Independent scholarship · Open accessThe monthly digest

Indexed text attachment

representation-theory-in-higher-categories.tex

Source for Representation Theory in Higher Categories

View paper · Download original file

Lines 2621–2720 of 3021. Source text is displayed literally.

2621  \dim(E_o\otimes V_\eta)^H2622  =\dim\operatorname{Hom}_H(V_\eta^*,E_o).2623\]2624Thus with integrated block convention \(\eta(F)=\int_H F(h)\eta(h)\,dh\), the rank of a descended bundle records the multiplicity of the dual type in its fiber. This specifies the diagram on actual bundles, and differences follow by additivity. Labelling the blocks by dual representations absorbs this indexing convention.2625 2626\section{The rotation action on the sphere}2627Take \(G=\mathrm{SO}(3)\) and \(H=\mathrm{SO}(2)\), the rotations about the north--south axis. The orbit is \(S^2\), and2628\[2629 K^0_{\mathrm{SO}(3)}(S^2)=\mathbf Z[z,z^{-1}],2630 \qquad2631 K_1(C(S^2)\rtimes\mathrm{SO}(3))=0,2632\]2633where \(z\) is the weight-one character of \(H\). Tensoring by a representation of \(G\) acts on the Laurent polynomial ring through its restriction to \(H\). This follows by taking the fiber of \(E\otimes (S^2\times W)\) at the north pole: the \(H\)-action there is the tensor product of the action on \(E_{1H}\) and the restricted action on \(W\).2634 2635Write \(w\) for the class of the complexified defining representation of \(G\). Its restriction is \(1+z+z^{-1}\). The degree-\(\ell\) irreducible \(V_\ell\) has weights \(-\ell,-\ell+1,\ldots,\ell\) on \(H\), each once, as in the rotation calculation of \cite[Sec.~10]{robert}. Restricted characters distinguish representations of \(G\): every rotation is conjugate to an element of \(H\), and characters distinguish compact representations. Multiplying the weight sum by \(1+z+z^{-1}\) gives the recurrence \(w[V_\ell]=[V_{\ell+1}]+[V_\ell]+[V_{\ell-1}]\) for \(\ell\geq1\). Thus \(w\) generates \(R(G)\), and its powers are independent since the Laurent polynomial \(w^n\) has highest exponent \(n\) with coefficient one. Over \(R(G)=\mathbf Z[w]\), the equivariant \(K\)-group of the sphere is free with basis \(1,z\). In fact \(z^{-1}=w-1-z\) and \(z^2=(w-1)z-1\) prove generation. If \(a(w)+b(w)z=0\), apply \(z\mapsto z^{-1}\) and subtract. Then \(b(w)(z-z^{-1})=0\) in an integral domain, so \(b(w)=a(w)=0\). This determines the action of every virtual rotation representation on the stable bundle classes.2636\par2637\chapter{Equivariant convolution and learning}2638\label{ch:learning}2639When a group acts on inputs, an equivariant layer takes the same action on its outputs. For rotation data, this might mean rotating a field of scalar intensities, or rotating both the base point and the components of a tangent vector field. The feature type specifies the representation on the output. With the input and output types fixed, Schur's lemma gives the full space of bounded linear layers. The descriptions by kernels in \cite{kondor-trivedi,cohen-geiger-weiler} are especially useful when a layer has finite spectral support or is defined on fields with an appropriate kernel class.2640 2641\section{The bounded linear classification}2642Let \(G\) be compact and let \(P,Q\) be unitary Hilbert representations. Write their Peter--Weyl decompositions as2643\[2644  P=\widehat{\bigoplus}_{\pi\in\widehat G} V_\pi\otimes M_\pi,2645  \qquad2646  Q=\widehat{\bigoplus}_{\pi\in\widehat G} V_\pi\otimes N_\pi .2647\]2648Each multiplicity space \(M_\pi\) or \(N_\pi\) carries the trivial action. The bounded equivariant linear maps \(T:P\to Q\) are exactly2649\begin{equation}2650  T=\widehat{\bigoplus}_{\pi\in\widehat G}(1_{V_\pi}\otimes A_\pi),2651  \quad A_\pi\in\mathcal L(M_\pi,N_\pi),2652  \quad \|T\|=\sup_\pi\|A_\pi\|<\infty .2653  \label{eq:learning-blocks}2654\end{equation}2655Here \(\mathcal L(M_\pi,N_\pi)\) denotes bounded linear maps of Hilbert spaces.2656 2657For the proof, the operator \(d_\pi\int_G\overline{\chi_\pi(g)}U_g\,dg\), with \(d_\pi=\dim V_\pi\), is the orthogonal projection onto the \(\pi\)-isotypic part: Schur's lemma makes the integral scalar on each irreducible, and character orthogonality determines its trace. A bounded intertwiner commutes with these projections on the two spaces. To compute the \(\pi\)-block, for \(m\in M_\pi\) and \(n\in N_\pi\) apply the block to \(v\otimes m\) and pair its \(N_\pi\)-coordinate with \(n\). The resulting endomorphism of \(V_\pi\) commutes with \(\pi\), so it is scalar. Equivalently, in an orthonormal basis of \(V_\pi\), the off-diagonal operator entries \(M_\pi\to N_\pi\) have every scalar matrix element zero and all diagonal entries agree. Their common value is \(A_\pi\). The block norm is \(\|A_\pi\|\). Conversely, a bounded family of those norms gives a bounded operator on the dense algebraic direct sum by summing the squared norm estimates, and its completion intertwines the two actions. Testing on a single isotypic part gives equality of the norms.2658 2659For finitely many types and finite multiplicities, the number of complex parameters is2660\[2661  \sum_{\pi\in\widehat G}(\dim M_\pi)(\dim N_\pi).2662\]2663For scalar \(L^2(S^2)\), \(M_\ell\) is one-dimensional for every integer \(\ell\geq0\), because the \(H=\mathrm{SO}(2)\)-fixed subspace of each \(V_\ell\) has dimension one. With \(c\) input channels and \(d\) output channels each spectral block is an arbitrary \(d\)-by-\(c\) matrix, applied in the same way to all \(2\ell+1\) coordinates of \(V_\ell\).2664 2665\section{Regular convolution and its scope}2666On the left regular representation \(L^2(G)\), a right convolution operator \(T_k f=f*k\) for \(k\in L^1(G)\) commutes with left translation by change of variables. With Haar probability, use2667\[2668 (f*k)(x)=\int_Gf(y)k(y^{-1}x)\,dy,\qquad2669 \widehat f(\pi)=\int_G f(g)\pi(g)^*\,dg.2670\]2671The Fourier multiplier is \(\widehat{f*k}(\pi)=\widehat k(\pi)\widehat f(\pi)\): substitute \(x=yz\) in the double integral and multiply \(\pi(z)^*\pi(y)^*\) in that order. Left translation by \(a\) acts on \(\widehat f(\pi)\) by right multiplication by \(\pi(a)^*\), so the position of \(\widehat k(\pi)\) agrees with \eqref{eq:learning-blocks}. Peter--Weyl orthogonality gives an actual continuous kernel for every finite-support family of matrix multipliers.2672 2673For \(k\in L^1(G)\) the convolution norm is bounded by \(\|k\|_1\). Indeed, rewriting the convolution as \(\int_G k(z)f(xz^{-1})\,dz\) expresses it as an integral of right translations, all \(L^2\)-isometries. Approximation of \(k\) in \(L^1\) by finite Peter--Weyl sums shows that every such \(T_k\) is compact on \(L^2(G)\). For an infinite compact group the identity is bounded and equivariant and has infinite-dimensional range, so it cannot have an \(L^1\) convolution kernel. It is convolution by the Dirac measure at the identity. Formula \eqref{eq:learning-blocks} applies to all bounded intertwiners independently of such kernel realizations.2674 2675\section{Pointwise layers and feature types}2676Let \(X=G/H\), and let \(V,W\) be finite-dimensional \(H\)-representations with actions \(\rho_V,\rho_W\). A pointwise activation sends the fiber coordinate \(v\) of \([g,v]\) to \([g,\sigma(v)]\). The relation \([gh,v]=[g,\rho_V(h)v]\) makes this rule independent of the representative exactly when2677\[2678   \sigma(\rho_V(h)v)=\rho_W(h)\sigma(v)\qquad(h\in H).2679\]2680The same formula then implies \(G\)-equivariance under left translation of \(g\). An arbitrary scalar activation works for scalar fields, since the stabilizer acts trivially on the fiber. A coordinatewise activation on a vector fiber need not satisfy the relation. For a unitary fiber a map \(v\mapsto a(\|v\|)v\) satisfies it. If the activation is Lipschitz, it induces a Lipschitz map on \(L^2\)-sections over the finite invariant base measure: the pointwise estimate integrates, and \(\|\sigma(v)\|\leq\|\sigma(0)\|+L\|v\|\) ensures square integrability. A nonlinear layer need not preserve the spectral truncation of its input.2681 2682\section{A rotation-invariant quadratic feature}2683Give \(S^2\) its rotation-invariant probability measure. For real scalar inputs define2684\[2685  (Tf)(x)=5\int_{S^2}P_2(x\cdot y)f(y)\,d\mu(y),2686  \qquad P_2(t)=\tfrac12(3t^2-1).2687\]2688The kernel uses only the dot product, so it is unchanged by simultaneous rotation of \(x\) and \(y\). For fixed \(y\), it is a traceless quadratic in \(x\) on the sphere. To determine its action, take \(f_A(y)=y^{\mathsf T}Ay\) for a real symmetric traceless matrix \(A\). The moments are2689\[2690 \int y_i y_j\,d\mu=\tfrac13\delta_{ij},\qquad2691 \int y_i y_j y_k y_l\,d\mu2692  =\tfrac1{15}(\delta_{ij}\delta_{kl}2693                 +\delta_{ik}\delta_{jl}+\delta_{il}\delta_{jk}).2694\]2695For the fourth moment, the polynomial in \(v\) obtained by integrating \((v\cdot y)^4\) is rotation invariant and homogeneous. Its value at a unit vector is \(1/5\), from the sphere's area measure in the height coordinate \(t\in[-1,1]\). Comparing coefficients, or polarizing \(\|v\|^4/5\), gives the tensor written above. The second moment is obtained from \(\|y\|^2=1\) and rotation symmetry. Consequently2696\[2697  \int f_A\,d\mu=0,\qquad2698  \int (x\cdot y)^2 f_A(y)\,d\mu(y)=\tfrac{2}{15}x^{\mathsf T}Ax.2699\]2700Thus \(Tf_A=f_A\). The kernel is symmetric and the range lies in the traceless quadratics, so \(T\) is their orthogonal projection.2701 2702Apply a scalar square after the projection and then pool by integration. The projection is finite rank with continuous outputs, so the square is defined and integrable for every \(L^2\) input. It is an equivariant pointwise layer followed by an invariant pooling layer.2703\[2704\begin{tikzcd}[column sep=large]2705 L^2(S^2;\mathbf R) \arrow[r,"T"]2706 & V_2^{\mathbf R}\arrow[r,"u\mapsto u^2"]2707 & C(S^2;\mathbf R)\arrow[r,"\int_{S^2}(-)\,d\mu"]2708 & \mathbf R2709\end{tikzcd}2710\]2711Its value on \(f_A\) is \(2\operatorname{tr}(A^2)/15\), from the fourth moment. For example, \(A=\operatorname{diag}(1,-1,0)\) gives \(4/15\), and a constant input gives zero. A channel version first projects each channel and then applies any matrix of linear coefficients to those projected channels, as allowed by \eqref{eq:learning-blocks}; each squared pooled output is invariant for every choice of those coefficients.2712 2713\section{Kernels for associated bundles}2714Take the spaces of sections of \(E=G\times_H V\) and \(F=G\times_H W\) with invariant measures and Hermitian metrics. Suppose an intertwiner \(T\) has finite spectral support. For each retained \(G\)-type there are only finitely many copies by compact Frobenius reciprocity, since \(V\) and \(W\) are finite dimensional. The retained subspaces consist of continuous sections, given by matrix coefficients as in Chapter~\ref{ch:synthesis}. For an orthonormal basis \(s_j\) of the retained input, define \(K(x,y):E_y\to F_x\) by2715\[2716   K(x,y)v=\sum_j(Ts_j)(x)\langle s_j(y),v\rangle.2717\]2718Use the Hermitian product linear in the second argument. The orthonormal expansion gives \((Ts)(x)=\int K(x,y)s(y)\,d\mu(y)\) and makes the formula independent of the chosen basis. Replacing \(s_j\) by the translated basis and using \(T U_a=U_a T\) shows \(K(ax,ay)=a K(x,y)a^{-1}\), with the two actions interpreted on the appropriate fibers. Conversely, a continuous kernel satisfying this covariance defines a bounded equivariant operator by change of variables and the uniform bound on the kernel over the compact base. For each output point one can transport the kernel at \(o\) to that point, and the result is independent of the transporting element precisely because the kernel at \(o\) intertwines the \(H\)-actions on the fibers over \(y\) and \(hy\). This describes how global rotation equivariance and local fiber coordinates constrain the same layer.2719\par2720\chapter{An orbit, its stabilizer, and its invariants}