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 2561–2660 of 3021. Source text is displayed literally.

2561\section{The inverse correspondence}2562A right Hilbert \(B\)-module is full if its inner products span a dense linear subspace of \(B\). An \(A\)-to-\(B\) correspondence that is full and whose left action identifies \(A\) with \(\mathcal K_B(E)\) is an imprimitivity bimodule. It implements strong Morita equivalence. For \eqref{eq:translation-compact}, let \(E=\mathcal H\) as a \(\mathcal K(\mathcal H)\)-to-\(\mathbf C\) correspondence. Its conjugate vector space \(\overline E\) has right action \(\bar e\cdot a=\overline{a^*e}\) and \(\mathcal K(\mathcal H)\)-valued inner product \(\langle\bar e,\bar f\rangle=\theta_{e,f}\). The left scalar action is ordinary multiplication on the conjugate space. The maps2563\[2564 \begin{aligned}2565 E\otimes_{\mathbf C}\overline E&\longrightarrow\mathcal K(\mathcal H),2566 & e\otimes\bar f&\longmapsto\theta_{e,f},\\2567 \overline E\otimes_{\mathcal K(\mathcal H)} E&\longrightarrow\mathbf C,2568 & \bar e\otimes f&\longmapsto\langle e,f\rangle2569 \end{aligned}2570\]2571preserve balances, inner products, and the left actions. For instance, in the first map the product of the adjoint of \(\theta_{e,f}\) with \(\theta_{e',f'}\) is \(\theta_{f\langle e,e'\rangle,f'}\), exactly the tensor inner product. Their images are dense by the rank-one formulas and the existence of a unit vector. The maps are therefore unitary. Composing twice with these correspondences recovers the starting module by the tensor associator. This yields the inverse equivalences on nondegenerate Hilbert-space representations and explains why a free transitive compact action and a point give the same stabilized invariants in Chapter~\ref{ch:equivariant-k}.2572 2573\section{Recovering a Hilbert-space representation}2574The Morita equivalence can also be seen inside an arbitrary nondegenerate representation of the kernel algebra. Let \(\rho:\mathcal K(\mathcal H)\to\mathcal L(\mathcal V)\) be such a representation and choose a unit vector \(e_0\in\mathcal H\). With \(p=\theta_{e_0,e_0}\) let \(M=\rho(p)\mathcal V\). Define2575\[2576  W:\mathcal H\odot M\longrightarrow\mathcal V,\qquad2577    W(e\otimes m)=\rho(\theta_{e,e_0})m .2578\]2579Using \(\theta_{e,e_0}^*\theta_{f,e_0}=\langle e,f\rangle p\) gives \(\langle W(e\otimes m),W(f\otimes n)\rangle=\langle e,f\rangle\langle m,n\rangle\). The range contains \(\rho(\theta_{e,f})v\) for every \(v\in\mathcal V\), since \(\theta_{e,f}=\theta_{e,e_0}\theta_{e_0,f}\) and \(\rho(\theta_{e_0,f})v\in M\). Nondegeneracy and density of rank-one operators therefore make \(W\) a unitary onto \(\mathcal V\). The rank-one identities also give \(W(ae\otimes m)=\rho(a)W(e\otimes m)\). Thus the multiplicity space \(M\) classifies \(\rho\) up to unitary equivalence. Choosing another unit vector changes the corner space by the unitary supplied by the corresponding rank-one operator. For the original action \(G\curvearrowright G\), the integrated operator includes both translation and multiplication. This result classifies that pair up to multiplicity, whereas a representation of translation alone has the many independent Peter--Weyl types.2580\par2581\chapter{Equivariant \texorpdfstring{\(K\)}{K}-theory and crossed products}2582\label{ch:equivariant-k}2583A crossed product retains the compatibility between functions on a space and motion by a group. Its \(K\)-theory expresses part of that compatibility in terms of stable classes. A projection \(p\) in a matrix algebra over \(C(X)\) determines a vector bundle by taking the range \(p(x)\) at every \(x\). The analogous definition for a general \(C^*\)-algebra uses the module \(pA^n\). For unital \(A\), \(K_0(A)\) is the Grothendieck group of stable equivalence classes of projections with addition by block sum. The group \(K_1(A)\) uses path components of unitaries after adjoining identity blocks. For a nonunital \(A\) one uses its unitization and takes the kernel of the map to the scalar summand. These constructions are stable under passage to compact operators on a full Hilbert module, by Morita invariance \cite{blackadar}.2584 2585\section{The compact comparison}2586Suppose \(G\) is a second-countable compact group and \(A\) a separable \(G\)-\(C^*\)-algebra with a continuous action. In the compact Green--Julg theorem the trivial algebra \(\mathbf C\) occurs in the first Kasparov variable. The natural isomorphisms are2587\[2588   KK^G_i(\mathbf C,A)\simeq K_i(A\rtimes G),\qquad i=0,1.2589\]2590One description descends a class to a correspondence from \(C^*(G)\) to \(A\rtimes G\) and takes its Kasparov product with the class of the projection for the trivial representation. In \(C^*(G)\) this is the continuous function \(g\mapsto 1\): convolution with itself remains \(1\) because Haar has mass one. The same compact theorem applies to the reduced product since compact groups are amenable. For \(A=C(X)\) with \(X\) compact metrizable and \(G\) compact Lie, \(KK^G_0(\mathbf C,C(X))\) identifies with the Grothendieck group \(K^0_G(X)\) of equivariant complex vector bundles, by the equivariant projection description. Odd topological \(K\)-theory is defined by suspension. The crossed product identification is an isomorphism of groups. Tensor multiplication of bundles is extra structure on the geometric side \cite[Sec.~11.7]{blackadar}\cite{segal}.2591 2592The hypotheses in this statement do work. For example, a general locally compact group has no Haar probability and the constant function need not furnish a projection in its group algebra. To see how the compact theorem behaves, it helps to compare the two transitive extremes, a point with stabilizer \(G\) and \(G\) itself with trivial stabilizer.2593 2594\section{Point, free orbit, and homogeneous orbit}2595At a point \(K^0_G(*)=R(G)\), the Grothendieck group of finite-dimensional continuous complex representations, with multiplication from tensor product. The full group \(C^*\)-algebra satisfies2596\[2597  C^*(G)\simeq\bigoplus_{\pi\in\widehat G}^{c_0}\operatorname{End}(V_\pi).2598\]2599Indeed, the integrated map uses \(\pi(F)=\int_G F(g)\pi(g)\,dg\). The full norm is the supremum of these matrix norms because every unitary representation decomposes into irreducibles. Peter--Weyl orthogonality supplies a continuous function for any single matrix block, with all other blocks zero. Finite sums of matrix coefficients are uniformly dense in continuous functions, hence dense in \(L^1(G)\). The estimate \(\|\pi(F)\|\leq\|F\|_1\) then gives precisely the \(c_0\)-completion. For finite sums of full matrix algebras the \(K_0\)-coordinate is matrix rank and the unitary groups are path connected. Passing to the limit over finite sets of blocks yields2600\[2601   K_0(C^*(G))=\bigoplus_{\pi\in\widehat G}\mathbf Z,2602   \qquad K_1(C^*(G))=0.2603\]2604The limit can also be seen from the norm topology. A projection in matrices over the \(c_0\)-sum has finitely many nonzero components since a nonzero projection has norm one. A unitary with scalar part \(1\) has tail uniformly near \(1\). On that tail one contracts it by the logarithm, while separately contracting each of the finitely many remaining finite-dimensional unitaries.2605 2606For the free orbit \(G\), an equivariant bundle \(E\) has canonical trivialization over the group given by \((g,v)\mapsto g v\) for \(v\in E_1\). Its inverse sends \(w\in E_g\) to \((g,g^{-1}w)\). The vector space \(E_1\) therefore determines its class and \(K^0_G(G)=\mathbf Z\). On the algebraic side, Chapter~\ref{ch:morita} identifies \(C(G)\rtimes G\) with the compacts, so this agrees with the compact-operator rank calculation.2607 2608For a closed subgroup \(H\) of a compact Lie group \(G\), evaluate a \(G\)-equivariant vector bundle on \(G/H\) at \(1H\). The fiber is an \(H\)-representation and an inverse constructs \(G\times_H V\) from \(V\). The defining relation is \((gh,v)\sim(g,hv)\); the repair of local coordinates on two sections of \(G\to G/H\) is consequently the action of the transition function on \(V\). For an input bundle \(E\), the formula \([g,v]\mapsto gv\) gives an equivariant linear bijection on fibers and a bundle isomorphism by local triviality. Since both operations preserve direct sum and tensor product, this proves \(K^0_G(G/H)\simeq R(H)\) as rings. On the algebra side, Green's imprimitivity theorem gives a Morita equivalence from \(C(G/H)\rtimes G\) to \(C^*(H)\). Thus either path computes the even group and \(K_1(C(G/H)\rtimes G)=0\) \cite[Thm.~4.22]{williams}.2609\[2610\begin{tikzcd}[column sep=2.4em, row sep=1.3em]2611 & R(H) \arrow[dr,"\mathrm{dual\ block\ ranks}"] &\\2612 K^0_G(G/H) \arrow[ur,"\mathrm{fiber}"] \arrow[dr,"\mathrm{Green\text{--}Julg}"]2613 && K_0(C^*(H))\\2614 & K_0(C(G/H)\rtimes G) \arrow[ur,"\mathrm{Morita}"] &2615\end{tikzcd}2616\]2617The block convention in this diagram can be checked using projections. Present an equivariant bundle \(E\) as the range of \(p(x)\) in a trivial bundle with finite-dimensional unitary fiber representation \(\rho\). The covariance equation is \(p(gx)=\rho(g)p(x)\rho(g)^*\). In matrices over the crossed product, the descended class comes from the kernel \(F(g,x)=p(x)\rho(g)\): it is the product of the bundle projection with the averaging projection for the diagonal action. Covariance and normalization of Haar give \(F*F=F=F^*\).2618 2619Under the homogeneous imprimitivity equivalence, the block indexed by an irreducible \(H\)-representation \(\eta\) acts on sections of \(E_\eta=G\times_H V_\eta\); it uses multiplication and the induced \(G\)-action \(U_g^\eta\). There \(F\) acts as \(p\) times the averaging operator for \(\rho(g)\otimes U_g^\eta\). These orthogonal projections commute by covariance. The range is the space of invariant sections of \(E\otimes E_\eta\), and averaging continuous sections has image in the finite-dimensional space determined by an \(H\)-invariant value at \(o\), since transitivity forces \(s(go)=g s(o)\). Continuous sections are dense, and the finite-dimensional image is closed. Thus the range in the Hilbert space has dimension2620\[2621  \dim(E_o\otimes V_\eta)^H2622  =\dim\operatorname{Hom}_H(V_\eta^*,E_o).2623\]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\[