_N would close it into a loop of ambient diffeomorphisms without changing the evaluated class. This contradicts centrality. The returned point is the same, while the endpoint ambient transformation has acquired a nontrivial component in the stabilizer.

The action groupoid of D0(N)\mathcal D_0(N) on Ei\mathcal E_i has classifying space equivalent to BKiBK_i. Indeed local triviality identifies Ei\mathcal E_i with D0(N)/Ki\mathcal D_0(N)/K_i. Using any contractible free model ED0(N)E\mathcal D_0(N), its homotopy quotient is the space in the middle of image The transitive action therefore gives a single orbit, while the groupoid retains the stabilizer. A problem about continuous choices of ambient extensions can be formulated as a lifting problem for qq.

Changing a seam by isotopy

Let M,NM,N have boundary components S⊂∂MS\subset\partial M and T⊂∂NT\subset\partial N with collars, and let φ:S→T\varphi:S\to T be a diffeomorphism. If utu_t is an isotopy of diffeomorphisms of SS beginning at the identity, put ψ=φ∘u1\psi=\varphi\circ u_1. The manifolds obtained by gluing along ψ\psi and along φ\varphi are diffeomorphic. To construct the map, choose a collar c:S×[0,ϵ)→Mc:S\times[0,\epsilon)\to M and a smooth β\beta equal to 11 near zero and to 00 near the far end. On the collar use the diffeomorphism c(x,s)⟼c(uβ(s)(x),s).c(x,s)\longmapsto c(u_{\beta(s)}(x),s). Extend it as the identity on the rest of MM and use the identity on NN. In the source gluing, xx and ψ(x)\psi(x) are identified. Their images are u1(x)u_1(x) and φ(u1(x))\varphi(u_1(x)), which agree in the target. Because the transformation is a product near the seam, the descended map is smooth in the signed collar coordinate and has a smooth inverse. The same construction works with disjoint seams by choosing disjoint collars, so their changes commute on the resulting manifold.

A basic isotopy of disks illustrates how local differential data can fix a model. For an embedding e:Dr→Rne:D^r\to\mathbb R^n smooth up to the boundary, the maps x⟼e(0)+e(tx)−e(0)t,0<t≤1,x\longmapsto e(0)+\frac{e(tx)-e(0)}{t}, \qquad 0<t\le1, are embeddings, and the integral formula for the difference quotient gives a smooth limit x↦e(0)+D0e(x)x\mapsto e(0)+D_0e(x) at t=0t=0. The limit is still embedded because the differential at zero is injective. This isotopy varies the image and need not preserve boundary values. Relative isotopy classes with fixed boundary can contain further information.

Surfaces and families of models

Surfaces illustrate two different classification problems. The first identifies an individual surface up to existence of a diffeomorphism. The second classifies families of these surfaces, including the diffeomorphisms used when parameters return around a loop. Handles answer the first problem, and the classifying space of the diffeomorphism group specifies the second.

Disk and band presentations

Let SS be a closed connected smooth surface. Start with a finite triangulation. Thicken vertices to disks, edges to bands, and faces to the complementary disks. This is the smooth 00, 11, and 22-handle description after rounding corners. It is also the type of decomposition produced by Morse theory. It can be simplified to a single starting disk and a single final disk as follows.

First merge along the bands of a spanning tree TT of the 11-skeleton. A pair of disks joined by a band between their distinct boundary circles forms a disk after rounding. As a result the union over the tree is one disk. Next take the dual graph on faces, using only edges which do not cross TT. This graph is connected. Otherwise a nonempty proper union of its components would give a mod-22 sum of faces whose boundary uses only edges of TT. That boundary is nonzero because the full dual graph is connected. It would be a nonzero 11-cycle in a tree. Choose a dual spanning tree and merge the final face disks across each crossed band in that tree. This leaves one final disk and some number rr of bands. Consequently χ(S)=2−r\chi(S)=2-r. In particular, if χ(S)=2\chi(S)=2 there are no bands and SS is a sphere. Two disks glued along their boundaries give the usual sphere because any diffeomorphism of their boundary circles extends to a disk. Indeed an orientation-preserving circle diffeomorphism is isotopic to the identity by linear interpolation of a monotone lift with the identity lift. The collar construction of Chapter [ch:isotopy] extends this isotopy. A reflection accounts for the other sign.

Cutting along simple curves

A mod-22 11-cycle on a surface can be represented by a disjoint collection of embedded smooth circles. For a simplicial representative there is an even number of incident selected edges at each vertex. In a small disk around the vertex, pair neighboring endpoints and round their arcs. The modification has the same mod-22 relative class in this disk because the set of endpoints is unchanged and the disk has zero first homology. The modified edges give the required collection.

Suppose first that SS is oriented and has χ(S)<2\chi(S)<2. Poincaré duality over F2\mathbb F_2 gives nonzero H1(S;F2)H_1(S;\mathbb F_2). In a representative collection for a nonzero class at least one circle is itself nonzero in homology, hence nonseparating. Otherwise it would bound one side modulo 22. An oriented surface gives each embedded circle an orientable normal line bundle, so such a circle has an annular neighborhood. Cutting along it leaves a connected surface with two additional boundary circles. Cap them with disks to obtain S′S', with χ(S′)=χ(S)+2\chi(S')=\chi(S)+2.

Recover SS from S′S' by adding a torus summand. To see the summand, choose an arc joining the capped disks in S′S'. A neighborhood of the arc together with the two disk caps is itself a disk. Removing the two caps from this neighborhood gives a disk with two holes, and the original annulus joins the two inner circles. This is a punctured torus, using the usual square model of the torus cut along one of its basic curves. Replacing the disk by the punctured torus is the connected sum. Thus iteration, terminating when χ=2\chi=2, gives a connected sum of tori, with the empty sum denoting the sphere.

When SS is nonorientable, its orientation character w1(TS):H1(S;F2)→F2w_1(TS):H_1(S;\mathbb F_2)\to\mathbb F_2 is nonzero. Apply the circle representation above to a class detected by this character. Some component is one-sided, since the tangent to a circle is orientable and the normal line has nontrivial orientation. Its neighborhood is a Möbius band. The complement is connected: every component must meet the new boundary, which is just one circle, since SS is connected. Removing the band and capping the new boundary gives S′S' with χ(S′)=χ(S)+1\chi(S')=\chi(S)+1. Recovering SS replaces the cap by a Möbius band, which is the punctured projective plane. Iteration leaves an orientable surface together with at least one projective-plane summand.

Lemma 1.1 (Exchanging a handle and a crosscap). The surface T2#RP2T^2\#\mathbb{RP}^2 is diffeomorphic to the connected sum of three projective planes.

Proof. Use mod-22 homology and let a,ba,b be the two basic torus classes and cc the core of the projective-plane band. Their only nonzero pairings under intersection are a⋅b=b⋅a=1a\cdot b=b\cdot a=1 and c⋅c=1c\cdot c=1. Here the self-intersection of the one-sided core is one, as can be computed by a perturbation in the nontrivial normal line. The orientation character is zero on a,ba,b and one on cc.

Join disjoint representatives of aa and cc by an embedded band. Replacing short subarcs by the opposite sides of the band gives an embedded circle of class x=a+cx=a+c. It is one-sided. Let CC denote the complement of the interior of its Möbius neighborhood. The Mayer–Vietoris sequence over F2\mathbb F_2 gives H1(S;F2)≃H1(C;F2)⊕⟨x⟩.H_1(S;\mathbb F_2)\simeq H_1(C;\mathbb F_2)\oplus\langle x\rangle. Indeed the sole gluing circle is zero in the homology of both sides, where it bounds the mod-22 fundamental class relative to that circle, and both sides are connected. The image of H1(C;F2)H_1(C;\mathbb F_2) is therefore the orthogonal complement of xx: it pairs trivially with xx, and x⋅x=1x\cdot x=1. The class b+cb+c belongs to this complement and has nonzero orientation character. Hence CC, and also its capped version, is nonorientable.

The capped version has Euler characteristic zero. Removing another one-sided band and capping leaves Euler characteristic one, which cannot occur for a closed orientable surface by the preceding torus reduction. One more such removal leaves Euler characteristic two and therefore a sphere. Reattaching the three bands proves the claim by the description of capping as connected sum. ◻

Theorem 1.2 (Classification of compact connected surfaces). Let SS have bb boundary components. Up to diffeomorphism, SS is a surface obtained by deleting bb open disks from one of the following closed models: Sg=#gT2(g≥0),Nh=#hRP2(h≥1).S_g=\#^g T^2\quad(g\ge 0),\qquad N_h=\#^h\mathbb{RP}^2\quad(h\ge 1). The type and its index are determined by orientability, bb, and χ(S)\chi(S). The formulas are χ(Sg)−b=2−2g−b\chi(S_g)-b=2-2g-b and χ(Nh)−b=2−h−b\chi(N_h)-b=2-h-b.

Proof. For closed surfaces, the reductions and Lemma 1.1 give the models. Connected sum subtracts two from the sum of Euler characteristics, since each puncture removes a disk and the gluing circle has Euler characteristic zero. This yields the displayed formulas, which distinguish different indices within each orientability type.

For a surface with boundary, cap all the circles. Orientability of the capped model is equivalent to orientability of SS, since an orientation on SS extends over each disk with the matching boundary convention. The positions of the disjoint cap disks carry no extra invariant for an unmarked surface. Shrink each disk inside a collar neighborhood, move the resulting disks along paths of centers in the complement of the other disks, and use diffeomorphisms in the coordinate neighborhoods of the disk images to match the final disks. This is the disk isotopy construction of [15, Chs. 8–9]. The set of boundary circles may be permuted because no boundary labels are part of the classification. ◻

Models in a family

For any fixed smooth surface SS, the groupoid with its diffeomorphisms as arrows has classifying space BDiff⁡(S)B\operatorname{Diff}(S) when the arrows have the C∞C^\infty topology. A locally trivial SS-bundle with smooth structure on the fibers has continuous transition functions valued in this C∞C^\infty group. On a paracompact base where the cover is numerable, such bundles are classified up to isomorphism by homotopy classes of maps to BDiff⁡(S)B\operatorname{Diff}(S). The classifying map of the principal frame bundle sends the fiber model to a copy of SS: image The square describes the pullback of the universal bundle in a model for which the classifying map is chosen. A proper smooth submersion of boundaryless manifolds with compact surface fibers provides such local trivializations. Choose a horizontal complement to the vertical tangent bundle and lift straight paths in a small coordinate ball in the base. Properness gives existence for the bounded time interval, and reversal of the paths gives the inverse maps. This identifies all nearby fibers smoothly with the chosen reference fiber.

For example, form a torus bundle over a circle by gluing the ends of a product interval with the torus diffeomorphism induced by (1101)\bigl(\begin{smallmatrix}1&1\\0&1\end{smallmatrix}\bigr) on R2/Z2\mathbb R^2/\mathbb Z^2. The homology groups of the fibers form a local system with this monodromy on H1(T2;Z)H_1(T^2;\mathbb Z). A product bundle would have identity monodromy. Conjugation cannot identify the displayed matrix with the identity. Thus the genus specifies the fiber of this family, while the classifying map records its nontrivial gluing over the parameter circle.

Smooth spaces and quotient stacks

A smooth manifold can be studied through the objects that map into it. A chart is a particularly useful test map, and a parametrized family is another. By allowing all smooth test manifolds, one retains families that pass through singular orbits and through changes of local trivialization. The resulting category also admits quotients with automorphisms.

Smooth test manifolds

Use the category Man\mathrm{Man} of finite-dimensional smooth manifolds with the open-cover topology. In this part the test manifolds have no boundary, and the underlying site is made small in a chosen universe. One can enlarge it to a site of manifolds with corners when boundary charts are required. A smooth space in this chapter is a sheaf of spaces on the site. Write E\mathcal E for the resulting ∞\infty-topos and M‾\underline M for the sheaf U↦C∞(U,M)U\mapsto C^\infty(U,M) represented by a manifold MM. Its values are sets regarded as discrete spaces. This differs from sending MM to its singular complex: the latter has paths and forgets the smooth structure, while M‾\underline M records smooth families.

The presheaf represented by MM is a sheaf. Given maps from an open cover of UU to MM that agree on pairwise intersections, their glued map is continuous and locally smooth, hence smooth. By Yoneda every morphism M‾→N‾\underline M\to\underline N comes from a unique smooth map M→NM\to N. Products are preserved: a map U→M×NU\to M\times N is exactly a pair of smooth maps. A fiber product in smooth spaces always exists. If the two maps to a representable target are transverse, the ordinary smooth fiber product represents it, because the ordinary submanifold cut out by the transverse equations has the universal lifting property of smooth maps. For singular equations the sheaf fiber product still makes sense. Derived smooth geometry carries additional information about those equations, as developed in Chapter [ch:derived-smooth].

Chapter [ch:manifolds] gives a smooth map that is a homeomorphism of its underlying spaces without being a diffeomorphism. The failure of invertibility is detected by the identity test family on its target. Thus recovering a space from all of its smooth probes also recovers the differential structure of its maps.

Locally represented geometry

Write X(U)\mathcal X(U) for the space of UU-families in a smooth space X\mathcal X. Paths in this space compare families, and higher simplices compare such paths. In a stack valued in groupoids these paths are isomorphisms. Descent replaces literal agreement on overlaps by compatible identifications. For an open cover UiU_i of UU, the comparison is X(U)≃lim⁡[n]∈Δ∏i0,…,inX(Ui0∩⋯∩Uin).\mathcal X(U)\simeq \lim_{[n]\in\boldsymbol\Delta} \prod_{i_0,\ldots,i_n} \mathcal X(U_{i_0}\cap\cdots\cap U_{i_n}). For a groupoid-valued example, local objects and isomorphisms on double overlaps must satisfy a cocycle condition on triple overlaps. Morphisms between global objects are computed by maps of descent data. A map from a manifold into a stack can therefore be locally represented by maps to an atlas with transitions between the local lifts.

An atlas for a differentiable stack is a representable surjective submersion from a manifold. Such a map admits local sections after pullback to a test manifold. It is therefore an effective epimorphism in E\mathcal E. Its Čech groupoid consists of the atlas, the space of arrows between its two lifts, and the higher fiber products encoding composition. The stack is the sheaf colimit of this groupoid [20, §6.1]. The manifold structure of the arrow space distinguishes a smooth groupoid from its classifying homotopy type.

An action groupoid as a quotient

Let a Lie group GG act smoothly on MM from the left. Its action groupoid has object space MM and arrow space G×MG\times M, with source (g,x)↦x(g,x)\mapsto x and target (g,x)↦gx(g,x)\mapsto gx. The quotient [M/G][M/G] is the colimit of the simplicial nerve in smooth spaces. It classifies local representatives of points modulo gauge transformations.

Proposition 1.1 (Families in an action quotient). A map from a test manifold UU to [M/G][M/G] can be described by a smooth principal right GG-bundle π:P→U\pi:P\to U and a smooth map a:P→Ma:P\to M satisfying a(pg)=g−1a(p)a(pg)=g^{-1}a(p). Arrows between such data are bundle isomorphisms preserving these maps.

Proof. Choose a local section sis_i of the torsor over each open UiU_i and write sj=sigijs_j=s_i g_{ij} on overlaps. The transition maps satisfy gijgjk=gikg_{ij}g_{jk}=g_{ik}. The points xi=a(si)x_i=a(s_i) obey xj=gij−1xix_j=g_{ij}^{-1}x_i, so the action arrows from xjx_j to xix_i carry exactly the transition data of the torsor. Conversely, descent data (xi,gij)(x_i,g_{ij}) glue Ui×GU_i\times G by (u,h)j∼(u,gij(u)h)i(u,h)_j\sim(u,g_{ij}(u)h)_i. The functions (u,h)i↦h−1xi(u)(u,h)_i\mapsto h^{-1}x_i(u) agree with this identification and give aa. Changing local sections changes the data by a groupoid isomorphism. Every bundle isomorphism preserving the map to MM is locally described by such changes. The constructions thus identify both objects and arrows in the quotient stack. ◻

The map [M/G]→[∗/G][M/G]\to[*/G] forgets the equivariant map and keeps its torsor. Pulling it back along the trivial torsor with a chosen trivialization gives back MM. image For a free proper action, M→M/GM\to M/G is a principal bundle. Writing a map to the manifold quotient as the pullback of this bundle with its tautological map to MM gives an inverse to the map [M/G]→M/G‾[M/G]\to\underline{M/G}. In general this equivalence fails at any point with a nontrivial stabilizer. The stabilizer is precisely the automorphism group of the corresponding family over a point.

The reflection quotient

For G=C2G=C_2 acting on R\mathbb R by reflection, the topological orbit space is the closed half-line, with coordinate t2t^2. The origin of the quotient stack has automorphism group C2C_2 while a nonzero orbit has none. A smooth family supported at the origin on a connected base is a principal C2C_2-bundle equipped with the zero map. Over S1S^1 the trivial cover and the connected double cover give distinct families, even though the same coarse function t2=0t^2=0 represents both.

The connected cover with total space R/(4πZ)\mathbb R/(4\pi\mathbb Z) admits a weight w([θ])=cos⁡(θ/2)w([\theta])=\cos(\theta/2) that is odd under the deck transformation. Its square descends to the smooth function (1+cos⁡θ)/2(1+\cos\theta)/2. There is no smooth real function on the base with that square. In a local coordinate centered at the unique zero, the square vanishes quadratically with positive quadratic coefficient. A smooth square root must have nonzero derivative and thus change sign there. The complement of the unique zero is connected, so a real nonzero function there could not change sign. The quotient stack permits the pair of local roots together with a sign transition. Smooth differential data see an obstruction absent from the coarse observable.

The next chapter keeps such transition data while forming spaces of maps and composing families. Chapter [ch:orientations] uses a related descent construction for orientations of bundles.

Mapping objects and families over a base

The space of maps between manifolds has several useful topologies, as in Chapter [ch:function-spaces]. Smooth families can also be specified without choosing a topology on the set of maps. This uses all smooth test objects and an evaluation morphism. The same construction makes sense for families over a base that is itself a stack.

Smooth mapping objects

For smooth spaces X,YX,Y in the ∞\infty-topos E\mathcal E, write [X,Y][X,Y] for their exponential. Thus Map⁡E(T,[X,Y])≃Map⁡E(T×X,Y)\operatorname{Map}_{\mathcal E}(T,[X,Y]) \simeq \operatorname{Map}_{\mathcal E}(T\times X,Y) naturally in TT. When X=M‾X=\underline M and Y=N‾Y=\underline N, Yoneda computes its value on a test manifold UU: [M‾,N‾](U)=C∞(U×M,N).[\underline M,\underline N](U)=C^\infty(U\times M,N). This specifies the regularity of derivatives in both the test and input variables. It has no requirement that the exponential be a finite-dimensional manifold. Even the smooth maps from a circle to a line have infinite-dimensional space of variations. The sheaf model also distinguishes changes of parameter that are smooth from continuous paths for a chosen function-space topology. The comparison with convenient calculus can be made in settings in which both sides’ smooth test plots agree [2, 17].

As a basic example, a family of real functions on MM is a map P→[M‾,R‾]P\to[\underline M,\underline{\mathbb R}]. Taking its ff at a point pp gives a function fpf_p. Restricting this family along a:P′→Pa:P'\to P gives fa(p′)f_{a(p')}. Differentiating with respect to the input gives the vertical one-form dMfd_M f on P×MP\times M, a section of the pullback of T∗MT^*M. The restriction of this section along aa is dMd_M of the restricted function, since the maps in the input directions are unchanged by varying aa. Differentiation with respect to the parameter instead pulls its covectors back along DaDa. The two directions have different transformation laws even for a trivial product.

Evaluation over a base

Let BB be any object of E\mathcal E. Products and exponentials in its slice are written ×B\times_B and [X,Y]B[X,Y]_B. This is the category of families varying over BB. The defining evaluation and its adjunction are ev⁡B:[X,Y]B×BX⟶Y,Map⁡B(P×BX,Y)≃Map⁡B(P,[X,Y]B).\begin{gathered} \operatorname{ev}_B:[X,Y]_B\times_B X\longrightarrow Y,\\ \operatorname{Map}_{B}(P\times_B X,Y) \simeq\operatorname{Map}_{B}(P,[X,Y]_B). \end{gathered} So a map f:P×BX→Yf:P\times_B X\to Y gives its classifying morphism rf:P→[X,Y]Br_f:P\to[X,Y]_B. A reparametrization includes a map between parameter objects over BB commuting with the classifying maps up to a specified homotopy. These reparametrizations and their higher comparisons form the slice E/[X,Y]B\mathcal E_{/[X,Y]_B}. The identity family has parameter BB. A family from XX to YY with parameter PP, followed by one from YY to ZZ with parameter QQ, has parameter P×BQP\times_B Q. Its evaluation is the same as applying the two evaluations in order. The associativity comparison comes from the product comparison on three parameters.

Proposition 1.1 (Base change for relative evaluation). For a:C→Ba:C\to B, pullback sends [X,Y]B[X,Y]_B to [a∗X,a∗Y]C[a^*X,a^*Y]_C by an equivalence preserving evaluations. It preserves composition and reparametrization of families.

Proof. The functor a!a_! left adjoint to pullback sends an object over CC to its structure map composed with aa. For T→CT\to C, a!(T×Ca∗X)≃(a!T)×BXa_!(T\times_C a^*X)\simeq (a_!T)\times_B X because each side is the same pullback equipped with the map to BB. Therefore Map⁡C(T,a∗[X,Y]B)≃Map⁡B(a!T,[X,Y]B)≃Map⁡B((a!T)×BX,Y)≃Map⁡C(T×Ca∗X,a∗Y).\begin{aligned} \operatorname{Map}_C(T,a^*[X,Y]_B) &\simeq \operatorname{Map}_B(a_!T,[X,Y]_B)\\ &\simeq \operatorname{Map}_B((a_!T)\times_B X,Y)\\ &\simeq \operatorname{Map}_C(T\times_C a^*X,a^*Y). \end{aligned} These comparisons are natural in TT. Yoneda gives the equivalence, and the construction evaluates the same map under the adjunctions. The product that defines composition is preserved by pullback. Naturality with respect to PP gives the last assertion. ◻

Proposition 1.2 (Descent of families). Let U→BU\to B be an effective epimorphism with Čech nerve U∙U_\bullet. Set Xn=X×BUnX_n=X\times_B U_n and Yn=Y×BUnY_n=Y\times_B U_n. Then the category of parametrized maps from XX to YY is equivalent to E/[X,Y]B≃lim⁡[n]∈ΔE/[Xn,Yn]Un.\mathcal E_{/[X,Y]_B}\simeq \lim_{[n]\in\boldsymbol\Delta} \mathcal E_{/[X_n,Y_n]_{U_n}}.

Proof. Descent for slices of an ∞\infty-topos identifies E/B\mathcal E_{/B} with the limit of E/Un\mathcal E_{/U_n} [20, §6.1.3]. Slicing at a compatible target commutes with taking that limit. Indeed an arrow into the target is a compatible family of local arrows into its components, with the same statement for all mapping spaces. Under the first equivalence the image of the target exponential is the compatible family of local exponentials by Proposition 1.1. ◻

In ordinary charts with transition equivalences chosen, this result glues local evaluations when the comparisons commute. If an input and output are bundles with changes of frames aija_{ij} and bijb_{ij}, the local maps transform as fi=bijfjaij−1f_i=b_{ij}f_j a_{ij}^{-1} when the parameter coordinates have already been transported. If the same parameter supplies two stages of a composite, the pullback along the diagonal P→P×BPP\to P\times_B P implements that sharing. In sheaves with higher morphisms, local comparisons themselves need coherence on multiple overlaps. The limit in the proposition records exactly those data.

Transport over a homotopy type

In the ∞\infty-category S\mathcal S of spaces, straightening identifies S/B\mathcal S_{/B} with Fun⁡(B,S)\operatorname{Fun}(B,\mathcal S), where BB is an ∞\infty-groupoid [20, §3.2]. A path γ:b→c\gamma:b\to c induces equivalences tX:Xb≃Xct_X:X_b\simeq X_c and tY:Yb≃Yct_Y:Y_b\simeq Y_c. A map on the fiber at bb is transported by h↦tYhtX−1h\mapsto t_Y h t_X^{-1}. The higher comparisons required for this formula come from the path composition of the original functors. The evaluation square image commutes with this transport. To check the formula for the exponential, first evaluate it at each bb and choose these conjugation maps along paths. A coherent map from PP to this family is a collection of maps Pb×Xb→YbP_b\times X_b\to Y_b commuting with all transports. Thus the family satisfies the exponential adjunction.

For an ordinary discrete group action this specializes to (g⋅h)(x)=g h(g−1x)(g\cdot h)(x)=g\,h(g^{-1}x), and a family f(p,x)f(p,x) obeys f(gp,gx)=gf(p,x)f(gp,gx)=g f(p,x). A particular fiber of the family can have less symmetry than the total family. It is equivariant under the stabilizer of pp. For smooth actions one may work over the smooth classifying stack of Chapter [ch:smooth-stacks] to retain the dependence on smooth generalized points. As an example, SO(2)SO(2) acts on R2\mathbb R^2 and its line directions. Orthogonal projection onto a line transforms by conjugation under rotation, so the whole family is equivariant with the moving line as parameter. Projection onto one fixed line commutes only with rotations preserving that line.

Orientation and characteristic classes in families

An oriented manifold has locally chosen positive frames whose transitions preserve orientation. For a family of manifolds there can instead be a local orientation on every fiber and a change of sign around a parameter loop. The orientation is then a double cover of the base or of the total space, as appropriate. This cover is itself descent data.

Determinants and orientation covers

For a real vector bundle EE of rank rr over MM, its determinant det⁡E\det E is ⋀rE\bigwedge^r E. The orientation cover Or⁡(E)→M\operatorname{Or}(E)\to M has fiber the two components of nonzero frames of det⁡E\det E. It is a principal C2C_2-bundle. Local frames of EE give transition matrices gij:Ui∩Uj→GL(r,R)g_{ij}:U_i\cap U_j\to GL(r,\mathbb R), with gijgjk=gikg_{ij}g_{jk}=g_{ik}. Their images under sgn⁡∘det⁡\operatorname{sgn}\circ\det are the transitions of the cover. An orientation of EE is a section of that cover. This is the same as a choice of frame component on each connected chart such that all transition signs are positive.

The determinant signs form a Čech 11-cocycle with values in C2≃Z/2C_2\simeq\mathbb Z/2. Replacing the local trivializations changes this cocycle by a coboundary. On a good cover of a manifold it represents the first Stiefel–Whitney class w1(E)w_1(E). In smooth spaces Or⁡(E)\operatorname{Or}(E) is classified by a map M‾→[∗/C2]\underline M\to[*/C_2] as in Proposition [prop:smooth-quotient]. For r≥1r\geq1, lifting a family of frames from GL(r,R)GL(r,\mathbb R) to the subgroup with positive determinant is equivalent to specifying a smooth trivialization of the sign torsor. The two descriptions give the same obstruction, and both commute with pullback.

For real line bundles the sign class is also complete. Choose a fiberwise positive inner product and take unit local frames. All transition functions now have values in {1,−1}\{1,-1\}. A sign cocycle determines a line by gluing, and two cocycles that differ by a coboundary give isomorphic bundles by changing unit frames. Thus orientation and triviality coincide for a real line. This reflects the contractibility of the subgroup of positive real scalars. On S1S^1, a line with sign monodromy −1-1 has no nowhere-zero section: a section on the interval parametrizing the circle has endpoint values of opposite sign unless both vanish. The intermediate value theorem then forces a zero. For the trivial bundle the constant unit section avoids this obstruction.

An exact sequence of relative tangent spaces

An exact sequence of vector bundles 0⟶K⟶E⟶F⟶00\longrightarrow K\longrightarrow E\longrightarrow F \longrightarrow 0 gives a natural isomorphism det⁡E≃det⁡K⊗det⁡F\det E\simeq\det K\otimes\det F. To construct it, start with a basis of KK and append lifts of a basis of FF. A change of lifts adds linear combinations of the earlier basis vectors and leaves the top exterior product unchanged. Changes of the two bases scale the wedge by the corresponding determinants. This constructs the isomorphism in local frames, where it is visibly natural under pullback. The orientation cover of EE is accordingly the tensor product of the two covers, and w1(E)=w1(K)+w1(F)w_1(E)=w_1(K)+w_1(F).

Let f:M→Nf:M\to N and g:Q→Ng:Q\to N be transverse smooth maps. At a point of their fiber product ZZ the derivative df−dgd f-d g is surjective. The preimage normal form of Chapter [ch:sard] identifies TZT Z with its kernel. If p:Z→Mp:Z\to M, q:Z→Qq:Z\to Q, and h=fp=gqh=f p=g q, this yields det⁡TZ≃p∗det⁡TM⊗q∗det⁡TQ⊗(h∗det⁡TN)∨.\det T Z \simeq p^*\det T M\otimes q^*\det T Q \otimes(h^*\det T N)^\vee . The convention uses the ordered sum TM⊕TQT M\oplus T Q and the displayed derivative. Choosing oriented frames of the three ambient bundles orients the intersection via this convention. When maps are replaced by parameter-dependent families transverse in the vertical directions, the same construction uses the vertical bundles. In a model for derived smooth intersections the two-term tangent complex continues to record this determinant when the derivative fails to be surjective. image This determinant relation is an orientation formula for the pullback, rather than a choice of a splitting of its normal sequence.

An oriented count over a parameter space

Suppose πM:M→B\pi_M:M\to B and πN:N→B\pi_N:N\to B are smooth submersions of boundaryless manifolds of equal relative dimension dd, with f:M→Nf:M\to N a smooth map over BB. Let s:B→Ns:B\to N be a smooth section, and suppose that for each bb, s(b)s(b) is a regular value of fbf_b. Put Z=M×NBZ=M\times_N B. If Z→BZ\to B is proper, it is a finite covering over each connected component of BB, allowing the empty covering. In local submersion coordinates the defining equations of ZZ have an invertible vertical Jacobian and can be solved as a smooth graph over BB. Thus Z→BZ\to B is a local diffeomorphism. Properness makes each fiber compact and discrete, hence finite, and prevents additional inverse images from entering a sufficiently small neighborhood. These local graphs form the sheets.

The derivative along each sheet is an isomorphism between the two vertical tangent bundles. If both are oriented globally, this determines a sign that is locally constant on the sheet. Monodromy permutes the sheets and preserves their signs. Consequently the sum of signs on a fiber is constant on each connected component of BB. For proper fiber maps of oriented manifolds this agrees with the regular-value formula for degree in Chapter [ch:degree], when the conventions for degree there apply. The unsigned count modulo two remains defined if orientations are absent. More generally one can keep the orientation line of each sheet. Locally changing a frame reverses its corresponding sign, and the local counts use the transitions of those lines. This description explains which orientation data must be fixed before adding signs as ordinary integers.

For example, let ξ\xi be the Möbius real line on S1S^1 and consider the fiber map from ξ\xi to itself given by multiplication by any fixed negative scalar. Each fiber has degree −1-1 when the same frame is used for both source and target. Neither of the vertical lines is orientable, but the tensor product of their orientation lines is trivial. The relative orientation is supplied by the identification of the two lines. By contrast, identifying ξ\xi with a trivial line by a fiberwise linear isomorphism would contradict its nonzero sign class.

Euler classes and pullback

For an oriented rank-rr bundle ξ\xi, the Thom class uξ∈Hr(E,E∖M;Z)u_\xi\in H^r(E,E\setminus M;\mathbb Z) restricts to the chosen positive generator on each fiber pair. It is natural for pullback bundles [26, 14]. The Euler class is e(ξ)=0ξ∗uξe(\xi)=0_\xi^*u_\xi, where the zero section maps MM to the total space. A nowhere-zero section has zero pullback of the relative class. Homotopy through sections to the zero section then gives e(ξ)=0e(\xi)=0. When a smooth section is transverse to zero, the local intersection computation identifies the Euler class with the cohomology class dual to its zero locus under the usual orientation and compactness hypotheses of Chapter [ch:intersection]. At a regular isolated zero this follows from identifying the local pullback of the Thom class with the local degree of the section in a bundle chart.

One can package these classes using a bundle classifying space and cohomology classes there. The smooth classifying stack retains smooth gauge data, whereas a topological classifying space classifies the bundle’s underlying topological class in the settings of Chapter [ch:bundles]. The resulting characteristic classes commute with pullback and therefore apply to families described by either local frames or their topological classification. The Euler class measures one aspect of the bundle and can coincide for distinct smooth families. Classifying maps retain the gluing data from which these invariants are extracted.

Discriminants and the topology of families

A smooth family can move through maps with singular fibers. The discriminant marks parameter values where the implicit function theorem fails in the input direction. Outside it, under a properness hypothesis, fibers form an ordinary bundle even though a drawing of two neighboring fibers may make their equivalence hard to see.

A regular total space with singular fibers

Let P,X,YP,X,Y be smooth finite-dimensional manifolds without boundary, and let F:P×X→YF:P\times X\to Y be smooth. Fix y∈Yy\in Y and suppose that FF is transverse to the point yy. Thus Z=F−1(y)Z=F^{-1}(y) is a manifold. Write π:Z→P\pi:Z\to P for the projection. For (p,x)∈Z(p,x)\in Z the tangent equation is DpF p˙+DxF x˙=0.D_pF\,\dot p+D_xF\,\dot x=0 . The derivative of π\pi sends (p˙,x˙)(\dot p,\dot x) to p˙\dot p. It is surjective if and only if DxFD_xF is onto. In one direction this follows by solving the tangent equation for every p˙\dot p. In the other direction, surjectivity of DπD\pi implies that the image of DpFD_pF is contained in the image of DxFD_xF. The sum of the two images is all of TyYT_yY by transversality of FF.

Let C⊂ZC\subset Z consist of the points where DπD\pi is not onto. This is closed by the rank criterion. Define the discriminant of this family at yy to be D=π(C)D=\pi(C). The corresponding map on the complementary locus is image The square is a pullback, and on the left the projection is a submersion. By Sard’s theorem the discriminant has measure zero in parameter charts. This conclusion requires the regular total space assumption; it makes no assertion about a family with an identically singular evaluation.

Proposition 1.1 (Bundles away from the discriminant). If π:Z→P\pi:Z\to P is proper, then DD is closed and over each connected component of P∖DP\setminus D in the image, the restricted projection is a smooth fiber bundle. On any component outside the image the fibers are all empty.

Proof. A proper map between manifolds is closed, so it sends CC to a closed set. Restriction to the open complement remains proper by base change. On that open set the image is both open (submersions have local sections near image points) and closed (properness). It is therefore a union of connected components.

Here is the local triviality argument for a proper submersion. Choose a Riemannian metric on its domain and define horizontal vectors to be orthogonal to ker⁡Dπ\ker D\pi. The derivative identifies this horizontal bundle with the pullback of the tangent bundle of the base. Take a coordinate ball centered at pp with compact closure contained in the base of the submersion. Lift each radial path in the ball horizontally starting at an arbitrary point of the fiber over pp. The lifted path stays in the inverse image of the closed ball, which is compact. The smooth ordinary differential equation therefore has a solution over the entire radial segment. Its dependence on endpoints and initial points is smooth. The reverse radial path gives the inverse identification with the fiber over pp. This gives a smooth trivialization over the ball. ◻

A different metric may give different transports, and a loop can give nontrivial bundle monodromy. The change of trivialization is part of the family, as in Chapter [ch:relative-maps]. Properness holds automatically for the projection of this closed preimage when XX is compact. With noncompact XX, roots can escape to infinity without ever becoming singular. For example F(t,x)=tx−1F(t,x)=tx-1 has one root when t≠0t\ne0 and none at 00, with DxF=tD_xF=t nonzero at every root. The entire zero locus is regular, but its projection is not proper, since the roots for t→0t\to0 are unbounded.

Fold and cusp calculations

For the family of maps F(t,x)=x2−tF(t,x)=x^2-t, the total zero set is a parabola, parametrized by xx. The projection to tt has one critical point at x=0x=0 and is proper. For t<0t<0 the fiber is empty, and for t>0t>0 its two points are x=±tx=\pm\sqrt{t}. The derivatives 2x2x give opposite local degree signs and their sum is zero. At t=0t=0 the vertical derivative vanishes but the parameter derivative is −1-1, so the total space remains smooth. This is the local fold mechanism by which a pair of solutions is created or lost. For a given one-dimensional family, the fold form follows locally if the parameter derivative and the second input derivative are nonzero at a critical root. Indeed one can solve the equation for tt as a function ϕ(x)\phi(x) by the implicit function theorem. Then ϕ′(0)=0\phi'(0)=0 and ϕ′′(0)=−Fxx/Ft≠0\phi''(0)=-F_{xx}/F_t\ne0. The one-dimensional Morse lemma gives the asserted normal form up to local coordinates and a possible reversal of parameter (Chapter [ch:morse]).

Now take parameters (a,b)(a,b) and equation F(a,b,x)=x3+ax+b=0.F(a,b,x)=x^3+ax+b=0 . The total locus is smooth since DbF=1D_bF=1. It projects properly to the plane of coefficients: a root with ∣x∣>1|x|>1 satisfies ∣x∣3≤∣a∣∣x∣+∣b∣|x|^3\le |a||x|+|b|, which bounds its size when (a,b)(a,b) lies in a compact set. A critical root satisfies 3x2+a=03x^2+a=0. In terms of that root ss the discriminant is parametrized by a=−3s2,b=2s3,4a3+27b2=0.a=-3s^2,\qquad b=2s^3,\qquad 4a^3+27b^2=0. For s≠0s\ne0 the critical point above (a,b)(a,b) is a fold of the projection. The root s=0s=0 is a triple root at the cusp. If a≥0a\ge0 a polynomial off the discriminant has one real root. If a<0a<0, evaluate it at the local maximum and minimum x=∓−a/3x=\mp\sqrt{-a/3}. Their values are b+2(−a/3)3/2b+2(-a/3)^{3/2} and b−2(−a/3)3/2b-2(-a/3)^{3/2}. There are three simple real roots when these have opposite signs, exactly when 4a3+27b2<04a^3+27b^2<0, and otherwise one off the discriminant. In the three-root region the ordered derivative signs are +,−,++,-,+. They have the same sum as the sign of the unique root elsewhere. Topology of the fiber distinguishes these regions although degree does not.

Transverse jet conditions in a family

The vertical rr-jet of a family is the map (p,x)⟼jxrFp(p,x)\longmapsto j_x^r F_p from P×XP\times X to Jr(X,Y)J^r(X,Y). A submanifold of the jet bundle can impose a rank condition, a root condition, or a constraint on higher derivatives. If this total jet map is transverse to a smooth submanifold SS, then its inverse image is smooth and the parametric transversality theorem of Chapter [ch:transversality] applies. The derivative test for the projection shows that a regular parameter has jrFpj^r F_p transverse to SS. If codim⁡S>dim⁡X\operatorname{codim} S> \dim X, such a parameter has empty pullback of SS. For a finite or countable collection of smooth strata, one can discard a countable union of measure-zero exceptional sets, provided total transversality holds to each. This test addresses the actual set of allowed parameters. It can fail for a constrained family even if a different unconstrained family contains all the needed perturbation directions.

Classical bundles describe the complementary locus pointwise and along paths. At a discriminant one can retain the original equation by a derived pullback, as in Chapter [ch:derived-intersection]. When the parameter space carries an appropriate stratification, one can also study transport adapted to the strata (Chapter [ch:exit]). In the cubic example the cusp stratum, fold stratum, and open regions record which regular configurations approach a given critical configuration. They preserve information absent from a single fiber count.

Derived smooth spaces

A smooth inverse image of a regular value has a manifold structure because its equations can be placed among the coordinates. The same equations at a critical value can define a singular set, a smooth set with excess dimension, or a point at which infinitesimal motions remain. Derived smooth geometry keeps the equations and their relations in the structure sheaf. Its pullback represents the compatibility of two maps even when their differentials fail to be transverse.

The local model of a zero set

Use derived manifolds in smooth geometry, with ordinary manifolds included by their smooth function rings. We work with finite local models of the form Z(F)=U×Rmh{0},F:U⟶Rm,Z(F)=U\times^{h}_{\mathbb R^m}\{0\}, \qquad F:U\longrightarrow\mathbb R^m, where UU is open in Rn\mathbb R^n and the superscript hh denotes a homotopy pullback in the category of derived manifolds. The underlying topological space is F−1(0)F^{-1}(0). Its functions form a sheaf of derived C∞C^\infty-rings. Spivak constructs derived manifolds by gluing such models [30]. Carchedi and Steffens characterize a finite-limit, idempotent-complete extension of the category of manifolds in which transverse pullbacks are preserved [5]. We use the finite local zero-locus case. More general iterated limits may have cotangent complexes of wider amplitude.

The ordered list of equations gives a concrete model. Write A\mathcal A for the sheaf of smooth germs from UU restricted to the topological zero set, and let F1,…,FmF_1,\ldots,F_m be the components of FF. Adjoin generators ηi\eta_i of cohomological degree −1-1: KF=A⊗R⋀R(η1,…,ηm),dηi=Fi.\mathcal K_F =\mathcal A\otimes_{\mathbb R} \bigwedge\nolimits_{\mathbb R}(\eta_1,\ldots,\eta_m), \qquad d\eta_i=F_i . The differential extends by the signed Leibniz rule. The degree-zero smooth structure is that of A\mathcal A. This is the local Koszul model inside the differential graded realization of derived smooth geometry [4]. One can see the origin of the formula by resolving the point in the coordinate target. In the smooth germs at the origin of Rm\mathbb R^m, the coordinate functions t1,…,tmt_1,\ldots,t_m form a regular sequence. Division by a coordinate for functions vanishing on its zero hyperplane follows by integrating the partial derivative along that coordinate. After quotienting by the first i−1i-1 coordinates, the same argument applies on their common zero plane. Multiplication by tit_i there is injective, since a smooth function annihilated by tit_i vanishes on the dense locus where ti≠0t_i\ne0. The Koszul resolution for (ti)(t_i) thus resolves the origin. In the differential graded model the resolution is formed by adjoining the degree −1-1 generators with differentials tit_i. Pullback transports these differentials to the corresponding components of FF. In particular, H0(KF)=A/(F1,…,Fm).H^0(\mathcal K_F)=\mathcal A/(F_1,\ldots,F_m). Negative cohomology describes relations of the equations modulo the boundaries of the Koszul complex.

If FF is submersive near a zero, the submersion chart writes it as the projection onto the first mm coordinates. The regular sequence argument then gives Hi(KF)=0H^i(\mathcal K_F)=0 for i<0i<0 and identifies H0H^0 with smooth functions on the usual level submanifold. This proves the agreement with the classical construction in a transverse coordinate chart. Another useful check is stabilization. Replacing F(u)F(u) by (F(u),s)(F(u),s) on U×RrU\times\mathbb R^r changes the local model by the independent regular-coordinate equations s1=⋯=sr=0s_1=\cdots=s_r=0. Successively resolving those equations reduces the model to KF\mathcal K_F. Equivalently, the two derived spaces agree by the composition and product laws for pullbacks.

The tangent complex

For a point xx of Z(F)Z(F) write TxZ(F)T_xZ(F) for the tangent complex, dual to the cotangent complex evaluated at xx. Its local model uses the cohomological degrees 0,10,1.

Proposition 1.1 (Tangent complex of equations). At a zero xx of FF there is a natural identification TxZ(F)≃[ TxU→DxFRm ],T_xZ(F)\simeq [\,T_xU\xrightarrow{D_xF}\mathbb R^m\,], where TxUT_xU has degree zero. In particular the cohomology spaces are ker⁡DxF\ker D_xF in degree zero and coker⁡DxF\operatorname{coker}D_xF in degree one.

Proof. In the differential graded chart, the cotangent module is generated over the function algebra by the differentials of the smooth coordinates in degree zero and the differentials of the ηi\eta_i in degree −1-1. The smooth chain rule gives d(dηi)=dFid(\mathrm{d}\eta_i)=\mathrm{d}F_i. This is the derived cotangent model because the coordinates and the Koszul generators form the smooth free chart with successively adjoined generators, with an open localization on UU. Evaluating at xx gives the two-term cotangent complex [ (Rm)∗→(DxF)∗Tx∗U ][\,(\mathbb R^m)^*\xrightarrow{(D_xF)^*}T_x^*U\,] in degrees −1,0-1,0. Dualizing this finite-dimensional complex gives the asserted two terms. Equivalently, the cotangent pushout formula at the point computes the same complex from the pullback of the two maps to Rm\mathbb R^m [30, §7]. ◻

The degree-zero cohomology is the space of compatible tangent vectors. Given a first-order right-hand side bb for F(u)=bF(u)=b, the degree-one cohomology records its class modulo the changes attainable by moving uu. The index dim⁡H0(TxZ(F))−dim⁡H1(TxZ(F))=n−m\dim H^0(T_xZ(F))-\dim H^1(T_xZ(F))=n-m is the virtual dimension at xx. The equality follows by subtracting the same rank from the two terms. Stabilization adds the acyclic complex [Rr→1Rr][\mathbb R^r\xrightarrow{1}\mathbb R^r], which leaves the index unchanged.

Example 1.2 (Tangent information and higher functions). Let RR be the ring of real smooth germs at 00. For F(t)=t2F(t)=t^2, the Koszul differential is multiplication by t2t^2. It is injective, since a smooth germ vanishing off the origin vanishes on a neighborhood. Taylor’s formula with smooth remainder gives R/(t2)≅R[ϵ]/(ϵ2)R/(t^2)\cong\mathbb R[\epsilon]/(\epsilon^2): every germ has a unique constant and linear term modulo t2t^2. Thus this example has a nonreduced ordinary function ring. Replacing F(t)F(t) by t3t^3 gives R[ϵ]/(ϵ3)\mathbb R[\epsilon]/(\epsilon^3) with the same tangent complex [R→0R][\mathbb R\xrightarrow{0}\mathbb R]. If FF is identically zero on R\mathbb R, the same complex appears at every point, and the degree −1-1 Koszul generator now survives in the function algebra. These descriptions distinguish the three models through functions, even though their pointwise tangent complexes agree at the selected zeros.

Example 1.3 (Redundant constraints). The map t↦(t,0)t\mapsto(t,0) to R2\mathbb R^2 has a single zero. Its Koszul algebra has one regular constraint and one generator of zero differential, so the germ is equivalent to ⋀R(η)\bigwedge_{\mathbb R}(\eta) with ∣η∣=−1|\eta|=-1. The tangent differential v↦(v,0)v\mapsto(v,0) is injective with one-dimensional cokernel. The virtual dimension is −1-1, accounting for the redundant target direction. The pair of constant maps {0}⇉R\{0\}\rightrightarrows\mathbb R also has a derived pullback whose function algebra is this exterior algebra. Its homotopy pullback in spaces is contractible, since the based loop space of the underlying real line is contractible. Here the extra information comes from the smooth structure on the target and is detected by the derived function algebra.

Compatibility as a mapping property

If X→W←YX\to W\leftarrow Y are ordinary smooth maps, the derived pullback has the universal property Map⁡(Q,X×WhY)≃Map⁡(Q,X)×Map⁡(Q,W)hMap⁡(Q,Y)\operatorname{Map}(Q,X\times^h_WY)\simeq \operatorname{Map}(Q,X)\times^h_{\operatorname{Map}(Q,W)} \operatorname{Map}(Q,Y) for derived test objects QQ. A point of the right-hand space consists of the two maps and a comparison between their composites to WW. The comparison is part of the universal property at all higher simplicial degrees. image Around a pair (x,y)(x,y) with f(x)=g(y)f(x)=g(y), choose a single chart on WW and express compatibility by F(u,v)=ϕ(f(u))−ϕ(g(v))F(u,v)=\phi(f(u))-\phi(g(v)). The diagonal of the coordinate space is its ordinary transverse zero set for the difference map. Pullback associativity therefore identifies this local derived model with Z(F)Z(F). The tangent complex in Proposition 1.1 becomes [ TxX⊕TyY→Df−DgTf(x)W ].[\,T_xX\oplus T_yY\xrightarrow{Df-Dg}T_{f(x)}W\,]. Changing target coordinates identifies the last term by the derivative of the chart transition. More intrinsically, the complex is the homotopy fiber of the difference of the tangent maps, a description that also makes sense after the local equations have been glued.

A further base change can be performed one arrow at a time or in a single pullback. Both comparisons represent the same space of maps from every QQ, by the repeated universal property and associativity of limits. This observation is useful for families of constraints: imposing incidence on a family and then restricting its parameter gives the same derived incidence space as restricting the family first. Surjectivity of a derivative is instead a chartwise test that identifies a particular derived intersection with an ordinary one. The next chapter uses this difference to calculate excess intersection classes.

Derived intersections and excess

Let AA and BB be embedded submanifolds of WW, with the manifolds taken without boundary. Their derived intersection represents the pullback of their inclusions. Its underlying points are precisely the points of A∩BA\cap B. The linearization at cc is the two-term complex [ TcA⊕TcB→(u,v)↦u−vTcW ],[\,T_cA\oplus T_cB\xrightarrow{(u,v)\mapsto u-v}T_cW\,], using the result of Chapter [ch:derived-smooth]. If AA is cut out locally by a submersion ss, the same derived intersection is the zero locus of s∣Bs|_B. The two presentations are equivalent by pasting pullback squares. In particular the normal equation model at cc has source TcBT_cB, target NA/W,cN_{A/W,c}, and differential induced by the inclusion into TcWT_cW.

Clean intersections

Definition 1.1. An intersection of embedded submanifolds is clean if C=A∩BC=A\cap B is an embedded submanifold and TcC=TcA∩TcBT_cC=T_cA\cap T_cB for all c∈Cc\in C.

This condition allows the tangent sum to be smaller than TcWT_cW. Since its dimension is locally constant on CC, the quotient E=TW∣C/(TA∣C+TB∣C)E=\left.TW\right|_C/ \big(\left.TA\right|_C+\left.TB\right|_C\big) is a smooth vector bundle. It is the excess bundle. The degree-zero tangent cohomology is TCTC by identifying a kernel pair with its common vector, and degree-one cohomology is EE. The natural map on normal bundles gives the exact sequence 0⟶NC/B⟶NA/W∣C⟶E⟶0. 0\longrightarrow N_{C/B} \longrightarrow \left.N_{A/W}\right|_C \longrightarrow E\longrightarrow 0 . Indeed, TB∣C→TW∣C/TA∣CTB|_C\to TW|_C/TA|_C has kernel TCTC, and its cokernel is the displayed quotient. Thus an excess direction can be read either from the derived tangent or from the normal bundles of the inclusion square. image For a transverse intersection the excess bundle has rank zero. If A=BA=B, the inclusion on the left of (1) has source zero and E=NA/WE=N_{A/W}. Both are clean cases. Even when TCTC and EE exhaust the tangent cohomology, the entire local function algebra of the derived intersection belongs to its geometric structure.

Local functions in the clean case

A clean zero set has a local simplification that explains the appearance of the excess bundle in the linearization. Let F:U→RmF:U\to\mathbb R^m have a smooth zero set CC with ker⁡DcF=TcC\ker D_cF=T_cC along CC. Near any chosen cc, select r=codim⁡UCr=\operatorname{codim}_U C components with independent differentials. Use them as normal coordinates u1,…,uru_1,\ldots,u_r. Their common zero set locally agrees with CC: the latter is an embedded submanifold of it of the same dimension, and hence open in it near cc. Shrinking to a product coordinate box, write each remaining function as Fr+j(a,u)=∑i=1rhji(a,u) ui.F_{r+j}(a,u)=\sum_{i=1}^r h_{ji}(a,u)\,u_i . The coefficients can be obtained by integrating ∂uiFr+j(a,tu)\partial_{u_i}F_{r+j}(a,t u) over 0≤t≤10\le t\le1. In the Koszul model the invertible change of generators ηr+j↦ηr+j−∑ihjiηi\eta_{r+j}\mapsto\eta_{r+j}-\sum_i h_{ji}\eta_i makes the new last m−rm-r generators closed. The first rr generators still resolve the coordinate plane. Taking their cohomology leaves the exterior algebra over the local functions of CC on m−rm-r generators of degree −1-1.

For the intersection of AA and BB, the degree −1-1 generators on CC span the dual of the cokernel of the normal derivative, namely E∗E^*. For example, a vector in the target dual annihilates the image of the derivative exactly when its linear combination of the differentials of the equations is zero on CC. The coordinate calculation then identifies the negative cohomology sheaves locally with exterior powers of E∗E^*. The calculation uses a splitting of the equations on the coordinate box. A splitting on each box does not by itself specify compatible choices on overlaps, so it supplies local models for gluing.

The Euler factor

Write uVu_V for the Thom class of an oriented real vector bundle VV, considered in the cohomology of its total space relative to the complement of the zero section. Its Euler class is e(V)=0V∗uVe(V)=0_V^*u_V under the zero section 0V0_V. Suppose the submanifolds AA and BB are closed as subsets of WW, and choose orientations of NC/BN_{C/B} and NA/WN_{A/W}. Orient EE by the exact sequence (1), taking the subbundle first. Let uA∈Hr(W,W∖A;Z)u_A\in H^r(W,W\setminus A;\mathbb Z) be the class obtained by excision from the oriented normal bundle of AA, where r=rank⁡(NA/W)r=\operatorname{rank}(N_{A/W}). Assume the ranks are constant on the components in question.

Theorem 1.2 (Clean excess intersection). Under excision to the normal bundle of CC in BB, the image of iB∗uAi_B^* u_A is uNC/B⌣p∗e(E),p:NC/B→C.u_{N_{C/B}}\smile p^*e(E), \qquad p:N_{C/B}\to C.

Proof. Choose tubular coordinates at AA in WW and at CC in BB as in Chapter [ch:tubular]. Shrink the latter tube to fit inside the preimage of the former under iBi_B. In these coordinates the inclusion is a map to NA/WN_{A/W} of the form (c,v)⟼(a(c,v), w(c,v)).(c,v)\longmapsto \big(a(c,v),\,w(c,v)\big). It sends v=0v=0 to the zero vector over cc, and w(c,v)≠0w(c,v)\ne0 for nonzero vv in the tube. Choose a radius disk bundle for the source tube, so that (c,tv)(c,tv) remains in the tube for 0≤t≤10\le t\le1. For t>0t>0 the maps (c,v)↦(a(c,tv),t−1w(c,tv))(c,v)\mapsto(a(c,tv),t^{-1}w(c,tv)) take values in the full normal bundle NA/WN_{A/W}. Taylor’s formula in local bundle frames extends these maps smoothly to t=0t=0, where the map is (c,v)↦(c,Jcv)(c,v)\mapsto(c,J_c v) for the inclusion JJ in (1). Each map sends the complement of the zero section into the complement. For t>0t>0 this uses the description of the preimage of AA. For t=0t=0 it uses the injectivity of JJ. Consequently the original pullback of the Thom class and its pullback along JJ agree as relative classes.

A bundle metric splits the exact sequence so that JJ is inclusion into NC/B⊕EN_{C/B}\oplus E as the first factor. The Thom class of a direct sum is the product of the Thom classes pulled back along the projections, with the given orientation order [3]. On the subbundle in question, the second projection is the zero section of EE after pp. Its pullback contributes p∗e(E)p^*e(E), giving the formula. ◻

The corresponding absolute class on BB is the Gysin image of e(E)e(E) under C↪BC\hookrightarrow B. The transverse formula is recovered since the Euler class of the zero bundle is the unit. Over F2\mathbb F_2, the same argument uses the canonical modulo-two Thom classes and dispenses with the orientation choices. The exact sequence still distinguishes a failure of transversality from the dimensions of the ordinary intersection.

Example 1.3 (Self-intersection of a diagonal). Let W=S2×S2W=S^2\times S^2 and let A=BA=B be the diagonal oriented by S2S^2. Its normal bundle is identified with TS2TS^2 by sending (v,w)(v,w) to w−vw-v. The orientation of this normal agrees with the coorientation induced by the product: in the ordered basis formed by (v,v)(v,v) followed by (0,w)(0,w), the change-of-basis matrix to the two factor bases has determinant one. The excess theorem gives the self-intersection number ⟨e(TS2),[S2]⟩\langle e(TS^2),[S^2]\rangle. For an explicit Euler calculation, project the constant unit vertical vector e3e_3 onto each tangent plane. The result is the section x↦e3−(e3⋅x)xx\mapsto e_3-(e_3\cdot x)x. Its two zeros are the poles. The tangent derivatives are −id⁡-\operatorname{id} at the north pole and id⁡\operatorname{id} at the south pole. Both preserve orientation in dimension two, giving total index 22. The ordinary set of self-intersection points is the whole diagonal. The normal excess determines its numerical contribution.

Tangencies and families

Consider the xx-axis and the graph of x2x^2 in R2\mathbb R^2. Their intersection consists of the origin, while the intersection of tangent spaces there is a line. The clean condition fails. Pulling the defining equation y=0y=0 back to the graph nevertheless gives the derived zero locus of t2t^2 in Example [ex:derived-smooth-square]. The function ring has a nonzero nilpotent linear coordinate, and the tangent cohomology has one dimension in each of degrees zero and one. The virtual dimension is zero.

The nearby family of graphs y=t2−λy=t^2-\lambda makes the variation visible. For λ>0\lambda>0 there are two transverse intersections, with opposite local signs in the normal coordinate yy, since the derivatives are 2λ2\sqrt{\lambda} and −2λ-2\sqrt{\lambda}. For λ<0\lambda<0 there are none. On a small interval in tt and for sufficiently small ∣λ∣|\lambda|, the total zero set meets neither vertical endpoint of the tt-interval. The equation on the parameter product is submersive, with derivative −1-1 in the λ\lambda direction. Its zero locus is the parabola λ=t2\lambda=t^2, and restriction to an interval about the critical parameter yields a compact oriented cobordism between the signed zero sets at the two regular endpoints. Hence the oriented number is zero on both sides while the cardinality changes. This is a direct local instance of the conservation of a signed intersection under a regular total-space deformation.

Here the total incidence manifold is smooth, but its map to the parameter has a critical point. The derived fiber at that point is still computed by the single equation after base change. This is a useful separation: transversality of the total incidence equation gives a smooth total space, and regularity of the projection governs which ordinary fibers vary smoothly. Derived base change retains a fiber at every parameter. The next chapter describes how the ordinary nearby topology can be recorded along different parameter strata.

Stratifications and exit paths

A parameter may cross a discriminant at which the regular fibers of an incidence map change. On each regular region one can transport nearby fibers along paths. At the discriminant one also needs comparisons from its strata into the nearby regions. Exit paths assemble both kinds of comparisons into a single category. The construction uses a stratified space as its source and can take values in spaces, retaining homotopies of the propagated data.

Conical neighborhoods and stratified simplices

Let II be a partially ordered set, with its open upward subsets, and let X→IX\to I be continuous. The fiber XaX_a over aa is the stratum with that label. We order the labels so that a point of a lower stratum may be approached from strata with higher labels. In the conical models used here, a neighborhood of a point of XaX_a is a product of a neighborhood in that stratum with the open cone on a link. The cone vertex has label aa, and the positive radial coordinates keep the labels of the link in I>aI_{>a}. For a compact link LL this cone is (L×[0,∞))/(L×{0})(L\times[0,\infty))/(L\times\{0\}). Its sections at fixed positive radius are copies of the link.

The exit-path simplicial set has simplices given by maps of a standard simplex to XX that follow a nondecreasing chain of labels a0≤⋯≤ana_0\le\cdots\le a_n. In barycentric coordinates (t0,…,tn)(t_0,\ldots,t_n) the point with largest nonzero index ii must map into XaiX_{a_i}. A one-simplex starts on its initial stratum and can enter a higher stratum continuously. A two-simplex supplies a homotopy compatible with such changes of stratum. Within one stratum this is the usual singular complex, and all its edges are invertible as paths. Between strata the allowed arrows are directional. We write Exit⁡(X)\operatorname{Exit}(X) for the resulting ∞\infty-category under conical hypotheses [19, Appendix A].

The cone on finitely many points is a useful first example. Give the vertex its own stratum and each positive open ray its own stratum. Between different rays there is no exit arrow because a path between them must return to the vertex. From the vertex to a fixed point of a ray there is a single exit arrow up to a contractible space of choices. Radial interpolation compares each such path to the straight one in that ray, preserving the initial vertex and the exit condition. In the stratified simplex model, interpolating radial coordinates preserves positivity on the prescribed positive faces. Thus the same comparison contracts choices through higher simplices. Each ray itself is contractible. This gives the category with an initial vertex and one object for each ray, and only the arrows from the vertex to those objects apart from identities. This is the finite discrete link case of the cone computation in [1, Lemma 3.3.5].

Constructible sheaves and their transport

A sheaf of spaces on XX is stratified constructible if its restriction to each stratum is locally constant. For clarity, one version of the exit-path theorem assumes that XX is paracompact, locally of singular shape, and conically stratified with respect to a poset satisfying the ascending chain condition. Locally of singular shape means that the shape of the sheaf topos on each open set is given by its singular homotopy type. In particular the usual manifold and finite conical examples below satisfy the hypotheses. In this setting the theorem gives an equivalence Shv⁡cstr(X;S)≃Fun⁡(Exit⁡(X),S)\operatorname{Shv}_{\mathrm{cstr}}(X;\mathcal S) \simeq \operatorname{Fun}\bigl(\operatorname{Exit}(X),\mathcal S\bigr) between constructible sheaves and functors to spaces [19, Theorem A.9.3]. One direction sends a constructible sheaf to its stalks, with transport along the exit arrows. The other assigns to an open set the space of coherent sections of the corresponding left fibration over its exit category. Ayala, Francis, and Rozenblyum develop a related theory using conically smooth stratifications and describe the descent of the exit construction itself [1]. The stated assumptions are part of the comparison.

For a single stratum, the exit category is the fundamental ∞\infty-groupoid, so the equivalence recovers local systems of spaces. For the cone on finitely many points, it gives a value VV at the vertex, a value WjW_j at each ray, and a map V→WjV\to W_j. In the ordinary sheaf case this follows directly using a basis of small stars at the origin and intervals in the rays. Restrictions between nested sufficiently small stars induce the same vertex stalk. Restrictions to each ray are the respective specialization maps and restrictions inside the ray are locally constant. Conversely, the indicated diagram defines the restrictions on a basis. On an overlap inside a ray sections agree precisely when their images in that ray agree, which is the ordinary sheaf gluing condition. The topological cone is contractible, but direction and specialization persist in its stratification.

A real fold and a complex branch

Let π:R→R\pi:\mathbb R\to\mathbb R be u↦u2u\mapsto u^2, with the target stratified as the origin and the two complementary rays. For a field kk, push the constant sheaf of kk-valued locally constant functions forward along π\pi. The resulting sheaf on the parameter line has stalks 00 on the negative ray and k2k^2 on the positive ray. This follows by counting components of the preimages of small intervals. A small interval containing the origin has connected preimage, so its stalk there is kk. Restricting from that interval to the positive side sends one constant to the same value on both components. Its exit diagram is image There is another constructible sheaf with the same three stalks and zero specialization on the positive side. The star calculation constructs it from that diagram. These two sheaves are inequivalent because the ranks of their specialization maps differ. Even matching the dimension of every nearby stalk leaves the gluing undetermined.

Over the complex line consider z↦z2z\mapsto z^2. On the punctured plane this is a double covering: a choice of one square root continues uniquely along a path and changes sign along one positive loop around the origin. The pushforward has stalk k2k^2 with the two factors exchanged by this monodromy. At the origin its value is kk, since the preimage of a small disk is again a disk. An exit along a radial segment again induces the diagonal map to k2k^2. Its image is fixed by the exchange, which gives compatibility of specialization with the loop. For example, the sections on the punctured disk have the same constant value on its two nearby sheets after transport around a full turn. The regular stalk dimension alone does not detect this condition.

In the real case the total incidence is the ordinary parabola λ=u2\lambda=u^2 from Chapter [ch:derived-intersection]. The exceptional derived fiber also remembers the relation u2=0u^2=0 in its local functions. The exit diagram of the pushforward above is instead computed from locally constant functions on sets of preimages. Both structures can be associated to the same family, with the exit diagram recording transport of nearby observations and the derived fiber retaining its infinitesimal equations.

Ordinary transport on a regular region

A condition for obtaining local systems from an incidence space is properness over its regular region. Let q:E→Vq:E\to V be a proper surjective submersion of smooth manifolds without boundary. For any Riemannian metric on EE, the orthogonal complement of ker⁡Dq\ker Dq is a horizontal bundle. Its projection to TVTV is a vector bundle isomorphism over EE. Take a small coordinate ball about v∈Vv\in V with compact closure in a larger coordinate chart. To transport a point of q−1(v)q^{-1}(v) to another fiber over the ball, lift the straight line from vv horizontally. This is an ordinary differential equation on EE using the horizontal lift of the coordinate velocity. Its projected path solves the chosen equation on the base. As long as its projected path remains in the closed ball, its lift stays in the compact inverse image of the ball’s closure. Compactness supplies existence for the whole time interval. Lifting the reversed straight line gives the inverse map to the original fiber. Smooth dependence on initial points and velocities identifies the restriction of qq to the ball with a product. This is the local trivialization in the proper submersion theorem.

Local trivializations supply transport of the homotopy types and cohomology of the fibers along regular paths. It is locally constant because each trivialization uses the same fiber throughout its ball. The fold restricts on the positive ray to such a locally trivial two-point family. The origin has no ordinary submersion chart and its comparison with that family uses the separate specialization map.

Properness is significant here. The disjoint union R⨿(R∖{0})\mathbb R\amalg(\mathbb R\setminus\{0\}) maps to R\mathbb R by the identity on each summand. The map is a surjective submersion, and its fiber cardinalities still change at zero. The missing point in the second summand prevents the horizontal transport argument from extending to that fiber. This example separates local regularity of individual points of the incidence space from control of a whole family of fibers.

Specialization between local systems

On S1×[0,∞)S^1\times[0,\infty), stratify by the lower circle and its open complement. In the stratified simplex construction, the label of a point is determined by its second coordinate. Consequently exit simplices in the product are pairs of ordinary singular simplices on S1S^1 and exit simplices of the half-line. The half-line has the exit category of the one-ray cone, hence is equivalent to the arrow category [1][1]. We obtain Exit⁡(S1×[0,∞))≃Π∞(S1)×[1].\operatorname{Exit}\bigl(S^1\times[0,\infty)\bigr) \simeq \Pi_\infty(S^1)\times[1]. A constructible sheaf of spaces on this cylinder specifies two local systems and a natural transformation from the lower one to the upper one. For vector spaces, writing their loop transport as AA and BB, a specialization ss satisfies Bs=sABs=sA. Indeed the two ways to move once around the circle and pass to the upper stratum form the corresponding commuting square. For sheaves of spaces the comparison is coherent along all paths and homotopies in the circle. This illustrates why even simple discriminant surfaces can require comparison maps between representations of fundamental groups, rather than separate lists of regular and singular fiber types.

Spaces of immersions and embeddings

The immersion condition depends on the first derivative at one point. The embedding condition also compares pairs of points. This distinction persists in families: rank conditions give a space of formal immersion data, while a theory of embeddings must also track configurations. We begin with the parametric form of immersion theory for manifolds without boundary.

Formal immersions and regular homotopy

For Mm,NnM^m,N^n, a formal immersion is a pair (f,F)(f,F) consisting of a smooth map f:M→Nf:M\to N and a smooth vector bundle monomorphism F:TM→f∗TNF:TM\to f^*TN. It is holonomic when F=DfF=Df. Write FImm⁡(M,N)\operatorname{FImm}(M,N) for its space with the compact-open smooth topology on the maps and their bundle data, and Imm⁡(M,N)\operatorname{Imm}(M,N) for the corresponding space of immersions. Equivalently the formal space is the space of sections over MM of the bundle whose fiber at xx consists of (y,L)(y,L) with L:TxM↪TyNL:T_xM\hookrightarrow T_yN. The differential gives the section corresponding to an immersion.

Theorem 1.1 (Smale–Hirsch). For smooth manifolds without boundary and n>mn>m, the map Imm⁡(M,N)⟶FImm⁡(M,N),f⟼(f,Df),\operatorname{Imm}(M,N)\longrightarrow \operatorname{FImm}(M,N),\qquad f\longmapsto(f,Df), is a weak homotopy equivalence.

This is the parametric immersion theorem [16, 33]. It identifies the homotopy types of families of immersions and formal monomorphisms in the stated dimension range. In particular it gives a regular homotopy classification on path components. A relative consequence for parameters follows from factoring the differential map through its mapping path fibration. The fibration has weakly contractible fibers. A family of genuine immersions on a subcomplex of a finite CW parameter space, with matching formal data on the full parameter space, lifts to that fibration. The lift extends from the subcomplex by successive disk lifts. Projection back gives genuine immersions on the full parameter space, together with a homotopy of their differentials to the prescribed formal family, fixed on the subcomplex. This conclusion is about continuous families for the topology in the theorem.

Example 1.2 (Plane curves). Trivializations of TS1TS^1 and TR2T\mathbb R^2 identify a formal derivative of a parametrized plane curve with a map S1→R2∖{0}S^1\to\mathbb R^2\setminus\{0\}. The underlying map to R2\mathbb R^2 can be contracted independently. Radial normalization and Theorem 1.1 classify path components of the immersion space by degrees of maps S1→S1S^1\to S^1, or tangent winding numbers. For instance t↦(sin⁡t,sin⁡2t)t\mapsto(\sin t,\sin 2t) is an immersion: cos⁡t=0\cos t=0 forces cos⁡2t=−1\cos 2t=-1. Its normalized tangent at t+πt+\pi is the reflection of its tangent at tt in the vertical axis. Translation of the parameter is orientation preserving, while reflection of the tangent circle reverses orientation. Thus the winding number equals its own negative and is zero. A smooth plane embedding of the circle has winding number 11 or −1-1 by the turning tangent theorem [29, Lecture 9]. The regular homotopy component of this figure eight therefore contains no embeddings.

Tests on pairs of points

Put M[2]=(M×M)∖ΔMM^{[2]}=(M\times M)\setminus\Delta_M. For a map f:M→Nf:M\to N, injectivity means that (x,y)↦(f(x),f(y))(x,y)\mapsto(f(x),f(y)) avoids ΔN\Delta_N on M[2]M^{[2]}. If MM is compact, an injective immersion is an embedding: it is a homeomorphism onto its image by compactness and the Hausdorff property, and its local immersion charts then give the smooth embedded structure. The deleted product expresses the pairwise test independently of the first-jet condition.

For compact MM the embedding locus is open in the C1C^1 topology. To see this, let ff be an embedding. At each point choose coordinates on a small source ball and on the target such that projection of the coordinate image of ff to its first mm coordinates has invertible derivative at the center. Shrink the source ball until this derivative is uniformly close to its center value. For every sufficiently C1C^1-close map gg, the projected coordinate derivative stays close to the same invertible linear map. Integrating along line segments in the ball then bounds the distance between the projected images of distinct points below by a positive multiple of their coordinate distance. This gives injectivity and full rank on that ball. Finitely many such balls cover MM. Pairs sharing a ball constitute a neighborhood of the diagonal in M×MM\times M. On the compact remaining set of pairs, the distances between their ff-images have a positive minimum. A small uniform perturbation keeps those images distinct as well. This proves openness and explains how local derivative margins and global pair separation give different stability conditions.

Proposition 1.3 (Generic embedding slices of a family). Let MM be compact of dimension mm and let f:P×M→Nnf:P\times M\to N^n be a smooth family, with P,M,NP,M,N finite-dimensional, second-countable manifolds without boundary. Assume n>2mn>2m. Suppose the vertical first jet (p,x)↦j1(fp)(x)(p,x)\mapsto j^1(f_p)(x) is transverse to each rank-deficient stratum in J1(M,N)J^1(M,N), and the map (p,x,y)⟼(fp(x),fp(y)) ⁣:P×M[2]⟶N×N(p,x,y)\longmapsto(f_p(x),f_p(y))\colon P\times M^{[2]}\longrightarrow N\times N is transverse to ΔN\Delta_N. Then fpf_p is an embedding outside a set of parameters of measure zero in local charts.

Proof. For each total map take the inverse image of the specified stratum or diagonal, a smooth submanifold by transversality. Sard’s theorem applied to its projection to PP gives a full-measure set of regular parameters. Regularity of this projection is equivalent to transversality in the slice at that parameter. To check the equivalence, write A,BA,B for the parameter and source parts of the derivative into the normal quotient to the specified submanifold. Total transversality means that im⁡A+im⁡B\operatorname{im}A+\operatorname{im}B is that quotient. Projection of the kernel onto the parameter tangent space is onto exactly when every AA-image lies in im⁡B\operatorname{im}B, which under total transversality is precisely surjectivity of BB.

There are finitely many rank strata. The stratum of rank k<mk<m in the space of linear maps Rm→Rn\mathbb R^m\to\mathbb R^n has codimension (m−k)(n−k)(m-k)(n-k): in a chart where a kk-minor is invertible, the remaining block is prescribed by its Schur complement. This also gives the corresponding codimension in jet coordinates. It is at least n−m+1n-m+1, which exceeds mm under the stated inequality. Therefore a transverse slice avoids every deficient stratum. The diagonal has codimension n>2mn>2m, so its transverse preimage in the 2m2m-dimensional deleted product is also empty. The slice is an injective immersion, and compactness gives the embedding. ◻

One can verify the diagonal hypothesis by showing that the parameter derivatives can move the difference of the two outputs in all normal directions at each collision. This requires independent control at the pair under consideration. A transversality theorem for the unrestricted smooth mapping space gives no such control for a prescribed family of finite dimension. For example the family fp(t)=(pcos⁡(2t),psin⁡(2t)),p∈R∖{0},f_p(t)=\bigl(p\cos(2t),p\sin(2t)\bigr), \qquad p\in\mathbb R\setminus\{0\}, immerses the circle for every pp, and it identifies tt with t+πt+\pi for every pp. Embedding the targets in a higher-dimensional Euclidean space by zero padding leaves this obstruction intact. Adding two output coordinates acos⁡t,asin⁡ta\cos t,a\sin t with a≠0a\ne0 yields an embedding, since the added coordinates separate points and already form an immersion. In an architecture using trigonometric features, the choice of features therefore changes the class of topological realizations before the weights are selected.

The configuration stages

Let O(U)\mathcal O(U) be the poset of open subsets of an open manifold domain U⊂MU\subset M, and let Ok(U)\mathcal O_k(U) be its subposet of opens diffeomorphic to unions of at most kk disjoint open balls. Embedding calculus studies the restriction functor V↦Emb⁡(V,N)V\mapsto\operatorname{Emb}(V,N). Its approximations can be expressed by the homotopy right Kan extensions TkEmb⁡(U,N)=holim⁡V∈Ok(U)opEmb⁡(V,N).T_k\operatorname{Emb}(U,N)= \operatorname*{holim}_{V\in\mathcal O_k(U)^{\mathrm{op}}} \operatorname{Emb}(V,N). Restriction from UU gives the comparison to each stage. Increasing kk adds simultaneous compatibility on configurations spread across more balls. In the range n>mn>m, the first stage is weakly equivalent to the immersion space [33]. One way to understand the first stage is that one-ball restrictions of embeddings and immersions have the same local homotopy information, while immersion spaces satisfy homotopy descent for the ordinary open covers used at that stage. Later stages retain compatibility beyond this one-point condition. image The convergence theorem in codimension at least three identifies the embedding space for a compact source without boundary with the homotopy inverse limit of its stages [12]. Both the codimension condition and the passage to the limit matter. The rank-and-pair test gives a criterion on an actual smooth map in a suitable dimension range. The tower compares entire spaces of embeddings and keeps their homotopies.

For families, an actual map from a parameter space into Emb⁡(M,N)\operatorname{Emb}(M,N) induces maps to all of these stages and to the formal immersion space. This expresses the comparison using functors between spaces of models. Applying the same construction to restrictions of the parameter commutes with the maps in the tower. Imposing an architecture supplies a particular subset or family of realizations inside these mapping spaces. The topology of its immersion and embedding loci depends on how that realization map meets the topology detected by differentials and configurations.

Compact group actions and orbit types

The stabilizer of a point determines which symmetries act on its tangent and normal spaces. This linear information also governs a neighborhood of the orbit. We use it to describe what remains when orbits are treated as objects with automorphisms, and to locate the strata on which ordinary smooth quotient charts suffice.

Throughout this chapter, GG is a compact Lie group acting smoothly on a Hausdorff second-countable manifold MM without boundary. We write Gx={g∈G:gx=x}G_x=\{g\in G:gx=x\} and Gx={gx:g∈G}Gx=\{gx:g\in G\}. An equivariant map f:M→Nf:M\to N satisfies f(gx)=gf(x)f(gx)=gf(x). Differentiation at xx gives Dgxf∘Dxg=Df(x)g∘Dxf.D_{gx}f\circ D_xg=D_{f(x)}g\circ D_xf. In particular, Gx⊆Gf(x)G_x\subseteq G_{f(x)}, and DxfD_xf is an intertwiner for GxG_x. This condition persists for smooth families, with the stabilizer taken in the total source when parameters themselves move.

Invariant metrics and fixed points

Choose a Riemannian metric hh on MM. With probability Haar measure dgdg, set hˉx(v,w)=∫Ghgx(Dxg(v),Dxg(w)) dg.\bar h_x(v,w)= \int_G h_{gx}\bigl(D_xg(v),D_xg(w)\bigr)\,dg . Compactness permits differentiation in xx in local charts. The average of the positive functions obtained by setting w=v≠0w=v\ne0 is positive, and right invariance of Haar measure makes hˉ\bar h invariant. Every element of GG therefore acts by isometries. Its action commutes with the exponential map wherever that map is defined.

Proposition 1.1 (Tangent spaces to fixed loci). Let KK be a closed subgroup of GG. The set MKM^K of points fixed by every element of KK is a closed smooth submanifold, allowing different dimensions on different connected components. At x∈MKx\in M^K its tangent space is (TxM)K(T_xM)^K.

Proof. An exponential chart at xx identifies an invariant small ball in TxMT_xM with a neighborhood of xx. The chart is KK-equivariant because KK fixes xx and acts isometrically. If a point of its image is fixed, injectivity implies that its tangent coordinate is fixed by KK. Conversely an exponential of a fixed vector is fixed. The fixed vectors form a linear subspace, so these charts prove both the local submanifold assertion and the tangent formula. Closedness follows by intersecting the closed fixed sets of individual elements. ◻

For example, S1S^1 acts on S2S^2 by rotation about its vertical axis. The two poles form the fixed locus and their entire tangent planes are rotating representations. The nonpolar points have trivial stabilizer. A tangent bundle on the orbit space alone therefore cannot account for the two polar tangent representations.

A slice from the normal bundle

Theorem 1.2 (Slice at a compact orbit). For x∈Mx\in M, put H=GxH=G_x and Vx=(TxGx)⊥⊂TxMV_x=(T_xGx)^\perp\subset T_xM for an invariant metric. There is a ball D⊂VxD\subset V_x about 00 and an equivariant diffeomorphism G×HD → ∼  U,[g,v]⟼gexp⁡x(v),G\times_H D \ \xrightarrow{\ \sim\ }\ U,\qquad [g,v]\longmapsto g\exp_x(v), where U⊂MU\subset M is an invariant open neighborhood of the orbit. The balanced product identifies (gh,v)(gh,v) with (g,hv)(g,hv).

Proof. The subgroup HH is closed. The orbit map induces an injective map G/H→MG/H\to M. The differential of G→MG\to M at the identity sends a Lie algebra element ξ\xi to the velocity of t↦exp⁡(tξ)xt\mapsto\exp(t\xi)x. A zero velocity means that the resulting fundamental vector field vanishes at xx. Uniqueness for the flow makes the entire one-parameter subgroup fix xx, so the kernel is the Lie algebra of HH. Translating establishes that G/H→MG/H\to M is an immersion, and compactness of G/HG/H makes it an embedding.

The normal spaces of the orbit at gxgx are the translates of VxV_x by DxgD_xg. Thus its normal bundle is G×HVxG\times_H V_x, via [g,v]↦Dxg(v)[g,v]\mapsto D_xg(v). The tubular-neighborhood construction of Chapter [ch:tubular] applies to the normal exponential. Compactness of GxGx allows a single small radius, and the metric makes the resulting diffeomorphism equivariant. Its fiber over xx is the ball DD. ◻

This proof is the compact-group case of the slice theorem. Proper Lie group actions admit a more general version [23]. In the slice above, the stabilizer of [g,v][g,v] is gHvg−1gH_vg^{-1} with Hv={h∈H:hv=v}H_v=\{h\in H:hv=v\}. Every nearby stabilizer is therefore conjugate to a subgroup of HH. Moreover, the points of the same orbit type as xx, meaning that their stabilizer is conjugate to HH, correspond locally to G×HDHG\times_H D^H. Indeed, if Hv⊆HH_v\subseteq H is conjugate in GG to HH, it has the same dimension and number of components as HH. The inclusion then identifies their identity components and reaches each component of HH, hence Hv=HH_v=H. The local orbit-type stratum is (G/H)×DH(G/H)\times D^H, with orbit quotient DHD^H. The entire neighborhood has orbit quotient D/HD/H. In particular a free action has charts with H=1H=1, giving a smooth quotient and a principal bundle over it.

There is also a reduction of the symmetry group along any fixed orbit type. Fix a closed subgroup HH which occurs as a stabilizer and put MH={x:Gx=H}M_H=\{x:G_x=H\}. In MHM^H, this set is open. To see this near x∈MHx\in M_H, a nearby stabilizer is conjugate to a subgroup of HH by the slice theorem. If it contains HH, equality follows by comparing dimensions and component counts as above. The normalizer NG(H)={g:gHg−1=H}N_G(H)=\{g:gHg^{-1}=H\} preserves MHM_H, and NG(H)/HN_G(H)/H acts freely on it. The GG-orbit of any point of type (H)(H) meets MHM_H, and two points of MHM_H in a common GG-orbit differ by an element of NG(H)N_G(H). It follows that M(H)/GM_{(H)}/G identifies with MH/(NG(H)/H)M_H/(N_G(H)/H). In the slice at a point of MHM_H, the exact-stabilizer locus is NG(H)×HDHN_G(H)\times_H D^H: a point [g,v][g,v] has stabilizer HH precisely when v∈DHv\in D^H and g∈NG(H)g\in N_G(H). Thus both quotient charts identify with DHD^H, proving that the identification is smooth for their quotient manifold structures. This describes the global quotient within one type using an action with free orbits.

The quotient stack in slice coordinates

Write [M/G][M/G] for the smooth quotient stack. Its objects over a test manifold TT are principal right GG-bundles P→TP\to T with a smooth map F:P→MF:P\to M satisfying F(pg)=g−1F(p)F(pg)=g^{-1}F(p). Morphisms are bundle isomorphisms commuting with these maps. Restricting to the neighborhood of Theorem 1.2 gives image To prove the lower equivalence, use the equivariant projection U→G/HU\to G/H. The inverse image of the coset HH under P→FU→G/HP\xrightarrow{F}U\to G/H is a principal right HH-bundle Q⊂PQ\subset P. Locally, a section of PP and a smooth lift of the resulting map into G/HG/H produce a section of QQ. The restriction F∣QF|_Q lands in the fiber DD. In the other direction a bundle QQ with equivariant u:Q→Du:Q\to D extends to Q×HGQ\times_H G with map [q,g]↦g−1[1,u(q)][q,g]\mapsto g^{-1}[1,u(q)]. The two constructions recover each other and their bundle isomorphisms. In particular they commute with pullback in TT.

The infinitesimal action at xx is the linear map g⟶TxM,ξ⟼ddt∣t=0exp⁡(tξ)x.\mathfrak g \longrightarrow T_xM,\qquad \xi\longmapsto \left.\frac{d}{dt}\right|_{t=0}\exp(t\xi)x . Its kernel is h=Lie⁡(H)\mathfrak h=\operatorname{Lie}(H) and its cokernel is the normal representation VxV_x. They are the cohomology of the tangent complex of the action groupoid in degrees −1-1 and 00, respectively. The reason for the complex is that an infinitesimal group element is an arrow from a displacement vv to v+ξM(x)v+\xi_M(x). Quotienting the displacement gives the cokernel, and arrows from a displacement to itself give the kernel. A finite stabilizer has zero Lie algebra, so the tangent complex can miss finite automorphisms even when the slice stack retains them.

Example 1.3 (An exceptional orbit with a disk quotient). Let S1S^1 act on S3⊂C2S^3\subset\mathbb C^2 by λ(z1,z2)=(λz1,λ2z2)\lambda(z_1,z_2)=(\lambda z_1,\lambda^2z_2). A point with z1≠0z_1\ne0 has trivial stabilizer. At x=(0,1)x=(0,1) the stabilizer is {1,−1}\{1,-1\} and the orbit tangent is the imaginary z2z_2-axis. Its orthogonal slice is the complex z1z_1-line, with −1-1 acting by multiplication by −1-1. The coarse chart D/{1,−1}D/\{1,-1\} is topologically a disk, by the map z↦z2z\mapsto z^2, whereas [D/{1,−1}][D/\{1,-1\}] assigns the order-two automorphism group to the central point. Both the Lie algebra of the stabilizer and the local coarse topology would leave that finite group undetected.

The slice converts a question about changing orbit types into one about linear representations. The next chapter uses this reduction for a local obstruction: the normal derivative of an equivariant map has to be surjective within an appropriate category of representations.

Equivariant transversality and fixed-locus constraints

The ordinary transversality argument perturbs a map in directions spanning the target normal space. Equivariance restricts those directions to intertwiners of stabilizer representations. At a fixed point even a tiny invariant target may require a normal representation absent from the source. We identify the obstruction at a single orbit, then show how the ordinary density argument returns for free source actions.

The normal map at a stabilizer

Let GG be a compact Lie group acting smoothly on manifolds M,NM,N without boundary. Let A⊂NA\subset N be a GG-invariant embedded submanifold and f:M→Nf:M\to N an equivariant smooth map. Fix x∈Mx\in M with y=f(x)∈Ay=f(x)\in A, set H=GxH=G_x, and let NyA=TyN/TyAN_yA=T_yN/T_yA denote the fiber of the normal bundle. The classical transversality condition at xx is surjectivity of TxM→DxfTyN→NyAT_xM\xrightarrow{D_xf}T_yN\to N_yA. This map is HH-equivariant. Its restriction to TxGxT_xGx vanishes, since f(Gx)⊆Af(Gx)\subseteq A. It therefore descends to an HH-map Dxf‾:TxM/TxGx⟶NyA.\overline{D_x f}:T_xM/T_xGx\longrightarrow N_yA .

Proposition 1.1 (Transverse germs at an orbit). Let x∈Mx\in M, y∈Ay\in A, and assume H=GxH=G_x fixes yy. There exists an equivariant map on an invariant neighborhood of GxGx taking xx to yy and transverse to AA along GxGx if and only if TxM/TxGxT_xM/T_xGx admits a surjective HH-linear map to NyAN_yA.

Proof. Any such map gives the stated normal derivative. Conversely choose invariant metrics and identify the two quotient spaces with orthogonal complements. Let L:Vx→(TyA)⊥L:V_x\to (T_yA)^\perp be a surjective HH-linear map. In the slice chart of Theorem [thm:compact-slice], use [g,v]⟼gexp⁡y(Lv).[g,v]\longmapsto g\exp_y(Lv). For a small enough ball this exponential is defined. Because yy is fixed by HH, the chart at yy commutes with the HH-action, making the formula well defined. Its normal derivative on the slice at v=0v=0 is LL, and translation carries surjectivity to each point of the orbit. ◻

There is a local perturbation version of the same criterion. Fix an equivariant ff with f(x)=yf(x)=y. On a smaller slice ball the expression u(v)=exp⁡y−1(f(exp⁡xv))u(v)=\exp_y^{-1}(f(\exp_x v)) is defined and is HH-equivariant. Choose an invariant bump χ\chi on this ball, equal to one near zero. Replacing u(v)u(v) by u(v)+tχ(v)Lvu(v)+t\chi(v)Lv, where LL is a surjection as in the proof, gives a smooth perturbation through yy with support near GxGx. Extend it by equivariance on the tube and by the old ff off the support. Its normal derivative at xx is Dxf‾+tL\overline{D_x f}+tL. A full-rank minor of LL provides a polynomial minor of this derivative with nonzero leading coefficient. Hence for arbitrarily small tt the new germ is transverse along GxGx. When the normal fiber is zero the original germ is already transverse. The construction concerns a single orbit and leaves global density as a separate question.

The fixed-point loci MH,NH,AHM^H,N^H,A^H are submanifolds by Chapter [ch:orbit-types]. Invariant metrics split the normal sequence, yielding TyNH/TyAH ≅ (NyA)H.T_yN^H/T_yA^H\ \cong\ (N_yA)^H . If ff is transverse at xx, its normal derivative is surjective on fixed vectors. In fact, averaging any preimage of an invariant vector gives an invariant preimage. Thus the restricted map f∣MH:MH→NHf|_{M^H}:M^H\to N^H is transverse to AHA^H at xx. Testing only the restriction for the full stabilizer HH tests only the trivial part of the normal representation.

Example 1.2 (A persistent singular zero). Let M=CM=\mathbb C have the S1S^1-action λz\lambda z, and N=CN=\mathbb C the action λ2w\lambda^2 w. Put A={0}A=\{0\}. Every equivariant map sends 00 to 00, since the only point of NN fixed by the full group is 00. The element −1-1 acts by −1-1 on T0MT_0M and by +1+1 on T0NT_0N. Hence D0f(−v)=D0f(v)D_0f(-v)=D_0f(v) and real linearity gives D0f=0D_0f=0. No equivariant perturbation achieves ordinary transversality at the origin. The fixed-locus restriction is a map from a point to a point and is transverse there. For the concrete map z↦z2z\mapsto z^2, the zero set is a single point, even though the linear normal test fails in both real directions.

Example 1.3 (Moving a parameter changes the constraint). Let P=RP=\mathbb R and N=RN=\mathbb R with the sign action of C2C_2 on each. The evaluation P→N, p↦pP\to N,\ p\mapsto p at a one-point input is an equivariant submersion. At the fixed parameter 00 the constant evaluation at the sole input hits {0}\{0\} and fails transversality. For p≠0p\ne0, the evaluation misses {0}\{0\}. Holding p≠0p\ne0 fixed leaves no sign symmetry of the parameter. That is why ordinary parameterized transversality can hold for the family while the restriction to fixed parameters has no transverse member.

Recovery for free source actions

Theorem 1.4 (Equivariant density for a free compact source). Let MM be compact without boundary with a free smooth GG-action, and let A⊂NA\subset N be invariant, closed, and embedded. In the smooth equivariant mapping space from MM to NN, the maps transverse to AA are dense and open in the C∞C^\infty topology.

Proof. Write B=M/GB=M/G. The slice theorem supplies the smooth principal bundle M→BM\to B. Equivariant maps M→NM\to N correspond to sections of the associated bundle E=(M×N)/G→BE=(M\times N)/G\to B, where GG acts diagonally. In a principal bundle chart an equivariant map is (g,b)↦g s(b)(g,b)\mapsto g\,s(b), with s(b)s(b) the fiber coordinate of the section. Its normal derivative to AA vanishes in the group directions and is the normal derivative of s(b)s(b) in the base directions, up to translation by gg. Thus transversality of the map to AA is equivalent to transversality of its section to EA=(M×A)/G⊂EE_A=(M\times A)/G\subset E.

Fix a section ss. Its image is compact. Cover the image by finitely many local bundle and fiber coordinate charts, with smaller charts still covering it and bump functions supported in the larger charts. Multiplying fiber coordinate vector fields by these bumps gives globally defined compactly supported vertical vector fields V1,…,VℓV_1,\ldots,V_\ell spanning the vertical space at every point of s(B)s(B). Their flows preserve the projection and can be composed to give sections sts_t for t∈Rℓt\in\mathbb R^\ell. The evaluation (b,t)↦st(b)(b,t)\mapsto s_t(b) has at t=0t=0 a differential whose parameter directions span the vertical tangent, while its base directions project onto TbBT_bB. It is a submersion on B×UB\times U for a sufficiently small open neighborhood UU of zero, by continuity and compactness. Parametric transversality from Chapter [ch:transversality] supplies transverse sections sts_t for parameters arbitrarily close to zero. The flow construction converges to ss in the smooth topology as t→0t\to0. Transversality is open in the C1C^1 topology on the compact source because AA is closed. Any sequence of failed normal-surjectivity tests would have base points with a convergent subsequence and would give a failed test for the limiting map. These statements transfer to equivariant maps by the local bundle charts. ◻

The proof uses perturbations of sections of a bundle on a manifold. When stabilizers occur, admissible perturbations have to respect their linear actions and the same submersion argument may become impossible. For related equivariant general position questions see [31]. The classical parametric argument used in the proof is developed in [15, Ch. 3].

Pullbacks of quotient stacks

For an everywhere transverse equivariant map, Z=f−1(A)Z=f^{-1}(A) is an invariant smooth submanifold. Its quotient stack fits into image This is a homotopy pullback of smooth stacks. Over a test manifold an object of the pullback consists of two principal bundles with equivariant maps to MM and AA, together with an isomorphism between the images in [N/G][N/G]. Use the isomorphism to identify the bundles. Their two maps then have the same composite to NN. The common bundle map lands in the ordinary transverse fiber product M×NAM\times_N A, identified with ZZ, and the construction also matches bundle isomorphisms. Taking quotient stacks hence retains the stabilizer at each point of the transverse solution set. The derived pullbacks of Chapter [ch:derived-intersection] give another question to ask when the normal map fails to be surjective: which tangent and obstruction directions belong to the intersection with its equations retained?

Factorization through an intermediate manifold

A smooth map f:X→Yf:X\to Y may be specified by two maps through an intermediate manifold LL. The factorization retains how the map is constructed. Rank bounds give local restrictions, and homotopy invariants can rule out a factorization even when every local rank test passes. We will also distinguish a pointwise choice of factorization from a continuous family of choices.

Differential constraints

Suppose f=d∘ef=d\circ e with e:X→Le:X\to L and d:L→Yd:L\to Y. The chain rule says that DxfD_xf factors as TxX→DxeTe(x)L→De(x)dTf(x)Y.T_xX\xrightarrow{D_xe}T_{e(x)}L\xrightarrow{D_{e(x)}d}T_{f(x)}Y. Consequently rank⁡Dxf≤dim⁡L\operatorname{rank}D_xf\le\dim L. The constant-rank normal form gives a local converse at points of an open region where the rank is some fixed rr: in suitable coordinates the map is the projection to rr coordinates followed by the inclusion of those coordinates in the target chart. This describes a local factorization through an open subset of Rr\mathbb R^r. The coordinate projections on two such neighborhoods need not provide one global intermediary.

Proposition 1.1 (An exact reconstruction embeds the source). If XX is compact and d∘e=idXd\circ e=\mathrm{id}_X with smooth e:X→Le:X\to L and d:L→Xd:L\to X, then ee is a smooth embedding. If XX and LL have the same dimension, both have no boundary, and LL is connected, then ee is a diffeomorphism for nonempty XX.

Proof. The left-inverse identity makes ee injective, and differentiating it makes DeDe injective at every point. Compactness of XX and the Hausdorff property of LL give a homeomorphism onto the image. Together with immersion, this is the embedding condition. In equal dimensions the inverse function theorem makes e(X)e(X) open. It is also compact, hence closed, and connectedness of LL implies e(X)=Le(X)=L. The local smooth inverses then give the global inverse. ◻

The derivative already imposes a representation constraint when a group acts on the spaces and both maps are equivariant. At a fixed input xx, the identity factors through Te(x)LT_{e(x)}L by two intertwiners. Thus TxXT_xX must occur as a split subrepresentation for the stabilizer of xx. This may exclude an intermediate manifold of sufficient ordinary dimension.

Conversely, every smooth embedding of compact XX in LL admits an exact reconstruction defined on a neighborhood of the encoded image. By the tubular-neighborhood theorem, a small neighborhood UU of e(X)e(X) is identified with a neighborhood of the zero section in its normal bundle. Projection from the bundle to e(X)e(X), followed by e−1e^{-1}, gives a smooth decoder U→XU\to X. The existence of a decoder on all of LL asks for an extension of a left inverse. A local normal form alone only supplies the neighborhood decoder, even when the embedded image is closed. For example, the usual inclusion of SnS^n in Rn+1\mathbb R^{n+1} has the radial decoder on an annulus, yet a global sphere-valued decoder would induce an impossible retraction on top cohomology.

Obstructions preserved by homotopy

Suppose now that d∘ed\circ e is homotopic to idX\mathrm{id}_X. For any commutative coefficient ring RR, functoriality in singular cohomology gives unital graded ring maps H∗(X;R)→d∗H∗(L;R)→e∗H∗(X;R),e∗d∗=id.H^*(X;R)\xrightarrow{d^*}H^*(L;R) \xrightarrow{e^*}H^*(X;R),\qquad e^*d^*=\mathrm{id}. A nonzero cup product in H∗(X;R)H^*(X;R) must remain nonzero after applying d∗d^*. Rank tests on the groups HkH^k alone can miss this restriction. Similarly, on homotopy groups a path supplied by the homotopy allows the base point of d∘ed\circ e to be compared with the starting point in XX. With that comparison in place, the map on each homotopy group induced by ee has a left inverse induced by dd. See [13] for the cohomology and base-point conventions.

For a closed connected oriented nn-manifold XX, with n>0n>0, a degree-one reconstruction of XX induces a nonzero map on Hn(X;Z)H^n(X;\mathbb Z). It cannot factor through a manifold of dimension less than nn, since that manifold has no cohomology in degree nn. Here one can use a triangulation of the intermediate smooth manifold and its dimension bound. This is a homotopy obstruction, so it also applies when the reconstruction has critical points. A factorization through a contractible LL makes the composite nullhomotopic by applying dd to a contraction after ee. Positive dimension allows degree to distinguish this case from the identity.

Example 1.2 (The cup product on a torus). Let X=T2X=T^2 and take LL to be the connected sum of two copies of S1×S2S^1\times S^2. The first cohomology groups of XX and LL both have rank two over Z\mathbb Z. The product of the two coordinate classes in H1(T2;Z)H^1(T^2;\mathbb Z) generates H2(T2;Z)H^2(T^2;\mathbb Z). On LL, all products of degree-one classes are zero. Indeed, collapsing the separating sphere of the connected sum gives a map into the wedge of the two closed summands, inducing an isomorphism in H1H^1 by the Mayer–Vietoris sequence. In the wedge, classes on different summands have zero product. Each summand has one degree-one generator pulled back from S1S^1, so its square vanishes too. Thus the ring retraction displayed above cannot exist, even though comparing H1H^1 alone would permit it.

A sharp uniform obstruction on the sphere

Proposition 1.3 (Reconstruction with a Euclidean intermediary). Let n>0n>0, and view SnS^n as the unit sphere in Rn+1\mathbb R^{n+1}. The following statements use the Euclidean norm on that ambient space.

  1. For every mm, if e:Sn→Rme:S^n\to\mathbb R^m and d:Rm→Snd:\mathbb R^m\to S^n are continuous, then sup⁡x∈Sn∥d(e(x))−x∥=2\sup_{x\in S^n}\|d(e(x))-x\|=2.

  2. If dd is allowed to land in Rn+1\mathbb R^{n+1} and m≤nm\le n, the minimum possible uniform reconstruction error is 11. For m≥n+1m\ge n+1, there is an exact ambient reconstruction.

Proof. In the first case d∘ed\circ e has degree zero by contractibility. If it had no antipodal coincidence d(e(x))=−xd(e(x))=-x, then the straight segment between xx and d(e(x))d(e(x)), normalized to the unit sphere, would be a homotopy to the identity. Such a segment passes through the origin only for antipodal endpoints. The degrees therefore force an antipodal coincidence and the claimed error.

For the second part, the Borsuk–Ulam theorem applied to ee, after including Rm\mathbb R^m into Rn\mathbb R^n, gives a pair x,−xx,-x with the same encoded value [22]. Both reconstructions are a common vector uu. The triangle inequality gives 2≤∥u−x∥+∥u+x∥2\le\|u-x\|+\|u+x\|, so at least one error is at least 11. The constant reconstruction 00 attains that bound. For m≥n+1m\ge n+1, include the sphere in the first n+1n+1 coordinates and decode by their projection. ◻

The distinction between the codomains matters. If a reconstruction lands close to the sphere, one can normalize along its realized image. This does not imply that the decoder can be normalized on the entire intermediate Rm\mathbb R^m, where it may take value zero.

Spaces of factorizations

In the category of spaces the factorizations of ff through LL, together with their chosen homotopies, form the homotopy pullback image A point is a triple (e,d,γ)(e,d,\gamma), with γ\gamma a path from d∘ed\circ e to ff. Paths between such points track changes of both maps and homotopies. This is a homotopy problem, while an exact smooth factorization requires equality d∘e=fd\circ e=f and smooth maps. For example, the smooth map eit↦exp⁡(isin⁡t)e^{it}\mapsto \exp(i\sin t) on S1S^1 is nullhomotopic and therefore factors through a point in the homotopy sense. It is nonconstant, so it does not admit an exact factorization through a point.

Families can have an obstruction even when every individual target value has a lift. With decoder d(z)=z2d(z)=z^2 on L=S1L=S^1, every single b∈S1b\in S^1 has a preimage. The family f(b)=bf(b)=b has no continuous choice of encoded value e(b)e(b) satisfying d(e(b))=bd(e(b))=b. Taking degrees would yield 2deg⁡(e)=12\deg(e)=1, which also rules out a choice accompanied by a homotopy to ff. Local choices are related on chart overlaps by the sign action. The obstruction is in the family of choices, invisible to the existence of each separate factorization.

Persistence of sublevel families

Morse theory studies a smooth function by following its sublevel sets across regular intervals and critical values. Keeping the sublevel sets together with their inclusion maps gives a diagram of spaces. Applying homology turns this diagram into a representation of the ordered set of thresholds. This chapter gives a direct stability estimate on the diagram and compares it with the stronger local triviality available for smooth parameter families at regular levels.

The filtered diagram

For a continuous function f:X→Rf:X\to\mathbb R, let Xf(a)={x∈X:f(x)≤a}.X_f(a)=\{x\in X:f(x)\le a\}. For a≤ba\le b there is an inclusion Xf(a)↪Xf(b)X_f(a)\hookrightarrow X_f(b). These data give a functor from the real line with its order arrows to topological spaces, and hence a diagram in the category of homotopy types. In particular, using singular chains before passing to homology keeps a chain-level diagram. For any coefficient field kk, the homology groups a↦Hj(Xf(a);k)a\mapsto H_j(X_f(a);k) form a persistence module. Its structure map is the homology map induced by inclusion.

Given ε≥0\varepsilon\ge0, translation takes a diagram FF to a↦F(a+ε)a\mapsto F(a+\varepsilon). A strict ε\varepsilon-interleaving of diagrams of spaces consists of natural maps αa:F(a)⟶E(a+ε),βa:E(a)⟶F(a+ε).\alpha_a:F(a)\longrightarrow E(a+\varepsilon),\qquad \beta_a:E(a)\longrightarrow F(a+\varepsilon). The two shifted composites must equal the respective structure maps for a step of length 2ε2\varepsilon. Passing from actual spaces to homotopy types sends these identities to coherent comparisons. A strict interleaving is enough for the applications below.

Proposition 1.1 (Uniform control of filtered spaces). If f,g:X→Rf,g:X\to\mathbb R are continuous and sup⁡x∈X∣f(x)−g(x)∣≤ε\sup_{x\in X}|f(x)-g(x)|\le\varepsilon, their sublevel diagrams have a strict ε\varepsilon-interleaving. It is natural under pullback of functions along continuous maps. When f,gf,g are invariant under a common group action, the maps of the interleaving are equivariant.

Proof. The inequalities g(x)≤f(x)+εg(x)\le f(x)+\varepsilon and f(x)≤g(x)+εf(x)\le g(x)+\varepsilon give inclusions in the indicated directions. All arrows are the restrictions of the identity map on XX. In particular the composition image is the structure inclusion, and the reverse composite is the analogous inclusion on gg-sublevels. For a continuous map u:Y→Xu:Y\to X, use uu on sublevel sets of the pulled-back functions. It commutes with the interleaving inclusions. Invariance makes each sublevel set a GG-space and makes the inclusions equivariant. ◻

This estimate survives functorial operations on the diagram. For instance one can take fixed spaces under a subgroup HH, followed by homology. The fixed sublevel Xf(a)HX_f(a)^H is the sublevel of f∣XHf|_{X^H} inside XHX^H. Alternatively, one can apply homology to EG×GXf(a)EG\times_G X_f(a), the homotopy quotient at threshold aa. The resulting invariants answer different questions about symmetry, and the same uniform bound at the diagram level applies to both. A bottleneck bound for their homology still requires a tameness hypothesis for the resulting modules.

When persistence diagrams are available

A persistence module is called qq-tame if every structure map over a strictly positive step has finite rank. This does not require finite-dimensional homology exactly at each threshold. When XX is the realization of a finite simplicial complex, the homology persistence modules of every continuous real-valued ff on XX are qq-tame. In fact, fix a<ba<b and subdivide so that the oscillation of ff on every simplex is less than b−ab-a, using compactness and uniform continuity. The union of simplices that meet Xf(a)X_f(a), together with their faces, is a finite subcomplex KK with Xf(a)⊆∣K∣⊆Xf(b).X_f(a)\subseteq |K|\subseteq X_f(b). The corresponding homology map factors through a finite-dimensional vector space Hj(∣K∣;k)H_j(|K|;k). This also applies to compact smooth manifolds by smooth triangulability.

The stability theorem for persistence diagrams bounds their bottleneck distance by the interleaving distance for qq-tame modules [6]. Consequently the bound in Proposition 1.1 gives a bottleneck bound ε\varepsilon in the finite-complex setting. One way to read the estimate is that a feature with lifetime exceeding 2ε2\varepsilon cannot simply be matched to the diagonal under a bottleneck matching of cost at most ε\varepsilon. See [7] for the classical stability result for sublevel-set persistence. The inclusions proved here are assertions at the level of spaces, independent of the tameness used to assign persistence diagrams. Applying homology forgets information, such as cup products, that can distinguish otherwise identical homology groups or homology persistence modules.

Example 1.2 (Equal homology modules with different filtered spaces). Equip S2∨S4S^2\vee S^4 and CP2\mathbb{CP}^2 with the constant function zero. In both cases the sublevel is empty before zero, and the entire space thereafter. The two spaces have one homology generator in each of degrees 0,2,40,2,4, with field coefficients, so the ordinary homology persistence modules agree in every degree. The degree-two generator on CP2\mathbb{CP}^2 has a nonzero square, whereas the degree-two class on the wedge is pulled back along the projection to S2S^2 and has square zero [13]. An equivalence of their filtered homotopy types would give a homotopy equivalence at every nonnegative threshold and would preserve these products. Thus the diagrams of spaces can distinguish filtrations with identical homology persistence.

Transport at finitely many regular thresholds

Let XX be a compact smooth manifold without boundary, PP a smooth parameter manifold, and F:P×X→RF:P\times X\to\mathbb R smooth. Write fp(x)=F(p,x)f_p(x)=F(p,x). The map p↦fpp\mapsto f_p is continuous for the uniform norm by compactness of XX. More quantitatively, along a smooth path cc of parameters of length ℓ(c)\ell(c), a bound LL on the norm of the partial derivative DpFD_pF over its image times XX gives ∥fc(1)−fc(0)∥∞≤Lℓ(c)\|f_{c(1)}-f_{c(0)}\|_\infty\le L\ell(c). This follows by differentiating F(c(t),x)F(c(t),x), integrating, and taking the supremum over xx. It gives an interleaving estimate even along paths where the topology at a chosen threshold changes.

Proposition 1.3 (Local trivialization of regular sublevel flags). If a1<⋯<ara_1<\cdots<a_r are regular values of fp0f_{p_0}, there is a neighborhood of p0p_0 on which the family of finite flags Xfp(a1)⊆⋯⊆Xfp(ar)X_{f_p}(a_1)\subseteq\cdots\subseteq X_{f_p}(a_r) is smoothly trivialized by diffeomorphisms of XX. An empty level set counts as regular.

Proof. Compactness allows the parameter neighborhood and small disjoint strips around the levels aia_i to be chosen so that ∇xF\nabla_x F is nonzero throughout each strip. Otherwise there would be sequences converging to a critical point of fp0f_{p_0} at one of those levels. Choose coordinates on a small ball in PP around p0p_0 and a metric on XX. To lift a constant coordinate vector vv on the ball, start with (v,0)(v,0) on P×XP\times X. Near each level modify its XX-component to − DpF(v)∥∇xF∥2∇xF.-\,\frac{D_pF(v)}{\|\nabla_xF\|^2}\nabla_xF. Use a cutoff in the strip equal to one at the level to extend this correction by zero. The resulting vector is tangent to each hypersurface F=aiF=a_i. Integrating these lifts along radial coordinate paths from p0p_0 to pp gives diffeomorphisms from the starting fiber to the ending fiber. Compactness of XX gives existence along every such path, and reversing the path gives the inverse. The flow preserves each level and its chosen sublevel side, since a trajectory cannot cross a hypersurface to which the field is tangent. Smooth dependence on the path endpoint gives the required local trivialization. ◻

Example 1.4 (A regular family with monodromy). Take P=X=S1P=X=S^1 and F(ϕ,θ)=cos⁡(2θ−ϕ)F(\phi,\theta)=\cos(2\theta-\phi), with both angles defined modulo 2π2\pi. Each level aa strictly between −1-1 and 11 is regular for every fϕf_\phi, and the sublevel is a pair of closed arcs. In a local lift of the angle ϕ\phi, translation θ↦θ+ϕ/2\theta\mapsto\theta+\phi/2 identifies the sublevels with those at ϕ=0\phi=0. A loop in ϕ\phi changes this translation by π\pi, interchanging the arcs. Thus the local system of connected components has nontrivial monodromy, although every sublevel in the family has the same homotopy type. Degree-zero homology records the same permutation on the two component generators.

Regularity controls a finite chosen flag. A moving critical value can cross a threshold and change its sublevel topology even under uniformly small perturbations. The interleaving estimate then compares shifted thresholds. A stratified description of the parameter and threshold space requires control of the critical locus and the hypotheses for stratified transport, as in Chapter [ch:exit]. A smooth parameter space by itself does not supply such a stratification. The local trivialization and the quantitative interleaving therefore give complementary information about a parametrized filtration.

Notation

The tangent bundle, cotangent bundle, and derivative of a smooth map. See Chapters [ch:tangent] and [ch:morse].

Spaces of smooth embeddings and immersions. The topology is specified when using continuity or homotopy of families. See Chapters [ch:function-spaces], [ch:isotopy], and [ch:parametric-topology].

The bundle of order-rr jets of maps and the jet of ff at xx. The map jrfj^r f is its holonomic section over MM. See Chapter [ch:jets].

Transversality to a submanifold and the fiber product. The ordinary smooth fiber product represents the transverse pullback. A superscript R\mathbf R or an explicit label distinguishes a derived pullback. See Chapters [ch:transversality] and [ch:derived-intersection].

The determinant line, orientation cover, and first Stiefel–Whitney class of a real vector bundle. See Chapter [ch:orientations].

Degree, Euler class, and Euler characteristic. An integral degree requires the appropriate orientation data. See Chapters [ch:degree] and [ch:intersection].

Thom space of a bundle and the Madsen–Tillmann spectrum for a tangential structure. See Chapters [ch:bordism] and [ch:bordism-cats].

A manifold represented as a smooth sheaf, a smooth quotient stack, and a slice over a base. See Chapters [ch:smooth-stacks] and [ch:relative-maps].

The internal function object and the mapping space in the slice. The former represents all parametrized maps over BB. See Chapter [ch:relative-maps].

A sublevel set and its degree-kk homology, with coefficients specified in context. See Chapters [ch:morse] and [ch:persistence].

Subject index

References point to the chapters where each term is developed.

References

  1. David Ayala, John Francis, and Nick Rozenblyum. A stratified homotopy hypothesis. Journal of the European Mathematical Society 21 (2019), 1071–1178. DOI.

  2. John C. Baez and Alexander E. Hoffnung. Convenient categories of smooth spaces. Transactions of the American Mathematical Society 363 (2011), 5789–5825. arXiv:0807.1704.

  3. Raoul Bott and Loring W. Tu. Differential Forms in Algebraic Topology. Graduate Texts in Mathematics 82, Springer, 1982. DOI.

  4. David Carchedi. Derived manifolds as differential graded manifolds. Preprint, 2023. arXiv:2303.11140.

  5. David Carchedi and Pelle Steffens. On the universal property of derived manifolds. arXiv:1905.06195.

  6. Frédéric Chazal, Vin de Silva, Marc Glisse, and Steve Oudot. The Structure and Stability of Persistence Modules. SpringerBriefs in Mathematics. Springer, 2016. DOI.

  7. David Cohen-Steiner, Herbert Edelsbrunner, and John Harer. Stability of persistence diagrams. Discrete & Computational Geometry 37 (2007), 103–120. DOI.

  8. Ralph L. Cohen, John D. S. Jones, and Graeme B. Segal. Morse Theory and Classifying Spaces. Preprint, 1995. Available manuscript.

  9. Daniel S. Freed. Bordism: Old and New. Lecture notes, 2013. Author’s notes.

  10. Søren Galatius, Ib Madsen, Ulrike Tillmann, and Michael Weiss. The homotopy type of the cobordism category. Acta Mathematica 202 (2009), 195–239. DOI.

  11. Martin Golubitsky and Victor Guillemin. Stable Mappings and Their Singularities. Graduate Texts in Mathematics 14. Springer, 1973. DOI.

  12. Thomas G. Goodwillie and Michael Weiss. Embeddings from the point of view of immersion theory II. Geometry & Topology 3 (1999), 103–118. DOI.

  13. Allen Hatcher. Algebraic Topology. Cambridge University Press, 2002. Author’s copy.

  14. Allen Hatcher. Vector Bundles and K-Theory. Version 2.2, 2017. Author’s notes.

  15. Morris W. Hirsch. Differential Topology. Graduate Texts in Mathematics 33. Springer, 1976. Publisher’s edition.

  16. Morris W. Hirsch. Immersions of manifolds. Transactions of the American Mathematical Society 93 (1959), 242–276. DOI.

  17. Andreas Kriegl and Peter W. Michor. The Convenient Setting of Global Analysis. Mathematical Surveys and Monographs 53. American Mathematical Society, 1997. DOI.

  18. John M. Lee. Introduction to Smooth Manifolds. Second edition. Graduate Texts in Mathematics 218. Springer, 2013. DOI.

  19. Jacob Lurie. Higher Algebra. Manuscript, 2017. Author’s manuscript.

  20. Jacob Lurie. Higher Topos Theory. Annals of Mathematics Studies 170. Princeton University Press, 2009. Author’s copy.

  21. Jacob Lurie. On the classification of topological field theories. Current Developments in Mathematics 2008, 129–280, 2009. arXiv:0905.0465.

  22. Jiří Matoušek. Using the Borsuk–Ulam Theorem. Universitext. Springer, 2003. DOI.

  23. Eckhard Meinrenken. Group Actions on Manifolds. Lecture notes. Author’s notes.

  24. John W. Milnor. Topology from the Differentiable Viewpoint. Princeton University Press, 1997 reissue. Publisher’s edition.

  25. John W. Milnor. Morse Theory. Annals of Mathematics Studies 51. Princeton University Press, 1963. Publisher’s edition.

  26. John W. Milnor and James D. Stasheff. Characteristic Classes. Annals of Mathematics Studies 76. Princeton University Press, 1974. Publisher’s edition.

  27. Richard S. Palais. Local triviality of the restriction map for embeddings. Commentarii Mathematici Helvetici 34 (1960), 305–312. Journal scan.

  28. Matthias Schwarz. Morse Homology. Progress in Mathematics 111. Birkhäuser, 1993. DOI.

  29. Paul Seidel. Differential Geometry. MIT OpenCourseWare 18.950, Fall 2008. Chapter 1, Lecture 9. Lecture notes.

  30. David I. Spivak. Derived smooth manifolds. Duke Mathematical Journal 153 (2010), 55–128. arXiv:0810.5174.

  31. Arthur G. Wasserman. Equivariant differential topology. Topology 8 (1969), 127–150. DOI.

  32. Katrin Wehrheim. Smooth structures on Morse trajectory spaces, featuring finite ends and associative gluing. Geometry & Topology Monographs 18 (2012), 369–450. DOI.

  33. Michael Weiss. Embeddings from the point of view of immersion theory I. Geometry & Topology 3 (1999), 67–101. DOI.

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}


\usepackage{enumitem}


\usepackage{caption,needspace,multicol}


\usepackage{titlesec}


\usepackage{fancyhdr}


\usepackage{xurl}


\usepackage{hyperref}


\usepackage{bookmark}



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

\hypersetup{
  pdftitle={Differential Topology in Higher Categories}, pdfauthor={Anonymous},
  pdfsubject={From regular values and cobordism to derived smooth families},
  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}
\newcommand{\Set}{\mathbf{Set}}
\newcommand{\FinSet}{\mathbf{FinSet}}
\newcommand{\Cat}{\mathcal{C}}

\setcounter{tocdepth}{1}
\DeclareMathOperator{\tr}{tr}
\DeclareMathOperator{\Tr}{Tr}
\DeclareMathOperator{\Ad}{Ad}
\DeclareMathOperator{\Map}{Map}
\DeclareMathOperator{\Tor}{Tor}
\DeclareMathOperator{\Ext}{Ext}
\DeclareMathOperator{\Rep}{Rep}
\DeclareMathOperator{\End}{End}

\tikzcdset{every label/.append style={/tikz/auto=false,description}}
\newcommand{\hrepPart}{}
\newcommand{\bookpart}[2]{  \addtocontents{toc}{\protect\hrepPartEntry{#1}{#2}}  \gdef\hrepPart{{\sffamily\small\color{bookaccent}Part #1\enspace #2\par\vspace{12pt}}}}
\newcommand{\hrepPartEntry}[2]{  \Needspace{6\baselineskip}  \par\addvspace{0.7\baselineskip}  \noindent\textcolor{bookaccent}{\textbf{#1\enspace #2}}\par\nopagebreak[4]}
\titleformat{\chapter}[display]
  {\normalfont\huge\bfseries\raggedright}
  {\hrepPart\sffamily\small\bfseries\color{bookaccent}\MakeUppercase{\chaptertitlename}\ \thechapter}
  {8pt}{}
  [\global\let\hrepPart\empty\vspace{4pt}{\color{bookrule}\titlerule}]



\begin{document}
\frontmatter
\begin{titlepage}
  \thispagestyle{empty}
  \vspace*{0.16\textheight}
  {\noindent\color{bookaccent}\rule{\textwidth}{1pt}\par}
  \vspace{16pt}
  {\fontsize{32}{38}\selectfont\bfseries\raggedright
   Differential Topology\\in Higher Categories\par}
  \vspace{13pt}
  {\noindent\Large From regular values and cobordism\\to derived smooth families\par}
  \vspace{32pt}
  {\noindent\large Anonymous\par}
  \vfill
  {\noindent\small November 2026\par}
\end{titlepage}

\tableofcontents
\mainmatter

\bookpart{I}{Smooth manifolds, maps, and jets}

\chapter{Smooth manifolds and their probes}
\label{ch:manifolds}

A smooth manifold may be specified by an atlas, reconstructed by
gluing its charts, or characterized by the smooth families of points
that it admits. The three descriptions support different operations.
Charts give access to calculus. Gluing exposes compatibility
conditions. Smooth families provide a test for morphisms that
continues to make sense even when a construction has no manifold
representative. We begin with the ordinary smooth objects of
Hirsch's first chapter \cite[Ch.~1]{hirsch}, and pass from these to
their functors of points.

\section{Coordinates and gluing}

A smooth $n$-manifold is a Hausdorff, second-countable topological
space with coordinate neighborhoods in $\mathbb R^n$ and smooth
transition diffeomorphisms. The local dimension is constant within
a connected component. For a manifold with boundary, charts instead
take values in relatively open subsets of the half-space
$[0,\infty)\times\mathbb R^{n-1}$, with the usual zero-dimensional
convention. Smoothness in these coordinates means local extension to
smooth maps on open subsets of the ambient Euclidean space. The
boundary is the locus whose charts have first coordinate zero. It
is intrinsic, as each transition and its inverse preserve the cone
of inward velocities. At a boundary point this cone is a half-space,
whereas at an interior point every direction occurs. The derivatives
of smooth extensions agree on overlaps by continuity from the
interior. Coordinate transitions therefore make the same smoothness
test in every chart \cite{lee-smooth}.

A map $f:M\to N$ is smooth if its coordinate expressions are smooth
on the open sets where they are defined. Composition is smooth by
the coordinate chain rule. The resulting category has a terminal
point, and maps out of the point recover the underlying sets. A map
of manifolds with boundary need not preserve the boundary. For
example, a map from a closed interval that is constant at an
interior point is smooth. When an argument requires a map of pairs
$(M,\partial M)\to(N,\partial N)$, the boundary condition is part of
the structure.

One can construct manifolds from coordinates given in advance.
Let $U_i$ be chart domains, let $U_{ij}\subset U_i$ be open subsets,
and let $\phi_{ji}:U_{ij}\to U_{ji}$ be smooth diffeomorphisms.
Assume the overlap domains and these maps obey the identity,
inverse, and triple-overlap conditions. The quotient of the
disjoint union $\coprod_i U_i$ by the resulting identifications
has each $U_i$ as an open chart. To see this, the saturate of an
open subset of $U_i$ is a union of open sets obtained by the
transition diffeomorphisms. The overlap identities ensure that
the quotient map is injective on each $U_i$. When the quotient
is Hausdorff and second countable, these charts give a smooth
manifold, with the manifold dimension and boundary behavior
specified on each chart.

Hausdorffness is a condition on this gluing. Take two copies of
the real line and identify them on $\mathbb R\setminus\{0\}$.
The resulting space has two origins. A neighborhood of either
contains the image of some punctured interval, so two neighborhoods
of the different origins cannot be disjoint. The coordinate changes
are smooth, but the result is outside the category of Hausdorff
manifolds. A setting for smooth families can still retain the
gluing diagram and its compatibility data.

\section{Smooth families and full faithfulness}

Fix a small site $\mathsf{U}$ of coordinate domains, with open covers
as its coverings. For boundary arguments we allow relatively open
subsets of quadrants $[0,\infty)^k\times\mathbb R^{n-k}$ with the
same convention of local smooth extension. Morphisms are all smooth
maps between the domains. Including quadrants allows finite products
of tests. The manifolds in a particular argument need only use the
boundary or boundaryless charts required there. Let $\mathcal E$
denote the $\infty$-category of sheaves of spaces on this site.

Associate to a manifold $M$ the discrete-valued functor
\[
 \underline M(U)=C^\infty(U,M),\qquad U\in\mathsf{U}.
\]
A map $U\to M$ can be recovered from compatible restrictions to
an open cover of $U$. Its assembled function is continuous because
continuity is local, and it is smooth because every point has a
smooth local expression. This proves the sheaf condition on
$\underline M$. If a test $U$ has coordinates $(u_1,\dots,u_d)$, a
section of $\underline M(U)$ is a smooth $d$-parameter family,
possibly with parameters on the faces.

\begin{proposition}[Reconstruction from test families]
\label{prop:smooth-yoneda}
The association $M\mapsto\underline M$ is fully faithful.
A natural transformation $\underline M\to\underline N$
is composition with a unique smooth map $M\to N$. As maps into
$\underline N$ in $\mathcal E$, these form a discrete mapping space.
\end{proposition}

\begin{proof}
Let $a_i:U_i\to M$ be open chart maps from coordinate domains
whose images cover $M$. A natural transformation $\tau$ assigns
the smooth map $\tau_{U_i}(a_i):U_i\to N$ to each of these families.
On an overlap, use an open subdomain of $U_i$ and its coordinate
change to $U_j$. Naturality identifies the two restrictions.
The maps thus glue to a smooth $f:M\to N$. Given any
$u:U\to M$, restrict its domain to open pieces on which it factors
through one of the $a_i$. Naturality on each piece shows that
$\tau_U(u)$ restricts to $f\circ u$. The sheaf condition then gives
$\tau_U(u)=f\circ u$. This also proves uniqueness. The representable
smooth objects under consideration are $0$-truncated sheaves.
By definition all mapping spaces into a $0$-truncated object are
$0$-truncated, giving the last assertion.
\end{proof}

The value of $\underline M$ at the point remembers only its
underlying set. The extra naturality requirement is substantial.
For instance, $x\mapsto x^3$ is bijective on points, but the
identity plot $u\mapsto u$ of its target has no smooth lift on
a neighborhood of $0$. The inverse is therefore absent even as
a local map of smooth sheaves. Test families detect a singularity
of a set-theoretic bijection.

\section{Atlases and interval homotopies}

An atlas takes a universal form in the category of sheaves.
Set $A=\coprod_i\underline U_i$ in $\mathcal E$ and let
$A\to\underline M$ be induced by the chart maps. Every test-section
of $\underline M$ locally lifts to $A$, by pulling back the
coordinate cover. In this sense the atlas is an effective
epimorphism, meaning that $\underline M$ is reconstructed as
the colimit of its \v{C}ech nerve. The first stages are
\[
 \begin{tikzcd}[column sep=large]
  A\times_{\underline M} A
  \arrow[r,shift left=.55ex,"\mathrm{pr}_1"]
  \arrow[r,shift right=.55ex,"\mathrm{pr}_2"'] &
  A\arrow[r] & \underline M .
 \end{tikzcd}
\]
The iterated fiber products of $A$ in this nerve record all
multiple overlaps. The reconstruction can be checked by testing
against a sheaf $F$ of spaces. A map $\underline M\to F$ is
given on plots into $M$ by sections of $F$. A compatible system
on the nerve restricts, for each plot, to compatible sections on
its pulled-back open cover, with the higher homotopies supplied
on repeated intersections. Sheaf descent glues these into a section
over the plot. This gluing is compatible with changes of plot and
has a contractible space of choices once the descent data are fixed.
Equivalently, mapping out of $\underline M$ gives the limit of the
mapping spaces out of the nerve.

Two different levels of smooth comparison are relevant.
Proposition~\ref{prop:smooth-yoneda} compares maps in the
unlocalized sheaf category. A smooth homotopy, on the other hand,
is a map $H:M\times[0,1]\to N$ restricting to specified maps
at the two endpoints. It is a family parametrized by the smooth
interval. For example, the maps $x\mapsto0$ and $x\mapsto x$
from $\mathbb R$ to $\mathbb R$ are joined by $(x,t)\mapsto tx$.
They are still distinct points of the discrete mapping space of
the corresponding unlocalized sheaves. If one subsequently
localizes at the projections $X\times[0,1]\to X$, the two
endpoint inclusions become equivalent, since both are sections
of an inverted arrow. Smooth homotopies then induce equal maps in
the homotopy category of the localization. This explains which
extra operation makes an interval family into a homotopical
comparison. The shape of a smooth object and its sheaf of plots
serve different purposes.

Finally, objects representing plot families can exist without
being manifolds of the chosen kind. For instance, the product
$\underline{[0,\infty)}\times\underline{[0,\infty)}$ is represented
on the chosen site by the quadrant. Its corner has an inward cone
with zero-dimensional lineality space, whereas an inward half-plane
has one-dimensional lineality and $\mathbb R^2$ has two-dimensional
lineality. Here lineality means the largest vector subspace contained
in the cone. An isomorphism of smooth coordinate models would induce
an invertible derivative carrying one cone onto the other.
Thus at the corner this product has no local model by an ordinary
two-dimensional manifold with boundary. The sheaf product,
with maps given by pairs of smooth maps, supplies the expected
universal property. Function objects and singular pullbacks
will use this enlargement in a similar way
\cite{baez-hoffnung}.
\par



\chapter{Tangent functors and smooth maps}
\label{ch:tangent}

Differentiation associates a linear map to a smooth map at each
point. Its functoriality comes from the fact that each linear map
describes the effect on first-order probes. Pointwise linear
algebra detects local immersion, submersion, and transversality
conditions. Global properties such as embedding also require
control of the topology of the image.

\section{First-order probes}

At a point $x$ of a smooth manifold $M$, let $C^\infty_{M,x}$ be
the real algebra of germs of smooth functions at $x$. A derivation
at $x$ is an $\mathbb R$-linear map $v:C^\infty_{M,x}\to\mathbb R$
satisfying
\[
 v(ab)=a(x)v(b)+v(a)b(x).
\]
Equivalently, it is a unital $\mathbb R$-algebra map from
$C^\infty_{M,x}$ to $\mathbb R[\epsilon]/(\epsilon^2)$ reducing
to evaluation at $x$. The map is $a\mapsto a(x)+\epsilon v(a)$.
Multiplication of such expressions gives the derivation rule
exactly because $\epsilon^2=0$. This is a description by
infinitesimal algebra. The ordinary site of Chapter~\ref{ch:manifolds}
uses smooth coordinate domains as probes. To use the dual numbers as
an actual probe object one enlarges the site or otherwise specifies
a theory of infinitesimals.

In a chart centered at $x$, write the coordinate germs as
$t_1,\ldots,t_n$. The identity
\[
 a(t)-a(0)
   =\sum_i t_i\int_0^1 \partial_i a(st)\,ds
\]
holds in a sufficiently small coordinate neighborhood, using a
smooth extension if $x$ is on the boundary. A derivation thus has
$v(a)=\sum_i v(t_i)\partial_i a(0)$. Conversely every choice of the
numbers $v(t_i)$ gives a derivation by this formula. This identifies
the tangent space $T_xM$ with $\mathbb R^n$ in coordinates. At an
interior point the same coordinates identify tangent vectors with
first derivatives of smooth curves. On the boundary the full tangent
space has dimension $n$. Velocities of curves defined only for
nonnegative times form its inward half-space, and two-sided curves
through the point have velocities tangent to the boundary if their
images lie in the manifold.

For smooth $f:M\to N$, define $D_xf(v)$ by applying $v$ after
pulling back the germ from $N$ to $M$. Composition of pullbacks
gives $D_x(gf)=D_{f(x)}g\circ D_xf$. For the identity the
derivative is the identity. Under coordinate changes the
derivation coordinates transform by the Jacobian matrices, so
gluing $U\times\mathbb R^n$ gives the smooth tangent bundle $TM$
and a smooth map $Tf:TM\to TN$. In bundle notation its defining
square is
\[
 \begin{tikzcd}[column sep=large]
 TM \arrow[r,"Tf"]\arrow[d] & TN\arrow[d]\\
 M  \arrow[r,"f"'] & N.
 \end{tikzcd}
\]
This construction preserves products, allowing corner charts if
necessary. A product chart separates the two lists of variables,
giving the canonical identification $T(M\times N)\cong TM\times TN$.
For example, the tangent map of multiplication
$m:\mathbb R^2\to\mathbb R$ satisfies
$D_{(x,y)}m(v,w)=yv+xw$. This formula follows as well by multiplying
$x+\epsilon v$ and $y+\epsilon w$ in the dual numbers.

\section{Rank and normal forms}

An immersion is a smooth map with injective derivative at every
point. A submersion has surjective derivative. We use the
constant-rank theorem for boundaryless coordinate neighborhoods in
the following form \cite[Ch.~1]{hirsch}.

\begin{proposition}[Local normal form]
Suppose $f$ is a smooth map between boundaryless manifolds, of
constant rank $r$ near $x$. There are source and target coordinates
in which $f(u,v)=(u,0)$, with $u\in\mathbb R^r$.
\end{proposition}

\begin{proof}
Work near the origin and choose coordinates so the first $r$
components have an invertible minor in the first $r$ variables.
Use $(f_1,\ldots,f_r,t_{r+1},\ldots,t_m)$ as a change of source
coordinates, by the inverse function theorem. The map is now
$(u,v)\mapsto(u,h(u,v))$. Since the derivative has rank exactly
$r$, its columns in the $v$ variables, whose first $r$ entries
vanish, must themselves vanish. Restricting to a coordinate box
shows that $h(u,v)=h(u,0)$ by integration along the $v$-segments.
Subtracting $h(u,0)$ from the last target coordinates gives the
stated form. Rank zero is the same argument with an empty first
minor.
\end{proof}

Full rank is open, so the hypothesis applies near every point of
an immersion or a submersion in the interior. For an immersion the
form is a coordinate inclusion. For a submersion it is a coordinate
projection. Boundary versions require hypotheses on the boundary
behavior. For instance $t\mapsto t^2$ on the half-line has a
critical boundary point even though it is a submersion on the
interior. Requiring a boundary-preserving map between half-spaces
does not by itself ensure a smooth inverse up to the boundary.

An embedding is an immersion that is a homeomorphism onto its
image with the subspace topology. The irrational-slope map
$\mathbb R\to\mathbb R^2/\mathbb Z^2$ given by
$t\mapsto (t,\alpha t)\bmod\mathbb Z^2$ illustrates the second
condition. For irrational $\alpha$ it is an injective immersion.
There are unbounded integers $n_j$ with $\alpha n_j$ tending to
zero modulo $\mathbb Z$. Indeed the pigeonhole principle applied
to the fractional parts of $0,\alpha,\ldots,N\alpha$ gives a
nonzero integer of size at most $N$ whose phase is within $1/N$
of zero. A fixed finite set of nonzero integers cannot do this
as $N$ grows. The images of $n_j$ tend to the image of $0$, so
the inverse on the image is discontinuous.

A proper injective immersion between the manifolds here is an
embedding. To check that it is closed, take $y$ in the closure
of the image of a closed set. In the first-countable target
there is a sequence $y_j\to y$ in that image. The sequence
together with its limit is compact, so the preimages of its
terms lie in a compact set by properness. A convergent subsequence
of the preimages has limit in the closed set and maps to $y$.
Thus the immersion is closed, and injectivity makes the inverse
on its image continuous. For an embedded $S\subset M$, the quotient
bundle $TM|_S/TS$ is its normal bundle. A metric identifies it
with the orthogonal complement, whereas the quotient comes
directly from the embedding.

At the boundary a useful class is given by \emph{neat embeddings}.
In this chapter this means an embedding $f:M\to N$ with
$f^{-1}(\partial N)=\partial M$ and with derivative transverse
to $\partial N$ at each point of $\partial M$. The boundary
normal form is a coordinate inclusion preserving the first
coordinate. In fact, let $t$ and $s$ denote source and target
boundary coordinates, with $t,s\ge0$. In a neighborhood of a
boundary point the first component of $f$ is $t b(t,u)$ for a
smooth $b$ by integrating its derivative in $t$, since the
component vanishes at $t=0$. Transversality and nonnegativity give
$b>0$ after shrinking. The other target coordinates can be
linearly recombined along $\partial N$ so that the first
$\dim M$ components, including $s$, have an invertible derivative.
They become source coordinates by the inverse function theorem
on smooth extensions. The sign of $t b(t,u)$ ensures this
coordinate change respects the half-spaces. Subtracting the graph
of the remaining components gives the coordinate inclusion by a
change that fixes $s$. This proves the claimed local form and
exhibits the boundary hypothesis needed for it.

\section{Transverse pullbacks}

Let $f:M\to Z$ and $g:N\to Z$ be smooth maps of boundaryless
manifolds. At a coincidence $(m,n)$ with $f(m)=g(n)=z$,
their derivatives define the linear map
\[
 (u,v)\longmapsto D_mf(u)-D_ng(v)\quad
       \text{from }T_mM\oplus T_nN\text{ to }T_zZ .
\]
When this is surjective at every coincidence, call the pair
transverse. In a target chart the coincidence condition is
the zero set of the difference of the coordinate expressions.
Its derivative on the zero set is precisely the displayed map,
in coordinates. The local submersion normal form gives the
solution set a smooth structure and its dimension
$\dim M+\dim N-\dim Z$ when nonempty. The change of target chart
does not change this embedded submanifold structure, since it is
the subspace preimage of the diagonal with the induced charts.
A smooth map into $M\times N$ meeting this set factors through it
smoothly by its coordinate inclusion. Hence this is the smooth
pullback $M\times_Z N$. Its tangent exact sequence at $(m,n)$ is
\[
 0\longrightarrow T_{(m,n)}(M\times_ZN)
 \longrightarrow T_mM\oplus T_nN
 \xrightarrow{\,Df-Dg\,}T_zZ\longrightarrow0 .
\]
The initial tangent term is the fiber product of the tangent
spaces over $T_zZ$. This explains the precise sense in which
differentiation preserves this pullback.

For $f(s,t)=s^2+t^2$ and $g$ the inclusion of $1$ in
$\mathbb R$, the pullback is the unit circle, with tangent line
$s u+t v=0$ at $(s,t)$. Replacing $1$ by $0$ leaves the singleton
$\{(0,0)\}$, while the linear equation given by the derivative has
the whole plane as its kernel. Both pullbacks exist in manifolds
in this particular example. At the singular value the ordinary
tangent to the solution set loses the first-order equation. Testing
with an infinitesimal point sends $(s,t)$ to $(a\epsilon,b\epsilon)$
and makes this difference explicit, since $s^2+t^2=0$ in the dual
numbers for all $a,b$. Later intersection constructions will
distinguish the agreement of smooth families from the infinitesimal
information of a singular equation.
\par



\chapter{Function spaces and smooth families}
\label{ch:function-spaces}

Convergence of smooth maps can be imposed on one compact set at a
time or simultaneously over the entire source. A smooth family of maps
has another kind of structure: it is evaluated jointly in its parameter
and source variables. Distinguishing these structures is especially
important for noncompact sources. We follow Hirsch's distinction
between the weak and strong topologies \cite[Ch.~2]{hirsch}, then
describe the object that represents jointly smooth families.

\section{Compact control and global control}

For a fixed nonnegative integer $r$, write $J^r(M,N)$ for the
bundle of order-$r$ jet data. A jet at a point consists of the value
and the coordinate derivatives through order $r$, interpreted
in all charts by the chain rule. Chapter~\ref{ch:jets} proves the
basic properties of this bundle. The weak $C^r$ topology is the
compact-open topology on jet prolongations. In particular
\[
 \{\,g:j^r g(K)\subset O\,\},\qquad
 K\subset M\text{ compact},\quad O\subset J^r(M,N)\text{ open},
\]
is a subbasic set. A neighborhood of $f$ in this topology
requires such a condition on finitely many compact sets. After
finitely many charts are chosen around the compact jet data, these
neighborhoods amount to uniform coordinate control through order
$r$ on compact subsets.

For the strong $C^r$ topology we allow a locally finite family
$(K_i)$ of compact subsets of $M$, with an open jet neighborhood
$O_i$ for each. The basic condition is $j^r g(K_i)\subset O_i$ for
every $i$. This is the same topology as that given by tubes about
jet graphs, namely sets of maps whose entire jet graph lies in
one open set of $J^r(M,N)$. To see one direction, let $\pi$ be the
source map of the jet bundle. The complement of
$\bigcup_i\bigl(\pi^{-1}(K_i)\setminus O_i\bigr)$ is open, since
the union is locally finite and each term is closed. Requiring a
graph to lie in that complement imposes exactly the compact-set
conditions. For the other direction, use a locally finite compact
cover of $M$ and impose the tube condition on each compact member.
Such a cover is available from a compact exhaustion or from a
locally finite refinement by precompact coordinate domains.
Throughout this discussion the order $r$ is fixed. A smooth map
can be tested at each finite order without committing to a single
topology that controls indefinitely many orders at infinity.

Consider a smooth bump function $b:\mathbb R\to[0,1]$ supported
in $(-1,1)$ with $b(0)=1$. The functions $g_j(x)=b(x-j)$ converge
weakly to zero through every finite order, since their supports
eventually miss any compact set. They miss the strong neighborhood
of zero given by the condition $|g(x)|<1/2$ at every $x$.
On a compact source every locally finite family of nonempty
compact sets is finite, so the two topologies agree. The bump
sequence shows why the compactness hypothesis cannot simply be
dropped.

Baire category arguments commonly use spaces in which an
exhaustion allows a completeness proof. For instance, on
$C^r(\mathbb R^m,\mathbb R^n)$ take the seminorm $p_j$ of uniform
control of all partial derivatives through order $r$ on the
closed ball of radius $j$. A compatible complete metric for the
weak topology is
\[
 d(f,g)=\sum_{j\geq1}2^{-j}\min\{1,p_j(f-g)\}.
\]
For a Cauchy sequence the functions and each derivative converge
uniformly on each ball. The fundamental theorem of calculus along
coordinate segments shows that the limits of the derivatives
are the corresponding derivatives of the function limit. Induction
on the order gives a global $C^r$ limit. This proves completeness and
the Baire property in this setting. On smooth Euclidean-valued
maps one can use the same construction with $p_j$ controlling
orders up to $j$, proving completeness in the weak smooth topology
by the same argument. Hirsch proves more generally that the
weak topology on $C^r(M,N)$ admits a complete metric, and every
weakly closed subset is a Baire space for the induced strong topology,
using closure of holonomic jets and completeness on the target
of jet maps \cite[Ch.~2, \S4]{hirsch}.
In later genericity arguments one has to
justify the topology and the Baire property of the actual class
of maps being perturbed.

\section{Approximation with fixed data}

Here is a useful approximation statement in a form that separates
the local analytic argument from the geometric target.

\begin{proposition}[Relative smoothing]
\label{prop:relative-smoothing}
Let $M,N$ be ordinary smooth manifolds without boundary and let
$f:M\to N$ be $C^r$ for finite $r$. If $f$ is smooth on a
neighborhood of a closed set $A\subset M$, it can be approximated
in any strong $C^r$ neighborhood by a smooth $g$ agreeing with
$f$ on a possibly smaller neighborhood of $A$. The approximation
can be joined to $f$ by a homotopy fixed near $A$.
\end{proposition}

\begin{proof}
For an $\mathbb R^L$-valued map $F$ take a locally finite smooth
partition of unity by functions $\psi_i$ with compact support in
coordinate domains. In each coordinate chart mollify $\psi_i F$
with sufficiently small support radius. The mollified functions
\(\widetilde F_i\) are smooth and supported in slightly larger
compact subsets of the same coordinate domains, chosen locally
finite from the start. Since convolution commutes with derivatives
through order $r$, their differences from $\psi_iF$ approach zero
in $C^r$ in the coordinate chart. Extend by zero and sum.

To meet a given strong neighborhood, use locally finite compact
controls on the sum. Each enlarged support meets only finitely
many of these control sets. On each meeting, derivatives in the
control charts are estimated by the chart derivatives using the
chain rule on compact subsets. Choose the smoothing radius for
the $i$th summand so its contribution to each relevant constraint
is at most $2^{-i-1}$ times the allowed error there. The sum
then lies in the prescribed neighborhood. If $F$ is smooth
near $A$, first choose a smooth cutoff $\chi$ with $\chi=1$
near $A$ and with support inside that smooth locus. Replace the
sum $\widetilde F$ by $\chi F+(1-\chi)\widetilde F$.
Multiplication by $1-\chi$ has a bounded derivative estimate
on each compact constraint. Choosing smaller local errors
gives the same approximation with the exact agreement required.

For a manifold target use a proper smooth embedding
$i:N\hookrightarrow\mathbb R^L$ and a tubular retraction
$\rho:V\to i(N)$ from an open neighborhood
\cite[Chs.~2, 4]{hirsch}. Apply the Euclidean argument to
$F=i\circ f$. Before smoothing, tighten the compact controls
so that the straight segments between $F$ and its approximation
remain in $V$. This is possible on each compact set since
$F$ maps it into a compact subset of the open set $V$.
The chain rule and compact control also allow the approximation
to be chosen small enough so that $\rho\circ\widetilde F$ meets
the original jet conditions. Composing the straight-line homotopy
with $i^{-1}\rho$ gives the assertion, including the fixed
neighborhood.
\end{proof}

Real-analytic approximation has another source. Local
partitions of unity cannot serve as analytic cutoffs, since an
analytic function on a connected domain that vanishes on a
nonempty open set vanishes identically. A compact interval
illustrates a different mechanism. If $f$ is $C^r$ on $[a,b]$,
approximate $f^{(r)}$ uniformly by a polynomial $q$.
For $r\geq1$ integrate $q$ successively $r$ times, choosing the
constants to match $f$ and its lower derivatives at $a$.
The resulting polynomial approximates the $j$th derivative
for $0\le j<r$ with error bounded by
\[
 \|q-f^{(r)}\|_\infty\,\frac{(b-a)^{r-j}}{(r-j)!}.
\]
This follows by integrating the derivative error at each step.
For $r=0$ it is the original polynomial approximation.
On general real-analytic manifolds the global approximation
theorem uses complex analytic methods. Hirsch states the
density of analytic maps in the strong topology for targets
and sources equipped with analytic structures
\cite[Ch.~2, \S5]{hirsch}. Density does not impose exact agreement
with arbitrary prescribed smooth data.

\section{The object representing jointly smooth families}

Let $\mathcal E$ be the category of sheaves of spaces from
Chapter~\ref{ch:manifolds}. It has an exponential
$[\underline M,\underline N]$, characterized by a natural
equivalence
\[
 \operatorname{Map}(S,[\underline M,\underline N])
 \simeq \operatorname{Map}(S\times\underline M,\underline N).
\]
If $S=\underline U$ is a coordinate probe, the right-hand side
is the discrete space of jointly smooth maps $U\times M\to N$.
To see this, the product is represented by the usual product
manifold, allowing corners for boundary probes, and test-family
reconstruction applies to it. The exponential therefore treats a
family and its evaluation under a single universal property.
Unlike a topology alone, it specifies which parameter tests
have smooth evaluation \cite{baez-hoffnung}.

A jointly smooth family $F:U\times M\to N$ is continuous into the
weak $C^r$ topology. For a point $u_0$ and compact set $K$, joint
continuity of the partial derivatives in the $M$ variables and
compactness of $K$ yield one parameter neighborhood for each
required compact control. This proof does not supply infinitely
many controls at once. For $F(s,x)=sx$ on $\mathbb R\times\mathbb R$,
the parameter family is not continuous at $s=0$ into the strong
topology: the strong neighborhood $|g(x)|<1$ on the whole source
contains $F(0,-)$ and no other member. Thus the smooth-family
structure does not implicitly choose the strong topology.

When $M$ is compact, familiar infinite-dimensional charts
explain the first-order geometry of this function object.
Choose a Riemannian metric on $N$. For $g$ sufficiently close to
$f$ in $C^0$, the inverse to the exponential map near the
zero section is defined for all pairs $(f(x),g(x))$, by
compactness of $f(M)$. The resulting section of $f^*TN$ is a
local coordinate for $g$, and its inverse sends a small section
$v$ to $x\mapsto\exp_{f(x)}v(x)$. For a smoothly evaluated curve
through $f$, differentiating in its curve parameter gives a
smooth section of $f^*TN$. Every such section occurs as a
velocity, since compactness permits one common small time
interval for this exponential construction. This gives the
first-order tangent model $\Gamma(M,f^*TN)$. A full manifold
structure on the space of maps requires an infinite-dimensional
calculus. For the compact-source convenient model and its
exponential law see \cite{kriegl-michor}.
\par



\chapter{Jets and differential constraints}
\label{ch:jets}

A map and its first derivative determine whether it is an
immersion or a submersion at a point. A second derivative
can distinguish a nondegenerate critical point from a more
singular one. A jet is the portion of a map needed for such
finite-order questions. Its construction turns differential
conditions into geometric conditions on a bundle, which can
then be tested on local families.

\section{Finite Taylor data}

Use boundaryless manifolds for the constructions here.
Two germs of smooth maps $M\to N$ at $x$ have the same $r$-jet
if their values are the same and their coordinate expressions
have equal partial derivatives through order $r$ at $x$.
In charts centered at $x$ and at that common value $y$,
this is equality of Taylor polynomials through order $r$.
The chain rule for compositions of Taylor polynomials,
discarding terms of degree $>r$, shows that the equivalence
relation is independent of the charts. The same calculation
shows that changing charts gives a smooth change of jet
coordinates. Such a change uses derivatives of the coordinate
transitions evaluated at $x$ and $y$, and polynomial expressions
in the given coefficients. Thus the classes form a smooth
bundle $J^r(M,N)$ over $M\times N$. At order zero the bundle is
$M\times N$. The first-jet bundle has fiber
$\operatorname{Hom}(T_xM,T_yN)$ at $(x,y)$.

For $r\geq1$ the truncation $J^r(M,N)\to J^{r-1}(M,N)$
is an affine bundle. If two jets lie over the same lower jet
at $(x,y)$, the difference of the $r$th-order coefficients
defines an element of
\[
 \operatorname{Sym}^r(T_x^*M)\otimes T_yN .
\]
Indeed, in a target coordinate change the linear term of
the change acts on the difference by its derivative at $y$.
Any nonlinear term that would contain a difference of order $r$
has another factor of positive order and contributes only
in order $>r$. Source coordinate change acts on the difference
by the $r$th symmetric power of the derivative of the source
change. Terms depending only on lower derivatives cancel.
The tensor difference is therefore intrinsic. In coordinates
one can add any homogeneous degree-$r$ term to a fixed Taylor
polynomial, and that polynomial is the germ of a smooth map
to the target chart after restricting the source. This gives
the free and transitive action on each fiber that defines the
affine structure. The resulting tower and its applications to
singularities are developed in \cite{golubitsky-guillemin}.

A map $f:M\to N$ has a prolongation $j^r f:M\to J^r(M,N)$,
given by its jet at each point. Over $M\times N$ it lies
over the graph of $f$:
\[
 \begin{tikzcd}[column sep=large]
  M\arrow[r,"j^r f"]\arrow[dr,"{(\mathrm{id}_M,f)}"'] &
  J^r(M,N)\arrow[d]\\
  &M\times N .
 \end{tikzcd}
\]
If $F:P\times M\to N$ is smooth, differentiating only in the
$M$ variables gives a smooth family $(p,x)\mapsto j^r_xF_p$.
It commutes with every smooth parameter change $P'\to P$.
This is the relative prolongation used when the parameters
sample different maps from the same source.

\section{Composition, formal data, and rank}

Write $j^r_x g$ for the jet of a germ with specified source
point $x$. For a composable pair of germs, truncated composition
gives
\[
 j^r_{f(x)}g\circ j^r_x f=j^r_x(g\circ f).
\]
This is well-defined for arbitrary representatives of the two
classes, by the Taylor argument above. Associativity follows
from composition of germs. Consequently postcomposition by a
fixed smooth map gives a map of jet bundles, also changing the
target value. Precomposition by a fixed $h:M'\to M$ first
pulls back jet data by its source map $h$ and then uses the
jet of $h$ at each point. The two operations commute. Between
points of a fixed manifold, the composable jets of germs of
local diffeomorphisms, for $r\geq1$, form a groupoid. Inverse
jets come from smooth local inverses. In coordinates, taking
the inverse of an invertible linear term and solving successively
in each higher degree gives smooth inversion on this groupoid.

A smooth section of $J^r(M,N)\to M$ is called formal
differential data. It is holonomic when it equals $j^r f$.
For $M=N=\mathbb R$, coordinates $(x,y,a)$ on $J^1$ designate
source, value, and derivative. A section $(x,y(x),a(x))$
is holonomic exactly when $dy-a\,dx$ pulls back to zero.
The choice $y(x)=0$ and $a(x)=1$ fails this condition.
Thus a bundle of possible derivatives does not by itself
supply a space of solutions by ordinary maps. A global
example is the first-jet section over $S^1$ with value
identically zero and derivative the angular $1$-form $d\theta$.
It specifies an injective linear map on each tangent line.
There can be no smooth immersion $S^1\to\mathbb R$, since
a global smooth real function on the circle has a critical
point. In particular this everywhere regular formal data
cannot be deformed through formal immersions to a holonomic
one between these fixed manifolds.

The rank strata of first jets provide useful smooth
submanifolds. For a linear map between spaces of dimensions
$m,n$, the locus of rank $\ell$ has codimension
$(m-\ell)(n-\ell)$ in the space of matrices. To prove this,
choose a nonvanishing $\ell\times\ell$ minor $A$ and decompose
a nearby matrix into blocks
\[
 \begin{pmatrix}A&B\\ C&D\end{pmatrix}.
\]
The rank is $\ell$ exactly when $D-CA^{-1}B=0$.
In these coordinates $(A,B,C,D-CA^{-1}B)$ is a smooth
coordinate change, so the displayed condition imposes
$(m-\ell)(n-\ell)$ independent equations. Every rank-$\ell$
matrix belongs to one of these charts. At rank zero the
condition is that all entries vanish. This proof applies
in local bundle frames, giving the strata in $J^1(M,N)$.

For example, the family $f_a(x)=x^3-a x$ has derivative
$3x^2-a$. The total parameter-source family meets the
zero-derivative stratum transversely, since differentiating
that equation in $a$ gives $-1$. Its critical-point locus
is therefore the smooth parabola $a=3x^2$. The projection
of this locus to parameters has a critical point at the
origin. For $a>0$ the fixed map has two critical points,
for $a<0$ it has none, and for $a=0$ the two branches meet
at a degenerate critical point. Regularity of a total
family and regularity of each parameter fiber are separate
questions even in this elementary example.

The construction admits a description by smooth families.
Let $\mathcal F_r(U)$ consist of the smooth maps
$U\times M\to J^r(M,N)$ over the projection to $M$.
These sets are a sheaf in $U$, and relative prolongation
is a natural map $[\underline M,\underline N]\to\mathcal F_r$.
Taking the value coordinate gives a retraction, since every
relative jet has the original map as its zeroth component.
The subfunctor of relative holonomic data consists precisely
of the image. Indeed, if formal data are holonomic over an
open cover of a parameter-source product, their local realizing
maps must be the value coordinate there. These maps consequently
glue, and their prolongations agree on overlaps. Holonomicity
is a local compatibility of derivatives with values, rather than
a restriction on the topology of the parameter alone.

\section{Jets at distinct source points}

For a positive integer $q$ let $M^{(q)}$ be the ordered
configuration space of $q$ distinct points of $M$. Over it
the order-$r$ multijet bundle has fiber at
$(x_1,\ldots,x_q)$ equal to the product of the jet fibers
at the specified source points, with values free in $N$.
The multijet $j^r_q f$ is evaluation of $j^r f$ in all the
slots. Its domain excludes repeated source points. The
purpose of this exclusion is that local perturbations can
be varied independently in the slots.

Here is the local independence argument. Fix a map $f$
and distinct points $x_i$. Take disjoint coordinate
neighborhoods of the $x_i$ with $f$ mapping each into a
target chart. In each source coordinate centered at $x_i$,
choose a bump function equal to one near the center and
supported compactly within the neighborhood. Add a
polynomial of degree at most $r$ times this bump to the
target coordinates of $f$. For small polynomial
coefficients all perturbed points still lie in the target
chart, since the change is made on a compact support.
Continue by $f$ outside the supports. At $x_i$ the
derivatives of the change are exactly the derivatives of
the chosen polynomial, and at all other chosen points the
jet is unchanged. Therefore the derivative of multijet
evaluation in the polynomial coefficients surjects onto
the product of the vertical tangent spaces of the jet
bundle at the configuration. This finite-dimensional
family is the input to multijet transversality arguments
\cite[Ch.~II]{golubitsky-guillemin}.

At order zero and $q=2$, the pair of values
$(f(x_1),f(x_2))$ can be tested against the diagonal of
$N\times N$. If this test is transverse, the double-point
locus away from the source diagonal is a submanifold of
dimension $2\dim M-\dim N$ when nonempty, by the transverse
pullback calculation of Chapter~\ref{ch:tangent}. The
rank condition is a one-point test. Accidental equality
of values is a two-point test. The distinction explains
why injectivity questions require multijets even when the
immersion condition is already settled.
\par



\bookpart{II}{Transversality and normal geometry}

\chapter{Critical values and smooth pullbacks}
\label{ch:sard}

A level equation \(f(x)=y\) asks for a pullback by the point
\(y:*\to N\). To identify this pullback with a smooth manifold
of the expected dimension, the crucial hypothesis is a rank
condition on the derivative along the fiber. Sard's theorem
makes that condition available for almost every choice of the
point. The point of view in this chapter is to keep the rank
test and the genericity argument distinct, so that each can
subsequently be applied to families.

\section{The fiber of a regular value}

Let \(M,N\) be finite-dimensional second-countable smooth
manifolds without boundary, with \(\dim M=m\) and \(\dim N=n\).
A point \(x\in M\) is \emph{critical} for a smooth map \(f\)
if \(df_x:T_xM\to T_{f(x)}N\) fails to be surjective.
The \emph{critical values} form the image of the critical
points. A regular value is a point outside this image, so a
point outside \(f(M)\) is a regular value with empty fiber.

\begin{theorem}[Regular preimages]
\label{thm:regular-preimage}
If \(y\) is a regular value, then \(f^{-1}(y)\) is an embedded
submanifold, of dimension \(m-n\) when nonempty. At a point
\(x\in f^{-1}(y)\) its tangent space is \(\ker df_x\).
More generally, for an embedded submanifold \(W\subseteq N\)
of codimension \(c\), suppose \(df_x(T_xM)+T_{f(x)}W=T_{f(x)}N\)
whenever \(f(x)\in W\). Then \(f^{-1}(W)\) is an embedded
submanifold of codimension \(c\) when nonempty, with tangent
space \(df_x^{-1}(T_{f(x)}W)\).
\end{theorem}

\begin{proof}
For the general statement choose a local submersion
\(q:U\to\mathbb R^c\) defining \(W\) near \(f(x)\), so that
\(W\cap U=q^{-1}(0)\). The displayed hypothesis says exactly
that \(d(qf)_x\) is surjective. In source coordinates whose
first \(c\) columns give an invertible minor, augment \(qf\)
by the remaining source coordinates. The derivative of the
augmented map is invertible, so by the inverse function theorem
its components are local coordinates. In these coordinates
the preimage is the zero set of the first \(c\) coordinates.
Differentiating the equations identifies its tangent space
with \(\ker d(qf)_x=df_x^{-1}(T_{f(x)}W)\). The case \(W=\{y\}\)
has \(c=n\). The local descriptions use the subspace topology
and thus assemble to the asserted embedded structures.
\end{proof}

The square in the theorem is a smooth pullback.
\[
\begin{tikzcd}[column sep=large]
 f^{-1}(W) \arrow[r] \arrow[d,hook] & W \arrow[d,hook]\\
 M \arrow[r,"f"'] & N .
\end{tikzcd}
\]
Indeed, a smooth test map into \(M\) whose image lands in
\(f^{-1}(W)\) factors smoothly through it. This follows from
the adapted coordinates in the proof. Testing against all
smooth manifolds gives the same pullback square of
representable sheaves. Such sheaves are \(0\)-truncated, so
their pullback in sheaves of spaces is also computed by the
ordinary fiber-product condition. Representability of that sheaf pullback requires a separate
argument. For
\(f(x)=x^2\) and \(y=0\), the fiber sheaf is represented by a
point, but \(df_0=0\) and its kernel is one-dimensional.
A derived enhancement of a smooth equation can record this
failure of the differential rank test.

\section{Sard's theorem in the smooth case}

A subset of a manifold is \emph{null} if its intersection
with each coordinate chart has Lebesgue outer measure zero
in that chart. This definition is invariant under smooth
change of coordinates because the change is locally Lipschitz
on a countable cover by compact coordinate boxes. A smooth
positive density induces the same local class of null sets.

\begin{theorem}[Sard]
\label{thm:sard-smooth}
For a smooth map between finite-dimensional second-countable
manifolds without boundary, the critical values are null.
In particular the regular values are dense.
\end{theorem}

\begin{proof}
The statement is local in source and target. Use countable
charts with the image of each source chart in a target chart,
and work with \(f:U\to\mathbb R^n\) for \(U\subseteq\mathbb R^m\)
open. The case \(n=0\) has no critical points. For \(n\geq1\),
induct on \(m\), with the case \(m=0\) giving at most a single
image point on each chart. Write \(C\) for the critical set
and \(C_j\) for the subset on which every partial derivative
of every component of \(f\), of orders \(1\) through \(j\),
vanishes. Thus \(C_1\subseteq C\).

\emph{Positive derivative rank.}
Near a point of \(C\setminus C_1\) there is a nonzero
first partial of some component. After permuting coordinates,
use that component in place of the first source coordinate.
The inverse function theorem makes this a local coordinate
change and the new expression for \(f\) has the form
\[
       f(t,u)=(t,F(t,u)).
\]
If \(n=1\), there are no critical points in this chart.
Otherwise the derivative has rank
\(1+\operatorname{rank}d_u F(t,u)\). At each \(t\), the
corresponding section of the image of the critical set
lies among the critical values of \(u\mapsto F(t,u)\).
It is null by induction. To apply Fubini, exhaust the
chart by countably many compact subsets and intersect the
critical set with each one. Since criticality is a closed
condition, the images of these intersections are compact,
and their slices are measurable. Fubini gives measure zero
on each compact image. Countably many such charts cover
\(C\setminus C_1\).

\emph{A first nonvanishing higher derivative.}
Near any \(x\in C_j\setminus C_{j+1}\), for \(j\geq1\),
some scalar partial derivative \(g\) of order \(j\)
has nonzero differential. It vanishes on \(C_j\), so
the portion of \(C_j\) in this neighborhood lies in the
smooth hypersurface \(g^{-1}(0)\) through \(x\). At every
point of \(C_j\) there, \(df=0\), and therefore the
restriction of \(f\) to the hypersurface is critical at each
of these points. Its source dimension is \(m-1\), so the
induction hypothesis gives a null image of this part of
\(C_j\). A countable collection of hypersurface charts
covers \(C_j\setminus C_{j+1}\).

\emph{High-order vanishing.}
Choose \(j\) with \((j+1)n>m\).
On a closed coordinate box \(K\subset U\), fix a bound for all
derivatives of order \(j+1\). Subdivide the box into
\(O(\epsilon^{-m})\) pieces of diameter at most a constant
times \(\epsilon\). For each piece meeting \(C_j\), expand at
one such point. Taylor's formula on the segment inside \(K\)
bounds the oscillation from that point by \(O(\epsilon^{j+1})\).
Thus \(f(C_j\cap K)\) is covered by
\(O(\epsilon^{-m})\) balls of radius \(O(\epsilon^{j+1})\).
Their total \(n\)-volume is \(O(\epsilon^{(j+1)n-m})\),
which tends to zero. A countable box cover handles all of
\(C_j\). The three cases exhaust \(C\).
\end{proof}

This smooth proof follows the induction underlying
\cite[Section 3]{milnor-differentiable}. Hirsch also treats finite
differentiability. For \(C^r\) maps, the sufficient threshold is
\(r>\max\{m-n,0\}\) \cite[Chapter 3]{hirsch}. In the smooth argument
above, choosing a high order of vanishing helps obtain the
last covering estimate. The sharper finite-differentiability threshold uses a different
argument. Sard's conclusion
concerns the image of the critical set. The set itself can
contain open regions, as for any map constant on an open set.

\section{Boundary strata and openness}

Suppose \(M\) has boundary and \(N\) has none. If a value is regular
for both \(f|_{\operatorname{int}M}\) and \(f|_{\partial M}\),
the fiber is a manifold with boundary, with boundary equal to
the fiber of the second restriction. At a boundary point write
a source chart as \((u,t)\) with \(t\geq0\). The surjectivity on
the boundary stratum lets us choose the invertible minor in
the \(u\)-directions. The coordinate change in the regular
preimage proof then leaves \(t\) unchanged. Apply Sard's
theorem to the two strata and take a union of null sets to
see that almost every value has this property. Testing only
the ambient tangent space at a boundary point is insufficient.
For the inclusion of \([0,\infty)\) in \(\mathbb R\), the
ambient derivative at \(0\) is surjective, but the boundary
restriction is critical there.

For a proper \(f\) without boundary, the critical set is closed
and its image is closed, so the regular values are open as
well as dense. Properness cannot be discarded in this statement.
For a connected example, let \(f:\mathbb R\to\mathbb R\) be
\(f(x)=e^{-x}(2+\sin(2x))\). It is strictly positive, so
\(0\) is a regular value with empty fiber. At \(x=j\pi\)
the derivative
\[
 f'(x)=e^{-x}\bigl(2\cos(2x)-2-\sin(2x)\bigr)
\]
vanishes, and the resulting critical values \(2e^{-j\pi}\)
approach zero as \(j\to\infty\). This illustrates why openness
statements for families typically include compactness or
control of the behavior at infinity.

\begin{example}[Completing the tangent exact sequence]
For the function \(f(x_1,\ldots,x_m)=\sum_i x_i^2\),
every positive real number \(a\) is regular and its fiber is
the sphere of radius \(\sqrt{a}\). At a point \(x\) of the sphere,
\[
 0\longrightarrow T_x f^{-1}(a)
 \longrightarrow \mathbb R^m
 \xrightarrow{\ v\mapsto 2\langle x,v\rangle\ }
 \mathbb R\longrightarrow 0
\]
is exact. Dividing by \(\|x\|^2\) constructs a splitting in the
radial direction, which varies smoothly for \(a>0\).
At \(a=0\) the ordinary fiber contracts to the origin and the
derivative has nontrivial cokernel. The comparison exhibits the
information in the linearized equation that its set of solutions
cannot retain.
\end{example}
\par



\chapter{Transversality in parameter families}
\label{ch:transversality}

The preceding chapter used regular values to vary a point in
the target. A family of maps allows another kind of variation.
The parameter can repair a normal derivative that fails to
be onto in the input variables. Parametric transversality
makes precise when this collective freedom gives individual
members with well-behaved pullbacks.

\section{The projection of an incidence manifold}

Let \(P,M,N\) be finite-dimensional second-countable smooth
manifolds without boundary, and let \(W\subset N\) be an embedded
submanifold. Say \(f\pitchfork W\) when \(f\) satisfies the
normal surjectivity condition in
Theorem~\ref{thm:regular-preimage}. Write \(F_p(x)=F(p,x)\)
for the members of a smooth family \(F:P\times M\to N\).

\begin{theorem}[Parametric transversality]
\label{thm:parametric-transversality}
If \(F\pitchfork W\), the incidence space
\(Z=F^{-1}(W)\) is an embedded submanifold. For
\(\pi:Z\to P\) the restricted projection, \(p\) is a regular
value of \(\pi\) exactly when \(F_p\pitchfork W\).
The set of such parameters is conull, dense, and residual.
If \(\pi\) is proper, it is also open.
\end{theorem}

\begin{proof}
Theorem~\ref{thm:regular-preimage} constructs \(Z\).
At \((p,x)\in Z\), write \(Q=T_{F(p,x)}N/T_{F(p,x)}W\).
The derivative followed by projection to \(Q\) gives a pair
of linear maps \(A:T_pP\to Q\), \(B:T_xM\to Q\), with
\(\operatorname{im}A+\operatorname{im}B=Q\). The tangent
space of \(Z\) consists of \((u,v)\) such that
\(Au+Bv=0\). The map \(d\pi\) sends this pair to \(u\), so
it is onto if and only if
\(\operatorname{im}A\subseteq\operatorname{im}B\). Joint
surjectivity makes this equivalent to surjectivity of \(B\).
That is the transverse condition for \(F_p\) at \(x\).
The criterion is vacuous when \(\pi^{-1}(p)\) is empty.

By Sard's theorem, the bad parameters form a null set.
The critical set of \(\pi\) is closed in a second-countable,
locally compact manifold \(Z\), hence is a countable union
of compact subsets. Each compact subset has compact, null
image in \(P\), which has empty interior. The bad set is
therefore a countable union of closed nowhere dense sets.
If \(\pi\) is proper its whole critical image is closed.
\end{proof}

The diagram compares the fibers with the total incidence space.
\[
\begin{tikzcd}[column sep=large]
 F_p^{-1}(W)\arrow[r,hook]\arrow[d] & Z\arrow[r,hook]\arrow[d,"\pi"] & P\times M\arrow[d]\\
 *\arrow[r,"p"'] & P\arrow[r,equal] & P .
\end{tikzcd}
\]
The left square gives a smooth manifold pullback at every
regular parameter. If \(M\) is compact and \(W\) is closed in
\(N\), the incidence space is a closed subset of
\(P\times M\), which makes its projection proper. Its restriction to the open regular locus is a proper
submersion and is smoothly locally trivial. Over a coordinate
ball choose a metric and lift radial paths from the center
horizontally. Properness keeps their lifts over a closed
smaller ball in a compact inverse image. The lifts can therefore
be integrated all the way along each path, and can be reversed.
Smooth dependence on the endpoints and initial values gives
the local trivialization. In particular, after any smooth
change of parameter into this regular locus, the incidence
manifold pulls back as an ordinary smooth manifold. This is
a representability result for that family pullback, proved by
the normal derivative criterion together with local transport.

\begin{example}[A constrained family on the circle]
Consider \(F((a,b),t)=\sin t+a\cos t+b\) on
\(\mathbb R^2\times S^1\), with \(W=\{0\}\).
The derivative with respect to \(b\) is \(1\), so the total
incidence manifold is smooth. The trigonometric summand
\(\sin t+a\cos t\) has amplitude \(R=\sqrt{1+a^2}\).
Its derivative vanishes exactly at its two extrema.
The parameters with a nonregular zero are therefore the two
curves \(b=\pm R\). In the region \(|b|<R\) each fiber has
two points. In the region \(|b|>R\) it is empty.
On a discriminant curve the vanished input derivative has a
nonzero second derivative. Although the full family is
transverse, freezing \(b\) and \(a\) at such a point gives a
nontransverse member. A different family
\(F(a,x)=ax^2\) fails even joint transversality along \(x=0\)
because both partial derivatives vanish there.
\end{example}

\section{Genericity in a space of maps}

Parametric transversality can also be used to construct dense
sets of transverse maps. Here the family is invented locally
for a given map, so the argument needs no global finite
dimensional chart on the space of maps.

\begin{proposition}[Transversality on a compact set]
\label{prop:transverse-compact}
Let \(W\subset N\) be closed and embedded, and let \(K\subset M\)
be compact. Smooth maps that are transverse to \(W\) at all
points of \(K\) form an open dense subset of
\(C^\infty(M,N)\) in the weak Whitney topology.
With a compact exhaustion of \(M\), global transversality
is residual in this topology, and dense by its Baire property.
\end{proposition}

\begin{proof}
For openness, cover \(K\) by finitely many relatively compact
source charts on which the original map either avoids \(W\)
or is tested against a target chart adapted to \(W\).
Avoidance on a compact subset persists because \(W\) is
closed. Where the map meets \(W\), the relevant minor of the
normal derivative is nonzero and persists after shrinking
the source chart and restricting to small \(C^1\) changes.
Together finitely many such conditions work on \(K\).

For density, start with a smooth \(f:M\to N\) and a compact
neighborhood \(K'\) of \(K\). Pick finitely many smooth
sections \(v_i\) of \(f^*TN\), supported in a compact
neighborhood of \(K'\), whose values span at every point
of \(K'\). To construct them, work in local bundle frames,
multiply frame vectors by cutoff functions that are
nonzero on smaller neighborhoods, and use a finite cover.
For small \(a=(a_1,\ldots,a_s)\) define
\[
 F(a,x)=\exp_{f(x)}
       \left(\sum_{i=1}^s a_i v_i(x)\right).
\]
Use a Riemannian metric on \(N\). Compact support lets
all these exponentials be defined uniformly for small \(a\).
At \(a=0\) the parameter derivative is surjective at every
point of \(K'\), and openness and compactness make it
surjective on \(K'\) for all small parameters.
Consequently the family is transverse on
the open manifold \(\operatorname{int}K'\), and
Theorem~\ref{thm:parametric-transversality} gives arbitrarily
small parameters whose maps are transverse on \(K\).
The maps converge smoothly on compacta to \(f\) as \(a\to0\),
so every weak neighborhood of \(f\) contains one of them.
Intersect the open dense tests for the members of a
countable compact exhaustion. The weak smooth mapping
space is Baire \cite[Chapter 2]{hirsch}, proving the last
assertion.
\end{proof}

The same local mechanism applies to jet conditions
from Chapter~\ref{ch:jets}. In coordinate neighborhoods of
a point \(x\), perturb a map by vector-valued polynomials
of degree at most \(r\) multiplied by a cutoff equal to one
near \(x\). Varying their coefficients spans all derivatives
through order \(r\) at any point of this smaller neighborhood:
translation of a polynomial is an invertible change of basis
of the finite-dimensional polynomial space. The source
coordinate of the jet is supplied by the input variable.
The evaluation of prolonged jets is thus a submersion there.
Applying parametric transversality to a smooth submanifold
in the jet bundle gives the local density step of the jet
transversality theorem \cite[Chapter 3]{hirsch}. In particular,
maps into a jet bundle arising from actual smooth functions
can be varied through such maps. Arbitrary perturbations
of a section need not remain a prolongation.

\section{Boundary and relative constraints}

When \(M\) has boundary, test transversality of the full family
on both \(P\times\operatorname{int}M\) and \(P\times\partial M\),
with \(P,N,W\) as above. Their incidence spaces are the
interior and boundary strata of \(Z\). At a boundary point one can choose a nonsingular normal minor in the boundary
variables, leaving the collar parameter unchanged. Apply
Sard to the two projections into \(P\). Outside their union
of critical images, both stratumwise restrictions of \(F_p\)
are transverse, and the fiber is a manifold with boundary
as in Chapter~\ref{ch:sard}. When the parameter manifold
itself has boundary, the same argument can be applied separately
to the parameter strata, using the appropriate total family
on each product stratum.

A relative problem can require the map to stay fixed on a
closed subset \(A\subset M\). If it is already transverse
on a neighborhood of \(A\), compact-set perturbations can
be supported away from \(A\). For the proof of
Proposition~\ref{prop:transverse-compact}, choose an open
neighborhood \(V\) of \(A\) with closure inside the region
where \(f\) is transverse. Then \(K\setminus V\) is compact
and disjoint from \(A\). Build the spanning sections using
supports in \(M\setminus A\) and apply the parametric argument
to an open neighborhood of \(K\setminus V\).
The existing transversality on \(K\cap\overline V\) persists
for sufficiently small perturbations, since this set is
compact. At a point constrained to remain fixed on an open
region, any preexisting failure of transversality in that
region persists as well. The support of the allowed perturbation
is therefore part of the geometric hypothesis.
\par



\chapter{Bundles, frames, and classifying objects}
\label{ch:bundles}

The tangent exact sequence of a transverse preimage is a
sequence of vector bundles rather than a list of separate
linear maps. Bundles make linear operations vary with the
base, and frame changes record how a local description varies.
They provide a useful first example of the distinction between
classifying objects up to existence of an isomorphism and
retaining the groupoid of their presentations and symmetries.

\section{Linear gluing and pullback}

Let \(X\) be a smooth manifold. A smooth rank-\(r\) real
vector bundle \(E\to X\) has smooth local trivializations
that are linear on fibers. Choose smooth frames \(s_i\) over
open sets \(U_i\subset X\), viewing \(s_i(x)\) as a linear
isomorphism \(\mathbb R^r\to E_x\).
On an overlap write \(s_j=s_i g_{ij}\), where
\(g_{ij}:U_i\cap U_j\to \mathrm{GL}_r(\mathbb R)\) is smooth.
On triple overlaps these matrices satisfy \(g_{ij}g_{jk}=g_{ik}\).
Conversely, such a cocycle, with \(g_{ii}=1\), glues the products
\(U_i\times\mathbb R^r\) by
\((x,v)_j\sim(x,g_{ij}(x)v)_i\). The cocycle makes the relations
consistent on all overlaps, and the inclusion of each product
gives a chart of the glued total space.
Changing frames by \(s'_i=s_i h_i\) replaces
\(g_{ij}\) by \(h_i^{-1}g_{ij}h_j\).
A section is a family of smooth functions \(v_i\) satisfying
\(v_i=g_{ij}v_j\).

A map \(a:Y\to X\) pulls the bundle back by pulling back its
frames and transitions. The diagram describes the pullback on total spaces.
\[
\begin{tikzcd}[column sep=large]
 a^*E \arrow[r] \arrow[d] & E \arrow[d]\\
 Y \arrow[r,"a"'] & X .
\end{tikzcd}
\]
If bundles \(E,F\) have transitions \(g_{ij},h_{ij}\) on
a common cover, the direct sum and tensor product have transitions
\(g_{ij}\oplus h_{ij}\) and \(g_{ij}\otimes h_{ij}\).
The bundle of linear maps \(E\to F\) has transition
\(A\mapsto h_{ij} A g_{ij}^{-1}\).
Thus a global linear map of bundles is a section of a bundle
defined by this action on matrices. This local transformation
law matters whenever a model takes a feature in one bundle
to a feature in another, because all coordinate matrices
must satisfy the overlap conditions.

A smooth positive definite fiber metric exists on any
bundle over a paracompact smooth manifold. Choose local
metrics and average them with a smooth partition of unity.
Each nonzero vector has positive length in every metric
with nonzero coefficient, so the average is again positive.
Orthonormalizing local frames for this metric gives
\(\mathrm O(r)\)-valued transitions. Orthogonal complements
then split smooth subbundles. For example, a short exact
sequence of smooth bundles splits by choosing a metric in
the middle term, but the splitting depends on that choice.

\section{Finite Grassmannians and stabilization}

Let \(\mathrm{Gr}_r(\mathbb R^N)\) be the Grassmannian of
\(r\)-planes with tautological bundle \(\gamma_{r,N}\). Its
fiber over a plane is the plane itself.

\begin{theorem}[A classifying map for a compact smooth base]
\label{thm:grassmann-classification}
Let \(X\) be a compact smooth manifold. A smooth rank-\(r\)
bundle over \(X\) is isomorphic to \(c^*\gamma_{r,N}\) for
some smooth map \(c:X\to\mathrm{Gr}_r(\mathbb R^N)\) and
some finite \(N\). Two such maps give isomorphic bundles
if and only if they are smoothly homotopic after inclusion
in a common larger Grassmannian.
\end{theorem}

\begin{proof}
Take a finite cover by bundle frames and smooth cutoff
functions \(\phi_i\) supported in the corresponding frame
domains, with at least one \(\phi_i(x)>0\) at each \(x\).
For \(v\in E_x\) define a linear embedding \(J_x\) by
concatenating the vectors \(\phi_i(x)s_i(x)^{-1}v\), extending
by zero outside their domains. One nonzero coefficient
suffices for injectivity. These embeddings are smooth over
\(X\), and their image planes form a smooth map \(c\) to the
Grassmannian. Locally the image is spanned by a smooth
full-rank matrix \(C(x)\). Its orthogonal projection is
\(C(x)(C(x)^{\mathsf T}C(x))^{-1}C(x)^{\mathsf T}\), which
also shows that the planes vary smoothly. The maps \(J_x\)
give a bundle isomorphism onto the tautological pullback.

If two maps classify isomorphic bundles, identify their
pullbacks with the same \(E\) and regard the resulting
fiberwise injections \(J_0,J_1\) as injections into the
same \(\mathbb R^N\) by padding with zeros. In the
direct sum of two copies of \(\mathbb R^N\), the family
\[
 v\longmapsto (\cos t\,J_0v,\ \sin t\,J_1v),
 \qquad 0\leq t\leq\pi/2 ,
\]
stays injective, since at least one coefficient is nonzero.
Its endpoints are \(J_0\) in the first block and \(J_1\)
in the second. A rotation of the two blocks moves the latter
to \(J_1\) in the first. Reparametrizing the two paths to be
stationary near their concatenation gives a smooth homotopy
of the stabilized classifying maps.

Conversely, a smooth homotopy pulls the tautological bundle
back to a bundle on \(X\times[0,1]\). A connection gives
parallel transport along each interval \(\{x\}\times[0,1]\).
One may take the connection induced by orthogonal projection of
the trivial connection in the ambient trivial bundle.
The resulting transport is linear, invertible by reverse
transport, and smooth in \(x\) by smooth dependence of linear
ordinary differential equations. It identifies the endpoint
bundles. Stabilizing a plane by inclusion of its ambient vector
space pulls the tautological bundle back to the original one,
so this is also an isomorphism before stabilization.
\end{proof}

Hirsch gives finite-dimensional bounds from the dimension
of the base \cite[Chapter 4]{hirsch}. The construction here
uses a finite cover and lets the ambient dimension grow.
The topological theorem for paracompact bases classifies
numerable bundles by homotopy classes of continuous maps
to an infinite Grassmannian, with the relevant topology on
that Grassmannian \cite[Chapter 1]{hatcher-vbkt}.
For a compact smooth base, a continuous map into a finite
Grassmannian can be smoothed up to homotopy, and continuous
homotopies between smooth maps into one can be smoothed
relative to their ends after making them stationary near the
ends. The smooth formulation retains a smooth bundle and
smooth changes of frames throughout its proof.

\section{Classifying the whole groupoid}

Consider the site of smooth test manifolds with covers by
open sets, using sheaves of spaces for the ambient category.
Let \(\mathcal G_r\) be the group sheaf
\(\mathcal G_r(U)=C^\infty(U,\mathrm{GL}_r(\mathbb R))\).
The geometric realization of the simplicial bar construction
of \(\mathcal G_r\), taken in sheaves, is its classifying stack
\(B\mathcal G_r\). Its groupoid of points on a test manifold
\(X\) is the groupoid of smooth rank-\(r\) bundles on \(X\)
and smooth bundle isomorphisms.

To see the classification, the frames of \(E\) form a
right \(\mathcal G_r\)-torsor. They are locally nonempty,
and any two frames differ by a unique smooth pointwise
matrix. Conversely a torsor \(P\) gives the associated bundle
with fibers \(P_x\times_{\mathrm{GL}_r(\mathbb R)}\mathbb R^r\).
Local sections of \(P\) reduce this construction to the
gluing equations above. These operations are inverse on
objects and on isomorphisms. The bar construction presents
the groupoid with one object and automorphism sheaf
\(\mathcal G_r\). Descent adds the torsors obtained by gluing
this object, with the changes of local frames giving their
isomorphisms. This computes the stack interpretation.

For example, the trivial line over a point has automorphism
group \(\mathbb R^\times\) as an ordinary group in this
groupoid. Its single isomorphism class says nothing about
these automorphisms. The topological classifying space
uses \(\mathrm{GL}_r(\mathbb R)\) with its Lie group topology
and classifies bundles through homotopy classes of maps.
It answers a corresponding isomorphism-class question
in a category with continuous homotopies. A map to the
smooth stack, in contrast, records actual smooth changes of
frame over each test manifold. The classifying map of
Theorem~\ref{thm:grassmann-classification} concerns only
the isomorphism class, and its equivalence relation is
stabilized homotopy.

\section{Orientation as reduction of structure group}

For \(r\geq1\), an orientation on a real rank-\(r\) bundle
is a coherent choice of orientation on its fibers. The functions
\(\operatorname{sgn}\det g_{ij}\) define a double cover with
fiber the two possible orientations. A section amounts to
choosing local signs that agree under these transitions,
equivalently to reducing the frame group to the positive
determinant subgroup. With a metric this is a reduction
from \(\mathrm O(r)\) to \(\mathrm{SO}(r)\).
On stacks the sign of determinant acts as a map to the
classifying stack of a two-element group, and an orientation
is a trivialization of the resulting torsor. This interpretation
also shows why the orientation pulls back with the bundle.

\begin{example}[Twisted coefficients on a circle]
Over \(S^1\) identify \((2\pi,v)\sim(0,-v)\) in
\([0,2\pi]\times\mathbb R\) to obtain the M\"obius line \(L\).
A section is represented by an antiperiodic real function
on the real line. A nowhere-zero section would have constant
sign by continuity, contradicting antiperiodicity. Thus \(L\)
has no orientation or global frame. The tensor-square line
has transition \((-1)^2=1\) and is trivial. A local feature
that is linear in a vector of \(L\) inherits the sign change.
A quadratic scalar in the same vector can be invariant
and global even when the frame for the original feature
does not exist.
\end{example}
\par



\chapter{Tubular neighborhoods and collars}
\label{ch:tubular}

An embedded submanifold determines normal directions by a
quotient of tangent bundles. Turning the quotient into
coordinates in the ambient manifold requires a choice.
Tubular neighborhoods provide this realization, and their
isotopy properties explain how subsequent constructions can
use these coordinates and still be independent up to homotopy.
Boundary collars play the same role for an inward half-line.

\section{Normal coordinates near an embedding}

Let \(i:A\hookrightarrow M\) be an embedded submanifold,
with \(A,M\) both without boundary, and set \(\nu_i=i^*TM/TA\).
A \emph{normalized tubular embedding} is an embedding
\(e:U\to M\) from an open neighborhood \(U\) of the zero
section in \(\nu_i\), with open image, such that \(e(0_x)=i(x)\)
and the induced derivative in the normal quotient at \(0_x\)
is the identity of \((\nu_i)_x\). The quotient property of
\(\nu_i\) makes the normalization meaningful without a metric.
The source projection determines a retraction on the image.
\[
\begin{tikzcd}[column sep=large]
 U \arrow[r,"e"]\arrow[d,"p"'] & e(U)\arrow[d,"p e^{-1}"]\\
 A\arrow[r,equal] & A .
\end{tikzcd}
\]

\begin{theorem}[Existence near a compact submanifold]
\label{thm:tubular-existence}
If \(A\) is compact, a normalized tubular embedding exists
on an open disk bundle of some positive radius.
\end{theorem}

\begin{proof}
Choose a Riemannian metric on \(M\), identifying \(\nu_i\) with
the orthogonal complement of \(TA\) in \(i^*TM\).
Restrict the Riemannian exponential to this complement.
At \(0_x\), the derivative on the tangent of the zero section
is \(di_x\), and the derivative along a fiber is the inclusion
of the normal complement into \(T_xM\). Their sum is an
isomorphism. It follows that the restriction is a local
diffeomorphism on some neighborhood of the zero section.
Because \(A\) is compact, a single small positive radius
works for definition of the exponential and invertibility of
its derivative on the resulting disk bundle.

If injectivity failed on every smaller disk bundle,
there would be pairs \(u_j,v_j\) of distinct vectors
of norms tending to zero with equal images. By compactness
their footpoints have convergent subsequences, with limits
\(x\) and \(y\). Continuity and injectivity of \(i\) force
\(x=y\). Both sequences then approach \(0_x\), contradicting
injectivity in a coordinate neighborhood where the map is
a diffeomorphism. An injective local diffeomorphism has
open image and smooth inverse on it. The derivative
calculation provides the normalization.
\end{proof}

For a closed noncompact embedded submanifold one can use
a positive radius depending on the footpoint
\cite[Chapter 4]{hirsch}. Uniform radius can fail at infinity.
For example, embed a countable discrete set in \(\mathbb R^2\)
as the points \((j,0)\) and \((j,1/j)\) for \(j\geq1\).
This is a closed zero-dimensional submanifold. Under the
Euclidean exponential its normal disks are ambient disks,
and no uniform positive radius makes all of them disjoint.
Choosing radii less than one third of the distance to the
nearest other point remedies the problem.

\section{Isotopy of normalized germs}

\begin{proposition}[Independence after shrinking]
\label{prop:tubular-germs}
Let \(A\) be compact, with ambient manifold without boundary.
Any two normalized tubular embeddings \(e_0,e_1\) of the same
embedding \(i\), defined on possibly different neighborhoods,
are joined by a smooth family of normalized tubular embeddings
on a common smaller neighborhood. The family is constant as
a germ wherever the two original germs agree.
\end{proposition}

\begin{proof}
After shrinking, form the transition \(g=e_0^{-1}e_1\)
near the zero section of \(\nu_i\). It fixes the zero section,
and its normal derivative is the identity. For \(s>0\) let
\(g_s=\delta_{1/s}\circ g\circ\delta_s\), where \(\delta_s\)
scales the fibers by \(s\). In a local trivialization
\(g\) has components \((b(x,v),w(x,v))\) and \(w(x,0)=0\).
The base component of \(g_s\) tends to \(x\), and its fiber
component has the formula
\[
 \frac{w(x,sv)}{s}
     =\int_0^1 D_vw(x,\tau s v)\,v\,d\tau .
\]
The equality supplies a smooth extension at \(s=0\),
with fiber component \(v\) by normalization. It also shows
that sufficiently small vectors, uniformly over the compact
zero section and \(s\in[0,1]\), have images in the domain
of \(e_0\). For positive \(s\) the transition is an embedding
on this smaller domain, by its definition as a composition
of embeddings. At zero it is the identity. Its derivative
on the quotient remains the identity, and \(e_0\circ g_s\)
gives the required family. On a region where the transition
is the identity as a germ, every rescaled transition is
the identity as a germ there.
\end{proof}

The formula uses compatible transition maps of the normal
bundle, so the extension shown in each trivialization glues
globally. This conclusion concerns tubular germs. An isotopy
on a prescribed radius without shrinkage requires control
away from the zero section. There is also a precise base
change statement for the quotient normal bundle. If
\(f:X\to M\) is transverse to \(A\) and \(B=f^{-1}(A)\), then
\(df\) followed by projection to \(\nu_i\) is onto along \(B\)
with kernel \(TB\). Consequently
\[
             TX|_B/TB\ \cong\ (f|_B)^*\nu_i .
\]
This identification is natural for repeated transverse base
change. An exponential map uses additional geometric data. Comparing
two chosen exponentials requires a corresponding comparison
of these choices.

\section{Collars and neat intersections}

A collar of a manifold with boundary \(M\) is an embedding
\(c:\partial M\times[0,\epsilon)\to M\) onto a neighborhood of
the boundary with \(c(x,0)=x\). For compact boundary such a
collar is obtained by choosing a smooth vector field near it
pointing strictly inward there. Use inward coordinate vectors
in local boundary charts, and glue with a partition of unity.
Positivity of the derivative of an inward boundary coordinate
is invariant under changing the defining coordinate, and is
preserved by positive sums of vector fields. Thus the glued
field remains inward. Its positive-time flow on the
boundary is defined for a uniform short interval by compactness.
At time zero its derivative has image the sum of the boundary
tangent space and the transverse inward direction. The local
inverse function theorem in half-space charts gives a collar
near each point. The same compactness argument as for tubular
neighborhoods gives a common interval on which this flow is
injective and has smooth inverse onto its image.

For a compact manifold \(M\) with boundary, suppose the
compact submanifold \(A\) is \emph{neat}. This means
\(\partial A=A\cap\partial M\) and near each of those points
there are half-space charts with coordinates \((u,v,t)\),
\(t\geq0\), in which \(A\) is defined by \(v=0\).
One can construct a collar of \(M\) whose restriction to
\(\partial A\) is a collar of \(A\). In the neat charts the
vector \(\partial_t\) is tangent to \(A\) along \(A\).
Choose all local inward fields meeting \(\partial A\)
in this form, and choose fields on the remaining patches
disjoint from \(A\) near those patches of the boundary.
A partition of unity preserves tangency to \(A\) and inwardness.
Uniqueness of solutions of the flow equation then gives
the compatible collar near \(\partial A\).

Use product metrics on a smaller common collar and double
\(M\) and \(A\) across their boundaries. The double is
smooth in the signed collar coordinate, and the doubles form
an embedded pair without boundary. Apply the normal
exponential construction to this pair. On the product collar,
the normal vectors to \(A\) have zero component in the collar
direction. Their short normal geodesics have constant collar
coordinate. Away from a smaller collar the feet of the
vectors are a compact set separated from the boundary.
After shrinking further, their geodesics remain in the
original half of the double. The resulting restriction is
a tubular neighborhood for the original neat pair, compatible
with its boundary product structure \cite[Chapter 4]{hirsch}.
The product condition avoids treating the inward half-line
as an unrestricted normal direction.

\begin{example}[Collapse to a Thom space]
Let \(A\) be compact inside a compact manifold \(M\) without
boundary, and choose a closed radius-\(r\) disk bundle
\(D_r(\nu_i)\) inside a tubular domain. The Thom space here
is \(D_r(\nu_i)/S_r(\nu_i)\), with the sphere bundle collapsed
to a base point. For positive normal rank, the tubular
embedding defines a map \(M\to D_r(\nu_i)/S_r(\nu_i)\):
on the embedded disk bundle use the inverse embedding
followed by the quotient, and on the complement of its
interior use the base point. The definitions agree on the
embedded sphere bundle, giving continuity by gluing over
two closed subsets. Shrinking the radius identifies Thom
spaces by fiberwise rescaling and gives homotopic collapse
maps by varying the radius continuously. Likewise, a family
of tubular embeddings on a common disk bundle gives a homotopy
of collapse maps. The inverse embedding on that family is
continuous jointly with the parameter by compactness of
the total closed disk bundle over the compact parameter
interval. Thus the collapse construction can use a tubular
choice and still define a homotopy class independent of it.
\end{example}
\par



\bookpart{III}{Degree, intersections, and Morse theory}

\chapter{Degree and oriented fibers}
\label{ch:degree}

A regular fiber of a map between manifolds of the same dimension is
a finite set when the source is compact. Orientation gives one sign
to each point of this set. The degree is the sum of these signs, and
its invariance comes from two compatible constructions. The first
uses the map on fundamental classes. The second compares regular
fibers by an oriented one-dimensional preimage. Both constructions
are useful in families. Here a closed manifold is a compact manifold
without boundary.

\section{Local orientations and the signed count}

Let $M$ and $N$ be closed oriented $n$-manifolds with $N$ connected.
For a smooth map $f:M\to N$, a value $y$ is regular if $D_xf$ is
surjective at every point of $f^{-1}(y)$. At such a point the tangent
map is an isomorphism. Define $\epsilon_x(f)$ to be $1$ or $-1$
according as it preserves or reverses orientation. The inverse
function theorem gives a discrete fiber. Its closedness in a compact
space makes it finite, including the possibility of an empty fiber.

The orientation at $x$ is equivalently a choice of generator of
$H_n(M,M\setminus\{x\};\mathbb Z)$. This follows by excision to a
coordinate ball and the relative homology of that ball punctured
at its center. A global fundamental class $[M]$ restricts to this
generator at each point. If $M$ is disconnected, $[M]$ sums the
fundamental classes of the oriented components. In particular,
if $f$ is surjective on some components and misses a given value
on others, every component can still be included in the same
formula.

\begin{theorem}[Regular-fiber degree]
\label{thm:regular-degree}
The homomorphism on top homology satisfies
\[
 f_*[M]=d[N], \qquad
 d=\sum_{x\in f^{-1}(y)}\epsilon_x(f)
\]
at every regular value $y$. Thus the signed count is independent
of the regular value and depends only on the homotopy class of $f$.
\end{theorem}

\begin{proof}
Since $H_n(N;\mathbb Z)$ is generated freely by $[N]$, the
first equation uniquely defines $d$. Put $Q=f^{-1}(y)$. Apply
homology to the maps from absolute to relative pairs.
\[
\begin{tikzcd}[column sep=large,row sep=large]
 H_n(M) \arrow{r}{f_*} \arrow{d} &
 H_n(N) \arrow{d} \\
 H_n(M,M\setminus Q) \arrow{r}{f_*} &
 H_n(N,N\setminus\{y\})
\end{tikzcd}
\]
All groups in the square use integer coefficients. Disjoint
coordinate balls about the finitely many points of $Q$ and
excision identify the lower left group with the direct sum
of the local orientation groups. The image of $[M]$ is the tuple
of their chosen generators. If $Q$ is empty, the relative group
is zero. On each summand the bottom arrow sends the local generator
to $\epsilon_x(f)$ times the local orientation at $y$.
Indeed, in small charts the map is a diffeomorphism, and its
action on local homology is its orientation sign. The right
vertical arrow sends $[N]$ to the local orientation at $y$.
The square proves the signed count. Homotopy invariance follows
from that of the induced homology map.
\end{proof}

This argument also explains why the signs must be computed
relative to both orientations. Reversing the orientation of $M$
changes $[M]$ by a minus sign, and the differential signs all
change accordingly. Reversing both orientations leaves the
degree unchanged. If $f:M\to N$ and $g:N\to L$ have connected
closed oriented targets of the same dimension, functoriality
gives $\deg(gf)=\deg(g)\deg(f)$. Disjoint union adds degrees
for maps to a common target. A map of nonzero degree is
surjective, since an omitted value would be regular with empty
fiber. The formula can apply even when the critical points
occupy part of the source. Only the chosen fiber needs an
isomorphism of tangent spaces at each of its points.

\section{Homotopy and the boundary of a fiber}

Suppose $f_0,f_1:M\to N$ have a common regular value $y$ and
are joined by a smooth homotopy $F$. Reparametrize near the
ends so that $F$ agrees with the end maps on their collars.
Relative transversality permits replacing the interior of
$F$ by a nearby homotopy transverse to $y$ without altering
these collars \cite[Chs.~3 and~5]{hirsch}. Its fiber over $y$ is
a compact one-manifold $W$ with boundary.
Orient $[0,1]\times M$ using the time coordinate followed by
the orientation on $M$. The preimage orientation on $W$ is given
by
\[
 0\longrightarrow TW\longrightarrow
 T([0,1]\times M)|_W
 \xrightarrow{\ DF\ } T_yN\longrightarrow 0
\]
using the kernel first. On each end collar a positive tangent
vector of $W$ points in the increasing-time direction exactly
when the corresponding local sign of $f_i$ is positive.
The boundary convention therefore gives
\[
 \partial W=f_1^{-1}(y)\amalg -f_0^{-1}(y)
\]
as oriented zero-manifolds. Each compact oriented interval
has one positive and one negative endpoint, and circle
components contribute none. This proves equality of the signed
cardinalities with a geometric comparison. Points of opposite signs can therefore be the two
boundary points of one oriented component of a transverse
interpolating fiber.

The same reasoning for a homotopy of targets can be expressed as
transversality to a graph. For a smooth path $\gamma:[0,1]\to N$,
one forms the preimage of its graph in $[0,1]\times N$ under
$(t,x)\mapsto(t,f(x))$. If this is transverse, oriented by the
normal orientation pulled back from $TN$, it has the corresponding
endpoint fibers. The cylinder argument and the graph argument
are instances of the same normal exact sequence. The top-homology
proof covers paths through arbitrary singular values without
choosing such a transverse comparison.

\section{Families of regular values}

Let $P$ be a smooth parameter manifold and $M$ compact.
Suppose $F:P\times M\to N$ is smooth and a smooth choice
$y:P\to N$ is regular for each $F_p=F(p,-)$, with
$\dim M=\dim N$. Consider the diagram
\[
\begin{tikzcd}[column sep=huge,row sep=large]
 Z \arrow{r} \arrow{d}{\pi} &
 P\times M \arrow{d}{(p,x)\mapsto(p,F(p,x))} \\
 P \arrow{r}{(1,y)} & P\times N .
\end{tikzcd}
\]
The diagram is a transverse pullback. In target coordinates the
equation is $F(p,x)=y(p)$, and its derivative in $x$ is invertible
at every solution. The implicit function theorem makes $\pi$
a local diffeomorphism. It is proper because $Z$ is closed in
$P\times M$ and $M$ is compact. On each component with nonempty
fibers it is a finite covering. To see the local triviality,
choose disjoint inverse branches about the points of a fiber.
Properness allows shrinking the base neighborhood until no
points remain outside these branches. Otherwise points outside
the branches over a convergent sequence would have a limit in
the original fiber.

For oriented $M,N$, the signs are locally constant on $Z$.
Transport around a loop can permute points, while preserving
their signs. Thus the signed fiber is a locally constant family
of finite sets with a map to $\{-1,1\}$. Forgetting transport
and taking signed cardinality yields a locally constant integer.
For $P=M=N=S^1$, let $F(p,z)=z^k$ and $y(p)=p$ with $k>0$.
The regular fiber has $k$ positive points. A single turn in $P$
cyclically permutes them. Its degree is $k$, and the permutation
is additional data of the family.

\section{Parity and examples}

Over $\mathbb F_2$, each closed $n$-manifold has a canonical
fundamental class even if it is nonorientable. The same excision
argument gives the parity formula
\[
 f_*[M]_2=\bigl(\# f^{-1}(y)\bmod 2\bigr)[N]_2
\]
for closed manifolds of dimension $n$ and connected $N$.
The mod-two degree is homotopy invariant. An unoriented compact
one-manifold has an even number of boundary points, so the
geometric comparison works with parity as well. If integral
orientations exist, coefficient reduction sends the integer
degree to the mod-two degree. An unbranched two-sheeted covering,
for instance the antipodal covering $S^2\to\mathbb{RP}^2$, has
mod-two degree zero, although the pullback of a point contains
two geometrically distinct points.

For a different example let $0<a<\pi$ and
$f(e^{it})=\exp(ia\sin t)$. The value $1$ has two preimages,
at $t=0$ and $t=\pi$, since $a\sin t$ lies in $(-\pi,\pi)$.
Their local angular derivatives are $a$ and $-a$. The count
is zero, in agreement with the homotopy
$\exp(isa\sin t)$ from the constant map to $f$. By contrast
the angle map $z\mapsto z^k$ has degree $k$ for every integer
$k$. For $k<0$ all $|k|$ regular preimages have negative local
derivative. For $k=0$ the map misses a value.

There is also a relative version for compact oriented manifolds
with boundary. If $f:(M,\partial M)\to(N,\partial N)$ is smooth
and $N$ is connected, use the map on $H_n(-,\partial-)$ and its
relative fundamental classes. A regular value in the interior
of $N$ has no preimage in the boundary of $M$. Excision gives
the same local formula by mapping the relative fundamental
classes into the local pairs. This version makes the behavior
on the boundary part of the hypotheses. Without it the
absolute top homology used for closed manifolds does not
supply the desired comparison.
\par



\chapter{Intersections and the Euler class}
\label{ch:intersection}

The sign in degree theory compares two vector spaces of equal
dimension. For intersections the relevant target is a normal
space. A transverse preimage is oriented using its tangent and
normal exact sequence. The Thom class expresses the same comparison
in cohomology, and it gives a formula when the intersection has
larger dimension than transversality predicts.

\section{Oriented transverse pullbacks}

Work with smooth manifolds without boundary in this chapter.
Let $A\hookrightarrow W$ be a closed embedded submanifold of
codimension $d$ with oriented normal bundle $\nu_{A/W}$.
Let $f:B\to W$ be transverse to $A$ and put $Z=f^{-1}(A)$.
The normal quotient of $D f$ is surjective on $Z$ and the
transverse preimage theorem yields the diagram
\[
\begin{tikzcd}[column sep=large,row sep=large]
 Z \arrow{r} \arrow{d}{j} & A \arrow{d} \\
 B \arrow{r}{f} & W .
\end{tikzcd}
\]
It is a pullback of smooth manifolds. Smoothness of
a factorization through $Z$ follows from the defining
local submersion coordinates, so the square has the
same universal property on all smooth test manifolds.
The differential identifies the normal bundle of $Z$ in $B$
with the pullback of $\nu_{A/W}$, by the exact sequence
\[
 0\longrightarrow TZ\longrightarrow TB|_Z
   \longrightarrow (f|_Z)^*\nu_{A/W}\longrightarrow 0 .
\]
If $B$ is oriented, we orient $TZ$ so that an oriented
basis of $TZ$, followed by lifts of a positive normal basis,
gives the orientation on $B$. When $\dim B=d$, each point
of $Z$ is oriented by the determinant sign of the normal
differential. For compact $B$ the intersection number is
the sum of these signs. Taking $A$ to be a point in an
oriented target recovers the regular-value formula of
Chapter~\ref{ch:degree}.

If $f_t$ is a homotopy between transverse maps with compact
source without boundary, a transverse replacement relative to
end collars has oriented preimage with boundary the two end
intersections, taken with opposite signs. The preimage is
compact by closedness of $A$. This proves invariance of the
intersection number and gives an oriented bordism in higher
dimension. The tangent sequence provides the sign convention
throughout. Specifying normal orientations also shows why
the order of intersection can matter. Interchanging two
independent oriented normal blocks of ranks $a$ and $b$
changes their combined orientation by $(-1)^{ab}$.

\section{Thom classes and zero sets}

For an oriented rank-$d$ real vector bundle $E\to B$ with
zero section $z$, the Thom class is the class
\[
 U_E\in H^d(E,E\setminus z(B);\mathbb Z)
\]
which restricts on each fiber to the chosen orientation
generator. Fiber restriction and the Thom isomorphism give
its uniqueness. One construction glues the local generator
on each oriented product chart, using the trivializations
and the orientation compatibility in the Thom isomorphism
\cite[\S9]{milnor-stasheff}. Its pullback along $z$, after
forgetting the relative condition, is the Euler class
$e(E)\in H^d(B;\mathbb Z)$.

Let $s$ be a section transverse to $z$ and let $Z=s^{-1}(0)$.
The class $s^*U_E$ belongs to $H^d(B,B\setminus Z;\mathbb Z)$.
Its absolute image is $e(E)$, by the homotopy
$t\mapsto ts$ of sections after forgetting the complements.
In a local bundle chart the derivative of the coordinate
vector of $s$ identifies $\nu_{Z/B}$ with $E|_Z$. This
identification preserves the induced normal orientation.
The Thom class $s^*U_E$ is consequently the normal Thom
class of $Z$. For compact oriented $B$ with $\dim B=d$,
evaluation on its fundamental class gives
\[
 \langle e(E),[B]\rangle
   =\sum_{x\in Z}\operatorname{sgn}\det(D_x^{\mathrm{vert}}s).
\]
Indeed, excise to disjoint oriented coordinate balls and
use the local fiber generators as in the degree proof.
This also proves independence of the signed zero count
from the transverse section. If $s$ is nowhere zero, the
pullback of the relative class is zero, so $e(E)$ vanishes.
The Euler class detects a necessary condition for a
nonvanishing section.

More generally, an oriented closed embedding
$j:Z\hookrightarrow B$ of codimension $k$ has a
cohomological pushforward
$j_!:H^r(Z)\to H^{r+k}(B)$ constructed by placing the orientation Thom class first in
the cup product with the pulled-back cohomology class.
Compute this in a tubular neighborhood, excise to the relative
pair in $B$, and forget the complement. We write $\operatorname{PD}_W[A]$ for the
class obtained by applying this operation to $1$ for
$A\hookrightarrow W$. These definitions give
$f^*\operatorname{PD}_W[A]=j_!1$ for the transverse
square. On a normal fiber the pullback is an orientation
isomorphism, which identifies the corresponding Thom classes.

\section{The excess bundle in a clean intersection}

Let $i:B\hookrightarrow W$ and $A\hookrightarrow W$ be
closed embedded submanifolds with a clean intersection.
This means that $Z=A\cap B$ is a submanifold and
$T_zZ=T_zA\cap T_z B$ for every $z\in Z$. For simplicity
assume the normal ranks are constant. Clean intersection
makes the normal derivative injective, giving
\[
 0\longrightarrow \nu_{Z/B}
   \longrightarrow \nu_{A/W}|_Z
   \longrightarrow E_{\mathrm{exc}}
   \longrightarrow 0 .
\]
The excess bundle $E_{\mathrm{exc}}$ measures the normal
directions of $A$ that are absent from the image of
$T_z B$. Suppose that $\nu_{A/W}$ and $\nu_{Z/B}$ are
oriented and orient $E_{\mathrm{exc}}$ by the displayed
order of the sequence.

\begin{theorem}[Clean intersection formula]
\label{thm:clean-intersection}
Under these hypotheses,
\[
 i^*\operatorname{PD}_W[A]=j_!e(E_{\mathrm{exc}})
 \quad\text{in }H^d(B;\mathbb Z),\qquad j:Z\hookrightarrow B .
\]
\end{theorem}

\begin{proof}
We compute the pullback near $Z$ in tubular neighborhoods.
A neighborhood of $Z$ in $B$ is represented by normal
vectors $\xi$ in $\nu_{Z/B}$. Over the corresponding
neighborhood of $A$ in $W$ the inclusion is a map into
$\nu_{A/W}$. Write it as
$\xi\mapsto (\alpha(\xi),v(\xi))$, where the first entry
lies in $A$ and the second in the normal fiber over
$\alpha(\xi)$. Here $v(0_z)=0$ and $\alpha(0_z)=z$.
Shrink the domain fiberwise if necessary. For $t>0$ the
rescaled map has value
\[
 \bigl(\alpha(t\xi),t^{-1}v(t\xi)\bigr).
\]
As $t\to 0$, it extends smoothly to the normal derivative
over $Z$, as is seen by integrating the first derivative
in a bundle chart along the line from $0_z$ to $\xi$.
For $t>0$ it is nonzero off the zero section since
$i^{-1}(A)=Z$. At $t=0$ it is nonzero there since the normal
derivative is injective. Thus the relative Thom pullback is
computed by the linear injection of normal bundles.

Choose a splitting so that the target normal bundle over $Z$
becomes $\nu_{Z/B}\oplus E_{\mathrm{exc}}$. The Thom class of
the sum is the product of the two Thom classes in the displayed
order. This follows by restricting to each fiber, where the
product is the orientation generator of the oriented direct
sum. Pulling back by the inclusion as the first summand
leaves its Thom class and pulls back the second class by its
zero section. The latter pullback is $e(E_{\mathrm{exc}})$.
Extending the resulting relative class to $B$ and forgetting
the complement gives precisely $j_!e(E_{\mathrm{exc}})$.
\end{proof}

An ordinary clean pullback displays the manifold $Z$, but
its tangent sequence also contains the cokernel bundle.
Transversality makes that cokernel vanish and recovers the
rank-zero Thom formula. A tangent-complex description keeps
it in a second degree, as developed later for derived
intersections. The formula above identifies a classical
contribution of that bundle without replacing the clean
intersection by a finite set.

\section{Self-intersection and tangent zeros}

Take $i$ to be the inclusion of $A$ itself. The intersection
inside $B=A$ is all of $A$, and the excess bundle is
$\nu_{A/W}$. Thus restricting $\operatorname{PD}_W[A]$
to $A$ gives the Euler class of the normal bundle. In
$\mathbb{CP}^2$, let $A$ be the projective line $z_0=0$.
The line $z_1=0$ is isotopic to $A$, through the rotations of
the first two homogeneous coordinates. The two lines meet
in one point $[0:0:1]$. In the affine chart at that point
they are complementary complex coordinate lines. Complex
linear identifications preserve the underlying real
orientation, so the intersection sign is positive. The
normal Euler number of $A$ is therefore $1$. The
self-intersection can be computed by a transverse motion
even though the original pullback is a two-dimensional
manifold.

When $E=TB$ for a closed oriented manifold, a vector field
with nondegenerate zeros computes the tangent Euler number
by the same formula. Take a Morse function and a metric,
using its gradient as the section. At a critical point
of index $\lambda$, the derivative of the gradient is the
inverse metric composed with the Hessian, so its determinant
sign is $(-1)^\lambda$. Each corresponding Morse handle
contributes that same sign to the Euler characteristic
by Chapter~\ref{ch:morse}. Thus
$\langle e(TB),[B]\rangle=\chi(B)$, the
Poincar\'e--Hopf equality in the transverse setting
\cite[Ch.~5]{hirsch}. For example, a height function of
$S^2$ has one minimum and one maximum. Both indices are
even, and the two contributions give $2$.
\par



\chapter{Morse functions and handle filtrations}
\label{ch:morse}

The sublevel sets of a smooth real-valued function form an
increasing family of spaces. A regular band contributes a
product, while an isolated nondegenerate critical point
adds a handle. This turns the real-valued function into a
description of the topology of its source. We use the regular
and critical level arguments of differential topology
\cite[Ch.~6]{hirsch} to describe the filtration and its
homotopy pushouts explicitly.

\section{The Hessian and the local form}

Let $f:M\to\mathbb R$ be smooth and let $p$ satisfy $df_p=0$.
Second partial derivatives in a chart define a symmetric
bilinear form on $T_pM$. Under a change of coordinates
the Hessian is conjugated by the Jacobian, since the term
involving the second derivatives of the coordinate change
is multiplied by $df_p=0$. A critical point is
\emph{nondegenerate} if this form is nonsingular, and its
\emph{index} is the dimension of a maximal negative definite
subspace. A Morse function has only nondegenerate critical
points. They are isolated by the inverse function theorem
applied to the coordinate components of $df$. For a compact
source the critical set is closed and therefore finite.

\begin{theorem}[Morse local form]
\label{thm:morse-lemma}
If $p$ has index $\lambda$ in an $n$-manifold, there are
local coordinates $(u,v)$ centered at $p$, of dimensions
$\lambda$ and $n-\lambda$, such that
\[
 f(u,v)=f(p)-\|u\|^2+\|v\|^2 .
\]
\end{theorem}

\begin{proof}
In coordinates $x$ near $p=0$, Taylor's integral formula
writes the difference as $x^{\mathsf T}A(x)x$ with
\[
 A(x)=\int_0^1(1-t)\,D^2 f(tx)\,dt .
\]
This is a smooth symmetric matrix and $A(0)$ is nonsingular.
A fixed linear change of basis makes $A(0)$ diagonal with
entries $-1$ and $1$. In a neighborhood of $0$ all successive
pivots in the symmetric elimination of $A(x)$ are nonzero.
Eliminating one row and column at a time is the completion
of a square, using only division by the current nonzero
pivot. The signs of the pivots are locally constant. Taking
square roots of their absolute values gives a smooth
factorization $A(x)=B(x)^{\mathsf T} D B(x)$, where $D$ is the
same diagonal sign matrix. The map $x\mapsto B(x)x$ has
invertible derivative $B(0)$ at $0$. Its components are the
required local coordinates by the inverse function theorem.
\end{proof}

The index is stable under small perturbations of the
Hessian through nonsingular forms. If $f_s$ is a smooth
family and $f_{s_0}$ has a nondegenerate critical point $p$,
the equation $df_s(x)=0$ near $(s_0,p)$ has a unique smooth
local solution $p(s)$ by the implicit function theorem.
The index is constant on this branch. This describes how
critical points move through regular parameter values.
When the Hessian becomes singular, the equation no longer
has to be locally the graph of a smooth section.

\section{Regular bands}

Write $M^a=\{x:f(x)\le a\}$. Suppose a closed band
$f^{-1}([a,b])$ is compact and contains no critical point.
Fix a Riemannian metric and use the vector field
\[
 V=-\nabla f/\|\nabla f\|^2
\]
on a neighborhood of the band. Along its flow
$f$ decreases at unit rate. A flow line starting at level $s$
therefore reaches level $a$ in time $s-a$.
Existence for this duration follows by compactness, since
the flow stays in the band until it reaches the designated
end. Flowing backward gives the diffeomorphism of the band
with $f^{-1}(a)\times[a,b]$. This proves that both boundaries
are regular copies of the same level. Flowing each point
of $M^b\setminus M^a$ for time $t(f(x)-a)$, for $0\le t\le1$,
and fixing $M^a$, gives a deformation retraction
$M^b\to M^a$. The formula is continuous at level $a$
because the travel time tends to zero.

This argument requires compactness of the band, or
another condition giving the same flow existence. A
nonproper function on a noncompact manifold can change
level behavior at infinity with no critical point. For
example, $f:\mathbb R\to\mathbb R$ given by $f(x)=e^x$
has empty levels at nonpositive values and single-point
levels at all positive values. Its derivative is nowhere
zero, and a band straddling zero is noncompact.

\section{Crossing one critical level}

Assume $M$ is closed and $p$ is the only critical point
in a band around $c=f(p)$. Set $a=c-\varepsilon$,
$b=c+\varepsilon$ with $\varepsilon>0$ small. Write
$D^j$ for the closed unit disk in $\mathbb R^j$, including
$D^0=\{0\}$ and $\partial D^0=\varnothing$. The topological
handle of index $\lambda$ is
$D^\lambda\times D^{n-\lambda}$ attached along
$\partial D^\lambda\times D^{n-\lambda}$.

\begin{theorem}[Sublevel attachment]
\label{thm:handle-pushout}
The upper sublevel $M^b$ has the homotopy type, relative
to $M^a$, of $M^a$ with one index-$\lambda$ handle attached.
Consequently there is a homotopy pushout of spaces
\[
\begin{tikzcd}[column sep=large,row sep=large]
 \partial D^\lambda \arrow{r} \arrow{d} & M^a \arrow{d} \\
 D^\lambda \arrow{r} & M^b .
\end{tikzcd}
\]
\end{theorem}

\begin{proof}
Work in the chart of Theorem~\ref{thm:morse-lemma}. Choose
a metric equal to the Euclidean metric there near $p$.
Choose $r>0$ with $r^2<\varepsilon$ and both small enough
that
\[
 H=\{(u,v):\|v\|\le r,\quad
             \|u\|\le\sqrt{\varepsilon+\|v\|^2}\}
\]
is contained in this metric neighborhood. Dividing $u$ by
$\sqrt{\varepsilon+\|v\|^2}$ and rescaling $v$ gives
the disk product. The inequality $f\ge a$ on $H$ is
an equality precisely on the attaching face. The
inequality $f\le c+r^2<b$ puts the handle inside the upper
sublevel. Hence $M^a\cap H$ is the attaching face.

Follow negative gradient flow outside $M^a\cup H$. There
are no critical points on the compact closure of the
region still to be added, since $p$ is interior to $H$
relative to $M^b$. The decrease rate of $f$ there is
bounded below in absolute value by a positive constant.
The flow reaches $M^a\cup H$ in finite time. At its
level-$a$ boundary the flow crosses strictly toward the
lower sublevel. At the other face of $H$, $\|v\|=r$,
the local gradient equation is $v'=-2v$, so the flow
crosses strictly inward there as well. These inequalities
give continuous first-entry time, using the first of the
two transverse crossings where the faces meet. Defining
time zero inside the union yields a deformation retraction
of $M^b$ onto $M^a\cup H$ by running a fixed fraction of
this time. If a disk factor has dimension zero, its empty
boundary requires no crossing argument.

The attaching face and its inclusions admit collars
in their respective spaces when the face is nonempty.
The resulting pushout computes the homotopy pushout.
Projection from the disk product to $D^\lambda$ is an
equivalence of pairs with the corresponding boundary.
Using the center $v=0$ for the attaching map therefore
gives the displayed square.
\end{proof}

The smooth handle attachment statement uses a smoothing
of the corners of such a model. It describes the boundary
surgery as well as the homotopy type
\cite[\S3]{milnor-morse}. For the homotopy filtration,
each interval between critical values can be replaced
by a single arrow. If the critical values coincide, a
small bump constant near each critical point separates
them. The transition regions of these bumps lie in a
compact set where $\|df\|$ has a positive lower bound,
so sufficiently small bumps create no new critical points.
The filtration then starts at the empty set and successively
attaches a cell of each critical index. The relative
homology of each step is free of rank one in the critical
index and zero in the other degrees. Additivity in the
long exact sequences of the pairs gives
\[
 \chi(M)=\sum_{p\in\operatorname{Crit}(f)}(-1)^{\operatorname{ind}(p)} .
\]

On the torus use $f(s,t)=\cos s+2\cos t$. The four
critical points occur where $s,t$ are $0$ or $\pi$.
The Hessian is diagonal with entries $-\cos s$ and
$-2\cos t$. The critical values in order are $-3,-1,1,3$,
with indices $0,1,1,2$. The filtration first creates a
component, then adds two one-dimensional cells, and
finally fills a two-dimensional cell. It gives
$\chi(T^2)=1-2+1=0$.

Indices in a family do not determine handle transport.
For $\theta\in\mathbb R/(2\pi\mathbb Z)$, let
$R(\phi)$ be the rotation of $\mathbb R^2$ by $\phi$ and
define the quadratic form
\[
 x^{\mathsf T}R(\theta/2)
 \begin{pmatrix}-1&0\\0&1\end{pmatrix}
 R(\theta/2)^{\mathsf T}x .
\]
It is well defined on the circle because adding $2\pi$
changes both rotation matrices by a sign. The origin
is a nondegenerate critical point of index one in each
fiber. Its negative eigenspace is spanned locally by
$R(\theta/2)(1,0)$, which reverses sign on a full turn.
The negative spaces form a M\"obius line, so the local
orientations of the index-one disks reverse under
transport. The critical point and its index are constant
throughout, while the relative homology generator of
the local handle has sign monodromy.
\par



\chapter{Trajectory spaces and Morse homology}
\label{ch:morsecx}

A handle filtration determines the ranks of the Morse chain
groups. To recover the boundary maps one needs information
about the attaching maps. Negative gradient trajectories
provide that information through compact spaces whose
boundary strata consist of successive trajectories.
Counting zero-dimensional spaces produces a chain complex.
Retaining all the spaces instead produces a category
enriched in spaces.

\section{Flow lines and the boundary operator}

Let $M$ be closed, $f:M\to\mathbb R$ Morse, and $g$
a metric. The negative gradient is complete by compactness.
For a critical point $p$, let $W^u(p)$ and $W^s(p)$ be
the unstable and stable manifolds. They consist of points
whose flow converges to $p$ at time $-\infty$ and
$+\infty$, respectively. The stable manifold theorem gives
their dimensions as $|p|=\operatorname{ind}(p)$ and
$\dim M-|p|$. These dimensions come from the positive
and negative eigenspaces of the linearized negative
gradient. Assume the pair is Morse--Smale, meaning all
$W^u(p)$ and $W^s(q)$ intersect transversely. For $p\ne q$,
the intersection, when nonempty, consists of nonstationary
trajectories and has dimension $|p|-|q|$. Quotienting by time translation
gives the moduli space $\mathcal M(p,q)$. When the space
is nonempty, $f(p)>f(q)$. Every trajectory crosses a fixed
regular level between the two values exactly once, so one
can realize the quotient as a smooth level slice.
Its dimension is $|p|-|q|-1$.

The Morse compactification theorem adds broken
trajectories to these spaces. The stratum with a sequence of
breaks through $p_1,\ldots,p_k$ is
\[
 \mathcal M(p,p_1)\times\mathcal M(p_1,p_2)
   \times\cdots\times\mathcal M(p_k,q).
\]
In particular, under a metric Euclidean in Morse coordinates
near the critical points, the spaces have smooth corner
models with these strata in codimension $k$
\cite[Thms.~2.3 and~2.6]{wehrheim-morse}.
To see the source of the boundary, cut trajectories at
regular levels between the critical values. The compact
slices control where segments can travel. If the time
between two slices tends to infinity, a limiting segment
can approach an intervening critical point. The
compactness theorem shows that only finitely many such
breaks occur and gives the resulting limiting tuple. The
gluing theorem supplies a nonnegative collar parameter
for each break, with zero recording the broken tuple.
The proofs below use both compactness and this description
of the boundary. A set of broken trajectories alone does
not establish the existence of the collars.

Let $C_j(f)$ be the $\mathbb F_2$-vector space on critical
points of index $j$. For a difference of indices equal to
one, each nonconstant segment drops the index by at least
one, so the compactification has no breaks. Thus $\mathcal M(p,q)$ is
compact and zero-dimensional, hence finite. Set
\[
 \partial p=
 \sum_{|q|=|p|-1}\#_2\mathcal M(p,q)\,q ,
\]
where $\#_2$ denotes parity. For $|p|-|r|=2$,
$\overline{\mathcal M}(p,r)$ is a compact one-manifold
whose boundary is the disjoint union of
$\mathcal M(p,q)\times\mathcal M(q,r)$ over
$|q|=|p|-1$. The coefficient of $r$ in $\partial^2p$
is the parity of that boundary. It is zero since each
interval has two endpoints. Over the integers, compatible
orientations of unstable manifolds give the trajectory
signs, and the oriented-boundary formula yields the
signed version \cite[Chs.~3--4]{schwarz-morse}.
The Morse homology theorem identifies the resulting
homology with singular homology. Its comparison uses
the generator of a handle pair from
Chapter~\ref{ch:morse} and identifies boundary
coefficients by flowing an attaching sphere to a
transverse normal slice at the next lower index
\cite[Ch.~4]{schwarz-morse}.

\section{Continuation and homotopies of continuations}

Let $(f_s,g_s)$, $s\in\mathbb R$, interpolate between two
Morse--Smale pairs and be constant for $|s|$ sufficiently
large. A continuation trajectory satisfies
\[
 \dot x(s)=-\nabla_{g_s}f_s(x(s)),\qquad
 \lim_{s\to-\infty}x(s)=p,\quad
 \lim_{s\to+\infty}x(s)=q .
\]
The fixed transition interval distinguishes a time
coordinate. In particular one does not divide out an
$\mathbb R$-translation action for this equation.
For regular interpolation data, the dimension is
$|p|-|q|$. Indeed, transfer $W^u(p)$ by the time-dependent
flow through the transition interval and intersect it
transversely with the terminal $W^s(q)$. Counting the
zero-dimensional continuation spaces gives a
degree-zero map $\Phi:C_*(f_-)\to C_*(f_+)$. Compactified
one-dimensional spaces have breaking at the two
stationary ends. The two types of boundary points
give the two terms in
$\partial_+\Phi+\Phi\partial_-=0$ over $\mathbb F_2$.

A regular one-parameter comparison of interpolations
has the additional parameter direction. For an index
difference of $-1$, counting it gives a degree-one
operator $K$. In index difference zero, its compactified
one-manifold has boundary from the two parameter ends
and from breaking at either stationary end. The
boundary parity therefore gives
\[
 \Phi_1+\Phi_0=\partial_+K+K\partial_- .
\]
Gluing along a long intermediate interval compares
composite continuation with a single interpolation.
Reversing the interpolation and comparing the
concatenation to the constant interpolation gives
inverse maps on homology \cite[Ch.~4]{schwarz-morse}.
These statements take regularity and compactified
gluing for the continuation equations as inputs.
The identities explain which boundaries are needed
for invariance and why a comparable chain map is
specified only up to chain homotopy.

\section{The enriched flow category}

The objects of the flow category $\mathcal C_f$ are
critical points. Its nonidentity morphism spaces are
$\overline{\mathcal M}(p,q)$. Concatenating broken
tuples is continuous and associative. In particular a
codimension-two corner for a three-stage trajectory
gives the same ordered tuple under either grouping.
The following diagram records the two face inclusions.
We abbreviate a compactified space by $C_{pq}$.
\[
\begin{tikzcd}[column sep=large,row sep=large]
 C_{pq}\times C_{qr}\times C_{rs}
    \arrow{r} \arrow{d} &
 C_{pr}\times C_{rs} \arrow{d} \\
 C_{pq}\times C_{qs} \arrow{r} & C_{ps}.
\end{tikzcd}
\]
The nerve is a simplicial space, with level $n$ given
by composable strings of $n$ morphisms. Faces compose
or forget arrows and degeneracies insert identities.
Applying the singular complex to each mapping space
gives a simplicial category, whose homotopy-coherent
nerve is an $\infty$-category. A path within one
trajectory space now gives a higher comparison of
arrows. The classifying-space result of
Cohen, Jones, and Segal recovers the homotopy type of
$M$ from the realization of the topological nerve in
this Morse--Smale case
\cite[Thms.~2.6 and~3.1]{cjs-flow}.
Composition here uses recorded break data. The
continuation calculation above explains the
one-dimensional coherence needed at the chain level.
An analogous construction of continuation on all the
enriched data would require compatible multiparameter
gluing.

\section{Two explicit flow categories}

On the circle take $f(e^{it})=\cos t$ and the standard
metric. The flow equation is $\dot t=\sin t$. The
maximum at $0$ and the minimum at $\pi$ are connected
by two unparametrized trajectories, one in each open
semicircle. There are no intermediate critical points
at which to break. The mod-two differential vanishes
because the trajectory count is two. The realization
of the flow category has two vertices and two edges
joining them, which gives a circle directly.

On the round sphere take the height function. The
maximum has index two, the minimum has index zero,
and the nonstationary trajectories are the meridians,
parametrized by their intersections with the equator.
There are no broken trajectories. The morphism space
from maximum to minimum is therefore $S^1$.
The nerve realization is obtained from
$S^1\times[0,1]$ by collapsing its two end circles
separately to the vertices. This is the suspension
$S^2$. If we retain only the path components of the
morphism spaces, the ordinary category has one
nonidentity arrow and its realization is an interval.
Thus taking components of the mapping spaces can
lose the topology of the manifold, even when all the
objects and their indices are retained.

The product torus gives an example with actual breaking.
Use the flat metric and $f(s,t)=\cos s+2\cos t$.
The negative gradient preserves the two coordinate equations,
and the stable and unstable manifolds are products of those
on the two circles. Transversality holds because it holds
in each factor. For index difference one, exactly one
coordinate moves, and the two semicircles in that factor
give parity zero. Thus the mod-two complex has one
generator in each of degrees zero and two, and two in
degree one, with zero differential.

For a trajectory from the maximum to the minimum both
coordinates move. Choosing one of the two semicircles
in each factor leaves a time shift for each of the two
one-dimensional motions. Modulo common translation
the difference of shifts is a real coordinate. Each
of the four choices therefore gives a real line in the
unbroken moduli space. In the two limits of that
coordinate, one coordinate transition occurs before the
other, giving the two orders of a broken trajectory.
The four closed intervals in the compactification
exhibit eight boundary points. The two possible
intermediate critical points account for four each.
\par



\bookpart{IV}{Cobordism, isotopy, and surfaces}

\chapter{Geometric bordism and the Thom construction}
\label{ch:bordism}

A closed manifold is a possible boundary, and a bordism asks whether
two closed manifolds can occur as the ends of a single manifold.
The question keeps track of tangent or normal structures when
these are part of the geometry. Its stable answer is a homotopy
group. The passage from a manifold to that group uses the same
two constructions that govern intersections: a tubular
neighborhood and a transverse inverse image.

\section{Bordisms and their elementary groups}

A closed $d$-manifold is compact and without boundary.
An unoriented bordism from $M_0$ to $M_1$ is a compact
$(d+1)$-manifold $W$ with designated ends
$\partial W\simeq M_0\amalg M_1$. In the oriented case the
boundary orientation identifies $\partial W$ with
$(-M_0)\amalg M_1$. The two conventions use an outward
normal first in the orientation of a boundary. One can choose
collars, as in Chapter~\ref{ch:tubular}, to glue successive
bordisms smoothly. The cylinder proves reflexivity, and
reversing the roles of the ends proves symmetry, using $-W$
in the oriented setting. Disjoint union gives an abelian
group of equivalence classes. The cylinder also gives
$[M]+[-M]=0$ in the oriented group, and $2[M]=0$ in the
unoriented group. Denote the respective groups by
$\Omega_d^{SO}$ and $\Omega_d^O$.

\begin{example}[Dimension zero]
An oriented closed zero-manifold is a finite list of positive
and negative points, with multiplicities. Each oriented interval
has one endpoint of each sign, and a circle has no endpoints.
The signed count consequently gives an invariant in $\mathbb Z$.
A finite collection with count zero bounds by pairing the
positive and negative points and taking one oriented interval
for each pair. Thus $\Omega_0^{SO}\simeq\mathbb Z$. Dropping
the orientations retains the parity of the number of points,
and the same interval construction proves
$\Omega_0^O\simeq\mathbb Z/2$.
\end{example}

Products endow the direct sum of the degree groups with graded
multiplication. A bordism $W$ for $M_0,M_1$ produces the
bordism $W\times N$ when $N$ is closed. In the oriented case
the factor order contributes the sign $(-1)^{de}$ on swapping
oriented manifolds of dimensions $d$ and $e$. A related
construction equips a manifold with a continuous map to a
space $X$, with both this map and its structure required to
extend across a bordism. This is a geometric way for a theory of
manifolds to run over variable spaces. We first consider the
coefficient groups where $X$ is a point.

\begin{example}[An unoriented surface that does not bound]
For a compact smooth $3$-manifold $W$, the double of $W$ is
a closed $3$-manifold. Its Euler characteristic is zero, by
Poincar\'e duality over $\mathbb F_2$. Additivity of Euler
characteristic gives
$\chi(\partial W)=2\chi(W)$, so no closed surface of odd Euler
characteristic can be its whole boundary. In particular
$\mathbb{RP}^2$ represents a nonzero element of $\Omega_2^O$.
This computation uses duality only as a vanishing statement on
the double. It applies regardless of orientability.
\end{example}

\section{Collapsing a normal neighborhood}

Embed $M^d$ in $\mathbb R^{d+k}$ and let $\nu$ be its normal
bundle of rank $k$. For the Thom space write
$\operatorname{Th}(\nu)=D(\nu)/S(\nu)$, using a metric on
$\nu$. A tubular identification of $D(\nu)$ with a closed
neighborhood of $M$ in the one-point compactification
$S^{d+k}$ gives a pointed collapse. The classifying map of
$\nu$ induces a bundle map to the universal rank-$k$ bundle
$\gamma_k$ on $BO(k)$. With compatible metrics it induces
the second arrow of
\[
 \begin{tikzcd}[column sep=3.5em]
 S^{d+k}\arrow[r,"\text{collapse}"] &
 \operatorname{Th}(\nu)\arrow[r,"\text{classifying map}"] &
 \operatorname{Th}(\gamma_k).
 \end{tikzcd}
\]
An oriented $M$ induces a normal orientation using the
ambient orientation and the order $TM\oplus\nu$. In that
case use the universal oriented bundle over $BSO(k)$.

The collapse sends the complement of the open normal
disk bundle to the base point. For a smaller radius the
annulus between the two boundaries can be absorbed by a
radial change of coordinates. An isotopy of tubular
neighborhoods gives a homotopy of collapse maps by applying
this construction over the interval. Increasing the ambient
dimension by one replaces the normal bundle by
$\nu\oplus\mathbb R$. Under this operation the collapse
suspends:
\[
 \begin{tikzcd}[column sep=3.6em,row sep=2.2em]
 \Sigma S^{d+k}\arrow[r,"\Sigma c_\nu"]\arrow[d,"\simeq"'] &
 \Sigma\operatorname{Th}(\gamma_k)\arrow[d]\\
 S^{d+k+1}\arrow[r,"c_{\nu\oplus\mathbb R}"'] &
 \operatorname{Th}(\gamma_{k+1}).
 \end{tikzcd}
\]
The right arrow comes from adjoining a trivial line to each
normal plane. Choices of embeddings disappear after
stabilization. Indeed, for embeddings $i_0,i_1$ placed in
independent coordinate summands, the family
$x\mapsto ((1-t)i_0(x),t i_1(x))$ stays embedded for
$0\le t\le 1$. Initial embeddings may therefore be compared
using a smooth cylinder in a common larger ambient space,
with orientation-preserving rotations of ambient coordinates
if an ordering of the normal directions has been specified.

\section{Transversality gives the inverse}

\begin{theorem}[Stable Pontryagin--Thom correspondence]
\label{thm:pontryagin-thom}
Collapse induces isomorphisms
\[
 \Omega_d^O\simeq\pi_d MO,
 \qquad
 \Omega_d^{SO}\simeq\pi_d MSO .
\]
The $k$th spaces in these Thom spectra are the Thom spaces
of the universal rank-$k$ bundles over $BO(k)$ and $BSO(k)$.
The stable group on the right is the colimit of the groups
in degrees $d+k$ under suspension and the Thom structure
maps.
\end{theorem}

\begin{proof}
We give the normal-bundle argument
\cite[Ch.~7]{hirsch}\cite[Lectures 2--5]{freed-bordism}.
Use finite Grassmannians for the bundle classifying maps,
enlarging their ambient dimension as needed. Representatives
with compact source, and their homotopies, can be taken at
finite stages by cellular approximation. This permits the
usual smooth transversality argument near the zero section.
The disk bundle away from the sphere identification is a
smooth total space, with the zero section a closed submanifold
of codimension $k$.

For a pointed map $a:S^{d+k}\to\operatorname{Th}(\gamma_k)$
represented at such a stage, approximate it smoothly near the
inverse image of the zero section and make it transverse there.
One may leave the map fixed near the base point of the source.
To see why local smoothing is enough, first choose a smaller
disk bundle about the zero section. The complement of its
inverse image maps away from zero and transversality needs
to be arranged only on a neighborhood of the smaller inverse
image. An arbitrarily small perturbation there, kept equal to
the original outside the larger inverse image, introduces no
zeros in the intervening closed annulus.

The transverse inverse image $M$ is a closed
$d$-manifold in the finite chart $\mathbb R^{d+k}$ of the
source. The differential identifies its normal bundle with
the pullback of $\gamma_k$ along $a|_M$. This is just the
isomorphism from the normal space to the quotient by the
tangent space of the zero section. In the oriented version
this gives its orientation by the stated convention.
The same argument applied to a homotopy, made transverse
relative to its endpoints, yields a bordism in
$S^{d+k}\times[0,1]$. It has the required collars after the
homotopy is made constant in the parameter near its ends.
Suspension appends a trivial direction to the normal bundle
of the inverse image.

Conversely, collapse along an embedded manifold is transverse
near the zero section and recovers that manifold with its
normal data. For a transverse $a$, shrink to a tubular
neighborhood of its inverse image by radially pushing the
complement of a smaller bundle neighborhood to the base
point. The resulting map is homotopic to normal collapse:
on a smaller tube one compares the normal-coordinate map
with its derivative along $M$. In local bundle coordinates
this uses the homotopy given by the rescaled normal component
$t^{-1}F(x,tv)$, which extends to the derivative at $t=0$.
Parallel transport in the target bundle identifies the nearby
fibers with the fibers over $a|_M$ for this calculation.
The derivative is a normal isomorphism. Compactness gives
a common tube on which the comparison is defined, and a
radial collapse sends its exterior to the base point. This
gives inverse operations on stable classes.

An abstract bordism has a neat embedding into a large
$\mathbb R^{d+k}\times[0,1]$ with product form near its ends.
This can be obtained by prescribing the collar coordinate
as the interval coordinate near the boundary and combining
it with a relative Euclidean embedding in the remaining
coordinates. After stabilization the end embeddings can
be adjusted by the embedding isotopies above. Collapsing
its normal neighborhood then gives the required homotopy
between end collapses. Finally, disjoint union corresponds to
addition by placing the two tubes in disjoint balls and
pinching the complement to the base point.
\end{proof}

\section{What stabilization remembers}

A normal framing is a specified ordered trivialization of
the normal bundle, stabilized by adding the standard last
coordinate. It gives a collapse to $S^k$, since the classifying
map of the trivialized bundle factors through the sphere model
of the framed theory. The same preimage operation recovers a
framing from a transverse inverse image of a regular point
of $S^k$. Stable framed bordism therefore has groups
$\pi_d^S=\varinjlim_k\pi_{d+k}(S^k)$. Forgetting a framing
and retaining its induced orientation gives the map from the
sphere spectrum to $MSO$ on coefficient groups. Distinct
framings on the same underlying manifold can give distinct
classes. In dimension zero, the relative orientation of an
ordered normal frame and the ambient frame contributes the
sign of the point, recovering the stable degree.

The theorem classifies manifolds after passing to bordism
classes. A space of embeddings and collars also retains
paths, homotopies of paths, and automorphisms of individual
geometric representatives. The collapse construction runs in
families as well as on points. The next chapter retains
composable geometric bordisms before taking a classifying
space and distinguishes the Thom spectrum in that setting
from the normal Thom spectrum used here.
\par



\chapter{Bordism as a higher category}
\label{ch:bordism-cats}

Collapsing a manifold records its bordism class. Bordisms also
form a system of composable arrows, and many of their
geometric constructions compare arrows rather than objects.
At the next level, a family of diffeomorphisms compares
such comparisons. A categorical model must specify which
differences are retained at each level.

\section{Gluing and its space of choices}

Fix $d$. At a real level $a$, an object is a closed
$(d-1)$-manifold embedded in $\{a\}\times\mathbb R^N$.
An arrow to an object at level $b>a$ is a compact
$d$-submanifold $W\subset[a,b]\times\mathbb R^N$ whose only
boundary consists of the designated objects and whose
embedding is a product in a neighborhood of each end.
At an interface level $b$, two consecutive arrows agree
on the object and their product neighborhoods glue. At the
level of these embedded representatives the composition is
union, so equality of the two composites of three
consecutive arrows follows from associativity of union.
This defines the positive-length part of an embedded
cobordism category $C_d$. One includes identities when
forming its category or uses the corresponding model with
nonidentity arrows and a semisimplicial classifying space.

Without fixed ambient levels one can glue two abstract
bordisms by identifying boundary collars. The overlap chart
on the union uses a signed coordinate on the two inward
half collars. Its smooth structure is independent up to
a diffeomorphism of collars, and repeated gluing keeps
the same signed coordinate along each seam. Retaining the
space of such gluings is useful for comparison with
embedded families.

For a fixed abstract bordism $W$ with its collared ends,
the embedding model has a space of embeddings of $W$
modulo its diffeomorphisms. A diffeomorphism acts
freely on parametrized embeddings by precomposition.
On a closed manifold, stable embeddings in increasing
Euclidean dimensions form a weakly contractible space:
a compact family at a finite stage can be deformed to
one fixed embedding using an independent block, as in
Chapter~\ref{ch:bordism}. For a sphere of parameters this
deformation fills the sphere by contracting it. In these
models one obtains classifying spaces of diffeomorphism
groups by taking the homotopy quotient of the embedding
space under reparametrization. The use of homotopy
quotients retains families with nontrivial monodromy.

\begin{example}[Parity and the classifying space]
Taking the nerve and geometric realization of $C_d$ gives
a space $BC_d$. An arrow becomes a path in this space
and can be traversed in either direction. On connected
components this identifies objects exactly up to
unoriented bordism. To verify the claim in one direction,
any abstract bordism can be embedded in a sufficiently
large Euclidean slab with collared ends. For the other
direction, paths in the space of object embeddings preserve
their diffeomorphism type, and one-simplices preserve the
bordism class. A finite zigzag with backward arrows gives
a bordism after reversing the corresponding slabs. In
particular $\pi_0 BC_1\simeq\mathbb Z/2$ via parity of points.
This calculation alone leaves the higher homotopy groups of
$BC_1$ undetermined.
\end{example}

\section{The spectrum of an embedded cobordism category}

The Pontryagin--Thom spectrum of the preceding chapter
stabilizes the normal rank of individual closed manifolds.
For the classifying space of a category with fixed
$d$-dimensional arrows one instead fixes the tangent rank
and varies the complementary normal directions. Over
$\operatorname{Gr}_d(\mathbb R^{d+n})$ let $\gamma_{d,n}$
be the tautological $d$-plane bundle and
$\gamma_{d,n}^{\perp}$ its rank-$n$ orthogonal complement.
The spectrum $MTO(d)$ is the Thom spectrum of the
virtual bundle $-\gamma_d$ on $BO(d)$. Its
$(d+n)$th space can be modeled by
$\operatorname{Th}(\gamma_{d,n}^{\perp})$. Adding an
ambient coordinate adds a trivial line to the complements
along the inclusion of Grassmannians and gives the
structure maps.

\begin{theorem}[Galatius--Madsen--Tillmann--Weiss]
The classifying space of the stable embedded unoriented
$d$-cobordism category is weakly equivalent to
$\Omega^{\infty-1} MTO(d)$. For the category with tangent
structures classified by a fibration $\theta:X\to BO(d)$,
the corresponding spectrum is the Thom spectrum
of $-\theta^*\gamma_d$.
\end{theorem}

Here the shifted infinite loop space is represented
by the stabilization of the spaces
$\Omega^{d+n-1}\operatorname{Th}(\gamma_{d,n}^{\perp})$.
The theorem of \cite{gmtw}, whose proof uses
parametrized geometry, starts with collapse near embedded
manifolds over their level coordinate. The tangent
$d$-plane specifies the Grassmannian coordinate of the
collapse. The remaining directions supply a point in
the complementary normal disk. Compatibility with a
cut recovers categorical composition at the level of
the classifying space. The theorem's converse compares
families of such local data to geometric bordisms, a
stronger statement than the inverse-image computation on
individual maps in Theorem~\ref{thm:pontryagin-thom}.
The difference between the spectra is visible in their
bundle ranks. In $MO$ the universal normal bundle's
rank increases in the stabilization. In $MTO(d)$ the
positive tangent plane has fixed rank $d$, and its
negative virtual class is stabilized.

An oriented version uses $BSO(d)\to BO(d)$. The
structure on a $(d-1)$-dimensional object is a lift of its
tangent bundle with a distinguished added interval direction.
This recovers the boundary orientation convention when
gluing two oriented arrows. In dimension zero the
positive-length category has only the empty manifold as
an object type, and its arrows are finite point configurations
in open slabs. The emptiness of the boundary in this case
is consistent with the shift: $MTO(0)$ is the sphere
spectrum and its shifted infinite loop space is connected.

\section{Corners, duals, and field theories}

A fully extended bordism category starts at dimension
zero. Its $j$-morphisms for $0\le j\le d$ are
$j$-dimensional bordisms with the relevant corners and
collars. Geometrically a corner has several independent
boundary coordinates, so that composition is gluing along
a selected face. As in the ordinary bordism category, one
retains diffeomorphisms and their homotopies above the
geometric dimensions. With disjoint union this gives a
symmetric monoidal $(\infty,d)$-category. In the
$d$-framed version $\mathrm{Bord}^{\mathrm{fr}}_d$, a
$j$-manifold is equipped with a compatible trivialization
of its tangent bundle with $d-j$ additional ordered
directions. Along collars these added directions identify
the boundary framing with the restriction of the framing
in the next dimension.

A dual for an object $A$ consists of an object $A^\vee$ and
coevaluation and evaluation arrows. They have composites
that are equivalent to the respective identities. One of
the two triangle identities has diagram
\[
 \begin{tikzcd}[column sep=small,row sep=2em]
 A \arrow[r,"\operatorname{coev}\otimes 1"]\arrow[dr,"1_A"']&
 A\otimes A^\vee\otimes A\arrow[d,"1\otimes\operatorname{ev}"]\\
 & A .
 \end{tikzcd}
\]
In dimension one the intervals with both ends outgoing or
both ends incoming supply these maps for the two framings
of a point. The triangle identities come from gluing
intervals and rectifying the result as a collared interval.
In higher dimensions the evaluation and coevaluation
morphisms have adjoints and the units and counits have
coherent adjointability data through the permitted levels.
This is the geometric structure behind full
dualizability. Full dualizability is defined in the target
symmetric monoidal $(\infty,d)$-category using these
successive levels of adjointability.

The framed cobordism hypothesis identifies the space of
symmetric monoidal functors out of
$\mathrm{Bord}^{\mathrm{fr}}_d$ with the maximal
$\infty$-groupoid of fully dualizable objects of the target,
by evaluation at the framed point
\cite[\S 2.4]{lurie-cobordism}. The functor space here uses
equivalences of functors. Other tangential structures require
compatible symmetry data on a fully dualizable object.
They do not result from taking the same target space and
merely relabeling the point. This universal property refers
to functors from the extended higher category. The
classifying-space theorem refers to a groupoid completion
of the embedded category with adjacent dimensions, so it
answers a different question about the geometry of
bordisms.

\begin{example}[A trace from gluing]
In dimension one take the target to be finite-dimensional
vector spaces over a field $k$ with the tensor product. If
the positively framed point is sent to $V$, its dual point
is sent to $V^*$. The evaluation and coevaluation
are the pairing and the tensor $\sum_i e_i\otimes e_i^*$.
The tensor is independent of basis because it corresponds
to the identity in $\operatorname{End}(V)$. Gluing the
outgoing ends of the coevaluation to the incoming ends of
the evaluation, with the symmetry exchanging the middle
factors, makes a circle. Its image is multiplication by
$\sum_i e_i^*(e_i)=\dim_k V$ viewed in $k$. More generally,
inserting an endomorphism on the strand gives its usual
trace in the target. The circle evaluation is therefore a
consequence of coherent gluing and the choice of dual.
\end{example}

\begin{example}[A surface cut into pairs of pants]
In dimension two the ordinary oriented bordism category
has circles and their disjoint unions as objects. In a
symmetric monoidal functor to finite-dimensional vector
spaces, put $A$ for the vector space assigned to one circle.
The twice-incoming pair of pants supplies a multiplication
$m:A\otimes A\to A$, and the outgoing disk is its unit.
There are two ways to glue two pairs of pants to get three
incoming circles and one outgoing circle. The resulting
labeled surfaces are diffeomorphic relative to their
boundary collars, so their images give associativity.
A diffeomorphism interchanging the two incoming circles
gives commutativity. Reversing the pants and the disk
gives a comultiplication and a counit $\epsilon$.

A cylinder drawn as a cup can be cut into an outgoing
disk and a one-incoming pair of pants. Its reverse gives
the pairing $(u,v)\mapsto\epsilon(m(u,v))$.
The cup and cap satisfy the cylinder identities, so this
pairing is nondegenerate. The two ways to cut a surface
from two circles to two circles into oppositely directed
pairs of pants give the Frobenius equality between
comultiplication after multiplication and multiplication
on one side of a comultiplication. These statements are
consequences of labeled gluing and the elementary
classification of the relevant surfaces
(Chapter~\ref{ch:surfaces}). They concern theories defined
on circles and surfaces. A fully extended theory also has
values on points and compatible additional gluing.
\end{example}

A geometric family also has a different source of higher
data: paths in an embedding space act through ambient
diffeomorphisms, with a stabilizer of the embedding.
The next chapter constructs these lifts, then describes
the distinction between a path of images and a chosen
lift through the ambient transformations.
\par



\chapter{Isotopies and ambient transformations}
\label{ch:isotopy}

An isotopy is a family of embeddings. An ambient
isotopy is a family of diffeomorphisms of the target
starting at its identity. The extension theorem turns
the first type of family into the second by extending
its time-dependent velocity. The resulting factorization
has a whole space of choices, which can be understood
as a stabilizer in a space of transformations.

\section{Extending velocities}

\begin{theorem}[Isotopy extension with smooth parameters]
\label{thm:isotopy-extension}
Let $A$ be a closed smooth manifold, and let $N$ be a
smooth manifold without boundary. Let $B$ be a compact
smooth parameter manifold. Suppose
$e:B\times[0,1]\times A\to N$ is smooth and every
$e_{b,t}$ is an embedding. There exists a smooth family
of compactly supported ambient isotopies $\Phi_{b,t}$ of
$N$ with
\[
 \Phi_{b,0}=1_N,\qquad
 \Phi_{b,t}\circ e_{b,0}=e_{b,t}.
\]
The support can be taken in one compact subset of $N$
independent of $b,t$. If $e_{b,t}$ is constant in $t$
for designated parameters $b$, then $\Phi_{b,t}$ can
be chosen equal to the identity for those parameters.
\end{theorem}

\begin{proof}
The velocity along $e_{b,t}$ is
$v_{b,t}(e_{b,t}(a))=\partial_t e(b,t,a)$. This is
well-defined because $e_{b,t}$ is injective. Choose a
metric on $N$. The normal spaces to the embedded slices
form a bundle over $B\times[0,1]\times A$. The normal
exponential map with $b,t$ unchanged is a tubular
coordinate map for a uniform sufficiently small
radius. In more detail, it has invertible derivative
along the zero section in the manifold with parameters.
Local inverse maps give neighborhoods near that section.
If there were intersections between distinct normal
disks at arbitrarily small radii, compactness would give
a limiting pair of points over the same $b,t$. The
injectivity of $e_{b,t}$ forces the pair to coincide,
and a local inverse then rules out the intersections
for the subsequence. This gives the required uniform
radius.

For a point in the tube at parameters $b,t$, transport
the slice velocity from the foot point along the normal
geodesic. Multiply by a cutoff in the normal radius
that is one near the zero section and zero near the
boundary of the tube. Extend by zero on the complement.
The resulting vector field $V_{b,t}$ on $N$ depends
smoothly on the parameters and has uniform compact
support. The extension uses the velocity at the same
$b,t$ and sends a zero velocity on a whole slice to
the zero field. The time-dependent flow exists for
$0\le t\le1$ by compact support, and smooth dependence
on initial conditions and parameters gives
$\Phi_{b,t}$. Solving the equation backward shows
each time map is a diffeomorphism. The two curves
$t\mapsto e_{b,t}(a)$ and
$t\mapsto\Phi_{b,t}(e_{b,0}(a))$ have the same initial
point and satisfy the same differential equation.
Uniqueness proves the result.
\end{proof}

The theorem gives a local extension principle as well.
An open subset of $N$ containing the track of the
isotopy contains the tubes of sufficiently small radius.
The vector field may therefore be supported there.
When the images miss a closed obstacle, this produces
an ambient isotopy fixed close to the obstacle.
Compactness is used to obtain a common tube and support.
Without it a pointwise prescription of velocities may
produce a field that escapes to infinity in finite time.
For sources with boundary or a target with boundary,
prescribed tangencies and compatible collar data require
a relative version. The classical extension and gluing
results for those settings are developed in \cite[Ch.~8]{hirsch}.

\section{Restriction and the stabilizer}

For this section take $N$ compact without boundary.
Use the $C^\infty$ topologies on spaces of smooth maps
and their subspaces. Write $i:A\hookrightarrow N$ for
one embedding, $\mathcal D_0(N)$ for the identity path
component of the diffeomorphism group, and
$\mathcal E_i$ for the path component of $i$ in the
embedding space. The restriction map
\[
 q:\mathcal D_0(N)\longrightarrow\mathcal E_i,\qquad
 q(F)=F\circ i
\]
has fiber $K_i=\{F\in\mathcal D_0(N)\mid F\circ i=i\}$
at $i$. In particular $K_i$ fixes the specified
parametrized embedding pointwise and may have many
path components.

\begin{proposition}[The restriction bundle]
\label{prop:restriction-bundle}
The map $q$ is a locally trivial principal $K_i$-bundle.
It yields a homotopy fiber sequence
\[
 K_i\longrightarrow\mathcal D_0(N)
      \longrightarrow\mathcal E_i .
\]
\end{proposition}

\begin{proof}
Here is the local-section argument behind the classical
restriction theorem \cite{palais-embeddings}. When a
smooth embedding $j$ is sufficiently $C^1$-close to $i$,
the vector
$v_j(a)=\exp_{i(a)}^{-1}(j(a))$ is defined and varies
continuously with $j$ in the $C^\infty$ topology. Extend
this section along $i$ to a vector field $V_j$ on $N$
by normal projection, transport, and a fixed cutoff.
For $j$ sufficiently close, the formula
$F_{j,t}(x)=\exp_x(tV_j(x))$ defines a local
diffeomorphism for each $t$, by $C^1$ closeness to the
identity. Compactness makes it a covering map on
each connected component. Being homotopic to the
identity, it induces an isomorphism on the fundamental
group of each component, so the covering has one sheet.
Thus it is a diffeomorphism, and $j\mapsto F_{j,1}$
is a continuous section near $i$ with values in
$\mathcal D_0(N)$.

The same construction based at every embedding shows
that all $\mathcal D_0(N)$-orbits in the full embedding
space are open. An orbit is path connected, and each orbit's complement
is open because it is a union of other open orbits.
A path cannot meet both an orbit and its complement. Consequently the orbit of $i$ is precisely
$\mathcal E_i$. Translation of the local section by
elements of $\mathcal D_0(N)$ gives sections throughout
that orbit. If $s$ is a section over $U$, the map
$(j,k)\mapsto s(j)k$ identifies $U\times K_i$ with
$q^{-1}(U)$, since any two diffeomorphisms sending $i$
to $j$ differ by a unique element of $K_i$ on the
right. The inverse is the restriction together with
$F\mapsto s(q(F))^{-1}F$. Local triviality gives the
stated homotopy fiber sequence.
\end{proof}

For example, an ambient lift of an embedding loop,
started at $1_N$, ends at an element of $K_i$. Its
component is independent of the chosen lift. This
follows because two lifts $F_t,F'_t$ starting at the
identity satisfy $F_t^{-1}F'_t\in K_i$, giving a path
from the identity that relates their endpoints.
With conventional base-point choices this is the
connecting map from $\pi_1(\mathcal E_i,i)$ to
$\pi_0(K_i)$. It detects when a loop cannot be
lifted to a loop of ambient diffeomorphisms.

\begin{example}[Point pushing on a surface]
Take $A$ to be a point and $N$ the closed oriented
surface of genus two. Then $\mathcal E_i=N$ and
$q$ evaluates a diffeomorphism at the designated point
$x_0$. Every loop $F_t$ in $\mathcal D_0(N)$ based at
the identity evaluates to a loop representing a central
element of $\pi_1(N,x_0)$. Indeed, for any based loop
$\gamma_s$ the map $(t,s)\mapsto F_t(\gamma_s)$ defines
a map from the torus to $N$. The two generators of the
torus group map to the evaluated loop and $\gamma$,
so their classes commute.

There exists a loop $\alpha$ in $N$ whose class is
noncentral. In the genus-two polygon presentation with
generators $a,b,c,d$ and relation $[a,b][c,d]=1$, sending
$b,d$ to $1$ realizes a surjection onto the free group
on $a,c$. The image of $a$ does not commute with $c$,
so one may take $\alpha$ to represent $a$.
Its ambient lift, supplied by the restriction bundle,
cannot end in the identity component of $K_i$.
If it did, a path in the stabilizer to $1_N$ would close
it into a loop of ambient diffeomorphisms without changing
the evaluated class. This contradicts centrality.
The returned point is the same, while the endpoint
ambient transformation has acquired a nontrivial
component in the stabilizer.
\end{example}

The action groupoid of $\mathcal D_0(N)$ on
$\mathcal E_i$ has classifying space equivalent to
$BK_i$. Indeed local triviality identifies
$\mathcal E_i$ with $\mathcal D_0(N)/K_i$. Using any
contractible free model $E\mathcal D_0(N)$, its
homotopy quotient is the space in the middle of
\[
 \begin{tikzcd}[column sep=3em]
 E\mathcal D_0(N)\times_{\mathcal D_0(N)}\mathcal E_i
 \arrow[r,"\simeq"] &
 E\mathcal D_0(N)/K_i
 \arrow[r,"\simeq"] &
 BK_i .
 \end{tikzcd}
\]
The transitive action therefore gives a single orbit,
while the groupoid retains the stabilizer. A problem
about continuous choices of ambient extensions can be
formulated as a lifting problem for $q$.

\section{Changing a seam by isotopy}

Let $M,N$ have boundary components $S\subset\partial M$
and $T\subset\partial N$ with collars, and let
$\varphi:S\to T$ be a diffeomorphism. If $u_t$ is an
isotopy of diffeomorphisms of $S$ beginning at the
identity, put $\psi=\varphi\circ u_1$. The manifolds
obtained by gluing along $\psi$ and along $\varphi$
are diffeomorphic. To construct the map, choose a
collar $c:S\times[0,\epsilon)\to M$ and a smooth
$\beta$ equal to $1$ near zero and to $0$ near the
far end. On the collar use the diffeomorphism
\[
 c(x,s)\longmapsto c(u_{\beta(s)}(x),s).
\]
Extend it as the identity on the rest of $M$ and use
the identity on $N$. In the source gluing, $x$ and
$\psi(x)$ are identified. Their images are $u_1(x)$
and $\varphi(u_1(x))$, which agree in the target.
Because the transformation is a product near the
seam, the descended map is smooth in the signed collar
coordinate and has a smooth inverse. The same
construction works with disjoint seams by choosing
disjoint collars, so their changes commute on the
resulting manifold.

A basic isotopy of disks illustrates how local
differential data can fix a model. For an embedding
$e:D^r\to\mathbb R^n$ smooth up to the boundary, the
maps
\[
 x\longmapsto e(0)+\frac{e(tx)-e(0)}{t},
 \qquad 0<t\le1,
\]
are embeddings, and the integral formula for the
difference quotient gives a smooth limit
$x\mapsto e(0)+D_0e(x)$ at $t=0$. The limit is still
embedded because the differential at zero is injective.
This isotopy varies the image and need not preserve
boundary values. Relative isotopy classes with fixed
boundary can contain further information.
\par



\chapter{Surfaces and families of models}
\label{ch:surfaces}

Surfaces illustrate two different classification problems.
The first identifies an individual surface up to existence
of a diffeomorphism. The second classifies families of
these surfaces, including the diffeomorphisms used when
parameters return around a loop. Handles answer the first
problem, and the classifying space of the diffeomorphism
group specifies the second.

\section{Disk and band presentations}

Let $S$ be a closed connected smooth surface. Start with a
finite triangulation. Thicken vertices to disks, edges to
bands, and faces to the complementary disks. This is
the smooth $0$, $1$, and $2$-handle description after
rounding corners. It is also the type of decomposition
produced by Morse theory. It can be simplified to a
single starting disk and a single final disk as follows.

First merge along the bands of a spanning tree $T$ of
the $1$-skeleton. A pair of disks joined by a band between
their distinct boundary circles forms a disk after rounding.
As a result the union over the tree is one disk. Next take
the dual graph on faces, using only edges which do not
cross $T$. This graph is connected. Otherwise a nonempty
proper union of its components would give a mod-$2$
sum of faces whose boundary uses only edges of $T$.
That boundary is nonzero because the full dual graph is
connected. It would be a nonzero $1$-cycle in a tree.
Choose a dual spanning tree and merge the final face
disks across each crossed band in that tree. This leaves
one final disk and some number $r$ of bands. Consequently
$\chi(S)=2-r$. In particular, if $\chi(S)=2$ there are
no bands and $S$ is a sphere. Two disks glued along
their boundaries give the usual sphere because any
diffeomorphism of their boundary circles extends to a
disk. Indeed an orientation-preserving circle
diffeomorphism is isotopic to the identity by linear
interpolation of a monotone lift with the identity lift.
The collar construction of Chapter~\ref{ch:isotopy} extends
this isotopy. A reflection accounts for the other sign.

\section{Cutting along simple curves}

A mod-$2$ $1$-cycle on a surface can be represented by
a disjoint collection of embedded smooth circles.
For a simplicial representative there is an even number
of incident selected edges at each vertex. In a small
disk around the vertex, pair neighboring endpoints and
round their arcs. The modification has the same mod-$2$
relative class in this disk because the set of
endpoints is unchanged and the disk has zero first
homology. The modified edges give the required
collection.

Suppose first that $S$ is oriented and has $\chi(S)<2$.
Poincar\'e duality over $\mathbb F_2$ gives nonzero
$H_1(S;\mathbb F_2)$. In a representative collection for
a nonzero class at least one circle is itself nonzero
in homology, hence nonseparating. Otherwise it would
bound one side modulo $2$. An oriented surface gives
each embedded circle an orientable normal line bundle,
so such a circle has an annular neighborhood. Cutting
along it leaves a connected surface with two additional
boundary circles. Cap them with disks to obtain $S'$,
with $\chi(S')=\chi(S)+2$.

Recover $S$ from $S'$ by adding a torus summand. To see
the summand, choose an arc joining the capped disks in
$S'$. A neighborhood of the arc together with the two
disk caps is itself a disk. Removing the two caps from
this neighborhood gives a disk with two holes, and the
original annulus joins the two inner circles. This is
a punctured torus, using the usual square model of the
torus cut along one of its basic curves. Replacing the
disk by the punctured torus is the connected sum. Thus
iteration, terminating when $\chi=2$, gives a connected
sum of tori, with the empty sum denoting the sphere.

When $S$ is nonorientable, its orientation character
$w_1(TS):H_1(S;\mathbb F_2)\to\mathbb F_2$ is nonzero.
Apply the circle representation above to a class
detected by this character. Some component is one-sided,
since the tangent to a circle is orientable and the
normal line has nontrivial orientation. Its neighborhood
is a M\"obius band. The complement is connected: every
component must meet the new boundary, which is just
one circle, since $S$ is connected. Removing the band and
capping the new boundary gives $S'$ with
$\chi(S')=\chi(S)+1$. Recovering $S$ replaces the cap
by a M\"obius band, which is the punctured projective plane.
Iteration leaves an orientable surface together with
at least one projective-plane summand.

\begin{lemma}[Exchanging a handle and a crosscap]
\label{lem:crosscap-exchange}
The surface $T^2\#\mathbb{RP}^2$ is diffeomorphic to the
connected sum of three projective planes.
\end{lemma}

\begin{proof}
Use mod-$2$ homology and let $a,b$ be the two basic
torus classes and $c$ the core of the projective-plane
band. Their only nonzero pairings under intersection
are $a\cdot b=b\cdot a=1$ and $c\cdot c=1$.
Here the self-intersection of the one-sided core is
one, as can be computed by a perturbation in the
nontrivial normal line. The orientation character is
zero on $a,b$ and one on $c$.

Join disjoint representatives of $a$ and $c$ by an
embedded band. Replacing short subarcs by the opposite
sides of the band gives an embedded circle of class
$x=a+c$. It is one-sided. Let $C$ denote the complement
of the interior of its M\"obius neighborhood. The
Mayer--Vietoris sequence over $\mathbb F_2$ gives
\[
 H_1(S;\mathbb F_2)\simeq
 H_1(C;\mathbb F_2)\oplus\langle x\rangle.
\]
Indeed the sole gluing circle is zero in the homology
of both sides, where it bounds the mod-$2$ fundamental
class relative to that circle, and both sides are connected.
The image of $H_1(C;\mathbb F_2)$ is therefore the
orthogonal complement of $x$: it pairs trivially
with $x$, and $x\cdot x=1$.
The class $b+c$ belongs to this complement and has
nonzero orientation character. Hence $C$, and also its
capped version, is nonorientable.

The capped version has Euler characteristic zero.
Removing another one-sided band and capping leaves
Euler characteristic one, which cannot occur for a
closed orientable surface by the preceding torus
reduction. One more such removal leaves Euler
characteristic two and therefore a sphere. Reattaching
the three bands proves the claim by the description
of capping as connected sum.
\end{proof}

\begin{theorem}[Classification of compact connected surfaces]
Let $S$ have $b$ boundary components. Up to diffeomorphism,
$S$ is a surface obtained by deleting $b$ open disks from
one of the following closed models:
\[
 S_g=\#^g T^2\quad(g\ge 0),\qquad
 N_h=\#^h\mathbb{RP}^2\quad(h\ge 1).
\]
The type and its index are determined by orientability,
$b$, and $\chi(S)$. The formulas are
$\chi(S_g)-b=2-2g-b$ and $\chi(N_h)-b=2-h-b$.
\end{theorem}

\begin{proof}
For closed surfaces, the reductions and
Lemma~\ref{lem:crosscap-exchange} give the models.
Connected sum subtracts two from the sum of Euler
characteristics, since each puncture removes a disk
and the gluing circle has Euler characteristic zero.
This yields the displayed formulas, which distinguish
different indices within each orientability type.

For a surface with boundary, cap all the circles.
Orientability of the capped model is equivalent to
orientability of $S$, since an orientation on $S$
extends over each disk with the matching boundary
convention. The positions of the disjoint cap disks
carry no extra invariant for an unmarked surface.
Shrink each disk inside a collar neighborhood, move
the resulting disks along paths of centers in the
complement of the other disks, and use diffeomorphisms
in the coordinate neighborhoods of the disk images
to match the final disks. This is the disk isotopy
construction of \cite[Chs.~8--9]{hirsch}. The set of
boundary circles may be permuted because no boundary
labels are part of the classification.
\end{proof}

\section{Models in a family}

For any fixed smooth surface $S$, the groupoid with its
diffeomorphisms as arrows has classifying space
$B\operatorname{Diff}(S)$ when the arrows have the
$C^\infty$ topology. A locally trivial $S$-bundle with smooth structure on the
fibers has continuous transition functions valued in this
$C^\infty$ group. On a paracompact base where the cover is
numerable, such bundles
are classified up to isomorphism by homotopy classes of
maps to $B\operatorname{Diff}(S)$. The classifying map
of the principal frame bundle sends the fiber model to a
copy of $S$:
\[
 \begin{tikzcd}[column sep=3.4em,row sep=2.3em]
 E\arrow[r]\arrow[d] &
 E\operatorname{Diff}(S)\times_{\operatorname{Diff}(S)}S\arrow[d]\\
 B\arrow[r,"\text{classifying map}"'] &
 B\operatorname{Diff}(S).
 \end{tikzcd}
\]
The square describes the pullback of the universal bundle
in a model for which the classifying map is chosen.
A proper smooth submersion of boundaryless manifolds
with compact surface fibers provides such local
trivializations. Choose a horizontal complement to the
vertical tangent bundle and lift straight paths in a
small coordinate ball in the base. Properness gives
existence for the bounded time interval, and reversal
of the paths gives the inverse maps. This identifies
all nearby fibers smoothly with the chosen reference
fiber.

For example, form a torus bundle over a circle by gluing
the ends of a product interval with the torus
diffeomorphism induced by
$\bigl(\begin{smallmatrix}1&1\\0&1\end{smallmatrix}\bigr)$
on $\mathbb R^2/\mathbb Z^2$. The homology groups of the
fibers form a local system with this monodromy on
$H_1(T^2;\mathbb Z)$. A product bundle would have identity
monodromy. Conjugation cannot identify the displayed
matrix with the identity. Thus the genus specifies the
fiber of this family, while the classifying map records
its nontrivial gluing over the parameter circle.
\par



\bookpart{V}{Smooth stacks and parametrized families}

\chapter{Smooth spaces and quotient stacks}
\label{ch:smooth-stacks}

A smooth manifold can be studied through the objects that map
into it. A chart is a particularly useful test map, and
a parametrized family is another. By allowing all smooth test
manifolds, one retains families that pass through singular
orbits and through changes of local trivialization. The
resulting category also admits quotients with automorphisms.

\section{Smooth test manifolds}

Use the category $\mathrm{Man}$ of finite-dimensional smooth
manifolds with the open-cover topology. In this part the test
manifolds have no boundary, and the underlying site is made
small in a chosen universe. One can enlarge it to a site of
manifolds with corners when boundary charts are required.
A smooth space in this chapter is a sheaf of spaces on the
site. Write $\mathcal E$ for the resulting $\infty$-topos and
$\underline M$ for the sheaf $U\mapsto C^\infty(U,M)$
represented by a manifold $M$. Its values are sets regarded
as discrete spaces. This differs from sending $M$ to its
singular complex: the latter has paths and forgets the smooth
structure, while $\underline M$ records smooth families.

The presheaf represented by $M$ is a sheaf. Given maps from
an open cover of $U$ to $M$ that agree on pairwise intersections,
their glued map is continuous and locally smooth, hence smooth.
By Yoneda every morphism $\underline M\to\underline N$ comes
from a unique smooth map $M\to N$. Products are preserved:
a map $U\to M\times N$ is exactly a pair of smooth maps.
A fiber product in smooth spaces always exists. If the two
maps to a representable target are transverse, the ordinary
smooth fiber product represents it, because the ordinary
submanifold cut out by the transverse equations has the
universal lifting property of smooth maps. For singular
equations the sheaf fiber product still makes sense. Derived
smooth geometry carries additional information about those
equations, as developed in Chapter~\ref{ch:derived-smooth}.

Chapter~\ref{ch:manifolds} gives a smooth map that is a
homeomorphism of its underlying spaces without being a
diffeomorphism. The failure of invertibility is detected by
the identity test family on its target. Thus recovering a
space from all of its smooth probes also recovers the
differential structure of its maps.

\section{Locally represented geometry}

Write $\mathcal X(U)$ for the space of $U$-families in a smooth
space $\mathcal X$. Paths in this space compare families,
and higher simplices compare such paths. In a stack valued in
groupoids these paths are isomorphisms. Descent replaces
literal agreement on overlaps by compatible identifications.
For an open cover $U_i$ of $U$, the comparison is
\[
 \mathcal X(U)\simeq
 \lim_{[n]\in\boldsymbol\Delta}
    \prod_{i_0,\ldots,i_n}
      \mathcal X(U_{i_0}\cap\cdots\cap U_{i_n}).
\]
For a groupoid-valued example, local objects and
isomorphisms on double overlaps must satisfy a cocycle
condition on triple overlaps. Morphisms between global
objects are computed by maps of descent data. A map from a
manifold into a stack can therefore be locally represented
by maps to an atlas with transitions between the local lifts.

An atlas for a differentiable stack is a representable
surjective submersion from a manifold. Such a map admits
local sections after pullback to a test manifold. It is
therefore an effective epimorphism in $\mathcal E$. Its
\v{C}ech groupoid consists of the atlas, the space of arrows
between its two lifts, and the higher fiber products encoding
composition. The stack is the sheaf colimit of this groupoid
\cite[\S6.1]{lurie-htt}. The manifold structure of the
arrow space distinguishes a smooth groupoid from its
classifying homotopy type.

\section{An action groupoid as a quotient}

Let a Lie group $G$ act smoothly on $M$ from the left.
Its action groupoid has object space $M$ and arrow space
$G\times M$, with source $(g,x)\mapsto x$ and target
$(g,x)\mapsto gx$. The quotient $[M/G]$ is the colimit of the
simplicial nerve in smooth spaces. It classifies local
representatives of points modulo gauge transformations.

\begin{proposition}[Families in an action quotient]
\label{prop:smooth-quotient}
A map from a test manifold $U$ to $[M/G]$ can be described
by a smooth principal right $G$-bundle $\pi:P\to U$ and
a smooth map $a:P\to M$ satisfying $a(pg)=g^{-1}a(p)$.
Arrows between such data are bundle isomorphisms preserving
these maps.
\end{proposition}

\begin{proof}
Choose a local section $s_i$ of the torsor over each open
$U_i$ and write $s_j=s_i g_{ij}$ on overlaps. The transition
maps satisfy $g_{ij}g_{jk}=g_{ik}$. The points
$x_i=a(s_i)$ obey $x_j=g_{ij}^{-1}x_i$, so the action arrows
from $x_j$ to $x_i$ carry exactly the transition data of the
torsor. Conversely, descent data $(x_i,g_{ij})$ glue
$U_i\times G$ by $(u,h)_j\sim(u,g_{ij}(u)h)_i$.
The functions $(u,h)_i\mapsto h^{-1}x_i(u)$ agree with
this identification and give $a$. Changing local sections
changes the data by a groupoid isomorphism. Every bundle
isomorphism preserving the map to $M$ is locally described
by such changes. The constructions thus identify both
objects and arrows in the quotient stack.
\end{proof}

The map $[M/G]\to[*/G]$ forgets the equivariant map and
keeps its torsor. Pulling it back along the trivial torsor
with a chosen trivialization gives back $M$.
\[
\begin{tikzcd}[row sep=large,column sep=large]
 \underline M \arrow[r] \arrow[d] & {[M/G]}\arrow[d]\\
 *\arrow[r,"\mathrm{trivial\ torsor}"] & {[*/G]}.
\end{tikzcd}
\]
For a free proper action, $M\to M/G$ is a principal bundle.
Writing a map to the manifold quotient as the pullback of
this bundle with its tautological map to $M$ gives an
inverse to the map $[M/G]\to\underline{M/G}$. In general
this equivalence fails at any point with a nontrivial
stabilizer. The stabilizer is precisely the automorphism
group of the corresponding family over a point.

\section{The reflection quotient}

For $G=C_2$ acting on $\mathbb R$ by reflection, the
topological orbit space is the closed half-line, with
coordinate $t^2$. The origin of the quotient stack has
automorphism group $C_2$ while a nonzero orbit has none.
A smooth family supported at the origin on a connected
base is a principal $C_2$-bundle equipped with the zero
map. Over $S^1$ the trivial cover and the connected double
cover give distinct families, even though the same coarse
function $t^2=0$ represents both.

The connected cover with total space
$\mathbb R/(4\pi\mathbb Z)$ admits a weight
$w([\theta])=\cos(\theta/2)$ that is odd under the deck
transformation. Its square descends to the smooth function
$(1+\cos\theta)/2$. There is no smooth real function on
the base with that square. In a local coordinate centered
at the unique zero, the square vanishes quadratically
with positive quadratic coefficient. A smooth square
root must have nonzero derivative and thus change sign
there. The complement of the unique zero is connected,
so a real nonzero function there could not change sign.
The quotient stack permits the pair of local roots
together with a sign transition. Smooth differential data
see an obstruction absent from the coarse observable.

The next chapter keeps such transition data while forming
spaces of maps and composing families. Chapter~\ref{ch:orientations}
uses a related descent construction for orientations of bundles.
\par



\chapter{Mapping objects and families over a base}
\label{ch:relative-maps}

The space of maps between manifolds has several useful
topologies, as in Chapter~\ref{ch:function-spaces}. Smooth
families can also be specified without choosing a
topology on the set of maps. This uses all smooth test
objects and an evaluation morphism. The same construction
makes sense for families over a base that is itself a stack.

\section{Smooth mapping objects}

For smooth spaces $X,Y$ in the $\infty$-topos $\mathcal E$,
write $[X,Y]$ for their exponential. Thus
\[
 \operatorname{Map}_{\mathcal E}(T,[X,Y])
 \simeq \operatorname{Map}_{\mathcal E}(T\times X,Y)
\]
naturally in $T$. When $X=\underline M$ and $Y=\underline N$,
Yoneda computes its value on a test manifold $U$:
\[
 [\underline M,\underline N](U)=C^\infty(U\times M,N).
\]
This specifies the regularity of derivatives in both the test
and input variables. It has no requirement that the exponential
be a finite-dimensional manifold. Even the smooth maps from a
circle to a line have infinite-dimensional space of variations.
The sheaf model also distinguishes changes of parameter that
are smooth from continuous paths for a chosen function-space
topology. The comparison with convenient calculus can be
made in settings in which both sides' smooth test plots agree
\cite{baez-hoffnung,kriegl-michor}.

As a basic example, a family of real functions on $M$ is a
map $P\to[\underline M,\underline{\mathbb R}]$. Taking its
$f$ at a point $p$ gives a function $f_p$. Restricting this
family along $a:P'\to P$ gives $f_{a(p')}$. Differentiating
with respect to the input gives the vertical one-form
$d_M f$ on $P\times M$, a section of the pullback of $T^*M$.
The restriction of this section along $a$ is $d_M$ of the
restricted function, since the maps in the input directions
are unchanged by varying $a$. Differentiation with respect to
the parameter instead pulls its covectors back along $Da$.
The two directions have different transformation laws even
for a trivial product.

\section{Evaluation over a base}

Let $B$ be any object of $\mathcal E$. Products and
exponentials in its slice are written $\times_B$ and
$[X,Y]_B$. This is the category of families varying over $B$.
The defining evaluation and its adjunction are
\[
 \begin{gathered}
 \operatorname{ev}_B:[X,Y]_B\times_B X\longrightarrow Y,\\
 \operatorname{Map}_{B}(P\times_B X,Y)
   \simeq\operatorname{Map}_{B}(P,[X,Y]_B).
 \end{gathered}
\]
So a map $f:P\times_B X\to Y$ gives its classifying
morphism $r_f:P\to[X,Y]_B$. A reparametrization includes
a map between parameter objects over $B$ commuting with
the classifying maps up to a specified homotopy.
These reparametrizations and their higher comparisons
form the slice $\mathcal E_{/[X,Y]_B}$. The identity
family has parameter $B$. A family from $X$ to $Y$
with parameter $P$, followed by one from $Y$ to $Z$
with parameter $Q$, has parameter $P\times_B Q$.
Its evaluation is the same as applying the two
evaluations in order. The associativity comparison
comes from the product comparison on three parameters.

\begin{proposition}[Base change for relative evaluation]
\label{prop:mapping-base-change}
For $a:C\to B$, pullback sends $[X,Y]_B$ to
$[a^*X,a^*Y]_C$ by an equivalence preserving evaluations.
It preserves composition and reparametrization of families.
\end{proposition}

\begin{proof}
The functor $a_!$ left adjoint to pullback sends an object
over $C$ to its structure map composed with $a$.
For $T\to C$,
\[
 a_!(T\times_C a^*X)\simeq (a_!T)\times_B X
\]
because each side is the same pullback equipped with
the map to $B$. Therefore
\[
\begin{aligned}
 \operatorname{Map}_C(T,a^*[X,Y]_B)
 &\simeq \operatorname{Map}_B(a_!T,[X,Y]_B)\\
 &\simeq \operatorname{Map}_B((a_!T)\times_B X,Y)\\
 &\simeq \operatorname{Map}_C(T\times_C a^*X,a^*Y).
\end{aligned}
\]
These comparisons are natural in $T$. Yoneda gives the
equivalence, and the construction evaluates the same
map under the adjunctions. The product that defines
composition is preserved by pullback. Naturality
with respect to $P$ gives the last assertion.
\end{proof}

\begin{proposition}[Descent of families]
\label{prop:descent-families}
Let $U\to B$ be an effective epimorphism with \v{C}ech nerve
$U_\bullet$. Set $X_n=X\times_B U_n$ and
$Y_n=Y\times_B U_n$. Then the category of parametrized
maps from $X$ to $Y$ is equivalent to
\[
 \mathcal E_{/[X,Y]_B}\simeq
 \lim_{[n]\in\boldsymbol\Delta}
 \mathcal E_{/[X_n,Y_n]_{U_n}}.
\]
\end{proposition}

\begin{proof}
Descent for slices of an $\infty$-topos identifies
$\mathcal E_{/B}$ with the limit of $\mathcal E_{/U_n}$
\cite[\S6.1.3]{lurie-htt}. Slicing at a compatible
target commutes with taking that limit. Indeed an
arrow into the target is a compatible family of local
arrows into its components, with the same statement
for all mapping spaces. Under the first equivalence
the image of the target exponential is the compatible
family of local exponentials by Proposition~\ref{prop:mapping-base-change}.
\end{proof}

In ordinary charts with transition equivalences chosen,
this result glues local evaluations when the
comparisons commute. If an input and output are
bundles with changes of frames $a_{ij}$ and $b_{ij}$,
the local maps transform as $f_i=b_{ij}f_j a_{ij}^{-1}$
when the parameter coordinates have already been
transported. If the same parameter supplies two stages
of a composite, the pullback along the diagonal
$P\to P\times_B P$ implements that sharing.
In sheaves with higher morphisms, local
comparisons themselves need coherence on
multiple overlaps. The limit in the proposition
records exactly those data.

\section{Transport over a homotopy type}

In the $\infty$-category $\mathcal S$ of spaces, straightening
identifies $\mathcal S_{/B}$ with $\operatorname{Fun}(B,\mathcal S)$,
where $B$ is an $\infty$-groupoid \cite[\S3.2]{lurie-htt}.
A path $\gamma:b\to c$ induces equivalences
$t_X:X_b\simeq X_c$ and $t_Y:Y_b\simeq Y_c$. A
map on the fiber at $b$ is transported by
$h\mapsto t_Y h t_X^{-1}$. The higher comparisons
required for this formula come from the path
composition of the original functors. The evaluation
square
\[
\begin{tikzcd}[column sep=large,row sep=large]
 {[X_b,Y_b]\times X_b} \arrow[r,"\mathrm{ev}"]
   \arrow[d,"\mathrm{transport}"'] & Y_b \arrow[d,"t_Y"]\\
 {[X_c,Y_c]\times X_c} \arrow[r,"\mathrm{ev}"] & Y_c
\end{tikzcd}
\]
commutes with this transport. To check the formula for
the exponential, first evaluate it at each $b$ and
choose these conjugation maps along paths. A coherent
map from $P$ to this family is a collection of maps
$P_b\times X_b\to Y_b$ commuting with all transports.
Thus the family satisfies the exponential adjunction.

For an ordinary discrete group action this specializes
to $(g\cdot h)(x)=g\,h(g^{-1}x)$, and a family
$f(p,x)$ obeys $f(gp,gx)=g f(p,x)$. A particular fiber
of the family can have less symmetry than the total
family. It is equivariant under the stabilizer of $p$.
For smooth actions one may work over the smooth
classifying stack of Chapter~\ref{ch:smooth-stacks} to retain
the dependence on smooth generalized points. As an
example, $SO(2)$ acts on $\mathbb R^2$ and its line
directions. Orthogonal projection onto a line transforms
by conjugation under rotation, so the whole family is
equivariant with the moving line as parameter. Projection
onto one fixed line commutes only with rotations
preserving that line.
\par



\chapter{Orientation and characteristic classes in families}
\label{ch:orientations}

An oriented manifold has locally chosen positive frames
whose transitions preserve orientation. For a family of
manifolds there can instead be a local orientation on every
fiber and a change of sign around a parameter loop. The
orientation is then a double cover of the base or of the total
space, as appropriate. This cover is itself descent data.

\section{Determinants and orientation covers}

For a real vector bundle $E$ of rank $r$ over $M$, its
determinant $\det E$ is $\bigwedge^r E$. The orientation cover
$\operatorname{Or}(E)\to M$ has fiber the two components of
nonzero frames of $\det E$. It is a principal $C_2$-bundle.
Local frames of $E$ give transition matrices
$g_{ij}:U_i\cap U_j\to GL(r,\mathbb R)$, with
$g_{ij}g_{jk}=g_{ik}$. Their images under
$\operatorname{sgn}\circ\det$ are the transitions of the cover.
An orientation of $E$ is a section of that cover. This is
the same as a choice of frame component on each connected
chart such that all transition signs are positive.

The determinant signs form a \v{C}ech $1$-cocycle with
values in $C_2\simeq\mathbb Z/2$. Replacing the local
trivializations changes this cocycle by a coboundary.
On a good cover of a manifold it represents the first
Stiefel--Whitney class $w_1(E)$. In smooth spaces
$\operatorname{Or}(E)$ is classified by a map
$\underline M\to[*/C_2]$ as in
Proposition~\ref{prop:smooth-quotient}. For $r\geq1$, lifting a family
of frames from $GL(r,\mathbb R)$ to the subgroup with
positive determinant is equivalent to specifying a smooth
trivialization of the sign torsor. The two descriptions
give the same obstruction, and both commute with pullback.

For real line bundles the sign class is also complete.
Choose a fiberwise positive inner product and take unit
local frames. All transition functions now have values in
$\{1,-1\}$. A sign cocycle determines a line by gluing,
and two cocycles that differ by a coboundary give
isomorphic bundles by changing unit frames. Thus
orientation and triviality coincide for a real line.
This reflects the contractibility of the subgroup of
positive real scalars. On $S^1$, a line with sign
monodromy $-1$ has no nowhere-zero section: a section
on the interval parametrizing the circle has endpoint
values of opposite sign unless both vanish. The
intermediate value theorem then forces a zero.
For the trivial bundle the constant unit section avoids
this obstruction.

\section{An exact sequence of relative tangent spaces}

An exact sequence of vector bundles
\[
 0\longrightarrow K\longrightarrow E\longrightarrow F
 \longrightarrow 0
\]
gives a natural isomorphism $\det E\simeq\det K\otimes\det F$.
To construct it, start with a basis of $K$ and append lifts
of a basis of $F$. A change of lifts adds linear combinations
of the earlier basis vectors and leaves the top exterior
product unchanged. Changes of the two bases scale the wedge
by the corresponding determinants. This constructs the
isomorphism in local frames, where it is visibly natural
under pullback. The orientation cover of $E$ is accordingly
the tensor product of the two covers, and
$w_1(E)=w_1(K)+w_1(F)$.

Let $f:M\to N$ and $g:Q\to N$ be transverse smooth maps.
At a point of their fiber product $Z$ the derivative
$d f-d g$ is surjective. The preimage normal form of
Chapter~\ref{ch:sard} identifies $T Z$ with its kernel.
If $p:Z\to M$, $q:Z\to Q$, and $h=f p=g q$, this yields
\[
 \det T Z \simeq
 p^*\det T M\otimes q^*\det T Q
       \otimes(h^*\det T N)^\vee .
\]
The convention uses the ordered sum $T M\oplus T Q$
and the displayed derivative. Choosing oriented frames of
the three ambient bundles orients the intersection via
this convention. When maps are replaced by parameter-dependent
families transverse in the vertical directions, the same
construction uses the vertical bundles. In a model for
derived smooth intersections the two-term tangent complex
continues to record this determinant when the derivative
fails to be surjective.
\[
\begin{tikzcd}[column sep=huge,row sep=large]
 Z\arrow[r,"q"]\arrow[d,"p"']& Q\arrow[d,"g"]\\
 M\arrow[r,"f"']& N
\end{tikzcd}
\]
This determinant relation is an orientation formula for
the pullback, rather than a choice of a splitting of its
normal sequence.

\section{An oriented count over a parameter space}

Suppose $\pi_M:M\to B$ and $\pi_N:N\to B$ are smooth
submersions of boundaryless manifolds of equal relative dimension
$d$, with $f:M\to N$
a smooth map over $B$. Let $s:B\to N$ be a smooth section,
and suppose that for each $b$, $s(b)$ is a regular value of
$f_b$. Put $Z=M\times_N B$. If $Z\to B$ is proper, it is
a finite covering over each connected component of $B$,
allowing the empty covering. In local submersion coordinates
the defining equations of $Z$ have an invertible vertical
Jacobian and can be solved as a smooth graph over $B$.
Thus $Z\to B$ is a local diffeomorphism. Properness makes
each fiber compact and discrete, hence finite, and prevents
additional inverse images from entering a sufficiently
small neighborhood. These local graphs form the sheets.

The derivative along each sheet is an isomorphism between the
two vertical tangent bundles. If both are oriented globally,
this determines a sign that is locally constant on the
sheet. Monodromy permutes the sheets and preserves their
signs. Consequently the sum of signs on a fiber is constant
on each connected component of $B$. For proper fiber maps of
oriented manifolds this agrees with the regular-value formula
for degree in Chapter~\ref{ch:degree}, when the conventions
for degree there apply. The unsigned count modulo two remains
defined if orientations are absent. More generally one can
keep the orientation line of each sheet. Locally changing
a frame reverses its corresponding sign, and the local
counts use the transitions of those lines. This description
explains which orientation data must be fixed before adding
signs as ordinary integers.

For example, let $\xi$ be the M\"obius real line on $S^1$ and
consider the fiber map from $\xi$ to itself given by multiplication
by any fixed negative scalar. Each fiber has degree $-1$ when
the same frame is used for both source and target. Neither
of the vertical lines is orientable, but the tensor product
of their orientation lines is trivial. The relative orientation
is supplied by the identification of the two lines.
By contrast, identifying $\xi$ with a trivial line by a
fiberwise linear isomorphism would contradict its nonzero
sign class.

\section{Euler classes and pullback}

For an oriented rank-$r$ bundle $\xi$, the Thom class
$u_\xi\in H^r(E,E\setminus M;\mathbb Z)$ restricts to the
chosen positive generator on each fiber pair.
It is natural for pullback bundles
\cite{milnor-stasheff,hatcher-vbkt}.
The Euler class is $e(\xi)=0_\xi^*u_\xi$, where the
zero section maps $M$ to the total space. A nowhere-zero
section has zero pullback of the relative class. Homotopy
through sections to the zero section then gives
$e(\xi)=0$. When a smooth section is transverse to zero,
the local intersection computation identifies the Euler
class with the cohomology class dual to its zero locus
under the usual orientation and compactness hypotheses
of Chapter~\ref{ch:intersection}. At a regular isolated
zero this follows from identifying the local pullback
of the Thom class with the local degree of the section
in a bundle chart.

One can package these classes using a bundle classifying
space and cohomology classes there. The smooth classifying
stack retains smooth gauge data, whereas a topological
classifying space classifies the bundle's underlying
topological class in the settings of
Chapter~\ref{ch:bundles}. The resulting characteristic
classes commute with pullback and therefore apply to
families described by either local frames or their
topological classification. The Euler class measures
one aspect of the bundle and can coincide for distinct
smooth families. Classifying maps retain the gluing data
from which these invariants are extracted.
\par



\chapter{Discriminants and the topology of families}
\label{ch:discriminants}

A smooth family can move through maps with singular fibers.
The discriminant marks parameter values where the implicit
function theorem fails in the input direction. Outside it,
under a properness hypothesis, fibers form an ordinary
bundle even though a drawing of two neighboring fibers may
make their equivalence hard to see.

\section{A regular total space with singular fibers}

Let $P,X,Y$ be smooth finite-dimensional manifolds without
boundary, and let $F:P\times X\to Y$ be smooth. Fix $y\in Y$
and suppose that $F$ is transverse to the point $y$. Thus
$Z=F^{-1}(y)$ is a manifold. Write $\pi:Z\to P$ for the
projection. For $(p,x)\in Z$ the tangent equation is
\[
 D_pF\,\dot p+D_xF\,\dot x=0 .
\]
The derivative of $\pi$ sends $(\dot p,\dot x)$ to $\dot p$.
It is surjective if and only if $D_xF$ is onto. In one
direction this follows by solving the tangent equation for
every $\dot p$. In the other direction, surjectivity of
$D\pi$ implies that the image of $D_pF$ is contained in
the image of $D_xF$. The sum of the two images is all of
$T_yY$ by transversality of $F$.

Let $C\subset Z$ consist of the points where $D\pi$ is
not onto. This is closed by the rank criterion. Define the
discriminant of this family at $y$ to be $D=\pi(C)$.
The corresponding map on the complementary locus is
\[
\begin{tikzcd}[column sep=large,row sep=large]
 \pi^{-1}(P\setminus D)\arrow[r,hook]\arrow[d,"\pi"']&
 Z\arrow[d,"\pi"]\\
 P\setminus D\arrow[r,hook]&P .
\end{tikzcd}
\]
The square is a pullback, and on the left the projection
is a submersion. By Sard's theorem the discriminant has
measure zero in parameter charts. This conclusion
requires the regular total space assumption; it makes
no assertion about a family with an identically singular
evaluation.

\begin{proposition}[Bundles away from the discriminant]
\label{prop:complement-bundles}
If $\pi:Z\to P$ is proper, then $D$ is closed and over each
connected component of $P\setminus D$ in the image, the
restricted projection is a smooth fiber bundle. On any
component outside the image the fibers are all empty.
\end{proposition}

\begin{proof}
A proper map between manifolds is closed, so it sends
$C$ to a closed set. Restriction to the open complement
remains proper by base change. On that open set the
image is both open (submersions have local sections
near image points) and closed (properness). It is therefore
a union of connected components.

Here is the local triviality argument for a proper
submersion. Choose a Riemannian metric on its domain
and define horizontal vectors to be orthogonal to $\ker D\pi$.
The derivative identifies this horizontal bundle with the
pullback of the tangent bundle of the base. Take a coordinate
ball centered at $p$ with compact closure contained in the
base of the submersion. Lift each radial path in the ball
horizontally starting at an arbitrary point of the fiber
over $p$. The lifted path stays in the inverse image of the
closed ball, which is compact. The smooth ordinary
differential equation therefore has a solution over the
entire radial segment. Its dependence on endpoints and
initial points is smooth. The reverse radial path gives
the inverse identification with the fiber over $p$. This
gives a smooth trivialization over the ball.
\end{proof}

A different metric may give different transports, and a
loop can give nontrivial bundle monodromy. The change of
trivialization is part of the family, as in
Chapter~\ref{ch:relative-maps}. Properness holds automatically
for the projection of this closed preimage when $X$ is
compact. With noncompact $X$, roots can escape to infinity
without ever becoming singular. For example $F(t,x)=tx-1$
has one root when $t\ne0$ and none at $0$, with $D_xF=t$
nonzero at every root. The entire zero locus is regular,
but its projection is not proper, since the roots for
$t\to0$ are unbounded.

\section{Fold and cusp calculations}

For the family of maps $F(t,x)=x^2-t$, the total zero set
is a parabola, parametrized by $x$. The projection to $t$
has one critical point at $x=0$ and is proper. For $t<0$
the fiber is empty, and for $t>0$ its two points are
$x=\pm\sqrt{t}$. The derivatives $2x$ give opposite local
degree signs and their sum is zero. At $t=0$ the vertical
derivative vanishes but the parameter derivative is $-1$,
so the total space remains smooth. This is the local fold
mechanism by which a pair of solutions is created or lost.
For a given one-dimensional family, the fold form follows
locally if the parameter derivative and the second input
derivative are nonzero at a critical root. Indeed one can
solve the equation for $t$ as a function $\phi(x)$ by the
implicit function theorem. Then $\phi'(0)=0$ and
$\phi''(0)=-F_{xx}/F_t\ne0$. The one-dimensional Morse
lemma gives the asserted normal form up to local coordinates
and a possible reversal of parameter (Chapter~\ref{ch:morse}).

Now take parameters $(a,b)$ and equation
\[
 F(a,b,x)=x^3+ax+b=0 .
\]
The total locus is smooth since $D_bF=1$. It projects
properly to the plane of coefficients: a root with
$|x|>1$ satisfies $|x|^3\le |a||x|+|b|$, which bounds its
size when $(a,b)$ lies in a compact set. A critical root
satisfies $3x^2+a=0$. In terms of that root $s$ the
discriminant is parametrized by
\[
 a=-3s^2,\qquad b=2s^3,\qquad 4a^3+27b^2=0.
\]
For $s\ne0$ the critical point above $(a,b)$ is a fold of the
projection. The
root $s=0$ is a triple root at the cusp. If $a\ge0$ a
polynomial off the discriminant has one real root.
If $a<0$, evaluate it at the local maximum and minimum
$x=\mp\sqrt{-a/3}$. Their values are
$b+2(-a/3)^{3/2}$ and $b-2(-a/3)^{3/2}$.
There are three simple real roots when these have
opposite signs, exactly when $4a^3+27b^2<0$, and otherwise
one off the discriminant. In the three-root region the
ordered derivative signs are $+,-,+$. They have the same
sum as the sign of the unique root elsewhere. Topology
of the fiber distinguishes these regions although degree
does not.

\section{Transverse jet conditions in a family}

The vertical $r$-jet of a family is the map
\[
 (p,x)\longmapsto j_x^r F_p
\]
from $P\times X$ to $J^r(X,Y)$. A submanifold of the jet
bundle can impose a rank condition, a root condition, or a
constraint on higher derivatives. If this total jet map is
transverse to a smooth submanifold $S$, then its inverse
image is smooth and the parametric transversality theorem
of Chapter~\ref{ch:transversality} applies. The derivative
test for the projection shows that a regular parameter
has $j^r F_p$ transverse to $S$. If $\operatorname{codim} S>
\dim X$, such a parameter has empty pullback of $S$.
For a finite or countable collection of smooth strata,
one can discard a countable union of measure-zero
exceptional sets, provided total transversality holds to
each. This test addresses the actual set of allowed
parameters. It can fail for a constrained family even
if a different unconstrained family contains all the
needed perturbation directions.

Classical bundles describe the complementary locus
pointwise and along paths. At a discriminant one can
retain the original equation by a derived pullback,
as in Chapter~\ref{ch:derived-intersection}. When the
parameter space carries an appropriate stratification,
one can also study transport adapted to the strata
(Chapter~\ref{ch:exit}). In the cubic example the cusp
stratum, fold stratum, and open regions record which
regular configurations approach a given critical
configuration. They preserve information absent from a
single fiber count.
\par



\bookpart{VI}{Derived geometry and stratified topology}

\chapter{Derived smooth spaces}
\label{ch:derived-smooth}

A smooth inverse image of a regular value has a manifold structure
because its equations can be placed among the coordinates. The same
equations at a critical value can define a singular set, a smooth set
with excess dimension, or a point at which infinitesimal motions remain.
Derived smooth geometry keeps the equations and their relations in
the structure sheaf. Its pullback represents the compatibility of
two maps even when their differentials fail to be transverse.

\section{The local model of a zero set}

Use derived manifolds in smooth geometry, with ordinary manifolds
included by their smooth function rings. We work with finite
local models of the form
\[
 Z(F)=U\times^{h}_{\mathbb R^m}\{0\},
 \qquad F:U\longrightarrow\mathbb R^m,
\]
where $U$ is open in $\mathbb R^n$ and the superscript $h$ denotes
a homotopy pullback in the category of derived manifolds.
The underlying topological space is $F^{-1}(0)$. Its functions
form a sheaf of derived $C^\infty$-rings. Spivak constructs
derived manifolds by gluing such models \cite{spivak-derived}.
Carchedi and Steffens characterize a finite-limit, idempotent-complete
extension of the category of manifolds in which transverse pullbacks
are preserved \cite{carchedi-steffens}. We use the finite local
zero-locus case. More general iterated limits may have cotangent
complexes of wider amplitude.

The ordered list of equations gives a concrete model. Write
$\mathcal A$ for the sheaf of smooth germs from $U$ restricted to the
topological zero set, and let $F_1,\ldots,F_m$ be the components
of $F$. Adjoin generators $\eta_i$ of cohomological degree $-1$:
\[
 \mathcal K_F
   =\mathcal A\otimes_{\mathbb R}
     \bigwedge\nolimits_{\mathbb R}(\eta_1,\ldots,\eta_m),
 \qquad d\eta_i=F_i .
\]
The differential extends by the signed Leibniz rule. The degree-zero
smooth structure is that of $\mathcal A$. This is the local Koszul
model inside the differential graded realization of derived
smooth geometry \cite{carchedi-dg}. One can see the origin of the
formula by resolving the point in the coordinate target. In the
smooth germs at the origin of $\mathbb R^m$, the coordinate functions
$t_1,\ldots,t_m$ form a regular sequence. Division by a coordinate for
functions vanishing on its zero hyperplane follows by integrating the
partial derivative along that coordinate. After quotienting by the
first $i-1$ coordinates, the same argument applies on their common
zero plane. Multiplication by $t_i$ there is injective, since a smooth
function annihilated by $t_i$ vanishes on the dense locus where
$t_i\ne0$. The Koszul resolution for $(t_i)$ thus resolves the
origin. In the differential graded model the resolution is formed by
adjoining the degree $-1$ generators with differentials $t_i$.
Pullback transports these differentials to the corresponding
components of $F$. In particular,
\[
 H^0(\mathcal K_F)=\mathcal A/(F_1,\ldots,F_m).
\]
Negative cohomology describes relations of the equations modulo
the boundaries of the Koszul complex.

If $F$ is submersive near a zero, the submersion chart writes it
as the projection onto the first $m$ coordinates. The regular
sequence argument then gives $H^i(\mathcal K_F)=0$ for $i<0$ and
identifies $H^0$ with smooth functions on the usual level
submanifold. This proves the agreement with the classical
construction in a transverse coordinate chart. Another useful check
is stabilization. Replacing $F(u)$ by $(F(u),s)$ on
$U\times\mathbb R^r$ changes the local model by the independent
regular-coordinate equations $s_1=\cdots=s_r=0$. Successively
resolving those equations reduces the model to $\mathcal K_F$.
Equivalently, the two derived spaces agree by the composition
and product laws for pullbacks.

\section{The tangent complex}

For a point $x$ of $Z(F)$ write $T_xZ(F)$ for the tangent complex,
dual to the cotangent complex evaluated at $x$. Its local model uses
the cohomological degrees $0,1$.

\begin{proposition}[Tangent complex of equations]
\label{prop:derived-smooth-tangent}
At a zero $x$ of $F$ there is a natural identification
\[
 T_xZ(F)\simeq [\,T_xU\xrightarrow{D_xF}\mathbb R^m\,],
\]
where $T_xU$ has degree zero. In particular the cohomology spaces
are $\ker D_xF$ in degree zero and $\operatorname{coker}D_xF$
in degree one.
\end{proposition}

\begin{proof}
In the differential graded chart, the cotangent module is generated
over the function algebra by the differentials of the smooth
coordinates in degree zero and the differentials of the
$\eta_i$ in degree $-1$. The smooth chain rule gives
$d(\mathrm{d}\eta_i)=\mathrm{d}F_i$. This is the derived cotangent
model because the coordinates and the Koszul generators form the
smooth free chart with successively adjoined generators, with
an open localization on $U$. Evaluating at $x$ gives the
two-term cotangent complex
\[
 [\,(\mathbb R^m)^*\xrightarrow{(D_xF)^*}T_x^*U\,]
\]
in degrees $-1,0$. Dualizing this finite-dimensional complex
gives the asserted two terms. Equivalently, the cotangent
pushout formula at the point computes the same complex from the
pullback of the two maps to $\mathbb R^m$
\cite[\S 7]{spivak-derived}.
\end{proof}

The degree-zero cohomology is the space of compatible tangent
vectors. Given a first-order right-hand side $b$ for
$F(u)=b$, the degree-one cohomology records its class modulo the
changes attainable by moving $u$. The index
\[
 \dim H^0(T_xZ(F))-\dim H^1(T_xZ(F))=n-m
\]
is the \emph{virtual dimension} at $x$. The equality follows by
subtracting the same rank from the two terms. Stabilization adds
the acyclic complex $[\mathbb R^r\xrightarrow{1}\mathbb R^r]$,
which leaves the index unchanged.

\begin{example}[Tangent information and higher functions]
\label{ex:derived-smooth-square}
Let $R$ be the ring of real smooth germs at $0$. For
$F(t)=t^2$, the Koszul differential is multiplication by $t^2$.
It is injective, since a smooth germ vanishing off the origin
vanishes on a neighborhood. Taylor's formula with smooth
remainder gives $R/(t^2)\cong\mathbb R[\epsilon]/(\epsilon^2)$:
every germ has a unique constant and linear term modulo $t^2$.
Thus this example has a nonreduced ordinary function ring.
Replacing $F(t)$ by $t^3$ gives
$\mathbb R[\epsilon]/(\epsilon^3)$ with the same tangent complex
$[\mathbb R\xrightarrow{0}\mathbb R]$. If $F$ is identically
zero on $\mathbb R$, the same complex appears at every point,
and the degree $-1$ Koszul generator now survives in the
function algebra. These descriptions distinguish the three
models through functions, even though their pointwise tangent
complexes agree at the selected zeros.
\end{example}

\begin{example}[Redundant constraints]
The map $t\mapsto(t,0)$ to $\mathbb R^2$ has a single zero.
Its Koszul algebra has one regular constraint and one
generator of zero differential, so the germ is equivalent
to $\bigwedge_{\mathbb R}(\eta)$ with $|\eta|=-1$.
The tangent differential $v\mapsto(v,0)$ is injective
with one-dimensional cokernel. The virtual dimension is
$-1$, accounting for the redundant target direction.
The pair of constant maps $\{0\}\rightrightarrows\mathbb R$
also has a derived pullback whose function algebra is this
exterior algebra. Its homotopy pullback in spaces is
contractible, since the based loop space of the underlying
real line is contractible. Here the extra information
comes from the smooth structure on the target and is
detected by the derived function algebra.
\end{example}

\section{Compatibility as a mapping property}

If $X\to W\leftarrow Y$ are ordinary smooth maps, the derived
pullback has the universal property
\[
 \operatorname{Map}(Q,X\times^h_WY)\simeq
 \operatorname{Map}(Q,X)\times^h_{\operatorname{Map}(Q,W)}
                       \operatorname{Map}(Q,Y)
\]
for derived test objects $Q$. A point of the right-hand space consists
of the two maps and a comparison between their composites to $W$.
The comparison is part of the universal property at all higher
simplicial degrees.
\[
\begin{tikzcd}[row sep=large,column sep=large]
 X\times^h_WY\arrow[r]\arrow[d]&Y\arrow[d,"g"]\\
 X\arrow[r,"f"']&W
\end{tikzcd}
\]
Around a pair $(x,y)$ with $f(x)=g(y)$, choose a single chart on
$W$ and express compatibility by
$F(u,v)=\phi(f(u))-\phi(g(v))$. The diagonal of the coordinate
space is its ordinary transverse zero set for the difference map.
Pullback associativity therefore identifies this local derived
model with $Z(F)$. The tangent complex in
Proposition~\ref{prop:derived-smooth-tangent} becomes
\[
 [\,T_xX\oplus T_yY\xrightarrow{Df-Dg}T_{f(x)}W\,].
\]
Changing target coordinates identifies the last term by the
derivative of the chart transition. More intrinsically, the
complex is the homotopy fiber of the difference of the tangent
maps, a description that also makes sense after the local
equations have been glued.

A further base change can be performed one arrow at a time
or in a single pullback. Both comparisons represent the same
space of maps from every $Q$, by the repeated universal
property and associativity of limits. This observation is useful
for families of constraints: imposing incidence on a family
and then restricting its parameter gives the same derived
incidence space as restricting the family first. Surjectivity
of a derivative is instead a chartwise test that identifies
a particular derived intersection with an ordinary one.
The next chapter uses this difference to calculate excess
intersection classes.
\par



\chapter{Derived intersections and excess}
\label{ch:derived-intersection}

Let $A$ and $B$ be embedded submanifolds of $W$, with the
manifolds taken without boundary. Their derived
intersection represents the pullback of their inclusions. Its
underlying points are precisely the points of $A\cap B$.
The linearization at $c$ is the two-term complex
\[
 [\,T_cA\oplus T_cB\xrightarrow{(u,v)\mapsto u-v}T_cW\,],
\]
using the result of Chapter~\ref{ch:derived-smooth}. If $A$ is cut
out locally by a submersion $s$, the same derived intersection is
the zero locus of $s|_B$. The two presentations are equivalent
by pasting pullback squares. In particular the normal equation
model at $c$ has source $T_cB$, target $N_{A/W,c}$, and differential
induced by the inclusion into $T_cW$.

\section{Clean intersections}

\begin{definition}
An intersection of embedded submanifolds is \emph{clean} if
$C=A\cap B$ is an embedded submanifold and
$T_cC=T_cA\cap T_cB$ for all $c\in C$.
\end{definition}

This condition allows the tangent sum to be smaller than
$T_cW$. Since its dimension is locally constant on $C$, the quotient
\[
 E=\left.TW\right|_C/
       \big(\left.TA\right|_C+\left.TB\right|_C\big)
\]
is a smooth vector bundle. It is the \emph{excess bundle}.
The degree-zero tangent cohomology is $TC$ by identifying a
kernel pair with its common vector, and degree-one cohomology
is $E$. The natural map on normal bundles gives the exact
sequence
\begin{equation}
 \label{eq:excess-bundle-sequence}
 0\longrightarrow N_{C/B}
   \longrightarrow \left.N_{A/W}\right|_C
   \longrightarrow E\longrightarrow 0 .
\end{equation}
Indeed, $TB|_C\to TW|_C/TA|_C$ has kernel $TC$, and its
cokernel is the displayed quotient. Thus an excess direction
can be read either from the derived tangent or from the
normal bundles of the inclusion square.
\[
\begin{tikzcd}[row sep=large,column sep=large]
 C\arrow[r,hook]\arrow[d,hook]&B\arrow[d,hook,"i_B"]\\
 A\arrow[r,hook]&W
\end{tikzcd}
\]
For a transverse intersection the excess bundle has rank zero.
If $A=B$, the inclusion on the left of
\eqref{eq:excess-bundle-sequence} has source zero and $E=N_{A/W}$.
Both are clean cases. Even when $TC$ and $E$ exhaust the tangent
cohomology, the entire local function algebra of the derived
intersection belongs to its geometric structure.


\section{Local functions in the clean case}

A clean zero set has a local simplification that explains the
appearance of the excess bundle in the linearization. Let
$F:U\to\mathbb R^m$ have a smooth zero set $C$ with
$\ker D_cF=T_cC$ along $C$. Near any chosen $c$, select
$r=\operatorname{codim}_U C$ components with independent
differentials. Use them as normal coordinates $u_1,\ldots,u_r$.
Their common zero set locally agrees with $C$: the latter
is an embedded submanifold of it of the same dimension, and
hence open in it near $c$. Shrinking to a product coordinate
box, write each remaining function as
\[
 F_{r+j}(a,u)=\sum_{i=1}^r h_{ji}(a,u)\,u_i .
\]
The coefficients can be obtained by integrating
$\partial_{u_i}F_{r+j}(a,t u)$ over $0\le t\le1$.
In the Koszul model the invertible change of generators
$\eta_{r+j}\mapsto\eta_{r+j}-\sum_i h_{ji}\eta_i$ makes
the new last $m-r$ generators closed. The first $r$
generators still resolve the coordinate plane. Taking their
cohomology leaves the exterior algebra over the local
functions of $C$ on $m-r$ generators of degree $-1$.

For the intersection of $A$ and $B$, the degree $-1$
generators on $C$ span the dual of the cokernel of the
normal derivative, namely $E^*$. For example, a vector in the
target dual annihilates the image of the derivative exactly
when its linear combination of the differentials of the
equations is zero on $C$. The coordinate calculation then
identifies the negative cohomology sheaves locally with
exterior powers of $E^*$. The calculation uses a splitting
of the equations on the coordinate box. A splitting on
each box does not by itself specify compatible choices on
overlaps, so it supplies local models for gluing.

\section{The Euler factor}

Write $u_V$ for the Thom class of an oriented real vector bundle
$V$, considered in the cohomology of its total space relative
to the complement of the zero section. Its Euler class is
$e(V)=0_V^*u_V$ under the zero section $0_V$. Suppose the
submanifolds $A$ and $B$ are closed as subsets of $W$, and choose
orientations of $N_{C/B}$ and $N_{A/W}$. Orient $E$ by the
exact sequence \eqref{eq:excess-bundle-sequence}, taking the
subbundle first. Let $u_A\in H^r(W,W\setminus A;\mathbb Z)$
be the class obtained by excision from the oriented normal
bundle of $A$, where $r=\operatorname{rank}(N_{A/W})$.
Assume the ranks are constant on the components in question.

\begin{theorem}[Clean excess intersection]
\label{thm:clean-excess}
Under excision to the normal bundle of $C$ in $B$,
the image of $i_B^* u_A$ is
\[
 u_{N_{C/B}}\smile p^*e(E),
 \qquad p:N_{C/B}\to C.
\]
\end{theorem}

\begin{proof}
Choose tubular coordinates at $A$ in $W$ and at $C$ in $B$ as
in Chapter~\ref{ch:tubular}. Shrink the latter tube to fit
inside the preimage of the former under $i_B$. In these
coordinates the inclusion is a map to $N_{A/W}$ of the form
\[
 (c,v)\longmapsto \big(a(c,v),\,w(c,v)\big).
\]
It sends $v=0$ to the zero vector over $c$, and $w(c,v)\ne0$
for nonzero $v$ in the tube. Choose a radius disk bundle for
the source tube, so that $(c,tv)$ remains in the tube for
$0\le t\le1$. For $t>0$ the maps
$(c,v)\mapsto(a(c,tv),t^{-1}w(c,tv))$ take values in the full
normal bundle $N_{A/W}$. Taylor's formula in local bundle
frames extends these maps smoothly to $t=0$, where the map is
$(c,v)\mapsto(c,J_c v)$ for the inclusion $J$ in
\eqref{eq:excess-bundle-sequence}. Each map sends the complement
of the zero section into the complement. For $t>0$ this
uses the description of the preimage of $A$. For $t=0$
it uses the injectivity of $J$. Consequently the original
pullback of the Thom class and its pullback along $J$ agree
as relative classes.

A bundle metric splits the exact sequence so that $J$ is
inclusion into $N_{C/B}\oplus E$ as the first factor.
The Thom class of a direct sum is the product of the
Thom classes pulled back along the projections, with the given
orientation order \cite{bott-tu}. On the subbundle in question,
the second projection is the zero section of $E$ after $p$.
Its pullback contributes $p^*e(E)$, giving the formula.
\end{proof}

The corresponding absolute class on $B$ is the Gysin image of
$e(E)$ under $C\hookrightarrow B$. The transverse formula is
recovered since the Euler class of the zero bundle is the unit.
Over $\mathbb F_2$, the same argument uses the canonical
modulo-two Thom classes and dispenses with the orientation
choices. The exact sequence still distinguishes a failure
of transversality from the dimensions of the ordinary
intersection.

\begin{example}[Self-intersection of a diagonal]
Let $W=S^2\times S^2$ and let $A=B$ be the diagonal
oriented by $S^2$. Its normal bundle is identified with
$TS^2$ by sending $(v,w)$ to $w-v$. The orientation of this
normal agrees with the coorientation induced by the product:
in the ordered basis formed by $(v,v)$ followed by $(0,w)$,
the change-of-basis matrix to the two factor bases has
determinant one. The excess theorem gives the self-intersection
number $\langle e(TS^2),[S^2]\rangle$. For an explicit
Euler calculation, project the constant unit vertical
vector $e_3$ onto each tangent plane. The result is the
section $x\mapsto e_3-(e_3\cdot x)x$. Its two zeros
are the poles. The tangent derivatives are
$-\operatorname{id}$ at the north pole and
$\operatorname{id}$ at the south pole. Both preserve
orientation in dimension two, giving total index $2$.
The ordinary set of self-intersection points is the whole
diagonal. The normal excess determines its numerical contribution.
\end{example}

\section{Tangencies and families}

Consider the $x$-axis and the graph of $x^2$ in $\mathbb R^2$.
Their intersection consists of the origin, while the
intersection of tangent spaces there is a line. The clean
condition fails. Pulling the defining equation $y=0$ back
to the graph nevertheless gives the derived zero locus
of $t^2$ in Example~\ref{ex:derived-smooth-square}.
The function ring has a nonzero nilpotent linear coordinate,
and the tangent cohomology has one dimension in each of
degrees zero and one. The virtual dimension is zero.

The nearby family of graphs $y=t^2-\lambda$ makes the
variation visible. For $\lambda>0$ there are two transverse
intersections, with opposite local signs in the normal
coordinate $y$, since the derivatives are $2\sqrt{\lambda}$
and $-2\sqrt{\lambda}$. For $\lambda<0$ there are none.
On a small interval in $t$ and for sufficiently small
$|\lambda|$, the total zero set meets neither vertical
endpoint of the $t$-interval. The equation on the parameter
product is submersive, with derivative $-1$ in the
$\lambda$ direction. Its zero locus is the parabola
$\lambda=t^2$, and restriction to an interval about the
critical parameter yields a compact oriented cobordism
between the signed zero sets at the two regular endpoints.
Hence the oriented number is zero on both sides while
the cardinality changes. This is a direct local instance of
the conservation of a signed intersection under a
regular total-space deformation.

Here the total incidence manifold is smooth, but its map to the
parameter has a critical point. The derived fiber at that
point is still computed by the single equation after base
change. This is a useful separation: transversality of the
total incidence equation gives a smooth total space, and
regularity of the projection governs which ordinary fibers
vary smoothly. Derived base change retains a fiber at every
parameter. The next chapter describes how the ordinary nearby
topology can be recorded along different parameter strata.
\par



\chapter{Stratifications and exit paths}
\label{ch:exit}

A parameter may cross a discriminant at which the regular
fibers of an incidence map change. On each regular region one
can transport nearby fibers along paths. At the discriminant
one also needs comparisons from its strata into the nearby
regions. Exit paths assemble both kinds of comparisons into a
single category. The construction uses a stratified space as
its source and can take values in spaces, retaining homotopies
of the propagated data.

\section{Conical neighborhoods and stratified simplices}

Let $I$ be a partially ordered set, with its open upward
subsets, and let $X\to I$ be continuous. The fiber $X_a$ over
$a$ is the stratum with that label. We order the labels so that
a point of a lower stratum may be approached from strata with
higher labels. In the conical models used here, a neighborhood
of a point of $X_a$ is a product of a neighborhood in that
stratum with the open cone on a link. The cone vertex has label
$a$, and the positive radial coordinates keep the labels of
the link in $I_{>a}$. For a compact link $L$ this cone is
$(L\times[0,\infty))/(L\times\{0\})$. Its sections at fixed
positive radius are copies of the link.

The exit-path simplicial set has simplices given by maps of a
standard simplex to $X$ that follow a nondecreasing chain of
labels $a_0\le\cdots\le a_n$. In barycentric coordinates
$(t_0,\ldots,t_n)$ the point with largest nonzero index $i$
must map into $X_{a_i}$. A one-simplex starts on its initial
stratum and can enter a higher stratum continuously. A
two-simplex supplies a homotopy compatible with such changes
of stratum. Within one stratum this is the usual singular
complex, and all its edges are invertible as paths. Between
strata the allowed arrows are directional. We write
$\operatorname{Exit}(X)$ for the resulting $\infty$-category
under conical hypotheses \cite[Appendix~A]{lurie-ha}.

The cone on finitely many points is a useful first example.
Give the vertex its own stratum and each positive open ray
its own stratum. Between different rays there is no exit
arrow because a path between them must return to the
vertex. From the vertex to a fixed point of a ray there is
a single exit arrow up to a contractible space of choices.
Radial interpolation compares each such path to the straight
one in that ray, preserving the initial vertex and the exit
condition. In the stratified simplex model, interpolating
radial coordinates preserves positivity on the prescribed
positive faces. Thus the same comparison contracts choices
through higher simplices. Each ray itself is contractible.
This gives the category with an initial vertex and one object
for each ray, and only the arrows from the vertex to those
objects apart from identities. This is the finite discrete
link case of the cone computation in
\cite[Lemma~3.3.5]{ayala-francis-rozenblyum}.

\section{Constructible sheaves and their transport}

A sheaf of spaces on $X$ is stratified constructible if its
restriction to each stratum is locally constant. For clarity,
one version of the exit-path theorem assumes that $X$ is
paracompact, locally of singular shape, and conically
stratified with respect to a poset satisfying the ascending
chain condition. Locally of singular shape means that the
shape of the sheaf topos on each open set is given by its
singular homotopy type. In particular the usual manifold
and finite conical examples below satisfy the hypotheses.
In this setting the theorem gives an equivalence
\[
 \operatorname{Shv}_{\mathrm{cstr}}(X;\mathcal S)
  \simeq
 \operatorname{Fun}\bigl(\operatorname{Exit}(X),\mathcal S\bigr)
\]
between constructible sheaves and functors to spaces
\cite[Theorem~A.9.3]{lurie-ha}. One direction sends a
constructible sheaf to its stalks, with transport along the
exit arrows. The other assigns to an open set the space
of coherent sections of the corresponding left fibration
over its exit category. Ayala, Francis, and Rozenblyum
develop a related theory using conically smooth
stratifications and describe the descent of the exit
construction itself \cite{ayala-francis-rozenblyum}.
The stated assumptions are part of the comparison.

For a single stratum, the exit category is the fundamental
$\infty$-groupoid, so the equivalence recovers local systems
of spaces. For the cone on finitely many points, it gives
a value $V$ at the vertex, a value $W_j$ at each ray, and
a map $V\to W_j$. In the ordinary sheaf case this follows
directly using a basis of small stars at the origin and
intervals in the rays. Restrictions between nested
sufficiently small stars induce the same vertex stalk.
Restrictions to each ray are the respective specialization
maps and restrictions inside the ray are locally constant.
Conversely, the indicated diagram defines the restrictions
on a basis. On an overlap inside a ray sections agree
precisely when their images in that ray agree, which is
the ordinary sheaf gluing condition. The topological cone
is contractible, but direction and specialization persist
in its stratification.

\section{A real fold and a complex branch}

Let $\pi:\mathbb R\to\mathbb R$ be $u\mapsto u^2$, with
the target stratified as the origin and the two complementary
rays. For a field $k$, push the constant sheaf of $k$-valued
locally constant functions forward along $\pi$. The
resulting sheaf on the parameter line has stalks $0$ on
the negative ray and $k^2$ on the positive ray. This follows
by counting components of the preimages of small intervals.
A small interval containing the origin has connected
preimage, so its stalk there is $k$. Restricting from that
interval to the positive side sends one constant to the
same value on both components. Its exit diagram is
\[
\begin{tikzcd}[column sep=large]
 0 & k\arrow[l]\arrow[r,"{a\mapsto(a,a)}"] & k^2 .
\end{tikzcd}
\]
There is another constructible sheaf with the same three
stalks and zero specialization on the positive side.
The star calculation constructs it from that diagram.
These two sheaves are inequivalent because the ranks of
their specialization maps differ. Even matching the
dimension of every nearby stalk leaves the gluing
undetermined.

Over the complex line consider $z\mapsto z^2$. On the
punctured plane this is a double covering: a choice of
one square root continues uniquely along a path and
changes sign along one positive loop around the origin.
The pushforward has stalk $k^2$ with the two factors
exchanged by this monodromy. At the origin its value is
$k$, since the preimage of a small disk is again a disk.
An exit along a radial segment again induces the diagonal
map to $k^2$. Its image is fixed by the exchange, which
gives compatibility of specialization with the loop.
For example, the sections on the punctured disk have
the same constant value on its two nearby sheets after
transport around a full turn. The regular stalk
dimension alone does not detect this condition.

In the real case the total incidence is the ordinary
parabola $\lambda=u^2$ from
Chapter~\ref{ch:derived-intersection}. The exceptional
derived fiber also remembers the relation $u^2=0$ in
its local functions. The exit diagram of the pushforward
above is instead computed from locally constant functions
on sets of preimages. Both structures can be associated to
the same family, with the exit diagram recording transport
of nearby observations and the derived fiber retaining
its infinitesimal equations.


\section{Ordinary transport on a regular region}

A condition for obtaining local systems from an incidence
space is properness over its regular region. Let
$q:E\to V$ be a proper surjective submersion of
smooth manifolds without boundary. For any Riemannian
metric on $E$, the orthogonal complement of $\ker Dq$
is a horizontal bundle. Its projection to $TV$ is
a vector bundle isomorphism over $E$. Take a small
coordinate ball about $v\in V$ with compact closure in
a larger coordinate chart. To transport a point of
$q^{-1}(v)$ to another fiber over the ball, lift the
straight line from $v$ horizontally. This is an
ordinary differential equation on $E$ using the
horizontal lift of the coordinate velocity.
Its projected path solves the chosen equation on the
base. As long as its projected path remains in the
closed ball, its lift stays in the compact inverse
image of the ball's closure. Compactness supplies
existence for the whole time interval. Lifting the
reversed straight line gives the inverse map to the
original fiber. Smooth dependence on initial points
and velocities identifies the restriction of $q$ to
the ball with a product. This is the local
trivialization in the proper submersion theorem.

Local trivializations supply transport of the
homotopy types and cohomology of the fibers along
regular paths. It is locally constant because each
trivialization uses the same fiber throughout its
ball. The fold restricts on the positive ray to
such a locally trivial two-point family. The origin
has no ordinary submersion chart and its comparison
with that family uses the separate specialization map.

Properness is significant here. The disjoint union
$\mathbb R\amalg(\mathbb R\setminus\{0\})$ maps
to $\mathbb R$ by the identity on each summand.
The map is a surjective submersion, and its fiber
cardinalities still change at zero. The missing point
in the second summand prevents the horizontal
transport argument from extending to that fiber.
This example separates local regularity of individual
points of the incidence space from control of a
whole family of fibers.

\section{Specialization between local systems}

On $S^1\times[0,\infty)$, stratify by the lower circle and
its open complement. In the stratified simplex construction,
the label of a point is determined by its second coordinate.
Consequently exit simplices in the product are pairs of
ordinary singular simplices on $S^1$ and exit simplices of
the half-line. The half-line has the exit category of the
one-ray cone, hence is equivalent to the arrow category $[1]$.
We obtain
\[
 \operatorname{Exit}\bigl(S^1\times[0,\infty)\bigr)
 \simeq \Pi_\infty(S^1)\times[1].
\]
A constructible sheaf of spaces on this cylinder specifies two
local systems and a natural transformation from the lower
one to the upper one. For vector spaces, writing their loop
transport as $A$ and $B$, a specialization $s$ satisfies
$Bs=sA$. Indeed the two ways to move once around the circle
and pass to the upper stratum form the corresponding
commuting square. For sheaves of spaces the comparison
is coherent along all paths and homotopies in the circle.
This illustrates why even simple discriminant surfaces
can require comparison maps between representations of
fundamental groups, rather than separate lists of regular
and singular fiber types.
\par



\chapter{Spaces of immersions and embeddings}
\label{ch:parametric-topology}

The immersion condition depends on the first derivative at
one point. The embedding condition also compares pairs of
points. This distinction persists in families: rank conditions
give a space of formal immersion data, while a theory of
embeddings must also track configurations. We begin with the
parametric form of immersion theory for manifolds without
boundary.

\section{Formal immersions and regular homotopy}

For $M^m,N^n$, a \emph{formal immersion} is a pair
$(f,F)$ consisting of a smooth map $f:M\to N$ and
a smooth vector bundle monomorphism $F:TM\to f^*TN$.
It is holonomic when $F=Df$. Write
$\operatorname{FImm}(M,N)$ for its space with the compact-open
smooth topology on the maps and their bundle data, and
$\operatorname{Imm}(M,N)$ for the corresponding space of
immersions. Equivalently the formal space is the space
of sections over $M$ of the bundle whose fiber at $x$
consists of $(y,L)$ with $L:T_xM\hookrightarrow T_yN$.
The differential gives the section corresponding to an
immersion.

\begin{theorem}[Smale--Hirsch]
\label{thm:smale-hirsch}
For smooth manifolds without boundary and $n>m$, the map
\[
 \operatorname{Imm}(M,N)\longrightarrow
 \operatorname{FImm}(M,N),\qquad f\longmapsto(f,Df),
\]
is a weak homotopy equivalence.
\end{theorem}

This is the parametric immersion theorem
\cite{hirsch-immersion,weiss-embeddings}. It identifies the
homotopy types of families of immersions and formal
monomorphisms in the stated dimension range. In particular
it gives a regular homotopy classification on path
components. A relative consequence for parameters follows
from factoring the differential map through its mapping
path fibration. The fibration has weakly contractible
fibers. A family of genuine immersions on a subcomplex
of a finite CW parameter space, with matching formal data
on the full parameter space, lifts to that fibration.
The lift extends from the subcomplex by successive disk
lifts. Projection back gives genuine immersions on the
full parameter space, together with a homotopy of their
differentials to the prescribed formal family, fixed on
the subcomplex. This conclusion is about continuous
families for the topology in the theorem.

\begin{example}[Plane curves]
Trivializations of $TS^1$ and $T\mathbb R^2$ identify a
formal derivative of a parametrized plane curve with a
map $S^1\to\mathbb R^2\setminus\{0\}$. The underlying
map to $\mathbb R^2$ can be contracted independently.
Radial normalization and Theorem~\ref{thm:smale-hirsch}
classify path components of the immersion space by
degrees of maps $S^1\to S^1$, or tangent winding numbers.
For instance $t\mapsto(\sin t,\sin 2t)$ is an immersion:
$\cos t=0$ forces $\cos 2t=-1$. Its normalized tangent
at $t+\pi$ is the reflection of its tangent at $t$ in
the vertical axis. Translation of the parameter is
orientation preserving, while reflection of the tangent
circle reverses orientation. Thus the winding number
equals its own negative and is zero. A smooth plane
embedding of the circle has winding number $1$ or $-1$
by the turning tangent theorem \cite[Lecture~9]{seidel-curves}.
The regular homotopy component of this figure eight
therefore contains no embeddings.
\end{example}

\section{Tests on pairs of points}

Put $M^{[2]}=(M\times M)\setminus\Delta_M$.
For a map $f:M\to N$, injectivity means that
$(x,y)\mapsto(f(x),f(y))$ avoids $\Delta_N$ on $M^{[2]}$.
If $M$ is compact, an injective immersion is an embedding:
it is a homeomorphism onto its image by compactness
and the Hausdorff property, and its local immersion charts
then give the smooth embedded structure. The deleted
product expresses the pairwise test independently of the
first-jet condition.


For compact $M$ the embedding locus is open in the $C^1$
topology. To see this, let $f$ be an embedding. At each
point choose coordinates on a small source ball and on
the target such that projection of the coordinate image
of $f$ to its first $m$ coordinates has invertible
derivative at the center. Shrink the source ball until
this derivative is uniformly close to its center value.
For every sufficiently $C^1$-close map $g$, the projected
coordinate derivative stays close to the same invertible
linear map. Integrating along line segments in the ball
then bounds the distance between the projected images of
distinct points below by a positive multiple of their
coordinate distance. This gives injectivity and full rank
on that ball. Finitely many such balls cover $M$.
Pairs sharing a ball constitute a neighborhood of the
diagonal in $M\times M$. On the compact remaining set of
pairs, the distances between their $f$-images have a
positive minimum. A small uniform perturbation keeps
those images distinct as well. This proves openness
and explains how local derivative margins and global
pair separation give different stability conditions.

\begin{proposition}[Generic embedding slices of a family]
\label{prop:generic-embedding-slices}
Let $M$ be compact of dimension $m$ and let
$f:P\times M\to N^n$ be a smooth family, with $P,M,N$
finite-dimensional, second-countable manifolds without
boundary. Assume $n>2m$. Suppose the vertical first jet
$(p,x)\mapsto j^1(f_p)(x)$ is transverse to each
rank-deficient stratum in $J^1(M,N)$, and the map
\[
 (p,x,y)\longmapsto(f_p(x),f_p(y))\colon
 P\times M^{[2]}\longrightarrow N\times N
\]
is transverse to $\Delta_N$. Then $f_p$ is an embedding
outside a set of parameters of measure zero in local charts.
\end{proposition}

\begin{proof}
For each total map take the inverse image of the
specified stratum or diagonal, a smooth submanifold by
transversality. Sard's theorem applied to its projection
to $P$ gives a full-measure set of regular parameters.
Regularity of this projection is equivalent to
transversality in the slice at that parameter. To check
the equivalence, write $A,B$ for the parameter and source
parts of the derivative into the normal quotient to the
specified submanifold. Total transversality means that
$\operatorname{im}A+\operatorname{im}B$ is that quotient.
Projection of the kernel onto the parameter tangent
space is onto exactly when every $A$-image lies in
$\operatorname{im}B$, which under total transversality
is precisely surjectivity of $B$.

There are finitely many rank strata. The stratum of rank
$k<m$ in the space of linear maps $\mathbb R^m\to\mathbb R^n$
has codimension $(m-k)(n-k)$: in a chart where a
$k$-minor is invertible, the remaining block is prescribed
by its Schur complement. This also gives the corresponding
codimension in jet coordinates. It is at least $n-m+1$,
which exceeds $m$ under the stated inequality. Therefore
a transverse slice avoids every deficient stratum.
The diagonal has codimension $n>2m$, so its transverse
preimage in the $2m$-dimensional deleted product is also
empty. The slice is an injective immersion, and compactness
gives the embedding.
\end{proof}

One can verify the diagonal hypothesis by showing that the
parameter derivatives can move the difference of the two
outputs in all normal directions at each collision.
This requires independent control at the pair under
consideration. A transversality theorem for the unrestricted
smooth mapping space gives no such control for a prescribed
family of finite dimension. For example the family
\[
 f_p(t)=\bigl(p\cos(2t),p\sin(2t)\bigr),
 \qquad p\in\mathbb R\setminus\{0\},
\]
immerses the circle for every $p$, and it identifies $t$
with $t+\pi$ for every $p$. Embedding the targets in a
higher-dimensional Euclidean space by zero padding leaves
this obstruction intact. Adding two output coordinates
$a\cos t,a\sin t$ with $a\ne0$ yields an embedding, since
the added coordinates separate points and already form
an immersion. In an architecture using trigonometric
features, the choice of features therefore changes the
class of topological realizations before the weights
are selected.

\section{The configuration stages}

Let $\mathcal O(U)$ be the poset of open subsets of an
open manifold domain $U\subset M$, and let
$\mathcal O_k(U)$ be its subposet of opens diffeomorphic to
unions of at most $k$ disjoint open balls.
Embedding calculus studies the restriction functor
$V\mapsto\operatorname{Emb}(V,N)$. Its approximations
can be expressed by the homotopy right Kan extensions
\[
 T_k\operatorname{Emb}(U,N)=
 \operatorname*{holim}_{V\in\mathcal O_k(U)^{\mathrm{op}}}
      \operatorname{Emb}(V,N).
\]
Restriction from $U$ gives the comparison to each stage.
Increasing $k$ adds simultaneous compatibility on
configurations spread across more balls. In the range
$n>m$, the first stage is weakly equivalent to the
immersion space \cite{weiss-embeddings}. One way to
understand the first stage is that one-ball restrictions
of embeddings and immersions have the same local homotopy
information, while immersion spaces satisfy homotopy descent
for the ordinary open covers used at that stage. Later
stages retain compatibility beyond this one-point condition.
\[
\begin{tikzcd}[column sep=large]
 \operatorname{Emb}(U,N)\arrow[r]\arrow[dr]&
 T_2\operatorname{Emb}(U,N)\arrow[d]&\\
 &T_1\operatorname{Emb}(U,N)\arrow[r,"\simeq"]
     &\operatorname{Imm}(U,N)
\end{tikzcd}
\]
The convergence theorem in codimension at least three
identifies the embedding space for a compact source without
boundary with the homotopy inverse limit of its stages
\cite{goodwillie-weiss}. Both the codimension condition
and the passage to the limit matter. The rank-and-pair
test gives a criterion on an actual smooth map in a
suitable dimension range. The tower compares entire
spaces of embeddings and keeps their homotopies.

For families, an actual map from a parameter space into
$\operatorname{Emb}(M,N)$ induces maps to all of these
stages and to the formal immersion space. This expresses
the comparison using functors between spaces of models.
Applying the same construction to restrictions of the
parameter commutes with the maps in the tower. Imposing
an architecture supplies a particular subset or family
of realizations inside these mapping spaces. The topology
of its immersion and embedding loci depends on how
that realization map meets the topology detected by
differentials and configurations.
\par



\bookpart{VII}{Equivariance, factorization, and persistence}

\chapter{Compact group actions and orbit types}
\label{ch:orbit-types}

The stabilizer of a point determines which symmetries act on its tangent
and normal spaces. This linear information also governs a neighborhood of
the orbit. We use it to describe what remains when orbits are treated as
objects with automorphisms, and to locate the strata on which ordinary
smooth quotient charts suffice.

Throughout this chapter, \(G\) is a compact Lie group acting smoothly on a
Hausdorff second-countable manifold \(M\) without boundary. We write
\(G_x=\{g\in G:gx=x\}\) and \(Gx=\{gx:g\in G\}\). An equivariant map
\(f:M\to N\) satisfies \(f(gx)=gf(x)\). Differentiation at \(x\) gives
\[
D_{gx}f\circ D_xg=D_{f(x)}g\circ D_xf.
\]
In particular, \(G_x\subseteq G_{f(x)}\), and \(D_xf\) is an intertwiner
for \(G_x\). This condition persists for smooth families, with the
stabilizer taken in the total source when parameters themselves move.

\section{Invariant metrics and fixed points}

Choose a Riemannian metric \(h\) on \(M\). With probability Haar measure
\(dg\), set
\[
\bar h_x(v,w)=
\int_G h_{gx}\bigl(D_xg(v),D_xg(w)\bigr)\,dg .
\]
Compactness permits differentiation in \(x\) in local charts. The average
of the positive functions obtained by setting \(w=v\ne0\) is positive,
and right invariance of Haar measure makes \(\bar h\) invariant. Every
element of \(G\) therefore acts by isometries. Its action commutes with
the exponential map wherever that map is defined.

\begin{proposition}[Tangent spaces to fixed loci]
Let \(K\) be a closed subgroup of \(G\). The set \(M^K\) of points fixed by
every element of \(K\) is a closed smooth submanifold, allowing different
dimensions on different connected components. At \(x\in M^K\) its
tangent space is \((T_xM)^K\).
\end{proposition}

\begin{proof}
An exponential chart at \(x\) identifies an invariant small ball in
\(T_xM\) with a neighborhood of \(x\). The chart is \(K\)-equivariant
because \(K\) fixes \(x\) and acts isometrically. If a point of its
image is fixed, injectivity implies that its tangent coordinate is
fixed by \(K\). Conversely an exponential of a fixed vector is
fixed. The fixed vectors form a linear subspace, so these charts
prove both the local submanifold assertion and the tangent formula.
Closedness follows by intersecting the closed fixed sets of
individual elements.
\end{proof}

For example, \(S^1\) acts on \(S^2\) by rotation about its vertical axis.
The two poles form the fixed locus and their entire tangent planes
are rotating representations. The nonpolar points have trivial
stabilizer. A tangent bundle on the orbit space alone therefore
cannot account for the two polar tangent representations.

\section{A slice from the normal bundle}

\begin{theorem}[Slice at a compact orbit]
\label{thm:compact-slice}
For \(x\in M\), put \(H=G_x\) and
\(V_x=(T_xGx)^\perp\subset T_xM\) for an invariant metric. There is a
ball \(D\subset V_x\) about \(0\) and an equivariant diffeomorphism
\[
G\times_H D \ \xrightarrow{\ \sim\ }\ U,\qquad
[g,v]\longmapsto g\exp_x(v),
\]
where \(U\subset M\) is an invariant open neighborhood of the orbit.
The balanced product identifies \((gh,v)\) with \((g,hv)\).
\end{theorem}

\begin{proof}
The subgroup \(H\) is closed. The orbit map induces an injective map
\(G/H\to M\). The differential of \(G\to M\) at the identity sends
a Lie algebra element \(\xi\) to the velocity of
\(t\mapsto\exp(t\xi)x\). A zero velocity means that the resulting
fundamental vector field vanishes at \(x\). Uniqueness for the
flow makes the entire one-parameter subgroup fix \(x\), so the
kernel is the Lie algebra of \(H\). Translating establishes that
\(G/H\to M\) is an immersion, and compactness of \(G/H\) makes it an
embedding.

The normal spaces of the orbit at \(gx\) are the translates of
\(V_x\) by \(D_xg\). Thus its normal bundle is \(G\times_H V_x\), via
\([g,v]\mapsto D_xg(v)\). The tubular-neighborhood construction of
Chapter~\ref{ch:tubular} applies to the normal exponential.
Compactness of \(Gx\) allows a single small radius, and the metric
makes the resulting diffeomorphism equivariant. Its fiber over \(x\)
is the ball \(D\).
\end{proof}

This proof is the compact-group case of the slice theorem. Proper
Lie group actions admit a more general version
\cite{meinrenken-actions}. In the slice above, the stabilizer of
\([g,v]\) is \(gH_vg^{-1}\) with \(H_v=\{h\in H:hv=v\}\).
Every nearby stabilizer is therefore conjugate to a subgroup of
\(H\). Moreover, the points of the same orbit type as \(x\), meaning
that their stabilizer is conjugate to \(H\), correspond locally to
\(G\times_H D^H\). Indeed, if \(H_v\subseteq H\) is conjugate in
\(G\) to \(H\), it has the same dimension and number of components
as \(H\). The inclusion then identifies their identity components
and reaches each component of \(H\), hence \(H_v=H\). The local
orbit-type stratum is \((G/H)\times D^H\), with orbit quotient
\(D^H\). The entire neighborhood has orbit quotient \(D/H\).
In particular a free action has charts with \(H=1\), giving
a smooth quotient and a principal bundle over it.

There is also a reduction of the symmetry group along any fixed
orbit type. Fix a closed subgroup \(H\) which occurs as a
stabilizer and put \(M_H=\{x:G_x=H\}\). In \(M^H\), this set is
open. To see this near \(x\in M_H\), a nearby stabilizer is
conjugate to a subgroup of \(H\) by the slice theorem. If it
contains \(H\), equality follows by comparing dimensions and
component counts as above. The normalizer
\(N_G(H)=\{g:gHg^{-1}=H\}\) preserves \(M_H\), and \(N_G(H)/H\)
acts freely on it. The \(G\)-orbit of any point of type \((H)\)
meets \(M_H\), and two points of \(M_H\) in a common \(G\)-orbit
differ by an element of \(N_G(H)\). It follows that
\(M_{(H)}/G\) identifies with \(M_H/(N_G(H)/H)\).
In the slice at a point of \(M_H\), the exact-stabilizer locus
is \(N_G(H)\times_H D^H\): a point \([g,v]\) has stabilizer \(H\)
precisely when \(v\in D^H\) and \(g\in N_G(H)\).
Thus both quotient charts identify with \(D^H\), proving that
the identification is smooth for their quotient manifold structures. This describes the global
quotient within one type using an action with free orbits.

\section{The quotient stack in slice coordinates}

Write \([M/G]\) for the smooth quotient stack. Its objects over a test
manifold \(T\) are principal right \(G\)-bundles \(P\to T\) with a smooth
map \(F:P\to M\) satisfying \(F(pg)=g^{-1}F(p)\). Morphisms are bundle
isomorphisms commuting with these maps. Restricting to the neighborhood
of Theorem~\ref{thm:compact-slice} gives
\[
\begin{tikzcd}[column sep=large]
D \arrow[r,hook] \arrow[d] &
G\times_H D \arrow[d] \\
{[D/H]} \arrow[r,"\simeq"] & {[U/G]} .
\end{tikzcd}
\]
To prove the lower equivalence, use the equivariant projection
\(U\to G/H\). The inverse image of the coset \(H\) under
\(P\xrightarrow{F}U\to G/H\) is a principal right \(H\)-bundle
\(Q\subset P\). Locally, a section of \(P\) and a smooth lift of the
resulting map into \(G/H\) produce a section of \(Q\). The
restriction \(F|_Q\) lands in the fiber \(D\). In the other
direction a bundle \(Q\) with equivariant \(u:Q\to D\) extends to
\(Q\times_H G\) with map \([q,g]\mapsto g^{-1}[1,u(q)]\). The
two constructions recover each other and their bundle
isomorphisms. In particular they commute with pullback in \(T\).

The infinitesimal action at \(x\) is the linear map
\[
\mathfrak g \longrightarrow T_xM,\qquad
\xi\longmapsto \left.\frac{d}{dt}\right|_{t=0}\exp(t\xi)x .
\]
Its kernel is \(\mathfrak h=\operatorname{Lie}(H)\) and its cokernel
is the normal representation \(V_x\). They are the cohomology of the tangent complex of the action
groupoid in degrees \(-1\) and \(0\), respectively. The reason for the complex is that an infinitesimal
group element is an arrow from a displacement \(v\) to
\(v+\xi_M(x)\). Quotienting the displacement gives the cokernel, and
arrows from a displacement to itself give the kernel. A finite
stabilizer has zero Lie algebra, so the tangent complex can miss
finite automorphisms even when the slice stack retains them.

\begin{example}[An exceptional orbit with a disk quotient]
Let \(S^1\) act on \(S^3\subset\mathbb C^2\) by
\(\lambda(z_1,z_2)=(\lambda z_1,\lambda^2z_2)\).
A point with \(z_1\ne0\) has trivial stabilizer. At \(x=(0,1)\)
the stabilizer is \(\{1,-1\}\) and the orbit tangent is the
imaginary \(z_2\)-axis. Its orthogonal slice is the complex
\(z_1\)-line, with \(-1\) acting by multiplication by \(-1\).
The coarse chart \(D/\{1,-1\}\) is topologically a disk, by the
map \(z\mapsto z^2\), whereas \([D/\{1,-1\}]\) assigns the
order-two automorphism group to the central point. Both the
Lie algebra of the stabilizer and the local coarse topology
would leave that finite group undetected.
\end{example}

The slice converts a question about changing orbit types into one
about linear representations. The next chapter uses this reduction
for a local obstruction: the normal derivative of an equivariant
map has to be surjective within an appropriate category of
representations.
\par



\chapter{Equivariant transversality and fixed-locus constraints}
\label{ch:equivariant-transversality}

The ordinary transversality argument perturbs a map in directions
spanning the target normal space. Equivariance restricts those
directions to intertwiners of stabilizer representations. At a fixed
point even a tiny invariant target may require a normal representation
absent from the source. We identify the obstruction at a single orbit,
then show how the ordinary density argument returns for free source
actions.

\section{The normal map at a stabilizer}

Let \(G\) be a compact Lie group acting smoothly on manifolds \(M,N\)
without boundary. Let \(A\subset N\) be a \(G\)-invariant embedded
submanifold and \(f:M\to N\) an equivariant smooth map.
Fix \(x\in M\) with \(y=f(x)\in A\), set \(H=G_x\), and let
\(N_yA=T_yN/T_yA\) denote the fiber of the normal bundle.
The classical transversality condition at \(x\) is surjectivity of
\(T_xM\xrightarrow{D_xf}T_yN\to N_yA\). This map is \(H\)-equivariant.
Its restriction to \(T_xGx\) vanishes, since \(f(Gx)\subseteq A\).
It therefore descends to an \(H\)-map
\[
\overline{D_x f}:T_xM/T_xGx\longrightarrow N_yA .
\]

\begin{proposition}[Transverse germs at an orbit]
\label{prop:equivariant-normal-map}
Let \(x\in M\), \(y\in A\), and assume \(H=G_x\) fixes \(y\).
There exists an equivariant map on an invariant neighborhood of \(Gx\)
taking \(x\) to \(y\) and transverse to \(A\) along \(Gx\) if and only if
\(T_xM/T_xGx\) admits a surjective \(H\)-linear map to \(N_yA\).
\end{proposition}

\begin{proof}
Any such map gives the stated normal derivative. Conversely choose
invariant metrics and identify the two quotient spaces with
orthogonal complements. Let \(L:V_x\to (T_yA)^\perp\) be a
surjective \(H\)-linear map. In the slice chart of
Theorem~\ref{thm:compact-slice}, use
\[
[g,v]\longmapsto g\exp_y(Lv).
\]
For a small enough ball this exponential is defined. Because \(y\)
is fixed by \(H\), the chart at \(y\) commutes with the \(H\)-action,
making the formula well defined. Its normal derivative on the slice
at \(v=0\) is \(L\), and translation carries surjectivity to each
point of the orbit.
\end{proof}

There is a local perturbation version of the same criterion.
Fix an equivariant \(f\) with \(f(x)=y\). On a smaller slice
ball the expression \(u(v)=\exp_y^{-1}(f(\exp_x v))\) is
defined and is \(H\)-equivariant. Choose an invariant bump
\(\chi\) on this ball, equal to one near zero. Replacing
\(u(v)\) by \(u(v)+t\chi(v)Lv\), where \(L\) is a surjection
as in the proof, gives a smooth perturbation through \(y\)
with support near \(Gx\). Extend it by equivariance on the
tube and by the old \(f\) off the support. Its normal derivative
at \(x\) is \(\overline{D_x f}+tL\). A full-rank minor of \(L\)
provides a polynomial minor of this derivative with nonzero
leading coefficient. Hence for arbitrarily small \(t\) the
new germ is transverse along \(Gx\). When the normal fiber
is zero the original germ is already transverse. The
construction concerns a single orbit and leaves global density
as a separate question.

The fixed-point
loci \(M^H,N^H,A^H\) are submanifolds by Chapter~\ref{ch:orbit-types}.
Invariant metrics split the normal sequence, yielding
\[
T_yN^H/T_yA^H\ \cong\ (N_yA)^H .
\]
If \(f\) is transverse at \(x\), its normal derivative is surjective
on fixed vectors. In fact, averaging any preimage of an invariant
vector gives an invariant preimage. Thus the restricted map
\(f|_{M^H}:M^H\to N^H\) is transverse to \(A^H\) at \(x\). Testing
only the restriction for the full stabilizer \(H\) tests only the
trivial part of the normal representation.

\begin{example}[A persistent singular zero]
Let \(M=\mathbb C\) have the \(S^1\)-action \(\lambda z\), and \(N=\mathbb C\)
the action \(\lambda^2 w\). Put \(A=\{0\}\). Every equivariant map sends
\(0\) to \(0\), since the only point of \(N\) fixed by the full group
is \(0\). The element \(-1\) acts by \(-1\) on \(T_0M\) and by \(+1\) on
\(T_0N\). Hence \(D_0f(-v)=D_0f(v)\) and real linearity gives
\(D_0f=0\). No equivariant perturbation achieves ordinary transversality
at the origin. The fixed-locus restriction is a map from a point
to a point and is transverse there. For the concrete map
\(z\mapsto z^2\), the zero set is a single point, even though the
linear normal test fails in both real directions.
\end{example}

\begin{example}[Moving a parameter changes the constraint]
Let \(P=\mathbb R\) and \(N=\mathbb R\) with the sign action of
\(C_2\) on each. The evaluation \(P\to N,\ p\mapsto p\) at a
one-point input is an equivariant submersion. At the fixed parameter
\(0\) the constant evaluation at the sole input hits \(\{0\}\) and
fails transversality. For \(p\ne0\), the evaluation misses \(\{0\}\).
Holding \(p\ne0\) fixed leaves no sign symmetry of the parameter.
That is why ordinary parameterized transversality can hold for the
family while the restriction to fixed parameters has no transverse
member.
\end{example}

\section{Recovery for free source actions}

\begin{theorem}[Equivariant density for a free compact source]
\label{thm:free-source-transversality}
Let \(M\) be compact without boundary with a free smooth \(G\)-action,
and let \(A\subset N\) be invariant, closed, and embedded. In the
smooth equivariant mapping space from \(M\) to \(N\), the maps
transverse to \(A\) are dense and open in the \(C^\infty\) topology.
\end{theorem}

\begin{proof}
Write \(B=M/G\). The slice theorem supplies the smooth principal
bundle \(M\to B\). Equivariant maps \(M\to N\) correspond to sections
of the associated bundle \(E=(M\times N)/G\to B\), where \(G\) acts
diagonally. In a principal bundle chart an equivariant map is
\((g,b)\mapsto g\,s(b)\), with \(s(b)\) the fiber coordinate of the
section. Its normal derivative to \(A\) vanishes in the group
directions and is the normal derivative of \(s(b)\) in the base
directions, up to translation by \(g\). Thus transversality of the
map to \(A\) is equivalent to transversality of its section to
\(E_A=(M\times A)/G\subset E\).

Fix a section \(s\). Its image is compact. Cover the image by
finitely many local bundle and fiber coordinate charts, with
smaller charts still covering it and bump functions supported in
the larger charts. Multiplying fiber coordinate vector fields by
these bumps gives globally defined compactly supported vertical
vector fields \(V_1,\ldots,V_\ell\) spanning the vertical space
at every point of \(s(B)\). Their flows preserve the projection
and can be composed to give sections \(s_t\) for
\(t\in\mathbb R^\ell\). The evaluation \((b,t)\mapsto s_t(b)\) has
at \(t=0\) a differential whose parameter directions span the
vertical tangent, while its base directions project onto
\(T_bB\). It is a submersion on \(B\times U\) for a sufficiently
small open neighborhood \(U\) of zero, by continuity and
compactness. Parametric transversality from
Chapter~\ref{ch:transversality} supplies transverse sections \(s_t\)
for parameters arbitrarily close to zero. The flow construction
converges to \(s\) in the smooth topology as \(t\to0\).
Transversality is open in the \(C^1\) topology on the compact source
because \(A\) is closed. Any sequence of failed normal-surjectivity
tests would have base points with a convergent subsequence and
would give a failed test for the limiting map. These statements
transfer to equivariant maps by the local bundle charts.
\end{proof}

The proof uses perturbations of sections of a bundle on a manifold.
When stabilizers occur, admissible perturbations have to respect
their linear actions and the same submersion argument may become
impossible. For related equivariant general position questions
see \cite{wasserman-equivariant}. The classical parametric argument
used in the proof is developed in \cite[Ch.~3]{hirsch}.

\section{Pullbacks of quotient stacks}

For an everywhere transverse equivariant map, \(Z=f^{-1}(A)\) is
an invariant smooth submanifold. Its quotient stack fits into
\[
\begin{tikzcd}[column sep=large]
{[Z/G]} \arrow[r] \arrow[d] &
{[M/G]} \arrow[d,"{[f]}"] \\
{[A/G]} \arrow[r] & {[N/G]} .
\end{tikzcd}
\]
This is a homotopy pullback of smooth stacks. Over a test manifold
an object of the pullback consists of two principal bundles with
equivariant maps to \(M\) and \(A\), together with an isomorphism
between the images in \([N/G]\). Use the isomorphism to identify
the bundles. Their two maps then have the same composite to \(N\).
The common bundle map lands in the ordinary transverse fiber
product \(M\times_N A\), identified with \(Z\), and the construction
also matches bundle isomorphisms. Taking quotient stacks hence
retains the stabilizer at each point of the transverse solution
set. The derived pullbacks of
Chapter~\ref{ch:derived-intersection} give another question to ask
when the normal map fails to be surjective: which tangent and
obstruction directions belong to the intersection with its
equations retained?
\par



\chapter{Factorization through an intermediate manifold}
\label{ch:factorizations}

A smooth map \(f:X\to Y\) may be specified by two maps through an
intermediate manifold \(L\). The factorization retains how the map
is constructed. Rank bounds give local restrictions, and homotopy
invariants can rule out a factorization even when every local
rank test passes. We will also distinguish a pointwise choice of
factorization from a continuous family of choices.

\section{Differential constraints}

Suppose \(f=d\circ e\) with \(e:X\to L\) and \(d:L\to Y\).
The chain rule says that \(D_xf\) factors as
\[
T_xX\xrightarrow{D_xe}T_{e(x)}L\xrightarrow{D_{e(x)}d}T_{f(x)}Y.
\]
Consequently \(\operatorname{rank}D_xf\le\dim L\). The constant-rank
normal form gives a local converse at points of an open region where
the rank is some fixed \(r\): in suitable coordinates the map is
the projection to \(r\) coordinates followed by the inclusion of
those coordinates in the target chart. This describes a local
factorization through an open subset of \(\mathbb R^r\). The
coordinate projections on two such neighborhoods need not provide
one global intermediary.

\begin{proposition}[An exact reconstruction embeds the source]
\label{prop:split-embedding}
If \(X\) is compact and \(d\circ e=\mathrm{id}_X\) with smooth
\(e:X\to L\) and \(d:L\to X\), then \(e\) is a smooth embedding.
If \(X\) and \(L\) have the same dimension, both have no boundary,
and \(L\) is connected, then \(e\) is a diffeomorphism for
nonempty \(X\).
\end{proposition}

\begin{proof}
The left-inverse identity makes \(e\) injective, and differentiating
it makes \(De\) injective at every point. Compactness of \(X\) and
the Hausdorff property of \(L\) give a homeomorphism onto the
image. Together with immersion, this is the embedding condition.
In equal dimensions the inverse function theorem makes \(e(X)\)
open. It is also compact, hence closed, and connectedness of \(L\)
implies \(e(X)=L\). The local smooth inverses then give the global
inverse.
\end{proof}

The derivative already imposes a representation constraint when
a group acts on the spaces and both maps are equivariant. At a
fixed input \(x\), the identity factors through \(T_{e(x)}L\) by
two intertwiners. Thus \(T_xX\) must occur as a split subrepresentation
for the stabilizer of \(x\). This may exclude an intermediate
manifold of sufficient ordinary dimension.

Conversely, every smooth embedding of compact \(X\) in \(L\)
admits an exact reconstruction defined on a neighborhood of the
encoded image. By the tubular-neighborhood theorem, a small
neighborhood \(U\) of \(e(X)\) is identified with a neighborhood
of the zero section in its normal bundle. Projection from the
bundle to \(e(X)\), followed by \(e^{-1}\), gives a smooth
decoder \(U\to X\). The existence of a decoder on all of \(L\)
asks for an extension of a left inverse. A local normal form
alone only supplies the neighborhood decoder, even when the
embedded image is closed. For example, the usual inclusion of
\(S^n\) in \(\mathbb R^{n+1}\) has the radial decoder on an
annulus, yet a global sphere-valued decoder would induce an
impossible retraction on top cohomology.

\section{Obstructions preserved by homotopy}

Suppose now that \(d\circ e\) is homotopic to \(\mathrm{id}_X\).
For any commutative coefficient ring \(R\), functoriality in singular cohomology
gives unital graded ring maps
\[
H^*(X;R)\xrightarrow{d^*}H^*(L;R)
 \xrightarrow{e^*}H^*(X;R),\qquad e^*d^*=\mathrm{id}.
\]
A nonzero cup product in \(H^*(X;R)\) must remain nonzero after
applying \(d^*\). Rank tests on the groups \(H^k\) alone can miss
this restriction. Similarly, on homotopy groups a path supplied
by the homotopy allows the base point of \(d\circ e\) to be compared
with the starting point in \(X\). With that comparison in place,
the map on each homotopy group induced by \(e\) has a left
inverse induced by \(d\). See \cite{hatcher-at} for the cohomology and
base-point conventions.

For a closed connected oriented \(n\)-manifold \(X\), with \(n>0\),
a degree-one reconstruction of \(X\) induces a nonzero map on
\(H^n(X;\mathbb Z)\). It cannot factor through a manifold of
dimension less than \(n\), since that manifold has no cohomology
in degree \(n\). Here one can use a triangulation of the
intermediate smooth manifold and its dimension bound. This is a
homotopy obstruction, so it also applies when the reconstruction
has critical points. A factorization through a contractible
\(L\) makes the composite nullhomotopic by applying \(d\) to a
contraction after \(e\). Positive dimension allows degree to
distinguish this case from the identity.

\begin{example}[The cup product on a torus]
Let \(X=T^2\) and take \(L\) to be the connected sum of two copies of
\(S^1\times S^2\). The first cohomology groups of \(X\) and \(L\) both
have rank two over \(\mathbb Z\). The product of the two coordinate
classes in \(H^1(T^2;\mathbb Z)\) generates \(H^2(T^2;\mathbb Z)\).
On \(L\), all products of degree-one classes are zero.
Indeed, collapsing the separating sphere of the connected sum gives
a map into the wedge of the two closed summands, inducing an
isomorphism in \(H^1\) by the Mayer--Vietoris sequence.
In the wedge, classes on different summands have zero product.
Each summand has one degree-one generator pulled back from \(S^1\),
so its square vanishes too. Thus the ring retraction displayed above
cannot exist, even though comparing \(H^1\) alone would permit it.
\end{example}

\section{A sharp uniform obstruction on the sphere}

\begin{proposition}[Reconstruction with a Euclidean intermediary]
\label{prop:sphere-reconstruction}
Let \(n>0\), and view \(S^n\) as the unit sphere in
\(\mathbb R^{n+1}\). The following statements use the Euclidean
norm on that ambient space.
\begin{enumerate}
\item For every \(m\), if \(e:S^n\to\mathbb R^m\) and
  \(d:\mathbb R^m\to S^n\) are continuous, then
  \(\sup_{x\in S^n}\|d(e(x))-x\|=2\).
\item If \(d\) is allowed to land in \(\mathbb R^{n+1}\) and
  \(m\le n\), the minimum possible uniform reconstruction error is \(1\).
  For \(m\ge n+1\), there is an exact ambient reconstruction.
\end{enumerate}
\end{proposition}

\begin{proof}
In the first case \(d\circ e\) has degree zero by contractibility.
If it had no antipodal coincidence \(d(e(x))=-x\), then the straight
segment between \(x\) and \(d(e(x))\), normalized to the unit sphere,
would be a homotopy to the identity. Such a segment passes through
the origin only for antipodal endpoints. The degrees therefore
force an antipodal coincidence and the claimed error.

For the second part, the Borsuk--Ulam theorem applied to \(e\),
after including \(\mathbb R^m\) into \(\mathbb R^n\), gives a pair
\(x,-x\) with the same encoded value \cite{matousek-borsuk}.
Both reconstructions are a common vector \(u\). The triangle
inequality gives \(2\le\|u-x\|+\|u+x\|\), so at least one error
is at least \(1\). The constant reconstruction \(0\) attains
that bound. For \(m\ge n+1\), include the sphere in the first
\(n+1\) coordinates and decode by their projection.
\end{proof}

The distinction between the codomains matters. If a reconstruction
lands close to the sphere, one can normalize along its realized
image. This does not imply that the decoder can be normalized on
the entire intermediate \(\mathbb R^m\), where it may take value zero.

\section{Spaces of factorizations}

In the category of spaces the factorizations of \(f\) through \(L\),
together with their chosen homotopies, form the homotopy pullback
\[
\begin{tikzcd}[column sep=large]
\operatorname{Fact}_L(f) \arrow[r] \arrow[d] &
\operatorname{Map}(X,L)\times\operatorname{Map}(L,Y)
  \arrow[d,"\circ"]\\
\{f\} \arrow[r,hook] & \operatorname{Map}(X,Y) .
\end{tikzcd}
\]
A point is a triple \((e,d,\gamma)\), with \(\gamma\) a path from
\(d\circ e\) to \(f\). Paths between such points track changes
of both maps and homotopies. This is a homotopy problem, while an
exact smooth factorization requires equality \(d\circ e=f\) and
smooth maps. For example, the smooth map
\(e^{it}\mapsto \exp(i\sin t)\) on \(S^1\) is nullhomotopic and
therefore factors through a point in the homotopy sense. It is
nonconstant, so it does not admit an exact factorization through a
point.

Families can have an obstruction even when every individual target
value has a lift. With decoder \(d(z)=z^2\) on \(L=S^1\), every
single \(b\in S^1\) has a preimage. The family \(f(b)=b\) has no
continuous choice of encoded value \(e(b)\) satisfying
\(d(e(b))=b\). Taking degrees would yield \(2\deg(e)=1\), which
also rules out a choice accompanied by a homotopy to \(f\).
Local choices are related on chart overlaps by the sign action.
The obstruction is in the family of choices, invisible to the
existence of each separate factorization.
\par



\chapter{Persistence of sublevel families}
\label{ch:persistence}

Morse theory studies a smooth function by following its sublevel sets
across regular intervals and critical values. Keeping the sublevel
sets together with their inclusion maps gives a diagram of spaces.
Applying homology turns this diagram into a representation of the
ordered set of thresholds. This chapter gives a direct stability
estimate on the diagram and compares it with the stronger local
triviality available for smooth parameter families at regular levels.

\section{The filtered diagram}

For a continuous function \(f:X\to\mathbb R\), let
\[
X_f(a)=\{x\in X:f(x)\le a\}.
\]
For \(a\le b\) there is an inclusion \(X_f(a)\hookrightarrow X_f(b)\).
These data give a functor from the real line with its order arrows to
topological spaces, and hence a diagram in the category of homotopy
types. In particular, using singular chains before passing to
homology keeps a chain-level diagram. For any coefficient field \(k\),
the homology groups \(a\mapsto H_j(X_f(a);k)\) form a persistence module.
Its structure map is the homology map induced by inclusion.

Given \(\varepsilon\ge0\), translation takes a diagram \(F\) to
\(a\mapsto F(a+\varepsilon)\). A strict \(\varepsilon\)-interleaving of
diagrams of spaces consists of natural maps
\[
\alpha_a:F(a)\longrightarrow E(a+\varepsilon),\qquad
\beta_a:E(a)\longrightarrow F(a+\varepsilon).
\]
The two shifted composites must equal the respective structure maps
for a step of length \(2\varepsilon\). Passing from actual spaces to
homotopy types sends these identities to coherent comparisons. A
strict interleaving is enough for the applications below.

\begin{proposition}[Uniform control of filtered spaces]
\label{prop:sublevel-interleaving}
If \(f,g:X\to\mathbb R\) are continuous and
\(\sup_{x\in X}|f(x)-g(x)|\le\varepsilon\), their sublevel diagrams
have a strict \(\varepsilon\)-interleaving. It is natural under
pullback of functions along continuous maps. When \(f,g\) are
invariant under a common group action, the maps of the
interleaving are equivariant.
\end{proposition}

\begin{proof}
The inequalities \(g(x)\le f(x)+\varepsilon\) and
\(f(x)\le g(x)+\varepsilon\) give inclusions in the indicated
directions. All arrows are the restrictions of the identity map
on \(X\). In particular the composition
\[
\begin{tikzcd}[column sep=large]
X_f(a) \arrow[r,hook,"\alpha_a"] &
X_g(a+\varepsilon) \arrow[r,hook,"\beta_{a+\varepsilon}"] &
X_f(a+2\varepsilon)
\end{tikzcd}
\]
is the structure inclusion, and the reverse composite is the
analogous inclusion on \(g\)-sublevels. For a continuous map
\(u:Y\to X\), use \(u\) on sublevel sets of the pulled-back functions.
It commutes with the interleaving inclusions. Invariance makes
each sublevel set a \(G\)-space and makes the inclusions equivariant.
\end{proof}

This estimate survives functorial operations on the diagram. For
instance one can take fixed spaces under a subgroup \(H\), followed
by homology. The fixed sublevel \(X_f(a)^H\) is the sublevel of
\(f|_{X^H}\) inside \(X^H\). Alternatively, one can apply homology
to \(EG\times_G X_f(a)\), the homotopy quotient at threshold \(a\).
The resulting invariants answer different questions about symmetry,
and the same uniform bound at the diagram level applies to both.
A bottleneck bound for their homology still requires a tameness
hypothesis for the resulting modules.

\section{When persistence diagrams are available}

A persistence module is called \(q\)-tame if every structure map
over a strictly positive step has finite rank. This does not
require finite-dimensional homology exactly at each threshold.
When \(X\) is the realization of a finite simplicial complex,
the homology persistence modules of every continuous real-valued
\(f\) on \(X\) are \(q\)-tame. In fact, fix \(a<b\) and subdivide
so that the oscillation of \(f\) on every simplex is less than
\(b-a\), using compactness and uniform continuity. The union of
simplices that meet \(X_f(a)\), together with their faces, is a
finite subcomplex \(K\) with
\[
X_f(a)\subseteq |K|\subseteq X_f(b).
\]
The corresponding homology map factors through a finite-dimensional
vector space \(H_j(|K|;k)\). This also applies to compact smooth
manifolds by smooth triangulability.

The stability theorem for persistence diagrams bounds their
bottleneck distance by the interleaving distance for \(q\)-tame
modules \cite{chazal-persistence}. Consequently the bound in
Proposition~\ref{prop:sublevel-interleaving} gives a bottleneck
bound \(\varepsilon\) in the finite-complex setting. One way to
read the estimate is that a feature with lifetime exceeding
\(2\varepsilon\) cannot simply be matched to the diagonal under
a bottleneck matching of cost at most \(\varepsilon\).
See \cite{cohen-steiner-stability} for the classical stability
result for sublevel-set persistence. The inclusions proved here
are assertions at the level of spaces, independent of the
tameness used to assign persistence diagrams. Applying homology
forgets information, such as cup products, that can distinguish
otherwise identical homology groups or homology persistence modules.

\begin{example}[Equal homology modules with different filtered spaces]
Equip \(S^2\vee S^4\) and \(\mathbb{CP}^2\) with the constant
function zero. In both cases the sublevel is empty before zero,
and the entire space thereafter. The two spaces have one
homology generator in each of degrees \(0,2,4\), with field
coefficients, so the ordinary homology persistence modules agree
in every degree. The degree-two generator on \(\mathbb{CP}^2\)
has a nonzero square, whereas the degree-two class on the wedge
is pulled back along the projection to \(S^2\) and has square zero
\cite{hatcher-at}. An equivalence of their filtered homotopy types
would give a homotopy equivalence at every nonnegative threshold
and would preserve these products. Thus the diagrams of spaces
can distinguish filtrations with identical homology persistence.
\end{example}

\section{Transport at finitely many regular thresholds}

Let \(X\) be a compact smooth manifold without boundary, \(P\) a smooth
parameter manifold, and \(F:P\times X\to\mathbb R\) smooth. Write
\(f_p(x)=F(p,x)\). The map \(p\mapsto f_p\) is continuous for the
uniform norm by compactness of \(X\). More quantitatively, along a
smooth path \(c\) of parameters of length \(\ell(c)\), a bound \(L\)
on the norm of the partial derivative \(D_pF\) over its image
times \(X\) gives
\(\|f_{c(1)}-f_{c(0)}\|_\infty\le L\ell(c)\).
This follows by differentiating \(F(c(t),x)\), integrating, and
taking the supremum over \(x\). It gives an interleaving estimate
even along paths where the topology at a chosen threshold changes.

\begin{proposition}[Local trivialization of regular sublevel flags]
\label{prop:regular-sublevel-family}
If \(a_1<\cdots<a_r\) are regular values of \(f_{p_0}\), there is a
neighborhood of \(p_0\) on which the family of finite flags
\(X_{f_p}(a_1)\subseteq\cdots\subseteq X_{f_p}(a_r)\)
is smoothly trivialized by diffeomorphisms of \(X\).
An empty level set counts as regular.
\end{proposition}

\begin{proof}
Compactness allows the parameter neighborhood and small disjoint
strips around the levels \(a_i\) to be chosen so that
\(\nabla_x F\) is nonzero throughout each strip. Otherwise there
would be sequences converging to a critical point of \(f_{p_0}\)
at one of those levels. Choose coordinates on a small ball in
\(P\) around \(p_0\) and a metric on \(X\). To lift a constant
coordinate vector \(v\) on the ball, start with \((v,0)\) on
\(P\times X\). Near each level modify its \(X\)-component to
\[
-\,\frac{D_pF(v)}{\|\nabla_xF\|^2}\nabla_xF.
\]
Use a cutoff in the strip equal to one at the level to extend
this correction by zero. The resulting vector is tangent to
each hypersurface \(F=a_i\). Integrating these lifts along
radial coordinate paths from \(p_0\) to \(p\) gives
diffeomorphisms from the starting fiber to the ending fiber.
Compactness of \(X\) gives existence along every such path,
and reversing the path gives the inverse. The flow preserves
each level and its chosen sublevel side, since a trajectory
cannot cross a hypersurface to which the field is tangent.
Smooth dependence on the path endpoint gives the required
local trivialization.
\end{proof}

\begin{example}[A regular family with monodromy]
Take \(P=X=S^1\) and \(F(\phi,\theta)=\cos(2\theta-\phi)\),
with both angles defined modulo \(2\pi\). Each level \(a\)
strictly between \(-1\) and \(1\) is regular for every
\(f_\phi\), and the sublevel is a pair of closed arcs.
In a local lift of the angle \(\phi\), translation
\(\theta\mapsto\theta+\phi/2\) identifies the sublevels with
those at \(\phi=0\). A loop in \(\phi\) changes this
translation by \(\pi\), interchanging the arcs. Thus the
local system of connected components has nontrivial monodromy,
although every sublevel in the family has the same homotopy
type. Degree-zero homology records the same permutation on
the two component generators.
\end{example}

Regularity controls a finite chosen flag. A moving critical value
can cross a threshold and change its sublevel topology even under
uniformly small perturbations. The interleaving estimate then
compares shifted thresholds. A stratified description of the
parameter and threshold space requires control of the critical
locus and the hypotheses for stratified transport, as in
Chapter~\ref{ch:exit}. A smooth parameter space by itself does
not supply such a stratification. The local trivialization and
the quantitative interleaving therefore give complementary
information about a parametrized filtration.
\par



\appendix

\chapter{Notation}
\label{app:notation}
\small\raggedright
\setlength{\parskip}{0pt}
\begin{description}[leftmargin=2em,style=nextline,itemsep=.30\baselineskip]
\item[$TM$, $T^*M$, $Df$] The tangent bundle, cotangent bundle,
  and derivative of a smooth map. See Chapters~\ref{ch:tangent} and~\ref{ch:morse}.
\item[$\operatorname{Emb}(M,N)$, $\operatorname{Imm}(M,N)$]
  Spaces of smooth embeddings and immersions.
  The topology is specified when using continuity or homotopy of families.
  See Chapters~\ref{ch:function-spaces}, \ref{ch:isotopy},
  and~\ref{ch:parametric-topology}.
\item[$J^r(M,N)$, $j^r_x f$] The bundle of order-$r$ jets of
  maps and the jet of $f$ at $x$. The map $j^r f$ is its holonomic
  section over $M$. See Chapter~\ref{ch:jets}.
\item[$f\pitchfork S$, $M\times_N Q$]
  Transversality to a submanifold and the fiber product. The ordinary
  smooth fiber product represents the transverse pullback. A superscript
  $\mathbf R$ or an explicit label distinguishes a derived pullback.
  See Chapters~\ref{ch:transversality} and~\ref{ch:derived-intersection}.
\item[$\det E$, $\operatorname{Or}(E)$, $w_1(E)$]
  The determinant line, orientation cover, and first Stiefel--Whitney
  class of a real vector bundle. See Chapter~\ref{ch:orientations}.
\item[$\deg f$, $e(E)$, $\chi(M)$]
  Degree, Euler class, and Euler characteristic. An integral degree
  requires the appropriate orientation data. See Chapters~\ref{ch:degree}
  and~\ref{ch:intersection}.
\item[$\operatorname{Th}(E)$, $MT\theta$] Thom space of a bundle
  and the Madsen--Tillmann spectrum for a tangential structure.
  See Chapters~\ref{ch:bordism} and~\ref{ch:bordism-cats}.
\item[{$\underline M$, $[M/G]$, $\mathcal E_{/B}$}]
  A manifold represented as a smooth sheaf, a smooth quotient stack,
  and a slice over a base. See Chapters~\ref{ch:smooth-stacks}
  and~\ref{ch:relative-maps}.
\item[{$[X,Y]_B$, $\operatorname{Map}_B(X,Y)$}]
  The internal function object and the mapping space in the slice.
  The former represents all parametrized maps over $B$.
  See Chapter~\ref{ch:relative-maps}.
\item[$M_{\le a}$, $H_k(M_{\le a})$]
  A sublevel set and its degree-$k$ homology, with coefficients
  specified in context. See Chapters~\ref{ch:morse} and~\ref{ch:persistence}.
\end{description}



\chapter{Subject index}
\label{app:index}
\small
\setlength{\parskip}{0pt}
\noindent References point to the chapters where each term is developed.
\smallskip

\begin{multicols}{2}\raggedright
\begin{description}[leftmargin=0pt,labelindent=0pt,itemsep=.08\baselineskip]
\item[Approximation] \ref{ch:function-spaces}.
\item[Atlas] Smooth, \ref{ch:manifolds}. Stack, \ref{ch:smooth-stacks}.
\item[Bordism] \ref{ch:bordism}, \ref{ch:bordism-cats}.
\item[Classifying space] \ref{ch:bundles}, \ref{ch:bordism-cats}.
\item[Collar] \ref{ch:tubular}, \ref{ch:bordism-cats}.
\item[Connected sum] \ref{ch:surfaces}.
\item[Degree] \ref{ch:degree}, \ref{ch:orientations}.
\item[Derived intersection] \ref{ch:derived-smooth}, \ref{ch:derived-intersection}.
\item[Descent] \ref{ch:smooth-stacks}, \ref{ch:relative-maps}, \ref{ch:orientations}.
\item[Discriminant] \ref{ch:discriminants}, \ref{ch:exit}.
\item[Embedding] \ref{ch:tangent}, \ref{ch:isotopy}, \ref{ch:parametric-topology}.
\item[Equivariance] \ref{ch:orbit-types}, \ref{ch:equivariant-transversality}.
\item[Euler class] \ref{ch:intersection}, \ref{ch:orientations}.
\item[Exit paths] \ref{ch:exit}.
\item[Factorization] \ref{ch:relative-maps}, \ref{ch:factorizations}.
\item[Fixed points] \ref{ch:orbit-types}, \ref{ch:equivariant-transversality}.
\item[Folds] \ref{ch:discriminants}.
\item[Handle attachment] \ref{ch:morse}, \ref{ch:surfaces}.
\item[Immersion] \ref{ch:tangent}, \ref{ch:parametric-topology}.
\item[Intersection number] \ref{ch:intersection}.
\item[Isotopy extension] \ref{ch:isotopy}.
\item[Jet] \ref{ch:jets}, \ref{ch:discriminants}.
\item[Mapping space] \ref{ch:function-spaces}, \ref{ch:relative-maps}.
\item[Monodromy] \ref{ch:smooth-stacks}, \ref{ch:orientations}.
\item[Morse theory] \ref{ch:morse}, \ref{ch:morsecx}.
\item[Orbit type] \ref{ch:orbit-types}.
\item[Orientation] \ref{ch:bundles}, \ref{ch:orientations}.
\item[Persistence] \ref{ch:persistence}.
\item[Pontryagin--Thom construction] \ref{ch:bordism}.
\item[Quotient stack] \ref{ch:smooth-stacks}, \ref{ch:orbit-types}.
\item[Regular value] \ref{ch:sard}, \ref{ch:discriminants}.
\item[Sard's theorem] \ref{ch:sard}, \ref{ch:transversality}.
\item[Slice theorem] \ref{ch:orbit-types}.
\item[Sublevel sets] \ref{ch:morse}, \ref{ch:persistence}.
\item[Tangent complex] \ref{ch:derived-smooth}, \ref{ch:derived-intersection}.
\item[Transversality] \ref{ch:transversality}, \ref{ch:equivariant-transversality}.
\item[Tubular neighborhood] \ref{ch:tubular}.
\item[Vector bundle] \ref{ch:bundles}, \ref{ch:orientations}.
\end{description}
\end{multicols}


\backmatter

\begin{thebibliography}{99}
\addcontentsline{toc}{chapter}{Bibliography}
\raggedright

\bibitem{ayala-francis-rozenblyum}
David Ayala, John Francis, and Nick Rozenblyum.
A stratified homotopy hypothesis.
\emph{Journal of the European Mathematical Society} 21 (2019), 1071--1178.
\href{https://doi.org/10.4171/JEMS/856}{DOI}.

\bibitem{baez-hoffnung}
John C. Baez and Alexander E. Hoffnung.
Convenient categories of smooth spaces.
\emph{Transactions of the American Mathematical Society} 363 (2011), 5789--5825.
\href{https://arxiv.org/abs/0807.1704}{arXiv:0807.1704}.

\bibitem{bott-tu}
Raoul Bott and Loring W. Tu.
\emph{Differential Forms in Algebraic Topology}.
Graduate Texts in Mathematics 82, Springer, 1982.
\href{https://doi.org/10.1007/978-1-4757-3951-0}{DOI}.

\bibitem{carchedi-dg}
David Carchedi. Derived manifolds as differential graded manifolds.
Preprint, 2023.
\href{https://arxiv.org/abs/2303.11140}{arXiv:2303.11140}.

\bibitem{carchedi-steffens}
David Carchedi and Pelle Steffens.
On the universal property of derived manifolds.
\href{https://arxiv.org/abs/1905.06195}{arXiv:1905.06195}.

\bibitem{chazal-persistence}
Fr\'ed\'eric Chazal, Vin de Silva, Marc Glisse, and Steve Oudot.
\emph{The Structure and Stability of Persistence Modules}.
SpringerBriefs in Mathematics. Springer, 2016.
\href{https://doi.org/10.1007/978-3-319-42545-0}{DOI}.

\bibitem{cohen-steiner-stability}
David Cohen-Steiner, Herbert Edelsbrunner, and John Harer.
Stability of persistence diagrams.
\emph{Discrete \& Computational Geometry} 37 (2007), 103--120.
\href{https://doi.org/10.1007/s00454-006-1276-5}{DOI}.

\bibitem{cjs-flow}
Ralph L. Cohen, John D. S. Jones, and Graeme B. Segal.
\emph{Morse Theory and Classifying Spaces}.
Preprint, 1995.
\href{https://abel.math.harvard.edu/~dafr/M392C-2018-MorseTheory/Readings/CJS.pdf}{Available manuscript}.

\bibitem{freed-bordism}
Daniel S. Freed. \emph{Bordism: Old and New}.
Lecture notes, 2013.
\href{https://abel.math.harvard.edu/~dafr/bordism.pdf}{Author's notes}.

\bibitem{gmtw}
S{\o}ren Galatius, Ib Madsen, Ulrike Tillmann, and Michael Weiss.
The homotopy type of the cobordism category.
\emph{Acta Mathematica} 202 (2009), 195--239.
\href{https://doi.org/10.1007/s11511-009-0036-9}{DOI}.

\bibitem{golubitsky-guillemin}
Martin Golubitsky and Victor Guillemin.
\emph{Stable Mappings and Their Singularities}.
Graduate Texts in Mathematics 14. Springer, 1973.
\href{https://doi.org/10.1007/978-1-4615-7904-5}{DOI}.

\bibitem{goodwillie-weiss}
Thomas G. Goodwillie and Michael Weiss.
Embeddings from the point of view of immersion theory II.
\emph{Geometry \& Topology} 3 (1999), 103--118.
\href{https://doi.org/10.2140/gt.1999.3.103}{DOI}.

\bibitem{hatcher-at}
Allen Hatcher. \emph{Algebraic Topology}.
Cambridge University Press, 2002.
\href{https://pi.math.cornell.edu/~hatcher/AT/at1-tc.html}{Author's copy}.

\bibitem{hatcher-vbkt}
Allen Hatcher. \emph{Vector Bundles and K-Theory}.
Version 2.2, 2017.
\href{https://pi.math.cornell.edu/~hatcher/VBKT/VBpage.html}{Author's notes}.

\bibitem{hirsch}
Morris W. Hirsch. \emph{Differential Topology}.
Graduate Texts in Mathematics 33. Springer, 1976.
\href{https://link.springer.com/book/10.1007/978-1-4684-9449-5}{Publisher's edition}.

\bibitem{hirsch-immersion}
Morris W. Hirsch. Immersions of manifolds.
\emph{Transactions of the American Mathematical Society} 93 (1959), 242--276.
\href{https://doi.org/10.1090/S0002-9947-1959-0119214-4}{DOI}.

\bibitem{kriegl-michor}
Andreas Kriegl and Peter W. Michor.
\emph{The Convenient Setting of Global Analysis}.
Mathematical Surveys and Monographs 53. American Mathematical Society, 1997.
\href{https://doi.org/10.1090/surv/053}{DOI}.

\bibitem{lee-smooth}
John M. Lee. \emph{Introduction to Smooth Manifolds}.
Second edition. Graduate Texts in Mathematics 218. Springer, 2013.
\href{https://doi.org/10.1007/978-1-4419-9982-5}{DOI}.

\bibitem{lurie-ha}
Jacob Lurie. \emph{Higher Algebra}. Manuscript, 2017.
\href{https://www.math.ias.edu/~lurie/papers/HA.pdf}{Author's manuscript}.

\bibitem{lurie-htt}
Jacob Lurie. \emph{Higher Topos Theory}.
Annals of Mathematics Studies 170. Princeton University Press, 2009.
\href{https://www.math.ias.edu/~lurie/papers/HTT.pdf}{Author's copy}.

\bibitem{lurie-cobordism}
Jacob Lurie. On the classification of topological field theories.
\emph{Current Developments in Mathematics 2008}, 129--280, 2009.
\href{https://arxiv.org/abs/0905.0465}{arXiv:0905.0465}.

\bibitem{matousek-borsuk}
Ji\v{r}\'i Matou\v{s}ek. \emph{Using the Borsuk--Ulam Theorem}.
Universitext. Springer, 2003.
\href{https://doi.org/10.1007/978-3-540-76649-0}{DOI}.

\bibitem{meinrenken-actions}
Eckhard Meinrenken. \emph{Group Actions on Manifolds}.
Lecture notes.
\href{https://www.math.utoronto.ca/~mein/teaching/LectureNotes/action.pdf}{Author's notes}.

\bibitem{milnor-differentiable}
John W. Milnor. \emph{Topology from the Differentiable Viewpoint}.
Princeton University Press, 1997 reissue.
\href{https://press.princeton.edu/books/paperback/9780691048338/topology-from-the-differentiable-viewpoint}{Publisher's edition}.

\bibitem{milnor-morse}
John W. Milnor. \emph{Morse Theory}.
Annals of Mathematics Studies 51. Princeton University Press, 1963.
\href{https://press.princeton.edu/books/paperback/9780691080086/morse-theory}{Publisher's edition}.

\bibitem{milnor-stasheff}
John W. Milnor and James D. Stasheff.
\emph{Characteristic Classes}.
Annals of Mathematics Studies 76. Princeton University Press, 1974.
\href{https://press.princeton.edu/books/paperback/9780691081229/characteristic-classes}{Publisher's edition}.

\bibitem{palais-embeddings}
Richard S. Palais. Local triviality of the restriction map for embeddings.
\emph{Commentarii Mathematici Helvetici} 34 (1960), 305--312.
\href{https://www.e-periodica.ch/cntmng?pid=com-001%3A1960%3A34%3A%3A24}{Journal scan}.

\bibitem{schwarz-morse}
Matthias Schwarz. \emph{Morse Homology}.
Progress in Mathematics 111. Birkh\"auser, 1993.
\href{https://doi.org/10.1007/978-3-0348-8577-5}{DOI}.

\bibitem{seidel-curves}
Paul Seidel. \emph{Differential Geometry}.
MIT OpenCourseWare 18.950, Fall 2008. Chapter 1, Lecture 9.
\href{https://ocw.mit.edu/courses/18-950-differential-geometry-fall-2008/f3ed7e720b456bf4defaf90a723b3edd_ch1_revised.pdf}{Lecture notes}.

\bibitem{spivak-derived}
David I. Spivak. Derived smooth manifolds.
\emph{Duke Mathematical Journal} 153 (2010), 55--128.
\href{https://arxiv.org/abs/0810.5174}{arXiv:0810.5174}.

\bibitem{wasserman-equivariant}
Arthur G. Wasserman. Equivariant differential topology.
\emph{Topology} 8 (1969), 127--150.
\href{https://doi.org/10.1016/0040-9383(69)90005-6}{DOI}.

\bibitem{wehrheim-morse}
Katrin Wehrheim. Smooth structures on Morse trajectory spaces,
featuring finite ends and associative gluing.
\emph{Geometry \& Topology Monographs} 18 (2012), 369--450.
\href{https://doi.org/10.2140/gtm.2012.18.369}{DOI}.

\bibitem{weiss-embeddings}
Michael Weiss. Embeddings from the point of view of immersion theory I.
\emph{Geometry \& Topology} 3 (1999), 67--101.
\href{https://doi.org/10.2140/gt.1999.3.67}{DOI}.

\end{thebibliography}


\end{document}

End of manuscript · Back to top ↑