Independent scholarship · Open accessThe monthly digest

Full manuscript · browser edition

Representation Theory of Parametrized Neural Models on Spaces

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

Contents · 19 sections

Representation Theory of Parametrized
Neural Models on Spaces

Statistical families, equivariant layers, and homotopy-coherent models

November 2026

Abstract. A parametrized network over a space is a family of evaluation maps. We represent its realizations and reparametrizations by a slice over an internal mapping object, prove base change and descent, and extend the construction to stochastic outputs for a strong probability monad. On smooth charts this gives deterministic manifold networks and conditional statistical models, together with pullback Fisher geometry. For compact symmetries we classify bounded layers linear in a moving parameter by multiplicity-space intertwiners. The classification exhibits frequency shifts that cannot occur for a fixed equivariant layer. For parameters with local symmetries we retain stabilizers and transition data in a quotient stack. A neural decoder on a twisted real line illustrates why global probability laws can lack a global smooth weight. Mapping spectra supply a stable version of equivariance with higher intertwiners.

Keywords. Parametrized networks, representation theory, equivariance, Fisher information, descent, quotient stacks.

From statistical families to representations

A supervised model commonly starts with a parameter manifold PP, a feature space XX, and a map f(p,x)f(p,x) valued in labels, predictions, or probability distributions. A convolutional network has additional structure. Features may live in bundles on an observation space, and a symmetry can transport both features and weights. Two local weights can also present the same model via a change of frame. These cases require tracking evaluation, transport, and equivalences of presentations.

Representation Theory in Higher Categories develops ordinary and coherent group actions, induced feature spaces, and equivariant convolutions [2, Chs. 17, 20, 27–28]. We start with its idea of encoding actions by functors and relate evaluation, symmetry, and descent. Evaluation records what a parameter computes. Representation theory controls how evaluations transform. Descent determines when local evaluations and parameter presentations assemble across a base. Theorem 2 gives the general comparison. Theorem 7 computes the moving-parameter part in a compact linear setting.

Parametrized composition and supervised learning have categorical formulations through learners, lenses, and reverse differentiation [6, 5]. Markov categories organize channels and sufficient statistics [7], while equivariant convolution on homogeneous spaces is described by induced representations [4]. Here we apply representation constraints to the evaluation families themselves. This permits the same typing of a manifold layer, a statistical decoder, and a locally presented system of layers.

Relative parametrized models

Let E\mathcal E be an ∞\infty-topos and BB an object of E\mathcal E. Possible choices are the ∞\infty-category S\mathcal S of spaces and the ∞\infty-topos of sheaves of spaces on a smooth site. Work first in E/B\mathcal E_{/B}. For X,YX,Y over BB, write [X,Y]B[X,Y]_B for their exponential in that slice. Its evaluation map is [X,Y]B×BX→Y[X,Y]_B\times_B X\to Y over BB. A choice of site specifies which families of sections are admitted. Smooth manifolds will be embedded by their smooth functors of points.

Definition 1 (A parametrized layer). A layer from XX to YY over BB consists of P→BP\to B and an evaluation f:P×BX→Yf:P\times_B X\to Y. A map to (Q,g)(Q,g) is a morphism u:P→Qu:P\to Q over BB with a specified homotopy f≃g(u×BidX)f\simeq g(u\times_B\mathop{\mathrm{id}}_X) over BB. Higher morphisms are the compatibilities in the mapping space of the resulting slice.

Write rf:P→[X,Y]Br_f:P\to[X,Y]_B for the transpose of ff. Given a second layer (Q,g):Y→Z(Q,g):Y\to Z, the series composite has parameter P×BQP\times_B Q and evaluation (p,q,x)↦g(q,f(p,x))(p,q,x)\mapsto g(q,f(p,x)). Here and below parametric formulas describe generalized points in the slice. A shared parameter PP can be imposed by reparametrization along the diagonal P→P×BPP\to P\times_B P. More general tying is expressed by a map P→∏i,BPiP\to\prod_{i,B}P_i into the product of layer parameters.

Theorem 2 (Relative representation and descent). The ∞\infty-category of layers from XX to YY is E/[X,Y]B\mathcal E_{/[X,Y]_B}. On these categories, series composition is induced by internal composition of exponentials and respects reparametrizations. The construction commutes with pullback in BB. For an effective epimorphism U→BU\to B with Čech nerve U∙U_\bullet, the global layer category is equivalent to lim⁡[n]∈ΔE/[Xn,Yn]Un,Xn=X×BUn,Yn=Y×BUn. \lim_{[n]\in\boldsymbol\Delta} \mathcal E_{/[X_n,Y_n]_{U_n}}, \qquad X_n=X\times_B U_n,\quad Y_n=Y\times_B U_n .

Proof. Currying yields the equivalence of mapping spaces MapB(P×BX,Y)≃MapB(P,[X,Y]B).\mathop{\mathrm{Map}}_B(P\times_B X,Y)\simeq\mathop{\mathrm{Map}}_B(P,[X,Y]_B). It is natural in PP. The two definitions of a morphism of layers therefore give the same homotopy fiber of a map of mapping spaces, including its higher simplices. The relative double slice (E/B)/[X,Y]B(\mathcal E_{/B})_{/[X,Y]_B} is E/[X,Y]B\mathcal E_{/[X,Y]_B}.

Evaluation in series transposes to a morphism [X,Y]B×B[Y,Z]B⟶[X,Z]B.[X,Y]_B\times_B[Y,Z]_B\longrightarrow [X,Z]_B. It sends the classifying maps of the two layers to the classifying map of their composite. The identity is classified by the transpose of idX\mathop{\mathrm{id}}_X with parameter BB. The associativity and unit compatibilities follow by uncurrying the iterated compositions. Both ways to combine a string yield evaluation in the same order on the product of its parameters. All product comparisons come from their limit universal properties.

Let a:C→Ba:C\to B. If T→CT\to C, the left adjoint a!a_! to a∗a^* is composition of structure maps. Using the pullback identification a!(T×Ca∗X)≃(a!T)×BXa_!(T\times_C a^*X)\simeq (a_!T)\times_B X gives MapC(T,a∗[X,Y]B)≃MapB(a!T,[X,Y]B)≃MapB(a!T×BX,Y)≃MapC(T×Ca∗X,a∗Y).\begin{aligned} \mathop{\mathrm{Map}}_C(T,a^*[X,Y]_B) &\simeq \mathop{\mathrm{Map}}_B(a_!T,[X,Y]_B)\\ &\simeq \mathop{\mathrm{Map}}_B(a_!T\times_B X,Y)\\ &\simeq \mathop{\mathrm{Map}}_C(T\times_C a^*X,a^*Y). \end{aligned} By Yoneda these equivalences identify a∗[X,Y]Ba^*[X,Y]_B with [a∗X,a∗Y]C[a^*X,a^*Y]_C, preserving evaluation. Pullback also preserves the products that define series composition.

Finally, the effective epimorphism is the colimit of its Čech nerve. Colimits in an ∞\infty-topos satisfy descent for slices [11, §6.1.3]. Hence E/B\mathcal E_{/B} is the limit of the categories E/Un\mathcal E_{/U_n}. Taking the slice at the compatible exponential commutes with this limit, since a slice is the category of arrows with fixed target. Base change identifies the local targets with the ones in (1). ◻

\includegraphics\includegraphics{hcs-diagram-8d72cd9d7b17cc9875cf812d5d5ecd14ee5e2c9be311f650ef39013ffdf438ce.png}

A reparametrization includes a commuting comparison of evaluations. In an ∞\infty-topos the comparison is a specified homotopy.

A point of [X,Y]B[X,Y]_B over bb is one evaluation on the fibers. The classifying map retains a distinction between a space of these evaluations and a space of weights mapping into it. An architecture selects a classifying map, for instance by composing affine maps, an equivariant convolution, and a nonlinearity. The ambient exponential also includes maps that the selected architecture cannot realize.

Transport and actions

In spaces, straightening identifies S/B\mathcal S_{/B} with Fun(B,S)\mathop{\mathrm{Fun}}(B,\mathcal S), where the base BB is regarded as an ∞\infty-groupoid. Transport acts by equivalences and the slice exponential is computed fiberwise [11, §3.2]. For a path γ:b→c\gamma:b\to c with transports tX,tYt_X,t_Y, exponentials transport by h↦tYhtX−1h\mapsto t_Y h t_X^{-1}. A classifying map is consequently a coherent family of maps Pb→[Xb,Yb]P_b\to [X_b,Y_b] commuting with transport. For an ordinary action of an abstract group GG on P,X,YP,X,Y, the resulting comparison reads f(gp,gx)=gf(p,x). f(gp,gx)=g f(p,x). It is a condition on the entire family. At a fixed pp, it forces equivariance under the stabilizer GpG_p. In a smooth site one can instead use the classifying stack of a Lie group. Sections of that stack are torsors with smooth local trivializations, and (2) then holds for smooth test families. The homotopy type of a smooth base retains path transport and may identify smoothly distinct data.

Realization on a quotient

Suppose a group object GG acts on PP and a realization r:P→Mr:P\to M is coherently invariant with trivial action on MM. Here MM can itself be a space of functions or probability laws. The action groupoid has simplicial levels P×GnP\times G^n. Its colimit is P//GP//G, and the colimit universal property gives Map(P//G,M)≃Tot Map(P×G∙,M).\mathop{\mathrm{Map}}(P//G,M)\simeq \mathop{\mathrm{Tot}}\,\mathop{\mathrm{Map}}(P\times G^\bullet,M). Thus rr factors through P//GP//G exactly when equipped with descent, whose first comparison identifies r(p)r(p) and r(gp)r(gp). Equivalences between local parameter presentations are part of a family in the quotient. Over S\mathcal S, the fiber of P//G→BGP//G\to BG over the base point is the parameter space PP, with transport the original action. For smooth PP and Lie GG, the same construction in sheaves gives the smooth quotient stack. The example in Section 6 distinguishes the quotient presentations from their common observable laws.

Stochastic evaluation

In a cartesian closed category, a strong monad TT has a natural strength stA,Y:A×TY→T(A×Y)\mathrm{st}_{A,Y}:A\times T Y\to T(A\times Y) compatible with products, unit η\eta, and multiplication μ\mu. When TT is a probability monad, the layer P×X→TYP\times X\to T Y assigns a probability-valued prediction. Currying represents it by P→[X,TY]P\to[X,T Y]. For f:P×X→TYf:P\times X\to T Y and g:Q×Y→TZg:Q\times Y\to T Z, their composite with parameter P×QP\times Q is P×Q×X⟶Q×TY→ st T(Q×Y)→ Tg T2Z→ μZ TZ,\begin{aligned} P\times Q\times X &\longrightarrow Q\times TY\\ &\xrightarrow{\ \mathrm{st}\ } T(Q\times Y) \xrightarrow{\ T g\ } T^2 Z \xrightarrow{\ \mu_Z\ } TZ , \end{aligned} with first map (p,q,x)↦(q,f(p,x))(p,q,x)\mapsto(q,f(p,x)). For a formal check, fix a context AA and form the category whose arrows X→YX\to Y are maps A×X→YA\times X\to Y. It is the co-Kleisli category of the comonad A×(−)A\times(-) with diagonal and projection. Strength is a mixed distributive law A×T(−)→T(A×(−))A\times T(-)\to T(A\times(-)) for this comonad and the monad TT. Its compatibility with the diagonal and projection is product coherence and naturality in AA. Monad unit and multiplication compatibility are the strength axioms. Thus TT acts as a monad on the category with context AA. Kleisli composition for that monad extends a map h:A×Y→TZh:A\times Y\to T Z using strength followed by ThT h and μZ\mu_Z. Taking A=P×QA=P\times Q and restricting the layers along its two projections gives the displayed formula. Context maps preserve the construction by naturality of strength. A compatible system of monads on slices gives the same construction over BB when their strengths and monad operations commute with base change.

For instance, the finitely supported distribution monad on sets gives finite channels by restricting the feature and output sets to be finite. If p,qp,q are fixed, the composite has law kp,q(z∣x)=∑ygq(z∣y)fp(y∣x).k_{p,q}(z\mid x)=\sum_y g_q(z\mid y) f_p(y\mid x). Dirac laws embed deterministic evaluations. Repeated sampling can share an unknown parameter by pulling a product family back along its diagonal. For a coin of bias 1/51/5 or 4/54/5 with equal prior probability, shared-parameter sampling gives probability 17/5017/50 of two successes. Independently drawing the parameter for each observation gives 1/41/4. The single-observation law therefore needs to be supplemented by a parameter sharing convention.

For smooth conditional models, the probability-valued evaluator can be defined on a represented family of probability laws. The next section uses finite outcome laws with positive probabilities, where the open simplex is a manifold and the Fisher tensor can be computed in ordinary coordinates.

Statistical and manifold models

A parametrized smooth map F:P×X→MF:P\times X\to M is a manifold-valued prediction, with a family of maps Fp:X→MF_p:X\to M. A deterministic network may also retain intermediate feature bundles by factoring FF through smooth maps of total spaces. With an ordinary parameter manifold and no prescribed equivalences of parameters, these are ordinary smooth families. For stochastic predictions on a finite set YY, a smooth readout is a map q:P×X⟶Δ∘(Y),(p,x)⟼(qi(p,x))i∈Y,q:P\times X\longrightarrow\Delta^\circ(Y), \qquad (p,x)\longmapsto(q_i(p,x))_{i\in Y}, where Δ∘(Y)\Delta^\circ(Y) is the simplex of strictly positive probability vectors. Evaluation at pp gives the conditional law of the outcome given xx. Taking X=∗X=* recovers the usual smooth finite statistical model. A deterministic prediction into a standard Borel space MM has the associated kernel x↦δF(p,x)x\mapsto\delta_{F(p,x)}. Dirac measures on a finite set with at least two elements lie on the boundary of the simplex. Strict positivity permits differentiation of the finite log likelihood.

A stochastic decoder can be composed with a succeeding stochastic layer by summing over the intermediate outcome. If Kji(p,x)K_{ji}(p,x) is the conditional probability of output jj given outcome ii, the new readout is rj=∑iKjiqir_j=\sum_i K_{ji}q_i. Associativity is the associativity of finite summation, and deterministic relabelings are the special case of Dirac kernels [7]. Smooth charts of PP and XX present these models locally: if the predictions agree on the overlaps after coordinate change, the coordinate functions qiq_i glue to a unique smooth readout. For a manifold target MM, the same assertion uses transition maps of MM on overlapping target charts.

The Fisher form and stochastic postprocessing

At a point qq of the positive simplex, write u,vu,v for tangent vectors with ∑iui=∑ivi=0\sum_i u_i=\sum_i v_i=0. The Fisher metric is gq(u,v)=∑i∈Yuiviqi.g_q(u,v)=\sum_{i\in Y}\frac{u_iv_i}{q_i}. At a fixed input xx, the score in direction v∈TpPv\in T_pP is sv(i)=dplog⁡qi(p,x)(v)s_v(i)=d_p\log q_i(p,x)(v). Its mean is zero by differentiation of ∑iqi=1\sum_i q_i=1. Thus the covariance of the scores in directions v,wv,w is Ip,x(v,w)=∑i∈Ydpqi(v) dpqi(w)qi(p,x).I_{p,x}(v,w)= \sum_{i\in Y}\frac{d_pq_i(v)\,d_pq_i(w)}{q_i(p,x)}. Equivalently, Ix=qx∗gI_x=q_x^*g, where qx(p)=q(p,x)q_x(p)=q(p,x). For a parameter-independent Borel probability measure μ\mu on XX, the joint experiment observes both the input and the label with law dμ(x) qi(p,x)d\mu(x)\,q_i(p,x). Its Fisher form is Ipμ=∫XIp,x dμ(x)I_p^\mu=\int_X I_{p,x}\,d\mu(x), assuming the scores are square integrable and the experiment is differentiable under the integral. These conditions hold for a smooth positive readout when XX is compact, locally in the parameter.

For a smooth reparametrization r:P′→Pr:P'\to P, the conditional Fisher form of q∘(r×id⁡X)q\circ(r\times\operatorname{id}_X) is r∗Ixr^*I_x by the chain rule. This holds for IμI^\mu as well. The identity gives the change-of-coordinates rule I′=Dr⊤I DrI'=Dr^\top I\,Dr on charts. On input charts the same joint form is computed by transporting μ\mu with the coordinates. Thus the tensor descends with the model and the chosen design measure. Its pointwise null space at xx is ker⁡dpqx\ker d_pq_x, because all qi>0q_i>0. The joint form vanishes on vv precisely when dpqi(v)d_pq_i(v) vanishes μ\mu-almost everywhere for every ii.

Proposition 3 (Finite Fisher contraction). A stochastic postprocessor KK, fixed with respect to pp, weakly decreases both the conditional Fisher form and its design average.

Proof. Omit output values jj that are unreachable under KK. For rj=∑iKjiqi>0r_j=\sum_i K_{ji}q_i>0, the score of rjr_j in direction vv is dplog⁡rj(v)=∑iKjiqirj sv(i).d_p\log r_j(v)=\sum_i\frac{K_{ji}q_i}{r_j}\,s_v(i). The weights are the conditional probabilities of ii given jj. Jensen’s inequality bounds the square by the weighted sum of squares. Multiplying by rjr_j and summing over jj yields Ip,xr(v,v)≤Ip,xq(v,v)I^{r}_{p,x}(v,v)\le I^{q}_{p,x}(v,v). Integrating yields the claim for IμI^\mu. At a fixed xx, equality in direction vv holds exactly when svs_v is constant on the intermediate outcomes that contribute to each reachable output. ◻

The parameter-independence hypothesis keeps the postprocessor from introducing its own score. Formula (3) is the finite instance of the parametrized-measure Fisher tensor [3].

Example 4 (Gaussian observations of a manifold family). Let m:P×X→Rkm:P\times X\to\mathbb R^k be a smooth readout and let the conditional outcome have distribution N(m(p,x),Σ(x))N(m(p,x),\Sigma(x)), where Σ(x)\Sigma(x) is positive definite and independent of pp. For an observed outcome zz, differentiation of the Gaussian log density at fixed xx gives the score dpm(v)⊤Σ(x)−1(z−m(p,x)).d_pm(v)^\top\Sigma(x)^{-1}\bigl(z-m(p,x)\bigr). The expectation of the product of two scores uses the covariance Σ(x)\Sigma(x) and gives Ip,x(v,w)=dpm(v)⊤Σ(x)−1dpm(w).I_{p,x}(v,w)=d_pm(v)^\top\Sigma(x)^{-1}d_pm(w). Thus an immersed mean manifold at fixed xx has a Riemannian metric by pullback of the Euclidean metric weighted by Σ(x)−1\Sigma(x)^{-1}. If the mean map has a kernel in its differential, the pullback has the same null directions. An invertible affine coordinate change of the output with linear part AA replaces Σ\Sigma by AΣA⊤A\Sigma A^\top and dpmd_pm by A dpmA\,d_pm, leaving the form invariant. For a general smooth output coordinate change, one transports the entire conditional law. The transformed density contains a Jacobian factor independent of pp, so its score is the transport of the original score even when the law exits the Gaussian family.

What the Fisher form forgets

Consider a two-feature decoder on S1S^1, with angle tt understood modulo 2π2\pi and parameters p=(a,b,c)∈R3p=(a,b,c)\in\mathbb R^3. Its logit is Lp(t)=acos⁡t+bcsin⁡tL_p(t)=a\cos t+bc\sin t and its Bernoulli success probability is sp(t)=(1+e−Lp(t))−1s_p(t)=(1+e^{-L_p(t)})^{-1}. For a tangent vector v=(va,vb,vc)v=(v_a,v_b,v_c), the score at label y∈{0,1}y\in\{0,1\} is (y−sp(t)) dLp(t)(v)(y-s_p(t))\,dL_p(t)(v). With any full-support design probability measure μ\mu on the circle, the joint Fisher form is Ipμ(v,v)=∫S1sp(t)(1−sp(t)) (δv(t))2 dμ(t),δv(t)=vacos⁡t+(cvb+bvc)sin⁡t.\begin{aligned} I^\mu_p(v,v) &=\int_{S^1}s_p(t)(1-s_p(t))\,(\delta_v(t))^2\,d\mu(t),\\ \delta_v(t)&=v_a\cos t+(c v_b+b v_c)\sin t. \end{aligned} Positivity of the Bernoulli variance and continuity in tt imply that the integral vanishes exactly when δv\delta_v vanishes identically. The independent harmonics give the two equations va=0v_a=0 and cvb+bvc=0c v_b+b v_c=0. The rank is two away from b=c=0b=c=0 and one along that locus. In particular the Fisher tensor is the pullback of a positive definite form on the two-dimensional logistic family with coordinates (a,d)(a,d), along the map (a,b,c)↦(a,bc)(a,b,c)\mapsto(a,bc). The strict monotonicity of the logistic function and the independent harmonics also show that this map labels exactly the distinct output laws. This is an elementary instance of singular parametrization [13].

The multiplicative group R×\mathbb R^\times acts by λ⋅(a,b,c)=(a,λb,λ−1c)\lambda\cdot(a,b,c)=(a,\lambda b,\lambda^{-1}c), preserving the readout. The action groupoid has stabilizer R×\mathbb R^\times at b=c=0b=c=0 and trivial stabilizer elsewhere. Its three distinct orbits over a given prediction with d=0d=0 correspond to the origin and the two punctured coordinate axes in the (b,c)(b,c)-plane. Thus the quotient groupoid of parameter symmetries contains information beyond the set of distinct output laws. Even the infinitesimal null space can include directions that change the output at higher order: along (a,ε,ε)(a,\varepsilon,\varepsilon) through (a,0,0)(a,0,0), d=ε2d=\varepsilon^2 changes the prediction to second order. Consequently equality of output laws, Fisher-null directions, and parameter equivalence answer three different questions.

A manifold decoder as a statistical model

Let ZZ be a latent manifold with fixed probability ν\nu and d:P×Z→Md:P\times Z\to M a smooth decoder. In the absence of observation noise the law at pp is (dp)∗ν(d_p)_*\nu. A noise channel N:M→YN:M\to Y then gives the observed law Pp(A)=∫ZN(dp(z),A) dν(z)\mathbb P_p(A)=\int_Z N(d_p(z),A)\,d\nu(z) for measurable A⊆YA\subseteq Y. This is ordinary composition of Markov kernels on these measurable spaces. The manifold evaluation dd can be represented by the smooth slice construction of Section 2. The pushforward additionally uses the chosen law ν\nu and noise NN. If a:Z→Za:Z\to Z preserves ν\nu, then replacing dpd_p by dp∘ad_p\circ a leaves Pp\mathbb P_p unchanged, by the change-of-variables identity for pushforwards. A law that observes zz and the output together can distinguish the two. Thus the observation protocol determines which parameter symmetries are invisible.

For an explicit case take Z=S1Z=S^1 with Haar probability and the ellipse e(t)=(acos⁡t,bsin⁡t)e(t)=(a\cos t,b\sin t) in R2\mathbb R^2 for a,b>0a,b>0 with a≠ba\ne b. A hidden phase θ\theta in t↦e(t+θ)t\mapsto e(t+\theta) is entirely unidentifiable after integrating out tt. Adding centered isotropic Gaussian noise with covariance σ2I\sigma^2 I preserves this invariance. The integrated Gaussian mixture density is independent of θ\theta, so its Fisher information in the phase direction is zero. In contrast, an output orientation in t↦Rθe(t)t\mapsto R_\theta e(t) rotates the signal covariance: Cov⁡θ(Y)=12Rθ(a200b2)Rθ⊤+σ2I.\operatorname{Cov}_\theta(Y)= \tfrac12 R_\theta \begin{pmatrix}a^2&0\\0&b^2\end{pmatrix}R_\theta^\top +\sigma^2 I . Because a2≠b2a^2\ne b^2, equality of these covariances forces coincidence of the principal axes and hence equality of θ\theta modulo π\pi. Conversely a rotation by π\pi is the latent shift t↦t+πt\mapsto t+\pi on this ellipse, so orientations differing by π\pi give the same law. A symmetry of latent coordinates and a rotation of the observed manifold therefore have distinct statistical effects even within the same two-dimensional harmonic decoder.

Homogeneous feature bundles

Let GG be a compact Lie group, H,K⊆GH,K\subseteq G closed subgroups, and PP a topological GG-space of parameters. Write X=G/HX=G/H, Y=G/KY=G/K, and let ρ\rho and σ\sigma be finite-dimensional unitary representations of HH and KK on VV and WW. Use E=G×HVE=G\times_H V and F=G×KWF=G\times_K W as input and output feature bundles. Our convention is [bh,v]=[b,ρ(h)v][bh,v]=[b,\rho(h)v]. An input field is therefore encoded by f:G→Vf:G\to V with f(bh)=ρ(h)−1f(b)f(bh)=\rho(h)^{-1}f(b), and a group element gg acts by left translation (gf)(b)=f(g−1b)(g f)(b)=f(g^{-1}b). Haar probability induces the invariant measures on X,YX,Y. The fixed-parameter kernel model is the familiar homogeneous-space convolution [4, 2]; the parameter action below allows the kernel itself to move with the symmetry.

Proposition 5 (Kernels with moving parameters). Let Lp(y,x):Ex→FyL_p(y,x):E_x\to F_y be jointly continuous in the parameter and the two base variables, with diagonal covariance Lgp(gy,gx)=gLp(y,x)g−1.L_{gp}(gy,gx)=g L_p(y,x)g^{-1}. Such families correspond bijectively to continuous functions B:P×G→Hom⁡(V,W)B:P\times G\to\operatorname{Hom}(V,W) satisfying B(kp,kzh)=σ(k)B(p,z)ρ(h).B(kp,kzh)=\sigma(k) B(p,z)\rho(h). Here k∈K,h∈Hk\in K,h\in H. The corresponding family of bounded linear layers is (Tpf)(a)=∫GB(a−1p,a−1b)f(b) db.(T_p f)(a)=\int_G B(a^{-1}p,a^{-1}b)f(b)\,db . It satisfies Tgp(gf)=g(Tpf)T_{gp}(g f)=g(T_p f).

Proof. Evaluate LpL_p at the output basepoint 1K1K and write Lp(1K,zH)[z,v]=[1,B(p,z)v]L_p(1K,zH)[z,v]=[1,B(p,z)v]. Changing input representative zz to zhzh introduces ρ(h)\rho(h). Acting by k∈Kk\in K fixes the output basepoint and acts there by σ(k)\sigma(k). Together these give (4). Conversely, define Lp(aK,bH)[b,v]=[a,B(a−1p,a−1b)v].L_p(aK,bH)[b,v] =[a,B(a^{-1}p,a^{-1}b)v]. If a,ba,b are replaced by ak,bhak,bh, then the arguments of BB are replaced by k−1a−1p,k−1a−1bhk^{-1}a^{-1}p,k^{-1}a^{-1}bh, and vv by ρ(h)−1v\rho(h)^{-1}v. The two associated-bundle relations and (4) cancel these changes. Continuity follows in local sections of G→X,YG\to X,Y, and diagonal covariance follows since (ga)−1gp=a−1p(ga)^{-1}gp=a^{-1}p. The two formulas are inverse.

Integrating LpL_p over XX gives (5). The integrand is independent of the input representative, and the output has the required right KK-transformation. For fixed pp the kernel is continuous on Y×XY\times X, hence square-integrable in the Hilbert–Schmidt norm on the finite-dimensional fibers. Cauchy–Schwarz gives ∥Tp∥≤∥Lp∥L2(Y×X)\|T_p\|\leq\|L_p\|_{L^2(Y\times X)}. Joint continuity and compactness of Y×XY\times X give uniform convergence of Lp′L_{p'} to LpL_p as p′→pp'\to p, so p↦Tpp\mapsto T_p is operator-norm continuous. Changing variables b↦gbb\mapsto gb proves the stated operator covariance. If two continuous kernels give the same operators, testing in local frames against smooth sections supported in coordinate charts identifies their distributional pairings. The resulting equality almost everywhere extends everywhere because the measures have full support. ◻

When GG fixes pp, each TpT_p is an intertwiner. Equation (4) then constrains the permissible value at zz to intertwiners for K∩zHz−1K\cap zHz^{-1}, where the input representation is transported by k↦z−1kzk\mapsto z^{-1}kz. This follows by setting h=z−1k−1zh=z^{-1}k^{-1}z for a kk in the intersection. When a parameter moves, the analogous constraint uses Kp∩zHz−1K_p\cap zHz^{-1}, where KpK_p fixes pp. Thus a parameter with a larger stabilizer can impose more equations on the kernel at an orbit representative.

Linear layers.

For a fixed parameter, Peter–Weyl and compact Frobenius reciprocity decompose L2(E)L^2(E) as the Hilbert direct sum of Uλ⊗MλEU_\lambda\otimes M_\lambda^E, where UλU_\lambda runs through irreducibles of GG and MλE=Hom⁡H(Uλ∣H,V)M_\lambda^E=\operatorname{Hom}_H(U_\lambda|_H,V). For FF use MλF=Hom⁡K(Uλ∣K,W)M_\lambda^F=\operatorname{Hom}_K(U_\lambda|_K,W). Give the multiplicity spaces the inner products that make the isotypic decompositions unitary. Schur’s lemma identifies bounded intertwiners with the blocks 1Uλ⊗Aλ1_{U_\lambda}\otimes A_\lambda, where Aλ:MλE→MλFA_\lambda:M_\lambda^E\to M_\lambda^F and sup⁡λ∥Aλ∥<∞\sup_\lambda\|A_\lambda\|<\infty. Finite spectral support gives smooth kernels by finite sums of matrix coefficients. A general bounded intertwiner has a distributional kernel by applying the Schwartz kernel theorem to Γ∞(E)→L2(F)→Γ−∞(F)\Gamma^\infty(E)\to L^2(F)\to\Gamma^{-\infty}(F) [9, Thm. 1.16]. If E=FE=F and XX is positive-dimensional with EE of positive rank, the identity has a diagonal delta kernel. Continuous kernels on the compact product give Hilbert–Schmidt operators, whereas the identity on this infinite-dimensional space fails that condition.

Example 6 (A learned axis). Take P=X=Y=S2P=X=Y=S^2 with the rotation action and scalar channels. For p∈Pp\in P define qp(x)=p⋅xq_p(x)=p\cdot x and choose the kernel Lp(y,x)=3qp(y)qp(x)L_p(y,x)=3q_p(y)q_p(x). Rotational invariance of Haar probability on the sphere gives ∫S2xixj dx=δij/3\int_{S^2} x_i x_j\,dx=\delta_{ij}/3, so 3∥qp∥22=13\|q_p\|_2^2=1. The layer TpT_p is the orthogonal projection onto Cqp\mathbb C q_p, a line of degree-one spherical harmonics. Every gg carries qpq_p to qgpq_{gp}, hence Tgp=gTpg−1T_{gp}=gT_pg^{-1}. A fixed-pp layer commutes with rotations preserving the axis Rp\mathbb{R}p. In particular T−p=TpT_{-p}=T_p, and its symmetry group in GG is strictly larger than the stabilizer of pp. The parameter evaluation identifies opposite poles and realizes the family of rank-one projections parameterized by RP2\mathbb{RP}^{2}. This is a model symmetry that arises from the map that realizes the parameters.

Spin channels.

For S2=SO(3)/SO(2)S^2=\mathrm{SO}(3)/\mathrm{SO}(2), let LmL_m correspond to ρm(eit)=eimt\rho_m(e^{it})=e^{imt}. The ℓ\ell-th spherical type UℓU_\ell restricts to the stabilizer with weights −ℓ,…,ℓ-\ell,\ldots,\ell, each of multiplicity one. Thus L2(Lm)L^2(L_m) contains one copy of UℓU_\ell for ℓ≥∣m∣\ell\geq |m|. The section representations at weights 11 and −1-1 are therefore unitarily equivalent. The oriented complex tangent line is L1L_1, and LmL_m is its mm-th tensor power, using duals for m<0m<0. The tangent Euler number 22 gives c1(Lm)=2m u,⟨u,[S2]⟩=1c_1(L_m)=2m\,u,\qquad \langle u,[S^2]\rangle=1 [8, Ch. 3]. The two opposite channels therefore carry inequivalent complex bundles. Fiberwise nonlinear layers detect the characters directly. A continuous fiber map induced by ϕ:C→C\phi:\mathbb C\to\mathbb C from L2L_2 to L3L_3 would obey ϕ(e2itv)=e3itϕ(v)\phi(e^{2it}v)=e^{3it}\phi(v). Taking t=πt=\pi forces ϕ=0\phi=0. A nonzero global intertwiner from L2L_2 to L3L_3 is obtained by retaining the common ℓ=3\ell=3 block. The equivariance constraints on the two operations arise at different levels of the representation.

Manifolds with local frames.

The kernel description also makes sense without a transitive global action. On a smooth manifold MM, let Q→MQ\to M be a principal HH-bundle and use ρ,σ\rho,\sigma as its two feature representations. Local frames of QQ identify a kernel Ex→FyE_x\to F_y with a matrix A(y,x)A(y,x). Replacing the frames independently by right translates by hx,hy∈Hh_x,h_y\in H transforms the matrix by A′(y,x)=σ(hy)−1A(y,x)ρ(hx).A'(y,x)=\sigma(h_y)^{-1} A(y,x)\rho(h_x). Indeed, new input coordinates must first be converted to the old ones by ρ(hx)\rho(h_x), and old output coordinates to new ones by σ(hy)−1\sigma(h_y)^{-1}. This relation is exactly the gluing condition for a section of the bundle with fibers Hom⁡(Ex,Fy)\operatorname{Hom}(E_x,F_y) on M×MM\times M. Integration against a chosen smooth positive density defines a global layer whenever the kernel satisfies the needed integrability conditions. Setting x=yx=y yields the corresponding frame change for a fiber map. A fiberwise nonlinearity can also be glued if it obeys the same change of coordinates in the sense of intertwining the two actions.

Quotients and descent.

For a smooth action on a manifold PP, the quotient stacks give  ≃ [P/H].\ \simeq\ [P/H]. An object over a test manifold consists of a right GG-torsor with a GG-equivariant map to P×G/HP\times G/H, with inverse action on the target in the right-torsor convention. The inverse image of 1H1H under the second component is an HH-reduction of the torsor. The first component restricts to an equivariant map from that reduction to PP. Conversely, extending the HH-torsor along H↪GH\hookrightarrow G produces the original object and both maps. The constructions commute with pullback and identify torsor isomorphisms [2, Chs. 20, 28]. For P=∗P=* this gives the smooth classifying stack BHBH. A representation H→U(V)H\to \mathrm{U}(V) then gives a morphism of smooth classifying stacks and a descended feature bundle. More generally a homotopy-coherent coefficient system on an action groupoid retains the diagrams of identifications over all composable arrows. Formulas (4) and (5) illustrate the ordinary continuous case with a chosen invariant measure. The descent data determine how to transport fibers, while the measure enters the integration operation separately.

Harmonic decomposition with moving weights

A fixed-parameter equivariant layer can mix multiplicities of an irreducible representation but preserves its isotypic component. A moving parameter can carry the difference between the input and output representation types. This section classifies that effect for layers linear in the parameter and in the features. We work with complex Hilbert spaces. The representation WW of the compact group GG is finite dimensional and unitary, and HX,HYH_X,H_Y have strongly continuous unitary actions. A linear filter family is a complex-linear map A:W→B(HX,HY)A:W\to\mathcal B(H_X,H_Y), with the equivariance condition A(gw)=ρY(g)A(w)ρX(g)−1.A(gw)=\rho_Y(g) A(w)\rho_X(g)^{-1}. Boundedness of the map into the operator norm is automatic for finite-dimensional WW.

Let G^\widehat G index the irreducible unitary representations VπV_\pi. Write MπXM_\pi^X for the Hilbert multiplicity space of VπV_\pi in HXH_X, and likewise MπYM_\pi^Y. Compact harmonic analysis gives the orthogonal decompositions HX≃⨁^π∈G^Vπ⊗MπX,HY≃⨁^τ∈G^Vτ⊗MτY.H_X\simeq\widehat{\bigoplus}_{\pi\in\widehat G} V_\pi\otimes M_\pi^X,\qquad H_Y\simeq\widehat{\bigoplus}_{\tau\in\widehat G} V_\tau\otimes M_\tau^Y . For intertwiners S,T:V→HS,T:V\to H with common irreducible source VV, we use the scalar inner product defined by S∗T=⟨S,T⟩idVS^*T=\langle S,T\rangle\mathop{\mathrm{id}}_V.

Theorem 7 (Classification of linear filter families). For each τ∈G^\tau\in\widehat G, set Cτ,π=HomG(Vτ,W⊗Vπ),Nτ=⨁^π∈G^Cτ,π⊗MπX.\begin{aligned} C_{\tau,\pi} &=\mathop{\mathrm{Hom}}_G(V_\tau,W\otimes V_\pi),\\ N_\tau &=\widehat{\bigoplus}_{\pi\in\widehat G} C_{\tau,\pi}\otimes M_\pi^X . \end{aligned} The families satisfying (7) are in bijection with families of bounded maps Bτ:Nτ→MτYB_\tau:N_\tau\to M_\tau^Y satisfying sup⁡τ∥Bτ∥<∞\sup_\tau\|B_\tau\|<\infty. The norm of the linearized operator W⊗HX→HYW\otimes H_X\to H_Y equals that supremum. In particular the transition from π\pi to τ\tau can contribute only if VτV_\tau is a summand of W⊗VπW\otimes V_\pi.

Proof. Send w⊗xw\otimes x to A(w)xA(w)x. For an orthonormal basis (wj)j=1d(w_j)_{j=1}^d of WW, Cauchy–Schwarz gives ∥∑jA(wj)xj∥≤(∑j∥A(wj)∥2)1/2(∑j∥xj∥2)1/2.\left\|\sum_j A(w_j)x_j\right\| \le \left(\sum_j\|A(w_j)\|^2\right)^{1/2} \left(\sum_j\|x_j\|^2\right)^{1/2}. Thus linearization extends to the Hilbert tensor product. Conversely, a bounded operator L:W⊗HX→HYL:W\otimes H_X\to H_Y gives A(w)x=L(w⊗x)A(w)x=L(w\otimes x). Applying (7) to a decomposable tensor is precisely the intertwining condition for LL with the diagonal action on its source.

The compact isotypic decomposition follows from complete reducibility of continuous unitary representations [2, Chs. 2–4]. Tensoring with WW and decomposing each finite-dimensional W⊗VπW\otimes V_\pi gives W⊗HX≃⨁^τ∈G^Vτ⊗Nτ.W\otimes H_X\simeq \widehat{\bigoplus}_{\tau\in\widehat G} V_\tau\otimes N_\tau . Evaluation Vτ⊗Cτ,π→W⊗VπV_\tau\otimes C_{\tau,\pi}\to W\otimes V_\pi identifies the image with the τ\tau-isotypic part: orthogonality of the summands is Schur’s lemma and the chosen Hilbert structure on Cτ,πC_{\tau,\pi}. Schur’s lemma applied using multiplicity coordinates then forces LL to have block idVτ⊗Bτ\mathop{\mathrm{id}}_{V_\tau}\otimes B_\tau and zero blocks between distinct types. The norm of each diagonal block is ∥Bτ∥\|B_\tau\|. Conversely the Hilbert sum of such blocks is bounded with norm their supremum, and commutes with GG on each block. ◻

With WW trivial of dimension one this gives the usual classification of a bounded intertwiner. A trivial parameter representation of larger dimension adds multiplicity but preserves the same restriction on which types are allowed. A nontrivial WW changes that restriction. In a network with two successive linear parameter families, the evaluation is bilinear in their weights. It linearizes in W2⊗W1W_2\otimes W_1, with its diagonal action. A path of intermediate irreducible types is allowed when each step obeys the tensor multiplicity condition. Its actual linear map is the composition dictated by the original layer maps. Not every block of a free W2⊗W1W_2\otimes W_1 model factors through a prescribed architecture.

Example 8 (A learned Fourier shift). Let G=S1G=S^1 act on L2(S1)L^2(S^1) by translation (ρ(θ)x)(t)=x(t−θ)(\rho(\theta)x)(t)=x(t-\theta). Let W=CW=\mathbb C transform by the character eimθe^{im\theta} for a nonzero integer mm. Then (A(w)x)(t)=w e−imtx(t)(A(w)x)(t)=w\,e^{-imt}x(t) has norm ∣w∣|w| and satisfies (7). For the Fourier vector en(t)=einte_n(t)=e^{int}, it maps ene_n to wen−mw e_{n-m}. Indeed, the representation character on w⊗enw\otimes e_n is e−i(n−m)θe^{-i(n-m)\theta}, which is the character on en−me_{n-m}. This is the only possible output frequency for input ene_n in any such family with scalar weight mm. More generally A(w)en=wanen−mA(w)e_n=w a_n e_{n-m} classifies these families by bounded sequences (an)n∈Z(a_n)_{n\in\mathbb Z}. If ww is frozen at a nonzero value without transforming, the displayed operator fails to commute with rotations that act nontrivially on that weight.

Pointwise nonlinearities must preserve equivariance at the feature representations on which they operate. For a complex line of a circle weight, the radial map z↦z ψ(∣z∣2)z\mapsto z\,\psi(|z|^2) for continuous ψ\psi is an equivariant activation. A scalar real activation applied separately to real and imaginary parts need not commute with rotations. For instance, coordinatewise rectification fails already at rotation by π\pi. Nonlinearities can mix Fourier modes even though the whole model respects (2). The theorem describes the linear layers and their allowed type transitions, rather than a Fourier diagonalization of the nonlinear network.

A sign-symmetric decoder on the circle

A tied-weight scalar neural decoder gives a direct comparison of observable predictions, parameter coordinates, and descent data. Let G={1,−1}G=\{1,-1\} act on a real weight by w↦−ww\mapsto -w. For input x∈Rx\in\mathbb R, define the hidden feature and Bernoulli decoder by hw(x)=tanh⁡(wx),ℓw(x)=whw(x),Kw(1∣x)=11+exp⁡(−ℓw(x)),Kw(0∣x)=1−Kw(1∣x).\begin{aligned} h_w(x)&=\tanh(wx),& \ell_w(x)&=w h_w(x), \\ K_w(1\mid x)&=\frac{1}{1+\exp(-\ell_w(x))},\notag\\ K_w(0\mid x)&=1-K_w(1\mid x). \notag \end{aligned} Under a sign change the hidden feature changes sign and the tied readout changes sign, so the logit and the two probabilities are invariant. The statistical model is a Markov kernel with a finite output space. Its local parameters encode a choice of frame for the one-dimensional hidden feature.

Proposition 9 (Observational equivalence of scalar weights). Two weights u,wu,w give the same kernel for every real input xx if and only if u=±wu=\pm w. The gauge stabilizer of a weight is GG at w=0w=0 and trivial elsewhere.

Proof. Equality of probabilities gives equality of logits by invertibility of the logistic function. Differentiation in the input gives ℓw′(0)=w2\ell'_w(0)=w^2, so equality implies u2=w2u^2=w^2. The sign invariance of (8) gives the converse. A sign fixes ww precisely when its product with ww is ww. ◻

Descent of weights and observables

For a smooth base BB, a family in the quotient stack [R/G][\mathbb R/G] consists of a principal GG-bundle P→BP\to B and a smooth equivariant function w:P→Rw:P\to\mathbb R. Morphisms over BB are principal-bundle isomorphisms compatible with ww. This is the torsor description of a quotient stack, using open covers and local smooth trivializations in this setting [12, cf. the algebraic version in Tag 04UV]. Equivalently, choose local frames over an open cover UiU_i and write the coordinates as wi:Ui→Rw_i:U_i\to\mathbb R. The locally constant signs εij\varepsilon_{ij} satisfy εijεjk=εik,wi=εijwj.\varepsilon_{ij}\varepsilon_{jk}=\varepsilon_{ik}, \qquad w_i=\varepsilon_{ij}w_j . Changing frames acts by a family of local signs αi:Ui→G\alpha_i:U_i\to G, replacing these data by αiεijαj−1\alpha_i\varepsilon_{ij}\alpha_j^{-1} and αiwi\alpha_iw_i. The same signs glue the hidden features hwih_{w_i}, and the signs cancel in wihwiw_i h_{w_i}. Hence the observable kernel is a single smooth family K:B×R→Prob⁡({0,1})K:B\times\mathbb R\to \operatorname{Prob}(\{0,1\}). The invariant function r(b)=w(p)2r(b)=w(p)^2, for p∈Pbp\in P_b, can be recovered from KK by differentiating the logit in xx at 00. Both constructions commute with pullback along smooth base maps. They give maps from the groupoid of families to the set of smooth nonnegative functions and then to the set of observational laws. For a fixed base, the second map is injective on the functions obtained from families by Proposition 9. The first map can identify nonisomorphic families, as the zero example below shows.

Monodromy in the zero fiber

Over S1=R/(2πZ)S^1=\mathbb R/(2\pi\mathbb Z), take the connected double cover P−P_- with total space R/(4πZ)\mathbb R/(4\pi\mathbb Z) and deck transformation θ↦θ+2π\theta\mapsto\theta+2\pi. The associated sign line is the Möbius bundle [8, §1.1]. Transport once around the circle switches sheets. For comparison, the trivial double cover P+=S1×GP_+=S^1\times G preserves each sheet. The connectedness of P−P_- distinguishes it from P+P_+. Both admit the zero equivariant weight, and both resulting decoders output probability 1/21/2 at every base point and input.

Proposition 10 (Families over the zero observable). The groupoid of smooth families on S1S^1 with constant observable K(1∣x)=1/2K(1\mid x)=1/2 is equivalent to BG⨿BGBG\amalg BG.

Proof. The recovery formula for rr forces the equivariant function ww to be zero, so the whole groupoid consists of GG-torsors and their isomorphisms. Cut the circle at one point and trivialize a torsor on the intervening interval. The two ends are glued by one element of GG acting by multiplication, and changing the trivialization conjugates that element. Here GG is abelian, so its two elements determine exactly the trivial and connected double covers. A torsor automorphism is multiplication by a locally constant sign on the base. The circle is connected, so each object has automorphism group GG. ◻

At a single base point the corresponding fiber is BGBG. The extra isomorphism class over the circle comes from sign monodromy. The statistic rr identifies orbits at each point, while the stack retains both their automorphisms and their compatibility around a loop.

A family with local weights

On P−P_-, the smooth function w([θ])=cos⁡(θ/2)w([\theta])=\cos(\theta/2) changes sign under the deck transformation, so it is an equivariant weight. The descended family has r([θ])=cos⁡2(θ/2)=12(1+cos⁡θ).r([\theta])=\cos^2(\theta/2)=\tfrac12(1+\cos\theta). Its logits and probabilities are smooth by the local gluing above. This family cannot be represented by a single smooth weight on the circle in the same architecture. Indeed, any such weight v:S1→Rv:S^1\to\mathbb R must satisfy v2=rv^2=r by Proposition 9. The sole zero of rr is at [π][\pi], where r([π+t])=14t2+O(t4).r([\pi+t])=\tfrac14 t^2+O(t^4). Thus smoothness forces v′([π])2=1/4v'([\pi])^2=1/4. A smooth weight with this derivative changes sign at its sole zero. The complement of that zero in the circle is connected, and a continuous nonvanishing real function on it has constant sign, a contradiction. The two chartwise signed weights do define the model in the quotient stack, with the nontrivial transition sign recording the absence of a global frame. The absolute value of the local weight yields a continuous scalar lift, which shows the role of the smooth category in the obstruction.

Information geometry is sensitive to the parameter chart at the fixed point. The Fisher information of one Bernoulli observation at fixed xx is Kw(1∣x)Kw(0∣x)(∂wℓw(x))2K_w(1\mid x)K_w(0\mid x)(\partial_w\ell_w(x))^2; at w=0w=0 it vanishes because ∂wℓw(x)=0\partial_w\ell_w(x)=0. For fixed xx, the expansion ℓw(x)=w2x−13w4x3+O(w6)\ell_w(x)=w^2x-\tfrac13 w^4x^3+O(w^6) gives Iw(x)=x2w2+O(w4)I_w(x)=x^2w^2+O(w^4). In the nonnegative coordinate r=w2r=w^2, the same logit starts as rx−13r2x3rx-\tfrac13r^2x^3. Its derivative from the right gives Fisher information Ir(x)=x2/4I_r(x)=x^2/4 at r=0r=0, which is positive for x≠0x\neq0. Away from zero the parameters satisfy the pullback identity Iw(x)=4w2Ir(x)I_w(x)=4w^2 I_r(x), with the right-hand information evaluated at r=w2r=w^2. This extends to zero by the displayed expansions. The sign automorphism at zero, the monodromy of P−P_-, and the Fisher degeneracy measure three different features of the same family. In particular, the finite gauge group has discrete orbits, while the map from smooth weights to observable laws loses its entire first derivative at the origin.

Equivariant comparison and stable features

At the level of an ∞\infty-category, an action of a discrete group GG on features takes the form of a functor from the groupoid BGBG. A layer carrying symmetry specifies a natural transformation from its source action to its target action. If its parameter and feature objects can be combined by a tensor product, the source is the corresponding diagonal action.

For a concrete stable target, take D(k)D(k) for a commutative ring kk. A linear layer from XX to YY with linear parameter WW is now a map W⊗kLX→YW\otimes_k^L X\to Y. A coherent GG-action on these objects equips the ordinary mapping spectrum map⁡D(k)(W⊗kLX,Y)\operatorname{map}_{D(k)}(W\otimes_k^L X,Y) with conjugation transport. In a strict chain model, an element gg sends φ\varphi to ρY(g)φ(ρW(g−1)⊗ρX(g−1))\rho_Y(g)\varphi(\rho_W(g^{-1})\otimes\rho_X(g^{-1})). The mapping spectrum of equivariant layers is map⁡Fun(BG,D(k))(W⊗kLX,Y)≃map⁡D(k)(W⊗kLX,Y)hG. \begin{aligned} &\operatorname{map}_{\mathop{\mathrm{Fun}}(BG,D(k))} (W\otimes_k^L X,Y)\\[-2pt] &\qquad\simeq \operatorname{map}_{D(k)}(W\otimes_k^L X,Y)^{hG}. \end{aligned} The superscript hGhG denotes the limit of a BGBG-action. The formula also holds without a parameter, by taking the trivial tensor unit W=kW=k.

To see (9), the homotopy end for natural transformations resolves a transformation along strings of group arrows. In cosimplicial degree nn the term is a product over GnG^n of copies of the underlying mapping spectrum. Transporting the source of each component along its associated group arrows converts the two outer cofaces into conjugation at one end and deletion at the other. The inner cofaces compose successive arrows. This is the cosimplicial replacement for the homotopy limit on the right-hand side. Taking the totalization imposes all the coherences of an equivariant layer [2, Chs. 17–18].

The formula compares intertwiners between representations in a stable setting. It can model layers between complexes of features, for example when grading records features and relations between them. It introduces no probability law by itself. A stochastic decoder still requires its own evaluation and law.

Example 11 (A split underlying comparison with an equivariant obstruction). Let G=C2G=C_2 and k=F2k=\mathbb F_2. Give kk the trivial action, and let R=k[C2]≃k[u]/(u2)R=k[C_2]\simeq k[u]/(u^2) with u=g−1u=g-1. The exact sequence of RR-modules 0⟶k→ 1↦u R→ ε k⟶00\longrightarrow k\xrightarrow{\,1\mapsto u\,} R\xrightarrow{\,\varepsilon\,}k\longrightarrow 0 splits over kk. An RR-linear splitting would choose a lift of 11 in RR killed by uu. Every such lift is 1+au1+au, and u(1+au)=u≠0u(1+au)=u\ne0. Hence the sequence gives a nonzero class of equivariant maps k→k[1]k\to k[1], becoming zero when the action is forgotten. The free resolution obtained by repeating multiplication by uu shows ExtRj(k,k)≃k\mathop{\mathrm{Ext}}^j_R(k,k)\simeq k for every j≥0j\ge 0. This gives the positive cohomological degrees in the homotopy fixed point spectrum of the endomorphism spectrum of kk.

The regular representation in the example and the trivial module are objects of the same stable coefficient category. To identify the action category with derived k[G]k[G]-modules, evaluation at the unique object of BGBG is conservative and preserves colimits. Its left adjoint is the free action, whose underlying monad is k[G]⊗kL(−)k[G]\otimes_k^L(-) with multiplication from GG. The same monad comes from extension and restriction of scalars for k→k[G]k\to k[G]. Monadicity gives the identification [10, §4.7.3]. Over a ring where ∣G∣|G| is a unit, the averaging idempotent makes the trivial module projective and the higher Ext groups in this example vanish. This clarifies the relation with classical averaging. With real or complex Hilbert features and continuous compact actions, averaging and decomposition follow instead from Haar integration under the analytic hypotheses of Section 5.

Losses, derivatives, and symmetry

The classifying map of a layer can be used for optimization when the parameter and feature objects carry differential structure. Work in smooth manifolds, or use smooth charts of bundles over a common base. Let h=f(p,x)h=f(p,x), y=g(q,h)y=g(q,h), and let a smooth loss have output covector λy=dyℓ\lambda_y=d_y\ell. The partial reverse derivatives of the two-layer realization are λq=(Dqg)∗λy,λh=(Dhg)∗λy,λp=(Dpf)∗λh,λx=(Dxf)∗λh.\begin{aligned} \lambda_q&=(D_qg)^*\lambda_y, &\lambda_h&=(D_hg)^*\lambda_y,\\ \lambda_p&=(D_pf)^*\lambda_h, &\lambda_x&=(D_xf)^*\lambda_h . \end{aligned} The dual maps in this display are cotangent pullbacks, with no metric identification. The chain rule proves that they glue across smooth changes of charts, using the usual pullback of a loss along the realization. Under a tying map u:P→∏iPiu:P\to\prod_iP_i the resulting parameter covector is (Du)∗(λp1,…,λpn)(D u)^*(\lambda_{p_1},\ldots,\lambda_{p_n}). In particular, tying two copies by a diagonal adds their covectors at the shared point. The same map specifies sharing in the forward composite of Section 2. This is the cotangent form of the compatibility exploited by categorical formulations of backpropagation [6, 5].

A scalar expected loss can be formed from the conditional laws of Section 3, using a finite dataset or a distribution for which integration and differentiation are justified. Gradient vectors additionally require a positive metric or another update rule. For example, take a GG-invariant statistical realization r:P→Mr:P\to\mathcal M, where M\mathcal M is a smooth space of laws with Fisher metric FMF_{\mathcal M}. The pullback FP=r∗FMF_P=r^* F_{\mathcal M} is positive semidefinite. It annihilates vectors tangent to GG-orbits and, more generally, ker⁡Dr\ker D r, which can be larger than the orbit tangent space. The degeneracies in Sections 3 and 6 arise even when the relevant symmetry group at the parameter is discrete or has tangent orbit of smaller dimension. The inverse-Fisher natural gradient is defined where the chosen statistical coordinates have nondegenerate information [1].

Proposition 12 (An equivariant gradient field). Suppose a compact Lie group GG acts smoothly on a finite-dimensional, paracompact parameter manifold PP and the smooth scalar loss LL is invariant. If FPF_P is a smooth invariant positive semidefinite tensor, there exists a GG-invariant Riemannian metric jj such that for every λ>0\lambda>0 the gradient of LL with respect to gλ=FP+λjg_\lambda=F_P+\lambda j is equivariant. Its local flow is equivariant. If the action is free, this field is horizontal for P→P/GP\to P/G in gλg_\lambda and projects to the gradient of the descended loss for the induced quotient metric.

Proof. Choose any Riemannian metric and average its pullbacks using normalized Haar measure. Integration of smooth functions over a compact Lie group gives smoothness. Positive definiteness follows by averaging positive values on a nonzero tangent vector. Invariance follows by the right invariance of compact Haar measure. This yields jj, and the sum gλg_\lambda remains invariant and positive definite.

For each a∈Ga\in G, differentiate L∘a=LL\circ a=L and apply metric invariance. Nondegeneracy of gλg_\lambda then gives Da gradL(p)=gradL(ap)D a\,\mathop{\mathrm{grad}}L(p)=\mathop{\mathrm{grad}}L(ap). Uniqueness for local ordinary differential equations supplies the flow compatibility. In the free case the compact action is proper, so P→P/GP\to P/G is a smooth principal bundle. The loss differential vanishes on vertical vectors. Hence gradL\mathop{\mathrm{grad}}L is horizontal, and the invariant horizontal metric induces a quotient metric by the differential of the projection. Pairing the projected field with arbitrary quotient tangent vectors by horizontal lift gives the gradient identity on P/GP/G. ◻

For a nonfree action the equivariant evolution carries presentations to presentations and induces local maps of the quotient stack on the invariant domains of the flow. For the scalar deck-symmetric model, a metric can make optimization well defined at the origin, but the first derivative of a smooth loss of its law still vanishes there. Regularization changes the identification of covectors with vectors. It does not enlarge the derivative of the realization.

Conclusion

Over an ordinary smooth base the construction recovers smooth networks and statistical families, with cotangent composition and Fisher geometry when present. Over a homogeneous feature space it recovers induced bundles and intertwining layers. Allowing the weights to move contributes additional harmonic types, determined exactly by the tensor multiplicities in Theorem 7. These mechanisms are compatible because composition and families are represented in the same slices.

The higher model distinguishes observable equality from equivalence of parameter presentations. Descent records the transition arrows and their coherence, while homotopy fixed points describe the comparison spaces for stable features. The circle decoder makes that distinction visible without changing the input or output law. For each specialization the relevant structures enter explicitly: a smooth site for smooth realizations, a probability interpretation for sampling, and a compact unitary action for the Hilbert decomposition.

References

  1. S.-i. Amari. Natural gradient works efficiently in learning. Neural Computation 10(2), 251–276, 1998. DOI.

  2. Anonymous. Representation Theory in Higher Categories: From Compact Harmonic Analysis to Coherent Group Actions. Higher Categorical Semantics, 2026. Chapters 2–4, 17–18, 20, and 27–28. Published edition.

  3. N. Ay, J. Jost, H. V. Lê, and L. Schwachhöfer. Information geometry and sufficient statistics. Probability Theory and Related Fields 162, 327–364, 2015. DOI.

  4. T. S. Cohen, M. Geiger, and M. Weiler. A general theory of equivariant CNNs on homogeneous spaces. Advances in Neural Information Processing Systems 32, 2019. Proceedings.

  5. G. S. H. Cruttwell, B. Gavranović, N. Ghani, P. Wilson, and F. Zanasi. Categorical foundations of gradient-based learning. In Programming Languages and Systems, ESOP 2022. DOI.

  6. B. Fong, D. I. Spivak, and R. Tuyéras. Backprop as functor: A compositional perspective on supervised learning. In Proceedings of LICS, 2019. Author manuscript.

  7. T. Fritz. A synthetic approach to Markov kernels, conditional independence and theorems on sufficient statistics. Advances in Mathematics 370, 107239, 2020. DOI.

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

  9. C. Kottke. Linear Analysis on Manifolds. Notes for Math 7376, 2016. Lecture notes.

  10. J. Lurie. Higher Algebra. Manuscript, 2017. Author’s manuscript.

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

  12. The Stacks Project Authors. The Stacks Project, section on quotients by group actions. Tag 04UV.

  13. S. Watanabe. Algebraic Geometry and Statistical Learning Theory. Cambridge University Press, 2009. DOI.

Complete flattened LaTeX source
\documentclass[10pt,letterpaper,twocolumn]{article}
\usepackage[T1]{fontenc}

\usepackage{lmodern}

\usepackage[
  letterpaper,left=.72in,right=.72in,top=.75in,bottom=.73in,
  headheight=14pt,headsep=14pt,footskip=26pt
]{geometry}

\usepackage{microtype}

\usepackage{mathtools,amssymb,amsthm}

\usepackage{xcolor}

\usepackage{tikz-cd}

\usepackage{titlesec}

\usepackage{fancyhdr}

\usepackage{caption}

\usepackage{booktabs}

\usepackage{xurl}

\usepackage{hyperref}

\usepackage{bookmark}


\definecolor{paperink}{HTML}{202824}
\definecolor{paperaccent}{HTML}{145A51}
\definecolor{papermuted}{HTML}{51655F}
\definecolor{paperrule}{HTML}{B6CBC4}
\hypersetup{
  pdftitle={Representation Theory of Parametrized Neural Models on Spaces},
  pdfsubject={Statistical models, manifold models, and homotopy-coherent actions},
  colorlinks=true,linkcolor=paperaccent,citecolor=paperaccent,urlcolor=paperaccent
}
\AtBeginDocument{\color{paperink}}
\setlength{\columnsep}{.29in}
\setlength{\parindent}{1em}
\setlength{\parskip}{1pt plus .5pt}
\setlength{\emergencystretch}{1.7em}
\linespread{1.035}
\allowdisplaybreaks[1]
\raggedbottom
\clubpenalty=1000
\widowpenalty=1000

\titleformat{\section}{\large\bfseries\raggedright}
  {\color{paperaccent}\thesection}{.55em}{}
\titlespacing*{\section}{0pt}{14pt plus 3pt minus 2pt}{5pt}
\titleformat{\subsection}{\normalsize\bfseries\raggedright}
  {\color{paperaccent}\thesubsection}{.55em}{}
\titlespacing*{\subsection}{0pt}{10pt plus 2pt minus 1pt}{4pt}
\captionsetup{font=small,labelfont={bf,color=paperaccent}}

\pagestyle{fancy}
\fancyhf{}
\fancyhead[L]{\footnotesize\sffamily\color{papermuted}Parametrized neural models on spaces}
\fancyfoot[R]{\footnotesize\sffamily\color{papermuted}\thepage}
\renewcommand{\headrulewidth}{0pt}
\fancypagestyle{plain}{  \fancyhf{}  \fancyfoot[R]{\footnotesize\sffamily\color{papermuted}\thepage}  \renewcommand{\headrulewidth}{0pt}}

\theoremstyle{plain}
\newtheorem{theorem}{Theorem}[section]
\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{\Map}{Map}
\DeclareMathOperator{\Tot}{Tot}
\DeclareMathOperator{\Fun}{Fun}
\DeclareMathOperator{\Hom}{Hom}
\DeclareMathOperator{\grad}{grad}
\DeclareMathOperator{\Ext}{Ext}
\DeclareMathOperator{\Res}{Res}
\DeclareMathOperator{\Kl}{Kl}
\DeclareMathOperator{\id}{id}
\DeclareMathOperator{\rank}{rank}
\newcommand{\E}{\mathcal E}
\newcommand{\Spc}{\mathcal S}
\newcommand{\R}{\mathbb R}
\newcommand{\C}{\mathbb C}
\newcommand{\Z}{\mathbb Z}
\newcommand{\overB}{\E_{/B}}
\tikzcdset{every label/.append style={font=\small}}
\begin{document}
\twocolumn[
\begin{@twocolumnfalse}
\noindent\textcolor{paperaccent}{\rule{\textwidth}{1pt}}\par
\vspace{11pt}
{\noindent\fontsize{25}{30}\selectfont\bfseries
Representation Theory of Parametrized\\
Neural Models on Spaces\par}
\vspace{8pt}
{\noindent\large\color{papermuted}
Statistical families, equivariant layers, and homotopy-coherent models\par}
\vspace{10pt}
{\noindent\small November 2026\par}
\vspace{9pt}
\noindent\begin{minipage}{\textwidth}\small
\textbf{Abstract.}
A parametrized network over a space is a family of evaluation maps.
We represent its realizations and reparametrizations by a slice over an
internal mapping object, prove base change and descent, and extend the
construction to stochastic outputs for a strong probability monad.
On smooth charts this gives deterministic manifold networks and
conditional statistical models, together with pullback Fisher geometry.
For compact symmetries we classify bounded layers linear in a moving
parameter by multiplicity-space intertwiners. The classification exhibits
frequency shifts that cannot occur for a fixed equivariant layer.
For parameters with local symmetries we retain stabilizers and transition
data in a quotient stack. A neural decoder on a twisted real line illustrates why
global probability laws can lack a global smooth weight. Mapping spectra
supply a stable version of equivariance with higher intertwiners.
\par\medskip
\noindent\textbf{Keywords.}\enspace
Parametrized networks, representation theory, equivariance, Fisher information,
descent, quotient stacks.
\end{minipage}
\vspace{17pt}
\end{@twocolumnfalse}
]
\thispagestyle{plain}

\section{From statistical families to representations}

A supervised model commonly starts with a parameter manifold $P$, a
feature space $X$, and a map $f(p,x)$ valued in labels, predictions, or
probability distributions. A convolutional network has additional
structure. Features may live in bundles on an observation space, and a
symmetry can transport both features and weights. Two local weights can
also present the same model via a change of frame. These cases require
tracking evaluation, transport, and equivalences of presentations.

\emph{Representation Theory in Higher Categories} develops ordinary
and coherent group actions,
induced feature spaces, and equivariant convolutions
\cite[Chs.~17, 20, 27--28]{reprbook}. We start with its idea of encoding
actions by functors and relate evaluation, symmetry, and descent.
Evaluation records what a parameter computes. Representation theory
controls how evaluations transform. Descent determines when local
evaluations and parameter presentations assemble across a base.
Theorem~\ref{thm:relative-representation} gives the general comparison.
Theorem~\ref{thm:harmonic} computes the moving-parameter part in a
compact linear setting.

Parametrized composition and supervised learning have categorical
formulations through learners, lenses, and reverse differentiation
\cite{fong-spivak-tuyeras,cruttwell}. Markov categories organize channels
and sufficient statistics \cite{fritz-2020}, while equivariant
convolution on homogeneous spaces is described by induced
representations \cite{cohen-geiger-weiler}. Here we apply representation
constraints to the evaluation families themselves. This permits the same
typing of a manifold layer, a statistical decoder, and a locally presented
system of layers.

\section{Relative parametrized models}
\label{sec:relative}

Let $\E$ be an $\infty$-topos and $B$ an object of $\E$.
Possible choices are the $\infty$-category $\Spc$ of spaces and the
$\infty$-topos of sheaves of spaces on a smooth site. Work first
in $\overB$. For $X,Y$ over $B$, write $[X,Y]_B$ for their
exponential in that slice. Its evaluation map is
$[X,Y]_B\times_B X\to Y$ over $B$. A choice of site specifies which
families of sections are admitted. Smooth manifolds will be embedded
by their smooth functors of points.

\begin{definition}[A parametrized layer]
A layer from $X$ to $Y$ over $B$ consists of $P\to B$ and an
evaluation $f:P\times_B X\to Y$. A map to $(Q,g)$ is a morphism
$u:P\to Q$ over $B$ with a specified homotopy
$f\simeq g(u\times_B\id_X)$ over $B$. Higher morphisms are the
compatibilities in the mapping space of the resulting slice.
\end{definition}

Write $r_f:P\to[X,Y]_B$ for the transpose of $f$. Given a second
layer $(Q,g):Y\to Z$, the series composite has parameter
$P\times_B Q$ and evaluation $(p,q,x)\mapsto g(q,f(p,x))$.
Here and below parametric formulas describe generalized points in the
slice. A shared parameter $P$ can be imposed by reparametrization along
the diagonal $P\to P\times_B P$. More general tying is expressed by a
map $P\to\prod_{i,B}P_i$ into the product of layer parameters.

\begin{theorem}[Relative representation and descent]
\label{thm:relative-representation}
The $\infty$-category of layers from $X$ to $Y$ is
$\E_{/[X,Y]_B}$. On these categories, series composition is
induced by internal composition of exponentials and respects
reparametrizations. The construction commutes with pullback in $B$.
For an effective epimorphism $U\to B$ with \v{C}ech nerve $U_\bullet$,
the global layer category is equivalent to
\begin{equation}
 \label{eq:layerdescent}
 \lim_{[n]\in\boldsymbol\Delta}
   \E_{/[X_n,Y_n]_{U_n}},
 \qquad X_n=X\times_B U_n,\quad Y_n=Y\times_B U_n .
\end{equation}
\end{theorem}

\begin{proof}
Currying yields the equivalence of mapping spaces
\[
 \Map_B(P\times_B X,Y)\simeq\Map_B(P,[X,Y]_B).
\]
It is natural in $P$. The two definitions of a morphism of
layers therefore give the same homotopy fiber of a map of mapping
spaces, including its higher simplices. The relative double slice
$(\overB)_{/[X,Y]_B}$ is $\E_{/[X,Y]_B}$.

Evaluation in series transposes to a morphism
\[
 [X,Y]_B\times_B[Y,Z]_B\longrightarrow [X,Z]_B.
\]
It sends the classifying maps of the two layers to the classifying
map of their composite. The identity is classified by the transpose
of $\id_X$ with parameter $B$. The associativity and unit
compatibilities follow by uncurrying the iterated compositions.
Both ways to combine a string yield evaluation in the same order
on the product of its parameters. All product comparisons come
from their limit universal properties.

Let $a:C\to B$. If $T\to C$, the left adjoint $a_!$ to $a^*$
is composition of structure maps. Using the pullback identification
$a_!(T\times_C a^*X)\simeq (a_!T)\times_B X$ gives
\begin{align*}
 \Map_C(T,a^*[X,Y]_B)
 &\simeq \Map_B(a_!T,[X,Y]_B)\\
 &\simeq \Map_B(a_!T\times_B X,Y)\\
 &\simeq \Map_C(T\times_C a^*X,a^*Y).
\end{align*}
By Yoneda these equivalences identify $a^*[X,Y]_B$
with $[a^*X,a^*Y]_C$, preserving evaluation. Pullback also
preserves the products that define series composition.

Finally, the effective epimorphism is the colimit of its
\v{C}ech nerve. Colimits in an $\infty$-topos satisfy descent for
slices \cite[\S6.1.3]{lurie-htt}. Hence $\overB$ is the limit of the
categories $\E_{/U_n}$. Taking the slice at the compatible
exponential commutes with this limit, since a slice is the
category of arrows with fixed target. Base change identifies
the local targets with the ones in \eqref{eq:layerdescent}.
\end{proof}

\begin{figure}[tbp]
\centering
\(
\begin{tikzcd}[column sep=2.4em,row sep=2.0em]
  P\times_B X \arrow[r,"f"] \arrow[d,"u\times_B\id_X"'] &
  Y \arrow[d,equal] \\
  Q\times_B X \arrow[r,"g"'] & Y
\end{tikzcd}
\)
\caption{A reparametrization includes a commuting comparison of
evaluations. In an $\infty$-topos the comparison is a specified homotopy.}
\label{fig:reparam}
\end{figure}

A point of $[X,Y]_B$ over $b$ is one evaluation on the fibers.
The classifying map retains a distinction between a space of these
evaluations and a space of weights mapping into it. An architecture
selects a classifying map, for instance by composing affine maps,
an equivariant convolution, and a nonlinearity. The ambient
exponential also includes maps that the selected architecture
cannot realize.

\subsection{Transport and actions}

In spaces, straightening identifies $\Spc_{/B}$ with
$\Fun(B,\Spc)$, where the base $B$ is regarded as an
$\infty$-groupoid. Transport acts by equivalences and the slice
exponential is computed fiberwise \cite[\S3.2]{lurie-htt}. For
a path $\gamma:b\to c$ with transports $t_X,t_Y$, exponentials
transport by $h\mapsto t_Y h t_X^{-1}$. A classifying map
is consequently a coherent family of maps
$P_b\to [X_b,Y_b]$ commuting with transport. For an ordinary
action of an abstract group $G$ on $P,X,Y$, the resulting
comparison reads
\begin{equation}
 \label{eq:equiv-eval}
 f(gp,gx)=g f(p,x).
\end{equation}
It is a condition on the entire family. At a fixed $p$, it
forces equivariance under the stabilizer $G_p$. In a smooth
site one can instead use the classifying stack of a Lie group.
Sections of that stack are torsors with smooth local
trivializations, and \eqref{eq:equiv-eval} then holds for
smooth test families. The homotopy type of a smooth base
retains path transport and may identify smoothly distinct data.

\subsection{Realization on a quotient}

Suppose a group object $G$ acts on $P$ and a realization
$r:P\to M$ is coherently invariant with trivial action on
$M$. Here $M$ can itself be a space of functions or probability
laws. The action groupoid has simplicial levels $P\times G^n$.
Its colimit is $P//G$, and the colimit universal property gives
\[
 \Map(P//G,M)\simeq \Tot\,\Map(P\times G^\bullet,M).
\]
Thus $r$ factors through $P//G$ exactly when equipped with
descent, whose first comparison identifies $r(p)$ and $r(gp)$.
Equivalences between local parameter presentations are part
of a family in the quotient. Over $\Spc$, the fiber of
$P//G\to BG$ over the base point is the parameter space $P$,
with transport the original action. For smooth $P$ and Lie
$G$, the same construction in sheaves gives the smooth quotient
stack. The example in Section~\ref{gauge:circle} distinguishes
the quotient presentations from their common observable laws.

\subsection{Stochastic evaluation}
\label{sec:stochastic}

In a cartesian closed category, a strong monad $T$ has a
natural strength $\mathrm{st}_{A,Y}:A\times T Y\to T(A\times Y)$
compatible with products, unit $\eta$, and multiplication $\mu$.
When $T$ is a probability monad, the layer $P\times X\to T Y$
assigns a probability-valued prediction. Currying represents it
by $P\to[X,T Y]$. For
$f:P\times X\to T Y$ and $g:Q\times Y\to T Z$, their
composite with parameter $P\times Q$ is
\[
 \begin{aligned}
 P\times Q\times X &\longrightarrow Q\times TY\\
 &\xrightarrow{\ \mathrm{st}\ } T(Q\times Y)
   \xrightarrow{\ T g\ } T^2 Z
   \xrightarrow{\ \mu_Z\ } TZ ,
 \end{aligned}
\]
with first map $(p,q,x)\mapsto(q,f(p,x))$.
For a formal check, fix a context $A$ and form the category
whose arrows $X\to Y$ are maps $A\times X\to Y$. It is the
co-Kleisli category of the comonad $A\times(-)$ with diagonal
and projection. Strength is a mixed distributive law
$A\times T(-)\to T(A\times(-))$ for this comonad and the monad
$T$. Its compatibility with the diagonal and projection is
product coherence and naturality in $A$. Monad unit and
multiplication compatibility are the strength axioms.
Thus $T$ acts as a monad on the category with context $A$.
Kleisli composition for that monad extends a map
$h:A\times Y\to T Z$ using strength followed by $T h$ and
$\mu_Z$. Taking $A=P\times Q$ and restricting the layers along
its two projections gives the displayed formula. Context maps
preserve the construction by naturality of strength.
A compatible system of monads on slices gives the same
construction over $B$ when their strengths and monad
operations commute with base change.

For instance, the finitely supported distribution monad on sets
gives finite channels by restricting the feature and output sets
to be finite. If $p,q$ are fixed, the composite has law
\[
 k_{p,q}(z\mid x)=\sum_y g_q(z\mid y) f_p(y\mid x).
\]
Dirac laws embed deterministic evaluations. Repeated sampling can
share an unknown parameter by pulling a product family back along
its diagonal. For a coin of bias $1/5$ or $4/5$ with equal prior
probability, shared-parameter sampling gives probability $17/50$
of two successes. Independently drawing the parameter for each
observation gives $1/4$. The single-observation law therefore
needs to be supplemented by a parameter sharing convention.

For smooth conditional models, the probability-valued evaluator
can be defined on a represented family of probability laws.
The next section uses finite outcome laws with positive
probabilities, where the open simplex is a manifold and the
Fisher tensor can be computed in ordinary coordinates.
\par
\section{Statistical and manifold models}
\label{stat:section}

A parametrized smooth map \(F:P\times X\to M\) is a manifold-valued
prediction, with a family of maps \(F_p:X\to M\). A deterministic network
may also retain intermediate feature bundles by factoring \(F\) through
smooth maps of total spaces. With an ordinary parameter manifold and no
prescribed equivalences of parameters, these are ordinary smooth families.
For stochastic predictions on a finite set \(Y\), a smooth readout is a map
\[
 q:P\times X\longrightarrow\Delta^\circ(Y),
 \qquad (p,x)\longmapsto(q_i(p,x))_{i\in Y},
\]
where \(\Delta^\circ(Y)\) is the simplex of strictly positive probability
vectors. Evaluation at \(p\) gives the conditional law of the outcome
given \(x\). Taking \(X=*\) recovers the usual smooth finite statistical
model. A deterministic prediction into a standard Borel space \(M\)
has the associated kernel \(x\mapsto\delta_{F(p,x)}\).
Dirac measures on a finite set with at least two elements lie on the
boundary of the simplex. Strict positivity permits differentiation of the finite log likelihood.

A stochastic decoder can be composed with a succeeding stochastic layer
by summing over the intermediate outcome. If \(K_{ji}(p,x)\) is the
conditional probability of output \(j\) given outcome \(i\), the new
readout is \(r_j=\sum_i K_{ji}q_i\). Associativity is the associativity
of finite summation, and deterministic relabelings are the special case
of Dirac kernels \cite{fritz-2020}. Smooth charts of \(P\) and \(X\)
present these models locally: if the predictions agree on the overlaps
after coordinate change, the coordinate functions \(q_i\) glue to a
unique smooth readout. For a manifold target \(M\), the same assertion
uses transition maps of \(M\) on overlapping target charts.

\subsection{The Fisher form and stochastic postprocessing}

At a point \(q\) of the positive simplex, write \(u,v\) for tangent
vectors with \(\sum_i u_i=\sum_i v_i=0\). The Fisher metric is
\[
 g_q(u,v)=\sum_{i\in Y}\frac{u_iv_i}{q_i}.
\]
At a fixed input \(x\), the score in direction \(v\in T_pP\) is
\(s_v(i)=d_p\log q_i(p,x)(v)\). Its mean is zero by differentiation of
\(\sum_i q_i=1\). Thus the covariance of the scores in directions
\(v,w\) is
\begin{equation}
 I_{p,x}(v,w)=
 \sum_{i\in Y}\frac{d_pq_i(v)\,d_pq_i(w)}{q_i(p,x)}.
 \label{stat:fisher}
\end{equation}
Equivalently, \(I_x=q_x^*g\), where \(q_x(p)=q(p,x)\).
For a parameter-independent Borel probability measure \(\mu\) on \(X\),
the joint experiment observes both the input and the label with law
\(d\mu(x)\,q_i(p,x)\). Its Fisher form is
\(I_p^\mu=\int_X I_{p,x}\,d\mu(x)\), assuming the scores are square
integrable and the experiment is differentiable under the integral.
These conditions hold for a smooth positive readout when \(X\) is
compact, locally in the parameter.

For a smooth reparametrization \(r:P'\to P\), the conditional Fisher
form of \(q\circ(r\times\operatorname{id}_X)\) is \(r^*I_x\)
by the chain rule. This holds for \(I^\mu\) as well. The identity gives
the change-of-coordinates rule \(I'=Dr^\top I\,Dr\) on charts.
On input charts the same joint form is computed by transporting
\(\mu\) with the coordinates. Thus the tensor descends with the
model and the chosen design measure. Its pointwise null space at
\(x\) is \(\ker d_pq_x\), because all \(q_i>0\). The joint form
vanishes on \(v\) precisely when \(d_pq_i(v)\) vanishes
\(\mu\)-almost everywhere for every \(i\).

\begin{proposition}[Finite Fisher contraction]
\label{stat:contraction}
A stochastic postprocessor \(K\), fixed with respect to \(p\), weakly
decreases both the conditional Fisher form and its design average.
\end{proposition}

\begin{proof}
Omit output values \(j\) that are unreachable under \(K\).
For \(r_j=\sum_i K_{ji}q_i>0\), the score of \(r_j\) in direction
\(v\) is
\[
 d_p\log r_j(v)=\sum_i\frac{K_{ji}q_i}{r_j}\,s_v(i).
\]
The weights are the conditional probabilities of \(i\) given \(j\).
Jensen's inequality bounds the square by the weighted sum of squares.
Multiplying by \(r_j\) and summing over \(j\) yields
\(I^{r}_{p,x}(v,v)\le I^{q}_{p,x}(v,v)\). Integrating yields the claim
for \(I^\mu\). At a fixed \(x\), equality in direction \(v\) holds
exactly when \(s_v\) is constant on the intermediate outcomes that
contribute to each reachable output.
\end{proof}

The parameter-independence hypothesis keeps the postprocessor from
introducing its own score. Formula~\eqref{stat:fisher} is the finite
instance of the parametrized-measure Fisher tensor
\cite{ay-jost-le-schwachhoefer-2015}.

\begin{example}[Gaussian observations of a manifold family]
Let \(m:P\times X\to\mathbb R^k\) be a smooth readout and let the
conditional outcome have distribution \(N(m(p,x),\Sigma(x))\), where
\(\Sigma(x)\) is positive definite and independent of \(p\). For an
observed outcome \(z\), differentiation of the Gaussian log density
at fixed \(x\) gives the score
\[
 d_pm(v)^\top\Sigma(x)^{-1}\bigl(z-m(p,x)\bigr).
\]
The expectation of the product of two scores uses the covariance
\(\Sigma(x)\) and gives
\[
 I_{p,x}(v,w)=d_pm(v)^\top\Sigma(x)^{-1}d_pm(w).
\]
Thus an immersed mean manifold at fixed \(x\) has a Riemannian
metric by pullback of the Euclidean metric weighted by
\(\Sigma(x)^{-1}\). If the mean map has a kernel in its differential,
the pullback has the same null directions. An invertible affine
coordinate change of the output with linear part \(A\) replaces
\(\Sigma\) by \(A\Sigma A^\top\) and \(d_pm\) by \(A\,d_pm\), leaving the
form invariant. For a general smooth output coordinate change,
one transports the entire conditional law. The transformed density
contains a Jacobian factor independent of \(p\), so its score is the
transport of the original score even when the law exits the Gaussian
family.
\end{example}

\subsection{What the Fisher form forgets}

Consider a two-feature decoder on \(S^1\), with angle \(t\) understood
modulo \(2\pi\) and parameters \(p=(a,b,c)\in\mathbb R^3\). Its logit
is \(L_p(t)=a\cos t+bc\sin t\) and its Bernoulli success probability
is \(s_p(t)=(1+e^{-L_p(t)})^{-1}\). For a tangent vector
\(v=(v_a,v_b,v_c)\), the score at label \(y\in\{0,1\}\) is
\((y-s_p(t))\,dL_p(t)(v)\). With any full-support design probability
measure \(\mu\) on the circle, the joint Fisher form is
\begin{align*}
 I^\mu_p(v,v)
  &=\int_{S^1}s_p(t)(1-s_p(t))\,(\delta_v(t))^2\,d\mu(t),\\
 \delta_v(t)&=v_a\cos t+(c v_b+b v_c)\sin t.
\end{align*}
Positivity of the Bernoulli variance and continuity in \(t\) imply
that the integral vanishes exactly when \(\delta_v\) vanishes
identically. The independent harmonics give the two equations
\(v_a=0\) and \(c v_b+b v_c=0\). The rank is two away from \(b=c=0\)
and one along that locus. In particular the Fisher tensor is the
pullback of a positive definite form on the two-dimensional logistic
family with coordinates \((a,d)\), along the map \((a,b,c)\mapsto(a,bc)\). The strict monotonicity of
the logistic function and the independent harmonics also show that
this map labels exactly the distinct output laws. This is an elementary instance of singular parametrization
\cite{watanabe-2009}.

The multiplicative group \(\mathbb R^\times\) acts by
\(\lambda\cdot(a,b,c)=(a,\lambda b,\lambda^{-1}c)\), preserving the
readout. The action groupoid has stabilizer \(\mathbb R^\times\) at
\(b=c=0\) and trivial stabilizer elsewhere. Its three distinct
orbits over a given prediction with \(d=0\) correspond to the
origin and the two punctured coordinate axes in the \((b,c)\)-plane.
Thus the quotient groupoid of parameter symmetries contains
information beyond the set of distinct output laws. Even the
infinitesimal null space can include directions that change the
output at higher order: along \((a,\varepsilon,\varepsilon)\) through
\((a,0,0)\), \(d=\varepsilon^2\) changes the prediction to second order.
Consequently equality of output laws, Fisher-null directions, and
parameter equivalence answer three different questions.

\subsection{A manifold decoder as a statistical model}

Let $Z$ be a latent manifold with fixed probability $\nu$ and
$d:P\times Z\to M$ a smooth decoder. In the absence of observation
noise the law at $p$ is $(d_p)_*\nu$. A noise channel $N:M\to Y$
then gives the observed law
\[
 \mathbb P_p(A)=\int_Z N(d_p(z),A)\,d\nu(z)
\]
for measurable $A\subseteq Y$. This is ordinary composition of
Markov kernels on these measurable spaces. The manifold
evaluation $d$ can be represented by the smooth slice
construction of Section~\ref{sec:relative}. The pushforward
additionally uses the chosen law $\nu$ and noise $N$.
If $a:Z\to Z$ preserves $\nu$, then replacing $d_p$ by
$d_p\circ a$ leaves $\mathbb P_p$ unchanged, by the
change-of-variables identity for pushforwards. A law that
observes $z$ and the output together can distinguish the two.
Thus the observation protocol determines which parameter
symmetries are invisible.

For an explicit case take $Z=S^1$ with Haar probability and
the ellipse $e(t)=(a\cos t,b\sin t)$ in $\mathbb R^2$ for
$a,b>0$ with $a\ne b$. A hidden phase $\theta$ in
$t\mapsto e(t+\theta)$ is entirely unidentifiable after
integrating out $t$. Adding centered isotropic Gaussian
noise with covariance $\sigma^2 I$ preserves this
invariance. The integrated Gaussian mixture density is
independent of $\theta$, so its Fisher information in
the phase direction is zero. In contrast, an output
orientation in $t\mapsto R_\theta e(t)$ rotates the
signal covariance:
\[
 \operatorname{Cov}_\theta(Y)=
 \tfrac12 R_\theta
 \begin{pmatrix}a^2&0\\0&b^2\end{pmatrix}R_\theta^\top
 +\sigma^2 I .
\]
Because $a^2\ne b^2$, equality of these covariances forces
coincidence of the principal axes and hence equality of
$\theta$ modulo $\pi$. Conversely a rotation by $\pi$ is
the latent shift $t\mapsto t+\pi$ on this ellipse, so
orientations differing by $\pi$ give the same law.
A symmetry of latent coordinates and a rotation of the
observed manifold therefore have distinct statistical effects
even within the same two-dimensional harmonic decoder.
\par
\section{Homogeneous feature bundles}
\label{geo:section}

Let \(G\) be a compact Lie group, \(H,K\subseteq G\) closed
subgroups, and \(P\) a topological \(G\)-space of parameters.
Write \(X=G/H\), \(Y=G/K\), and let \(\rho\) and \(\sigma\)
be finite-dimensional unitary representations of \(H\) and \(K\)
on \(V\) and \(W\). Use \(E=G\times_H V\) and \(F=G\times_K W\)
as input and output feature bundles. Our convention is
\([bh,v]=[b,\rho(h)v]\). An input field is therefore encoded by
\(f:G\to V\) with \(f(bh)=\rho(h)^{-1}f(b)\), and a group element \(g\)
acts by left translation \((g f)(b)=f(g^{-1}b)\). Haar probability
induces the invariant measures on \(X,Y\). The fixed-parameter
kernel model is the familiar homogeneous-space convolution
\cite{cohen-geiger-weiler,reprbook}; the parameter action below
allows the kernel itself to move with the symmetry.

\begin{proposition}[Kernels with moving parameters]
\label{geo:kernels}
Let \(L_p(y,x):E_x\to F_y\) be jointly continuous in the
parameter and the two base variables, with diagonal covariance
\[
 L_{gp}(gy,gx)=g L_p(y,x)g^{-1}.
\]
Such families correspond bijectively to continuous functions
\(B:P\times G\to\operatorname{Hom}(V,W)\) satisfying
\begin{equation}
 B(kp,kzh)=\sigma(k) B(p,z)\rho(h).
 \label{geo:bi}
\end{equation}
Here \(k\in K,h\in H\). The corresponding family of bounded
linear layers is
\begin{equation}
 (T_p f)(a)=\int_G B(a^{-1}p,a^{-1}b)f(b)\,db .
 \label{geo:conv}
\end{equation}
It satisfies \(T_{gp}(g f)=g(T_p f)\).
\end{proposition}

\begin{proof}
Evaluate \(L_p\) at the output basepoint \(1K\) and write
\(L_p(1K,zH)[z,v]=[1,B(p,z)v]\). Changing input representative
\(z\) to \(zh\) introduces \(\rho(h)\). Acting by \(k\in K\)
fixes the output basepoint and acts there by \(\sigma(k)\).
Together these give \eqref{geo:bi}. Conversely, define
\[
 L_p(aK,bH)[b,v]
   =[a,B(a^{-1}p,a^{-1}b)v].
\]
If \(a,b\) are replaced by \(ak,bh\), then the arguments of \(B\)
are replaced by \(k^{-1}a^{-1}p,k^{-1}a^{-1}bh\), and \(v\) by
\(\rho(h)^{-1}v\). The two associated-bundle relations and
\eqref{geo:bi} cancel these changes. Continuity follows in local
sections of \(G\to X,Y\), and diagonal covariance follows since
\((ga)^{-1}gp=a^{-1}p\). The two formulas are inverse.

Integrating \(L_p\) over \(X\) gives \eqref{geo:conv}. The integrand
is independent of the input representative, and the output has
the required right \(K\)-transformation. For fixed \(p\) the
kernel is continuous on \(Y\times X\), hence square-integrable
in the Hilbert--Schmidt norm on the finite-dimensional fibers.
Cauchy--Schwarz gives
\(\|T_p\|\leq\|L_p\|_{L^2(Y\times X)}\).
Joint continuity and compactness of \(Y\times X\) give
uniform convergence of \(L_{p'}\) to \(L_p\) as \(p'\to p\),
so \(p\mapsto T_p\) is operator-norm continuous.
Changing variables \(b\mapsto gb\) proves the stated operator
covariance. If two continuous kernels give the same operators,
testing in local frames against smooth sections supported in
coordinate charts identifies their distributional pairings.
The resulting equality almost everywhere extends everywhere
because the measures have full support.
\end{proof}

When \(G\) fixes \(p\), each \(T_p\) is an intertwiner.
Equation~\eqref{geo:bi} then constrains the permissible value at
\(z\) to intertwiners for \(K\cap zHz^{-1}\), where the input
representation is transported by \(k\mapsto z^{-1}kz\).
This follows by setting \(h=z^{-1}k^{-1}z\) for a \(k\) in the
intersection. When a parameter moves, the analogous constraint
uses \(K_p\cap zHz^{-1}\), where \(K_p\) fixes \(p\). Thus a
parameter with a larger stabilizer can impose more equations on
the kernel at an orbit representative.

\paragraph{Linear layers.}
For a fixed parameter, Peter--Weyl and compact Frobenius
reciprocity decompose \(L^2(E)\) as the Hilbert direct sum of
\(U_\lambda\otimes M_\lambda^E\), where \(U_\lambda\) runs through
irreducibles of \(G\) and
\(M_\lambda^E=\operatorname{Hom}_H(U_\lambda|_H,V)\).
For \(F\) use \(M_\lambda^F=\operatorname{Hom}_K(U_\lambda|_K,W)\).
Give the multiplicity spaces the inner products that make the
isotypic decompositions unitary. Schur's lemma identifies bounded
intertwiners with the blocks \(1_{U_\lambda}\otimes A_\lambda\),
where \(A_\lambda:M_\lambda^E\to M_\lambda^F\) and
\(\sup_\lambda\|A_\lambda\|<\infty\). Finite spectral support
gives smooth kernels by finite sums of matrix coefficients.
A general bounded intertwiner has a distributional kernel by
applying the Schwartz kernel theorem to
\(\Gamma^\infty(E)\to L^2(F)\to\Gamma^{-\infty}(F)\)
\cite[Thm.~1.16]{kottke-lam}. If \(E=F\) and \(X\) is
positive-dimensional with \(E\) of positive rank, the identity
has a diagonal delta kernel. Continuous kernels on the compact
product give Hilbert--Schmidt operators, whereas the identity on
this infinite-dimensional space fails that condition.

\begin{example}[A learned axis]
\label{geo:axis}
Take \(P=X=Y=S^2\) with the rotation action and scalar
channels. For \(p\in P\) define \(q_p(x)=p\cdot x\) and
choose the kernel \(L_p(y,x)=3q_p(y)q_p(x)\).
Rotational invariance of Haar probability on the sphere gives
\(\int_{S^2} x_i x_j\,dx=\delta_{ij}/3\), so \(3\|q_p\|_2^2=1\).
The layer \(T_p\) is the orthogonal projection onto
\(\mathbb C q_p\), a line of degree-one spherical harmonics.
Every \(g\) carries \(q_p\) to \(q_{gp}\), hence
\(T_{gp}=gT_pg^{-1}\). A fixed-\(p\) layer commutes with
rotations preserving the axis \(\mathbb{R}p\).
In particular \(T_{-p}=T_p\), and its symmetry group in \(G\)
is strictly larger than the stabilizer of \(p\).
The parameter evaluation identifies opposite poles and realizes
the family of rank-one projections parameterized by
\(\mathbb{RP}^{2}\). This is a model symmetry that arises from
the map that realizes the parameters.
\end{example}

\paragraph{Spin channels.}
For \(S^2=\mathrm{SO}(3)/\mathrm{SO}(2)\), let \(L_m\) correspond
to \(\rho_m(e^{it})=e^{imt}\). The \(\ell\)-th spherical type
\(U_\ell\) restricts to the stabilizer with weights
\(-\ell,\ldots,\ell\), each of multiplicity one.
Thus \(L^2(L_m)\) contains one copy of \(U_\ell\) for
\(\ell\geq |m|\). The section representations at weights \(1\)
and \(-1\) are therefore unitarily equivalent. The oriented
complex tangent line is \(L_1\), and \(L_m\) is its
\(m\)-th tensor power, using duals for \(m<0\).
The tangent Euler number \(2\) gives
\[
 c_1(L_m)=2m\,u,\qquad
 \langle u,[S^2]\rangle=1
\]
\cite[Ch.~3]{hatcher-vbkt}. The two opposite channels therefore
carry inequivalent complex bundles. Fiberwise nonlinear layers
detect the characters directly. A continuous fiber map induced
by \(\phi:\mathbb C\to\mathbb C\) from \(L_2\) to \(L_3\) would
obey \(\phi(e^{2it}v)=e^{3it}\phi(v)\). Taking \(t=\pi\) forces
\(\phi=0\). A nonzero global intertwiner from \(L_2\) to \(L_3\)
is obtained by retaining the common \(\ell=3\) block.
The equivariance constraints on the two operations arise at
different levels of the representation.

\paragraph{Manifolds with local frames.}
The kernel description also makes sense without a transitive
global action. On a smooth manifold \(M\), let \(Q\to M\) be
a principal \(H\)-bundle and use \(\rho,\sigma\) as its two
feature representations. Local frames of \(Q\) identify a
kernel \(E_x\to F_y\) with a matrix \(A(y,x)\). Replacing
the frames independently by right translates by
\(h_x,h_y\in H\) transforms the matrix by
\[
 A'(y,x)=\sigma(h_y)^{-1} A(y,x)\rho(h_x).
\]
Indeed, new input coordinates must first be converted to the
old ones by \(\rho(h_x)\), and old output coordinates to new
ones by \(\sigma(h_y)^{-1}\). This relation is exactly the
gluing condition for a section of the bundle with fibers
\(\operatorname{Hom}(E_x,F_y)\) on \(M\times M\). Integration
against a chosen smooth positive density defines a global
layer whenever the kernel satisfies the needed integrability
conditions. Setting \(x=y\) yields the corresponding frame
change for a fiber map. A fiberwise nonlinearity can also be
glued if it obeys the same change of coordinates in the sense
of intertwining the two actions.

\paragraph{Quotients and descent.}
For a smooth action on a manifold \(P\), the quotient stacks give
\begin{equation}
 [\,(P\times G/H)/G\,]\ \simeq\ [P/H].
 \label{geo:stack}
\end{equation}
An object over a test manifold consists of a right \(G\)-torsor
with a \(G\)-equivariant map to \(P\times G/H\), with inverse
action on the target in the right-torsor convention. The inverse
image of \(1H\) under the second component is an \(H\)-reduction
of the torsor. The first component restricts to an equivariant
map from that reduction to \(P\). Conversely, extending the
\(H\)-torsor along \(H\hookrightarrow G\) produces the original
object and both maps. The constructions commute with pullback
and identify torsor isomorphisms \cite[Chs.~20, 28]{reprbook}.
For \(P=*\) this gives the smooth classifying stack \(BH\). A representation \(H\to
\mathrm{U}(V)\) then gives a morphism of smooth classifying
stacks and a descended feature bundle. More generally a
homotopy-coherent coefficient system on an action groupoid
retains the diagrams of identifications over all composable
arrows. Formulas \eqref{geo:bi} and \eqref{geo:conv} illustrate
the ordinary continuous case with a chosen invariant measure.
The descent data determine how to transport fibers, while the
measure enters the integration operation separately.
\par
\section{Harmonic decomposition with moving weights}
\label{sec:harmonics}

A fixed-parameter equivariant layer can mix multiplicities of an
irreducible representation but preserves its isotypic component.
A moving parameter can carry the difference between the input
and output representation types. This section classifies that
effect for layers linear in the parameter and in the features.
We work with complex Hilbert spaces. The representation $W$
of the compact group $G$ is finite dimensional and unitary,
and $H_X,H_Y$ have strongly continuous unitary actions.
A linear filter family is a complex-linear map
$A:W\to\mathcal B(H_X,H_Y)$, with the equivariance condition
\begin{equation}
 A(gw)=\rho_Y(g) A(w)\rho_X(g)^{-1}.
 \label{eq:filter-covariance}
\end{equation}
Boundedness of the map into the operator norm is automatic
for finite-dimensional $W$.

Let $\widehat G$ index the irreducible unitary
representations $V_\pi$. Write $M_\pi^X$ for the Hilbert
multiplicity space of $V_\pi$ in $H_X$, and likewise $M_\pi^Y$.
Compact harmonic analysis gives the orthogonal decompositions
\[
 H_X\simeq\widehat{\bigoplus}_{\pi\in\widehat G}
         V_\pi\otimes M_\pi^X,\qquad
 H_Y\simeq\widehat{\bigoplus}_{\tau\in\widehat G}
         V_\tau\otimes M_\tau^Y .
\]
For intertwiners $S,T:V\to H$ with common irreducible source $V$,
we use the scalar inner product defined by
$S^*T=\langle S,T\rangle\id_V$.

\begin{theorem}[Classification of linear filter families]
\label{thm:harmonic}
For each $\tau\in\widehat G$, set
\[
 \begin{aligned}
 C_{\tau,\pi}
    &=\Hom_G(V_\tau,W\otimes V_\pi),\\
 N_\tau
    &=\widehat{\bigoplus}_{\pi\in\widehat G}
         C_{\tau,\pi}\otimes M_\pi^X .
 \end{aligned}
\]
The families satisfying \eqref{eq:filter-covariance} are in
bijection with families of bounded maps
$B_\tau:N_\tau\to M_\tau^Y$ satisfying
$\sup_\tau\|B_\tau\|<\infty$. The norm of the linearized operator
$W\otimes H_X\to H_Y$ equals that supremum. In particular the
transition from $\pi$ to $\tau$ can contribute only if
$V_\tau$ is a summand of $W\otimes V_\pi$.
\end{theorem}

\begin{proof}
Send $w\otimes x$ to $A(w)x$. For an orthonormal basis
$(w_j)_{j=1}^d$ of $W$, Cauchy--Schwarz gives
\[
 \left\|\sum_j A(w_j)x_j\right\|
 \le
 \left(\sum_j\|A(w_j)\|^2\right)^{1/2}
 \left(\sum_j\|x_j\|^2\right)^{1/2}.
\]
Thus linearization extends to the Hilbert tensor product.
Conversely, a bounded operator $L:W\otimes H_X\to H_Y$
gives $A(w)x=L(w\otimes x)$. Applying \eqref{eq:filter-covariance}
to a decomposable tensor is precisely the intertwining
condition for $L$ with the diagonal action on its source.

The compact isotypic decomposition follows from complete
reducibility of continuous unitary representations
\cite[Chs.~2--4]{reprbook}. Tensoring with $W$ and
decomposing each finite-dimensional $W\otimes V_\pi$ gives
\[
 W\otimes H_X\simeq
   \widehat{\bigoplus}_{\tau\in\widehat G} V_\tau\otimes N_\tau .
\]
Evaluation $V_\tau\otimes C_{\tau,\pi}\to W\otimes
V_\pi$ identifies the image with the $\tau$-isotypic part:
orthogonality of the summands is Schur's lemma and the
chosen Hilbert structure on $C_{\tau,\pi}$. Schur's lemma applied
using multiplicity coordinates then forces $L$ to have block
$\id_{V_\tau}\otimes B_\tau$ and zero blocks between distinct
types. The norm of each diagonal block is $\|B_\tau\|$.
Conversely the Hilbert sum of such blocks is bounded
with norm their supremum, and commutes with $G$ on each block.
\end{proof}

With $W$ trivial of dimension one this gives the usual
classification of a bounded intertwiner. A trivial parameter
representation of larger dimension adds multiplicity but
preserves the same restriction on which types are allowed.
A nontrivial $W$ changes that restriction. In a network with
two successive linear parameter families, the evaluation is
bilinear in their weights. It linearizes in $W_2\otimes W_1$,
with its diagonal action. A path of intermediate irreducible
types is allowed when each step obeys the tensor multiplicity
condition. Its actual linear map is the composition dictated
by the original layer maps. Not every block of a free
$W_2\otimes W_1$ model factors through a prescribed architecture.

\begin{example}[A learned Fourier shift]
Let $G=S^1$ act on $L^2(S^1)$ by translation
$(\rho(\theta)x)(t)=x(t-\theta)$. Let $W=\C$ transform by the
character $e^{im\theta}$ for a nonzero integer $m$. Then
\[
 (A(w)x)(t)=w\,e^{-imt}x(t)
\]
has norm $|w|$ and satisfies \eqref{eq:filter-covariance}.
For the Fourier vector $e_n(t)=e^{int}$, it maps $e_n$
to $w e_{n-m}$. Indeed, the representation character on
$w\otimes e_n$ is $e^{-i(n-m)\theta}$, which is the character
on $e_{n-m}$. This is the only possible output frequency
for input $e_n$ in any such family with scalar weight $m$.
More generally $A(w)e_n=w a_n e_{n-m}$ classifies these
families by bounded sequences $(a_n)_{n\in\Z}$.
If $w$ is frozen at a nonzero value without transforming,
the displayed operator fails to commute with rotations that
act nontrivially on that weight.
\end{example}

Pointwise nonlinearities must preserve equivariance at the
feature representations on which they operate. For a complex
line of a circle weight, the radial map
$z\mapsto z\,\psi(|z|^2)$ for continuous $\psi$ is an
equivariant activation. A scalar real activation applied
separately to real and imaginary parts need not commute with
rotations. For instance, coordinatewise rectification fails
already at rotation by $\pi$. Nonlinearities can mix Fourier
modes even though the \emph{whole} model respects
\eqref{eq:equiv-eval}. The theorem describes the linear layers
and their allowed type transitions, rather than a Fourier
diagonalization of the nonlinear network.
\par
\section{A sign-symmetric decoder on the circle}
\label{gauge:circle}

A tied-weight scalar neural decoder gives a direct comparison of
observable predictions, parameter coordinates, and descent data. Let
\(G=\{1,-1\}\) act on a real weight by \(w\mapsto -w\). For input
\(x\in\mathbb R\), define the hidden feature and Bernoulli decoder by
\begin{align}
 h_w(x)&=\tanh(wx),&
 \ell_w(x)&=w h_w(x), \label{gauge:logit}\\
 K_w(1\mid x)&=\frac{1}{1+\exp(-\ell_w(x))},\notag\\
 K_w(0\mid x)&=1-K_w(1\mid x). \notag
\end{align}
Under a sign change the hidden feature changes sign and the tied
readout changes sign, so the logit and the two probabilities are
invariant. The statistical model is a Markov kernel with a finite
output space. Its local parameters encode a choice of frame for the
one-dimensional hidden feature.

\begin{proposition}[Observational equivalence of scalar weights]
\label{gauge:pointwise}
Two weights \(u,w\) give the same kernel for every real input \(x\)
if and only if \(u=\pm w\). The gauge stabilizer of a weight is \(G\) at \(w=0\)
and trivial elsewhere.
\end{proposition}

\begin{proof}
Equality of probabilities gives equality of logits by invertibility
of the logistic function. Differentiation in the input gives
\(\ell'_w(0)=w^2\), so equality implies \(u^2=w^2\).
The sign invariance of \eqref{gauge:logit} gives the converse.
A sign fixes \(w\) precisely when its product with \(w\) is \(w\).
\end{proof}

\subsection{Descent of weights and observables}

For a smooth base \(B\), a family in the quotient stack
\([\mathbb R/G]\) consists of a principal \(G\)-bundle \(P\to B\)
and a smooth equivariant function \(w:P\to\mathbb R\). Morphisms
over \(B\) are principal-bundle isomorphisms compatible with \(w\).
This is the torsor description of a quotient stack, using open
covers and local smooth trivializations in this setting
\cite[cf.\ the algebraic version in Tag 04UV]{stacks-quotients}.
Equivalently, choose local frames over an open cover \(U_i\) and
write the coordinates as \(w_i:U_i\to\mathbb R\). The locally
constant signs \(\varepsilon_{ij}\) satisfy
\[
 \varepsilon_{ij}\varepsilon_{jk}=\varepsilon_{ik},
 \qquad w_i=\varepsilon_{ij}w_j .
\]
Changing frames acts by a family of local signs
\(\alpha_i:U_i\to G\), replacing these data by
\(\alpha_i\varepsilon_{ij}\alpha_j^{-1}\) and \(\alpha_iw_i\).
The same signs glue the hidden features \(h_{w_i}\), and the
signs cancel in \(w_i h_{w_i}\). Hence the observable kernel is a
single smooth family \(K:B\times\mathbb R\to
\operatorname{Prob}(\{0,1\})\). The invariant function
\(r(b)=w(p)^2\), for \(p\in P_b\), can be recovered from \(K\)
by differentiating the logit in \(x\) at \(0\).
Both constructions commute with pullback along smooth base maps.
They give maps from the groupoid of families to the set of smooth
nonnegative functions and then to the set of observational laws.
For a fixed base, the second map is injective on the functions
obtained from families by Proposition~\ref{gauge:pointwise}.
The first map can identify nonisomorphic families, as the zero
example below shows.

\subsection{Monodromy in the zero fiber}

Over \(S^1=\mathbb R/(2\pi\mathbb Z)\), take the connected double
cover \(P_-\) with total space \(\mathbb R/(4\pi\mathbb Z)\) and
deck transformation \(\theta\mapsto\theta+2\pi\). The associated
sign line is the M\"obius bundle
\cite[\S 1.1]{hatcher-vbkt}. Transport once around the circle
switches sheets. For comparison, the trivial double cover
\(P_+=S^1\times G\) preserves each sheet. The connectedness of
\(P_-\) distinguishes it from \(P_+\). Both admit the zero
equivariant weight, and both resulting decoders output probability
\(1/2\) at every base point and input.

\begin{proposition}[Families over the zero observable]
\label{gauge:zero}
The groupoid of smooth families on \(S^1\) with constant
observable \(K(1\mid x)=1/2\) is equivalent to
\(BG\amalg BG\).
\end{proposition}

\begin{proof}
The recovery formula for \(r\) forces the equivariant function
\(w\) to be zero, so the whole groupoid consists of \(G\)-torsors
and their isomorphisms. Cut the circle at one point and trivialize
a torsor on the intervening interval. The two ends are glued by
one element of \(G\) acting by multiplication, and changing the
trivialization conjugates that element. Here \(G\) is abelian,
so its two elements determine exactly the trivial and connected
double covers. A torsor automorphism is multiplication by a
locally constant sign on the base. The circle is connected, so
each object has automorphism group \(G\).
\end{proof}

At a single base point the corresponding fiber is \(BG\).
The extra isomorphism class over the circle comes from sign
monodromy. The statistic \(r\) identifies orbits at each point,
while the stack retains both their automorphisms and their
compatibility around a loop.

\subsection{A family with local weights}

On \(P_-\), the smooth function \(w([\theta])=\cos(\theta/2)\)
changes sign under the deck transformation, so it is an
equivariant weight. The descended family has
\[
 r([\theta])=\cos^2(\theta/2)=\tfrac12(1+\cos\theta).
\]
Its logits and probabilities are smooth by the local gluing
above. This family cannot be represented by a single smooth
weight on the circle in the same architecture. Indeed, any such
weight \(v:S^1\to\mathbb R\) must satisfy \(v^2=r\) by
Proposition~\ref{gauge:pointwise}. The sole zero of \(r\) is at
\([\pi]\), where
\[
 r([\pi+t])=\tfrac14 t^2+O(t^4).
\]
Thus smoothness forces \(v'([\pi])^2=1/4\). A smooth weight with
this derivative changes sign at its sole zero. The complement
of that zero in the circle is connected, and a continuous
nonvanishing real function on it has constant sign, a
contradiction. The two chartwise signed weights do define the
model in the quotient stack, with the nontrivial transition
sign recording the absence of a global frame. The absolute
value of the local weight yields a continuous scalar lift,
which shows the role of the smooth category in the obstruction.

Information geometry is sensitive to the parameter chart at
the fixed point. The Fisher information of one Bernoulli
observation at fixed \(x\) is
\(K_w(1\mid x)K_w(0\mid x)(\partial_w\ell_w(x))^2\);
at \(w=0\) it vanishes because \(\partial_w\ell_w(x)=0\).
For fixed \(x\), the expansion
\[
 \ell_w(x)=w^2x-\tfrac13 w^4x^3+O(w^6)
\]
gives \(I_w(x)=x^2w^2+O(w^4)\). In the nonnegative coordinate
\(r=w^2\), the same logit starts as \(rx-\tfrac13r^2x^3\).
Its derivative from the right gives Fisher information
\(I_r(x)=x^2/4\) at \(r=0\), which is positive for \(x\neq0\).
Away from zero the parameters satisfy the pullback identity
\(I_w(x)=4w^2 I_r(x)\), with the right-hand information evaluated at
\(r=w^2\). This extends to zero by the displayed expansions.
The sign automorphism at zero, the monodromy of \(P_-\), and the
Fisher degeneracy measure three different features of the same
family. In particular, the finite gauge group has discrete orbits,
while the map from smooth weights to observable laws loses its
entire first derivative at the origin.
\par
\section{Equivariant comparison and stable features}
\label{sec:stable}

At the level of an $\infty$-category, an action of a
discrete group $G$ on features takes the form of a functor
from the groupoid $BG$. A layer carrying symmetry specifies
a natural transformation from its source action to its
target action. If its parameter and feature objects can be
combined by a tensor product, the source is the corresponding
diagonal action.

For a concrete stable target, take $D(k)$ for a commutative
ring $k$. A linear layer from $X$ to $Y$ with linear parameter
$W$ is now a map $W\otimes_k^L X\to Y$. A coherent $G$-action
on these objects equips the ordinary mapping spectrum
$\operatorname{map}_{D(k)}(W\otimes_k^L X,Y)$ with conjugation
transport. In a strict chain model, an element $g$ sends
$\varphi$ to
$\rho_Y(g)\varphi(\rho_W(g^{-1})\otimes\rho_X(g^{-1}))$.
The mapping spectrum of equivariant layers is
\begin{equation}
 \label{eq:homotopy-fixed}
 \begin{aligned}
 &\operatorname{map}_{\Fun(BG,D(k))}
                (W\otimes_k^L X,Y)\\[-2pt]
 &\qquad\simeq
     \operatorname{map}_{D(k)}(W\otimes_k^L X,Y)^{hG}.
 \end{aligned}
\end{equation}
The superscript $hG$ denotes the limit of a $BG$-action.
The formula also holds without a parameter, by taking the
trivial tensor unit $W=k$.

To see \eqref{eq:homotopy-fixed}, the homotopy end for natural
transformations resolves a transformation along strings of
group arrows. In cosimplicial degree $n$ the term is a product
over $G^n$ of copies of the underlying mapping spectrum.
Transporting the source of each component along its
associated group arrows converts the two outer cofaces into
conjugation at one end and deletion at the other.
The inner cofaces compose successive arrows. This is the
cosimplicial replacement for the homotopy limit on the
right-hand side. Taking the totalization imposes all the
coherences of an equivariant layer
\cite[Chs.~17--18]{reprbook}.

The formula compares intertwiners between representations
in a stable setting. It can model layers between complexes of
features, for example when grading records features and relations
between them. It introduces no probability law by itself.
A stochastic decoder still requires its own evaluation and law.

\begin{example}[A split underlying comparison with an
equivariant obstruction]
Let $G=C_2$ and $k=\mathbb F_2$. Give $k$ the trivial action,
and let $R=k[C_2]\simeq k[u]/(u^2)$ with $u=g-1$.
The exact sequence of $R$-modules
\[
 0\longrightarrow k\xrightarrow{\,1\mapsto u\,}
 R\xrightarrow{\,\varepsilon\,}k\longrightarrow 0
\]
splits over $k$. An $R$-linear splitting would choose a lift
of $1$ in $R$ killed by $u$. Every such lift is $1+au$, and
$u(1+au)=u\ne0$. Hence the sequence gives a nonzero class of
equivariant maps $k\to k[1]$, becoming zero when the
action is forgotten. The free resolution obtained by repeating
multiplication by $u$ shows
$\Ext^j_R(k,k)\simeq k$ for every $j\ge 0$.
This gives the positive cohomological degrees in the homotopy
fixed point spectrum of the endomorphism spectrum of $k$.
\end{example}

The regular representation in the example and the trivial
module are objects of the same stable coefficient category.
To identify the action category with derived $k[G]$-modules,
evaluation at the unique object of $BG$ is conservative and
preserves colimits. Its left adjoint is the free action, whose
underlying monad is $k[G]\otimes_k^L(-)$ with multiplication
from $G$. The same monad comes from extension and
restriction of scalars for $k\to k[G]$. Monadicity gives the
identification \cite[\S4.7.3]{lurie-ha}. Over a ring where
$|G|$ is a unit, the averaging idempotent makes the trivial
module projective and the higher Ext groups in this example
vanish. This clarifies the relation with classical averaging.
With real or complex Hilbert features and continuous compact
actions, averaging and decomposition follow instead from Haar
integration under the analytic hypotheses of
Section~\ref{sec:harmonics}.
\par
\section{Losses, derivatives, and symmetry}

The classifying map of a layer can be used for optimization
when the parameter and feature objects carry differential
structure. Work in smooth manifolds, or use smooth charts of
bundles over a common base. Let
$h=f(p,x)$, $y=g(q,h)$, and let a smooth loss have output
covector $\lambda_y=d_y\ell$. The partial reverse derivatives
of the two-layer realization are
\[
 \begin{aligned}
 \lambda_q&=(D_qg)^*\lambda_y,
   &\lambda_h&=(D_hg)^*\lambda_y,\\
 \lambda_p&=(D_pf)^*\lambda_h,
   &\lambda_x&=(D_xf)^*\lambda_h .
 \end{aligned}
\]
The dual maps in this display are cotangent pullbacks, with
no metric identification. The chain rule proves that they glue
across smooth changes of charts, using the usual pullback of
a loss along the realization. Under a tying map
$u:P\to\prod_iP_i$ the resulting parameter covector is
$(D u)^*(\lambda_{p_1},\ldots,\lambda_{p_n})$. In particular,
tying two copies by a diagonal adds their covectors at the
shared point. The same map specifies sharing in the forward
composite of Section~\ref{sec:relative}. This is the
cotangent form of the compatibility exploited by categorical
formulations of backpropagation
\cite{fong-spivak-tuyeras,cruttwell}.

A scalar expected loss can be formed from the conditional laws
of Section~\ref{stat:section}, using a finite dataset or a
distribution for which integration and differentiation are
justified. Gradient vectors additionally require a positive
metric or another update rule. For example, take a
$G$-invariant statistical realization $r:P\to\mathcal M$,
where $\mathcal M$ is a smooth space of laws with Fisher metric
$F_{\mathcal M}$. The pullback $F_P=r^* F_{\mathcal M}$ is
positive semidefinite. It annihilates vectors tangent to
$G$-orbits and, more generally, $\ker D r$, which can be
larger than the orbit tangent space. The degeneracies in
Sections~\ref{stat:section} and~\ref{gauge:circle} arise even
when the relevant symmetry group at the parameter is discrete
or has tangent orbit of smaller dimension. The inverse-Fisher
natural gradient is defined where the chosen statistical
coordinates have nondegenerate information \cite{amari}.

\begin{proposition}[An equivariant gradient field]
\label{prop:gradient}
Suppose a compact Lie group $G$ acts smoothly on a
finite-dimensional, paracompact parameter manifold $P$ and
the smooth scalar loss $L$ is invariant. If $F_P$ is a smooth
invariant positive semidefinite tensor, there exists a
$G$-invariant Riemannian metric $j$ such that for every
$\lambda>0$ the gradient of $L$ with respect to
$g_\lambda=F_P+\lambda j$ is equivariant. Its local flow
is equivariant. If the action is free, this field is
horizontal for $P\to P/G$ in $g_\lambda$ and projects to the
gradient of the descended loss for the induced quotient metric.
\end{proposition}

\begin{proof}
Choose any Riemannian metric and average its pullbacks using
normalized Haar measure. Integration of smooth functions
over a compact Lie group gives smoothness. Positive
definiteness follows by averaging positive values on a
nonzero tangent vector. Invariance follows by the right
invariance of compact Haar measure. This yields $j$, and the
sum $g_\lambda$ remains invariant and positive definite.

For each $a\in G$, differentiate $L\circ a=L$ and apply
metric invariance. Nondegeneracy of $g_\lambda$ then gives
$D a\,\grad L(p)=\grad L(ap)$. Uniqueness for local ordinary
differential equations supplies the flow compatibility.
In the free case the compact action is proper, so $P\to P/G$
is a smooth principal bundle. The loss differential vanishes
on vertical vectors. Hence $\grad L$ is horizontal, and the
invariant horizontal metric induces a quotient metric by
the differential of the projection. Pairing the projected
field with arbitrary quotient tangent vectors by horizontal
lift gives the gradient identity on $P/G$.
\end{proof}

For a nonfree action the equivariant evolution carries
presentations to presentations and induces local maps of the
quotient stack on the invariant domains of the flow. For the
scalar deck-symmetric model, a metric can make optimization
well defined at the origin, but the first derivative of a
smooth loss of its law still vanishes there. Regularization
changes the identification of covectors with vectors. It
does not enlarge the derivative of the realization.

\section{Conclusion}

Over an ordinary smooth base the construction recovers
smooth networks and statistical families, with cotangent
composition and Fisher geometry when present. Over a
homogeneous feature space it recovers induced bundles and
intertwining layers. Allowing the weights to move contributes
additional harmonic types, determined exactly by the tensor
multiplicities in Theorem~\ref{thm:harmonic}. These mechanisms
are compatible because composition and families are
represented in the same slices.

The higher model distinguishes observable equality from
equivalence of parameter presentations. Descent records the
transition arrows and their coherence, while homotopy fixed
points describe the comparison spaces for stable features.
The circle decoder makes that distinction visible without
changing the input or output law. For each specialization
the relevant structures enter explicitly: a smooth site
for smooth realizations, a probability interpretation for
sampling, and a compact unitary action for the Hilbert
decomposition.
\par
\begin{thebibliography}{99}
\small\raggedright

\bibitem{amari}
S.-i. Amari. Natural gradient works efficiently in learning.
\emph{Neural Computation} 10(2), 251--276, 1998.
\href{https://doi.org/10.1162/089976698300017746}{DOI}.

\bibitem{reprbook}
Anonymous. \emph{Representation Theory in Higher Categories:
From Compact Harmonic Analysis to Coherent Group Actions}.
Higher Categorical Semantics, 2026. Chapters 2--4, 17--18, 20, and 27--28.
\href{https://higher-categorical-semantics.pages.dev/research/representation-theory-in-higher-categories-1d29df27}{Published edition}.

\bibitem{ay-jost-le-schwachhoefer-2015}
N. Ay, J. Jost, H. V. L\^e, and L. Schwachh\"ofer.
Information geometry and sufficient statistics.
\emph{Probability Theory and Related Fields} 162, 327--364, 2015.
\href{https://doi.org/10.1007/s00440-014-0574-8}{DOI}.

\bibitem{cohen-geiger-weiler}
T. S. Cohen, M. Geiger, and M. Weiler.
A general theory of equivariant CNNs on homogeneous spaces.
\emph{Advances in Neural Information Processing Systems} 32, 2019.
\href{https://proceedings.neurips.cc/paper_files/paper/2019/hash/b9cfe8b6042cf759dc4c0cccb27a6737-Abstract.html}{Proceedings}.

\bibitem{cruttwell}
G. S. H. Cruttwell, B. Gavranovi\'c, N. Ghani, P. Wilson, and F. Zanasi.
Categorical foundations of gradient-based learning.
In \emph{Programming Languages and Systems}, ESOP 2022.
\href{https://doi.org/10.1007/978-3-030-99336-8_1}{DOI}.

\bibitem{fong-spivak-tuyeras}
B. Fong, D. I. Spivak, and R. Tuy\'eras.
Backprop as functor: A compositional perspective on supervised learning.
In \emph{Proceedings of LICS}, 2019.
\href{https://hdl.handle.net/1721.1/123513}{Author manuscript}.

\bibitem{fritz-2020}
T. Fritz. A synthetic approach to Markov kernels, conditional
independence and theorems on sufficient statistics.
\emph{Advances in Mathematics} 370, 107239, 2020.
\href{https://doi.org/10.1016/j.aim.2020.107239}{DOI}.

\bibitem{hatcher-vbkt}
A. 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{kottke-lam}
C. Kottke. \emph{Linear Analysis on Manifolds}.
Notes for Math 7376, 2016.
\href{https://ckottke.ncf.edu/neu/7376_sp16/lacm.pdf}{Lecture notes}.

\bibitem{lurie-ha}
J. Lurie. \emph{Higher Algebra}. Manuscript, 2017.
\href{https://www.math.ias.edu/~lurie/papers/HA.pdf}{Author's manuscript}.

\bibitem{lurie-htt}
J. 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{stacks-quotients}
The Stacks Project Authors. \emph{The Stacks Project}, section on
quotients by group actions.
\href{https://stacks.math.columbia.edu/tag/04UV}{Tag 04UV}.

\bibitem{watanabe-2009}
S. Watanabe. \emph{Algebraic Geometry and Statistical Learning Theory}.
Cambridge University Press, 2009.
\href{https://doi.org/10.1017/CBO9780511800474}{DOI}.
\end{thebibliography}
\end{document}

End of manuscript · Back to top ↑