Differential Topology of Parametrized Neural Networks
Continuous, single-column reading. Equations, tables, proofs, references and appendices stay in document order.
Contents · 19 sections
Differential Topology of
Parametrized Neural Networks
Jet representations, decision boundaries, and reconstruction
November 2026
Abstract. A smooth neural architecture represents a space of weights by evaluations on a data manifold. We give architecture-level conditions for controlling first jets, decision boundaries, and reconstruction. An affine readout of immersed features is submersive on first jets, while an arbitrary smooth feature map with free final biases still gives generic transverse classification boundaries. We compute a fold discriminant for a quadratic readout on the circle and a regular family whose decision boundaries have nontrivial monodromy. For equivariant layers the stabilizer representation can force an excess intersection even when the total evaluation is submersive. An explicit torus calculation exhibits the associated obstruction complex. For autoencoders, exact ambient recovery requires an embedding, and a spherical example gives a sharp uniform error bound below the required latent dimension. Relative mapping objects and descent specify how these results apply to local parameter and feature presentations.
A parametrized neural network is a family of maps realized by a specified architecture. Differential topology studies these maps through their derivatives, their fibers, and their behavior along paths of parameters. This makes it possible to ask which local perturbations a layer actually supplies and which global topological constraints remain even if local perturbations are plentiful. A basic example is an encoder with a smooth, trainable readout. The readout can have enough derivatives to move a decision boundary, while a narrow intermediate space still obstructs reconstruction on the whole data manifold.
The starting points are Hirsch’s treatment of jets and transversality [8, Chs. 2–4] and their extension to relative families, derived pullbacks, and equivariant representations in [1, Chs. 18–22, 25–28]. Earlier categorical models of learning represent layers by parametrized maps, including their compositions and shared weights [6, 2]. Here the map from a prescribed architecture to the function object is subjected to differential-topological tests. Related empirical work investigates changes in data topology across network layers [10]. Our statements concern smooth realizations and their included perturbations.
Relative evaluation and representations
Let E be the ∞-topos of sheaves of spaces on smooth test manifolds, and let B be a coefficient or geometry base. For X,Y→B write [X,Y]B for the internal mapping object in E/B. A model with weights P→B is an evaluation F:P×BX⟶Y over B, or its transpose P→[X,Y]B. For example, if B=∗ and X,Y are represented by smooth manifolds, an evaluation on P is a smooth map P×X→Y on ordinary representable parameter charts. The classifying map records the realizations available to that architecture. Two different points of P can specify the same evaluation, and a change of coordinates acts on the presenting parameter space. A parameter shared by multiple factors of the network is a pullback along the corresponding diagonal into the product of their parameter spaces.
Internal exponentials commute with base change in B. For an effective epimorphism U→B with Čech nerve U∙, we have the descent equivalence E/[X,Y]B≃[n]∈ΔlimE/[X∣Un,Y∣Un]Un from [1, Ch. 18]. Thus compatible evaluations and parameter presentations on geometry charts glue. On charts of smooth bundles, a change of input and output coordinates transforms the vertical derivative by the derivative of the output change on the left and the inverse input change on the right. It preserves rank, so rank strata and regularity conditions are intrinsic to the glued model. For an action, the same formula is conjugation by the linear representations on tangent spaces. Its restriction to stabilizers becomes a concrete selection rule in Section 3.
Which kind of realization is measured
A differentiable classifier with real logit fp has a decision boundary fp−1(0) and sublevel sets fp−1((−∞,c]). A scalar strictly monotone link, such as the logistic link, preserves the boundary when the corresponding threshold is transported. A classifier with several logits admits a stratification by the indices realizing the maximum. An autoencoder, by contrast, is a factorization of an approximation to the identity on a data manifold. Its topology depends on the encoder’s image in the latent space and on the codomain of the decoder. We use smooth activations for all jet and regular-value assertions. The uniform reconstruction obstruction requires only continuous realizations, so it also applies to continuous piecewise-linear activations.
Readouts and the available jet directions
Let M be an n-dimensional smooth manifold without boundary. Fix a smooth feature extractor ϕ:M→Rd and use the affine readout fA,b(x)=Aϕ(x)+b,(A,b)∈A=Hom(Rd,Rq)×Rq. The choice to hold the feature extractor fixed specifies a subfamily of whatever full architecture realizes ϕ. This separates the derivatives supplied by the readout from those provided by perturbing upstream layers.
Theorem 1 (First-jet control by an affine readout). If ϕ is an immersion, then J:A×M→J1(M,Rq), (A,b,x)↦jx1fA,b, is a submersion. Without the immersion assumption, the corresponding map to J0(M,Rq) remains a submersion. In either case, for a finite or countable family of smooth jet submanifolds at the indicated order, the parameters for which the corresponding jet section is transverse to all of them form a conull residual set.
Proof. At fixed x, use jet coordinates (y,L) with L:TxM→Rq. The differential in the readout directions is (A˙,b˙)⟼(A˙ϕ(x)+b˙,A˙Dϕx). Given any L0:TxM→Rq and immersive ϕ, extend L0(Dϕx)−1 linearly from Dϕx(TxM) to Rd to obtain A˙. Then any value variation y0 is realized by b˙=y0−A˙ϕ(x). This proves surjectivity in the vertical jet coordinates. Base coordinates are covered because J lies over the projection to M. For J0 only the value variation is needed, and free biases supply it at every x.
A submersion is transverse to each smooth jet submanifold S. Its preimage is smooth, and regular values of the projection of the preimage to A are exactly the parameters whose jet extensions are transverse to S. This is the parametric transversality argument [1, Ch. 6]. Sard’s theorem gives a null exceptional set. The critical locus has a countable compact exhaustion because it is closed in a second-countable manifold. Its images under the projection are compact null sets and therefore nowhere dense. Taking the countable intersection of the good parameter sets proves the claim. ◻
For a binary classifier, Theorem 1 implies that generically the decision boundary is empty or a codimension-one smooth submanifold, even if ϕ has singular points. The free final bias alone suffices. A fixed-bias family can have additional constraints. For example, if a feature identically vanishes on a region, a bias-free readout may vanish on that entire region. For vector outputs, immersion of ϕ supplies every first-order normal perturbation at a given input. It makes no assertion that independent perturbations at multiple inputs can be realized by the same A˙.
Several classes and several points
Consider q logits and the open part of the tie stratum where a subset I of r classes is exactly the set of maximizers. Fix i0∈I. Equalities yi−yi0=0 for i∈I∖{i0} are r−1 independent linear equations, and strict inequalities separate the other coordinates. This is a smooth codimension-(r−1) submanifold of Rq. Consequently, the inverse image for a generic readout is empty or has codimension r−1. No r-way maximum tie occurs if r>n+1. The conclusion also holds when the bias space is restricted to vectors with coordinate sum zero. The differences of any bias vector are unchanged by subtracting a constant vector so the variation still surjects onto the normal directions to each tie. Thus the ordinary redundancy of adding the same constant to every logit can be removed without losing the relevant perturbations.
There is a complementary global result when ϕ is an embedding. For different points x,x′ the map (A,x,x′)↦A(ϕ(x)−ϕ(x′)) is a submersion to Rq. Indeed, ϕ(x)−ϕ(x′) is nonzero and a linear map can assign it any value. The source in the input variables is M×M∖Δ, of dimension 2n. When q>2n, parametric transversality forces absence of off-diagonal coincidences for a conull residual set of A. By the first-jet argument, r-rank derivative strata of codimension (n−r)(q−r) are avoided when that codimension exceeds n. This formula follows by choosing an invertible r-minor and using block elimination: the remaining (q−r)×(n−r) block after elimination must be zero. In particular for q≥2n all nonimmersive strata are avoided. For compact M and q≥2n+1 the generic readout of an embedding is thus an embedding. An affine bias has no effect on this two-point test. This is a version of the projection argument for embeddings applied to the allowed neural readouts.
A two-parameter boundary discriminant
Use the unit circle as a data manifold with coordinate x=(cost,sint). A network with one square unit and affine readout can realize F(a,b,t)=cost+acos(2t)−b because cos(2t)=2(cost)2−1. Its zero locus in (a,b,t) is smooth since the last bias derivative is −1. Projection to (a,b) is proper because the circle is compact. A parameter belongs to the discriminant when a root also satisfies ∂tF=−sint(1+4acost)=0. Thus the discriminant consists of the two lines b=a+1 and b=a−1 and the curve b=−a−1/(8a) for ∣a∣≥1/4. At ∣a∣=1/4 the latter joins a line with a degenerate critical point. At a=1/4 and t=π the first nonzero variation of cost+acos2t in t−π is quartic, with coefficient 1/8.
Boundary discriminant of the quadratic circle score. The lines come from t=0,π and the curved branches from cost=−1/(4a). The black points mark degenerate critical points. The dashed slice gives the regular counts in the text.
For example, fix a=1/2. With u=cost, the logit equals u2+u−1/2−b. The minimum as a function of u∈[−1,1] occurs at u=−1/2, while the endpoint values before subtracting b are −1/2 and 3/2. Every interior root in u yields two roots on the circle and an endpoint root yields one. The number of boundary points at regular thresholds is therefore
Regular threshold b
Boundary points
(−∞,−3/4)
0
(−3/4,−1/2)
4
(−1/2,3/2)
2
(3/2,∞)
0
At the lower threshold two critical points enter simultaneously because the network score is even in t. At the other thresholds there is a single nondegenerate critical point. Each isolated nondegenerate event has the fold normal form, so it adds or removes an opposite-sign pair of local roots [1, Ch. 20]. The unsigned boundary topology changes, while the signed count at every regular threshold is zero.
More generally, for a compact M and a scalar evaluation F:P×M→R with 0 a regular value, the projection of the zero locus is a proper submersion off its discriminant. It is a smooth bundle over each regular chamber in its image [1, Ch. 20]. Over the other chambers the fibers are empty. Local triviality concerns a whole chamber. It leaves open the possibility of monodromy around a loop in that chamber, even if the point count is fixed. Section 4 computes that effect in a different network.
A stabilizer obstruction to regular readouts
A model with moving parameters can be equivariant even when most of its fixed-parameter maps have few symmetries. Let a compact Lie group G act smoothly on manifolds P,X, and linearly on a finite-dimensional real space V. Suppose a smooth readout F:P×X→V satisfies F(gp,gx)=gF(p,x). A single realization Fp(x)=F(p,x) is equivariant for K=Gp. At a zero x of Fp, the subgroup governing its derivative is H=Kx=Gp∩Gx. The general relation between stabilizers and normal derivatives is developed in [1, Chs. 25–26].
Proposition 2 (Regularity at a parameter stabilizer). If Fp is transverse to zero at x, then TxX/Tx(Kx) admits a surjective H-linear map to V. In particular its σ-isotypic summand has dimension at least that of Vσ for each real irreducible σ of H. Conversely, a surjection of these representations is realizable as the normal derivative of a smooth K-equivariant map defined near the orbit Kx.
Proof. For h∈H, differentiate Fp(hx′)=hFp(x′) at x. The result is DxFp∘Dxh=h∘DxFp. The derivative vanishes on Tx(Kx) since Fp is zero along the orbit. Surjectivity is preserved when the derivative factors through this quotient. Because H is compact, an averaged inner product gives an H-invariant complement of the kernel of a surjection. Therefore a copy of the target sits in the source representation, with the asserted inequality on each isotypic part.
For the converse, identify TxX/Tx(Kx) with the normal representation Nx for an invariant metric. The compact slice theorem gives an invariant neighborhood of Kx modeled on K×HS, where S is an invariant neighborhood of 0 in Nx [1, Ch. 25]. For an H-linear surjection L:Nx→V, the assignment [k,s]↦kL(s) is well defined by equivariance of L. Its normal derivative at [1,0] is L, and translation by K transports this surjectivity along the orbit. ◻
The converse concerns arbitrary equivariant smooth germs. An architecture restricts which such germs can occur. In the total space the parameter derivative contributes an additional H-map TpP→V. Surjectivity at (p,x) tests the sum of its image and the image of the feature derivative. The two regularity tests correspond to different domains of the same evaluation.
Two sign readouts on a torus
Take G=C2 and X=(R/2πZ)2, with the reflection r(θ,ϕ)=(−θ,ϕ). Write R+ and R− for the trivial and sign representations. Use two real sign readouts, so that V=R−2. The fixed locus consists of the two circles θ=0,π. A G-equivariant readout vanishes there because VG=0. On each fixed circle the source tangent representation is R−⊕R+: the normal angle changes sign and the angle along the circle is fixed. Its derivative can have rank at most one by Proposition 2. Every smooth equivariant map of this feature and output type therefore has a critical zero at every point of the fixed locus. Increasing the number of trainable coefficients within this equivariance constraint leaves the representation bound in force. A single sign readout can be regular along the circles, as is demonstrated by the first output below by itself. The obstruction concerns the simultaneous pair of constraints. Increasing output multiplicity makes their joint zero set a different regularity problem.
Consider the specific realization f(θ,ϕ)=(sinθ,sinθcosϕ). The coordinate features u=sinθ and v=cosϕ are respectively sign and trivial features. A square activation realizes their product using uv=((u+v)2−(u−v)2)/4. Under reflection the two inputs u+v,u−v are swapped and simultaneously negated, so their squares are swapped and the difference changes sign. Both readouts therefore have the required transformation law. The zero set is exactly the pair of fixed circles.
Functions and infinitesimal equations
Near a fixed circle, t=sinθ is a transverse coordinate, since its derivative in the angular chart is 1 at θ=0 and −1 at θ=π. Write c=cosϕ and A for the smooth germ ring on the ambient chart at a point of the circle. The equations are t=0 and tc=0. The ordinary ideal is (t), so the ordinary fiber has a smooth one-dimensional function ring on this chart.
For the derived fiber, use the smooth Koszul model [1, Chs. 21–22]. Its cochain algebra is A⊗⋀(η1,η2) with ∣ηi∣=−1, dη1=t, and dη2=tc. The invertible change η2′=η2−cη1 gives dη2′=0. Multiplication by t on smooth germs is injective because t=0 on a dense open subset of the chart. The first generator thus resolves A/(t) and the second gives cohomology H0=A/(t),H−1=(A/(t))η2′. The remaining negative cohomology vanishes. The generator η2′ transforms by sign. The tangent calculation gives the same excess on the dual side: Txder=[R−⊕R+(a,b)↦(a,ca)R−2]. The terms have cohomological degrees zero and one. The kernel is the trivial tangent line of the fixed circle. The quotient of the target by the image is the sign line, explicitly identified by (u1,u2)↦u2−cu1. The virtual dimension is zero. The ordinary dimension is one, and the derived equation record specifies the excess direction and its stabilizer action.
The squared readout loss provides another comparison. With the usual invariant Euclidean metric on the outputs, its local expression is ℓ(t,ϕ)=21∥f(t,ϕ)∥2=21t2(1+cos2ϕ). At t=0 the Hessian has normal eigenvalue 1+cos2ϕ and tangential eigenvalue zero in the chosen coordinate basis. Thus each fixed circle is a smooth critical manifold with positive-definite Hessian on its normal line. This loss detects movement away from the circle. The excess target line survives even though the loss controls every normal displacement. At a zero, the general identity D2ℓ=(Df)∗Df identifies the null directions of the Hessian with kerDf, whereas the derived tangent complex encodes the cokernel in a separate degree.
A moving bias and the symmetric fiber
Take a bias b∈P=V transforming by the same sign action, and set F(b,x)=f(x)−b. The total zero equation is regular because DbF=−IV. Its ordinary total incidence space Z={(b,x):b=f(x)} is the smooth graph of f. Projection Z→P identifies with f:X→V and is singular along its fiber over zero. The parameter b=0 has stabilizer G, and the readout on this slice is exactly the original rank-deficient equivariant map. In contrast, if 0<∣b1∣<1 and ∣b2/b1∣<1, the fiber has four points. The equations prescribe sinθ=b1 and cosϕ=b2/b1, with two choices for each angle. The derivative determinant is −cosθsinθsinϕ, which is nonzero at all four points. At such a bias the parameter stabilizer is trivial.
The quotient also retains the type of the parameter. For this finite group, the smooth action stacks give a map [Z/G]→[P/G] as in [1, Ch. 17]. The zero-parameter orbit in the base is BG=[{0}/G]. In the classical smooth setting the pullback over this orbit is [C/G], where C is the pair of fixed circles. It has the product description C×BG. In derived smooth stacks the homotopy pullback over that orbit is [Z0der/G], with Z0der the derived zero fiber computed above. Indeed, pulling back along the atlas of BG trivializes the torsor and gives the derived incidence square on the manifolds. Gluing this calculation under the torsor’s changes of frame recovers its G-action. Thus the total evaluation can be made regular by the bias, while its symmetric fiber keeps both the stabilizer and the excess equation of the fixed-parameter model.
Decision boundaries with parameter monodromy
The regular fibers of a score can carry a permutation representation of the parameter space’s fundamental group. An invariant computed separately at each parameter can miss this information. We apply the signed-fiber, relative mapping, discriminant, and persistence constructions of [1, Chs. 9, 18, 20, and 28] to one exact network whose decision boundary has this property.
An exact rotation-covariant score
Fix an integer k≥2. Let the input and parameter spaces both be the unit circle in C, and consider s(p,z)=Im(pzk). The real and imaginary coordinates of zk can be computed by a finite network with affine maps and the activation x↦x2. Indeed, ab=41((a+b)2−(a−b)2) implements a real multiplication, hence the real and imaginary parts of a complex multiplication. Iterating this operation computes the indicated power. If a scalar needs to pass through an activation, the identity is x=((x+1)2−(x−1)2)/4. With p=p1+ip2 the final affine layer reads out p1Im(zk)−p2Re(zk). Thus p1,p2 may be treated as the only trainable weights, constrained to the parameter circle. The classifier thresholds the score at zero.
For g∈S1, act on inputs by z↦gz and on parameters by p↦gkp. The feature zk transforms by the character g↦gk, with its real two-dimensional form in the network. The final expression pairs that feature with a parameter of the same character, giving s(gkp,gz)=s(p,z). At a fixed p the rotations stabilizing the parameter are the kth roots of unity, denoted μk. They act on the entire zero fiber of the classifier.
Oriented sheets and their transport
Orient the input circle by increasing angle. In the coordinates p=eiθ and z=eiφ, the score is sin(kφ−θ), with input derivative kcos(kφ−θ). A boundary point is therefore regular and has positive or negative derivative according as zk=p or zk=−p. Write B±={(p,z):zk=±p},B=B+⨿B−, with projection π:B→P=S1.
Proposition 3 (Permutation of boundary points). Each of B+ and B− is a connected k-sheeted covering of P. The positive generator of π1(P) acts on each signed fiber by a k-cycle. Neither covering admits a continuous section.
Proof. Projection onto z identifies each component with a circle. The projection onto p is then z↦zk on the positive component and z↦−zk on the negative one. Over an angular interval the roots can be written as zj±(θ)=exp(ikθ+c±+2πj),c+=0,c−=π, for j reduced modulo k. Increasing θ by 2π sends label j to j+1. A continuous section would select a point fixed by this permutation. Equivalently, its map into the input circle would have a degree which, when multiplied by k, gives 1, contradicting integral degree. ◻
Pullback along the parameter cover r(α)=αk trivializes the positive component. Here h(α,ζ)=(αk,αζ). Choosing a fixed kth root of −1 gives the corresponding trivialization for B−. More generally, pullback along α↦αr for a positive integer r replaces the permutation by j↦j+r. Its cycles have length k/gcd(k,r), and there are gcd(k,r) of them. Thus the cover with r=k is the least positive winding cover that permits a global labeling of all the sheets. In each fiber the signed count is k−k=0. The transport still gives two copies of the cyclic permutation representation. This persists on a range of perturbed networks. Suppose a smooth perturbed score s satisfies, for some 0<η<1, ∥s−s∥∞<η,∥∂φ(s−s)∥∞<k1−η2. At a zero of any linear interpolation between the scores we have ∣s∣<η, hence ∣∂φs∣≥k1−η2. Thus zero remains regular in the input direction and its derivative sign is preserved at the interpolated zeros. The relative zero set is a proper local diffeomorphism over the cylinder of interpolation times and parameters. Path lifting along the interpolation identifies the two signed covers at the ends of the cylinder.
Persistence over the parameter space
Use the closed sublevels Fp(t)={z:s(p,z)≤t}. At t<−1 the fiber is empty. At t=−1 it consists of the k minima. For −1<t<1 it is a disjoint union of k closed arcs, each containing one minimum. At t=1 it becomes the whole circle. Over a field K, the degree-zero barcode therefore has one copy of [−1,∞) and k−1 copies of [−1,1). Degree one has one copy of [1,∞). These statements follow because inclusions before 1 preserve the components, while the inclusion into the circle sends all component classes to the same class. Each p has the same barcodes.
The total sublevels F(t)={(p,z):s(p,z)≤t}→P, with their inclusions for increasing t, define a diagram of spaces over P. A local lift α(p)k=p simultaneously trivializes the diagram by z=α(p)u, since the score becomes Im(uk). The transitions rotate u by elements of μk, and transport around the base permutes the k components for every −1≤t<1. Fiberwise H0 is therefore a local system with permutation module K[μk] in that range. At t=1 the structure map is the augmentation K[μk]→K, where K carries trivial transport. Its kernel gives the local system of reduced degree-zero classes that die. If the characteristic divides k, this augmentation has no equivariant section. An invariant element is a multiple of the sum of the k basis vectors, whose augmentation is zero. This extension is invisible in the barcode of each fiber, where the augmentation is a surjective map of vector spaces. Over C, let ω=e2πi/k, with ej the component basis of H0. The character vector ∑jω−ℓjej acquires the factor ωℓ under monodromy. The trivial character survives the merge at 1 and each nontrivial character dies there. The k−1 numerically equal finite bars consequently carry distinct transport characters.
For parameters p,q, the chord distance δ=∣p−q∣ bounds the uniform score change, since ∣s(p,z)−s(q,z)∣≤∣p−q∣∣zk∣=δ. Thus Fp(t)⊆Fq(t+δ) and Fq(t)⊆Fp(t+δ). The successive inclusion maps are exactly the shift by 2δ, giving a δ-interleaving both before and after taking homology. The same proof applies to any uniform score perturbation. For q-tame persistent homology the stability theorems bound the bottleneck distance by that uniform change [4, 5]. In this rotation family the distance between fiber barcodes is zero. The sublevel diagram over P additionally records which components are identified by going around a loop. For example, the constant family of scores s0(p,z)=Im(zk) has the same barcode at every parameter. Its component local system is trivial, whereas the generator of the parameter loop acts on the components for s by the nonidentity permutation. The two filtered diagrams therefore cannot be equivalent over P. For two general scores on the same parameter and input spaces, a uniform bound on their difference also gives the shifted inclusions of the total sublevels over P. This makes the sublevel diagram over the parameter space a stable comparison object while still retaining the representation.
Reconstruction through a Euclidean latent space
Let X be a nonempty compact smooth manifold without boundary, and fix an embedding i:X↪RD. An encoder e:X→Rm and ambient decoder d:Rm→RD realize the reconstruction x↦d(e(x)). Here the latent width m is the dimension of the intermediate Euclidean space. The encoder and decoder may be evaluation maps of neural architectures at selected parameters. We first allow arbitrary smooth maps to isolate the topological constraints on any such architecture. This applies the tube and factorization constructions of [1, Chs. 8, 24, 27]. For related effects of data topology on autoencoding, see [3].
Exact recovery and smooth families
Theorem 4 (Smooth ambient reconstruction). A smooth encoder e:X→Rm admits a smooth ambient decoder on all of Rm satisfying d∘e=i if and only if e is a smooth embedding. Moreover e extends to a smooth encoder on all of RD agreeing with it along i(X).
Proof. If d∘e=i, injectivity of i forces injectivity of e. Differentiating gives Dxi=De(x)dDxe, so Dxe is injective. The compact-to-Hausdorff criterion makes e a homeomorphism onto its image. That and immersion give an embedding.
Conversely let U⊂Rm be a tubular neighborhood of e(X) with smooth retraction π:U→e(X). On U the map d0=i∘e−1∘π satisfies d0∘e=i. Choose a smooth bump χ equal to one near e(X) and with compact support in U. Set d=χd0 on U and d=0 off U. The support condition makes this extension smooth, and the equality holds along the image. To extend e from i(X) to RD, apply the same construction to e∘i−1 using a tube about i(X). ◻
Consequently the least possible exact latent width in unrestricted smooth ambient reconstruction is the smooth Euclidean embedding dimension of X. Derivative rank alone does not determine that width. For example, the double traversal z↦z2 from S1 to the unit circle in R2 is a smooth immersion, but a decoder cannot separate the two points at any encoded value. The tube construction supplies a smooth map in the ambient model class. Exact realization by a prescribed finite neural architecture requires a separate representability argument.
Corollary 5 (A decoder varying smoothly with parameters). Let P be a smooth manifold and e:P×X→Rm a smooth family. There exists a jointly smooth d:P×Rm→RD with d(p,e(p,x))=i(x) for all p,x if and only if each ep is a smooth embedding.
Proof. Necessity follows by restricting to any p. For sufficiency, the map j(p,x)=(p,e(p,x)) is an injective immersion, since a vector in its derivative kernel has zero parameter component and then zero X-component. The preimage of a compact set K is closed in the compact set prP(K)×X. Thus j is proper and is a closed embedding. Take a smooth tubular retraction π to its image in P×Rm. On this tube use i∘prX∘j−1∘π. A smooth cutoff equal to one near the closed image and supported in the tube lets us extend this map by zero. The cutoff can be chosen by a partition of unity subordinate to the tube and the complement of the image, with locally finite supports when P is noncompact. Its value on j(p,x) is i(x), as required. ◻
Uniform approximation and a sharp spherical bound
The existence of a uniform error gap extends beyond exact recovery. Suppose dimX=n>0. Choose a tubular neighborhood U⊂RD of i(X) and a number δ>0 such that all points at distance less than δ from i(X) belong to U. Such a number exists by compactness.
Proposition 6 (A closed-manifold error gap). If m≤n, every continuous encoder and ambient decoder satisfy supx∈X∥d(e(x))−i(x)∥≥δ.
Proof. If the supremum were less than δ, the straight segment from i(x) to d(e(x)) would lie in U for every x. Put V=d−1(U), an open subset of Rm containing e(X), and let r:U→X be the tubular retraction with r∘i=idX. Applying r to those segments gives a homotopy to the identity for the composite XeVd∣VUrX. With coefficients in F2, the identity on Hn(X) would then factor through Hn(V). For m>0, every component of V is a noncompact m-manifold, whose ordinary homology vanishes in degrees at least m [7, Prop. 3.29]. For m=0 its positive-degree homology also vanishes. The mod-two fundamental class of any component of X is nonzero, yielding a contradiction. ◻
The use of the open latent subset V is essential because the decoder may leave the tubular neighborhood elsewhere in latent space. The argument applies to continuous layers, including piecewise smooth realizations. It bounds uniform error for an entire data manifold and therefore also bounds each eligible member of a parameter family.
Proposition 7 (Precise bounds on the unit sphere). Take X=Sn and let i:Sn↪Rn+1 be the standard unit sphere, with n≥1. For m≤n, the optimal uniform error among continuous encoders and ambient decoders is 1. For m≥n+1, exact ambient recovery is possible. If the decoder is constrained to take values in Sn on all of Rm, the uniform error is 2 for every width m.
Proof. For m≤n, apply Borsuk–Ulam to the composite of the encoder and the coordinate inclusion of Rm in Rn. It gives an x with e(x)=e(−x) [9]. If u is their common decoded value, 2≤∥x−u∥+∥−x−u∥. The zero decoder attains error 1. For m≥n+1, a coordinate inclusion and projection give exact recovery. Finally a sphere-valued reconstruction through all of Rm is nullhomotopic. If it avoided antipodal coincidences, its straight segments to the identity, normalized to unit length, would give a homotopy to the identity. This would contradict its zero degree. Hence it reaches error 2, the largest distance between points of the sphere. ◻
The optimal errors for the unit sphere are summarized below.
Decoder range
m≤n
m≥n+1
Ambient Rn+1
1
0
Sphere Sn
2
2
For a sphere of radius R>0, rescaling replaces the positive entries by R and 2R, respectively. The bound holds for every parameter of an appropriate family. Its proof chooses an antipodal pair separately at each parameter and needs no continuous choice of witnesses.
Uniform and average losses have different implications. For example, at width n choose stereographic coordinates σ from the complement of the north pole of Sn to Rn. Multiply σ by a smooth cutoff equal to one outside a small polar cap C and equal to zero near the pole, extending the encoder by zero there. The smooth decoder d=σ−1 gives exact recovery off C. For normalized spherical measure μ, the squared error averaged over the sphere is at most 4μ(C), which can be made arbitrarily small while the uniform error is 2. Thus the uniform obstruction does not force a positive lower bound on this average across all smooth maps of width n.
A uniform Lipschitz bound prevents this concentration. If the reconstruction has Lipschitz constant at most L with respect to chord distance, its error function has Lipschitz constant at most 1+L. For ambient width m≤n, some point has error at least 1, so every point within chord distance 1/(2(1+L)) of it has error at least 1/2. Its mean squared error is at least one quarter of the normalized measure of such a cap, a positive bound at fixed L. This estimate states the additional analytic hypothesis under which the pointwise obstruction controls a population loss.
Conclusion
The analysis separates three conditions on a parametrized network. The available weight directions determine where the total jet map is transverse. The symmetry representation determines which normal derivatives a fixed realization can have at a stabilizer. The topology of the domain and latent factorization determines which global reconstructions are possible. A model can satisfy one of these conditions while exhibiting an obstruction governed by another.
Examples on the circle and torus make these distinctions explicit. Ordinary point counts jump at the fold discriminant. They can also remain constant while boundary sheets are permuted along a regular parameter loop. The resulting persistence module over the parameter retains that permutation representation. In the equivariant torus model the squared residual controls normal movement, while the tangent complex records an excess target direction. The autoencoder estimates bound uniform reconstruction error and give a separate mean-square bound under a Lipschitz constraint.
Relative evaluation and descent explain how these calculations pass between local presentations of a model. The assumptions used for the calculations are geometric: smoothness of the relevant charts, specified perturbation directions, properness for topology of fibers, and an explicit choice of decoder range. They apply to a realized architecture once those hypotheses have been checked.
References
Anonymous. Differential Topology in Higher Categories: From Regular Values and Cobordism to Derived Smooth Families. Submitted manuscript, 2026.
Anonymous. Representation theory of parametrized neural models on spaces. Higher Categorical Semantics, 2026. Published article.
J. Batson, C. G. Haaf, Y. Kahn, and D. A. Roberts. Topological obstructions to autoencoding. Journal of High Energy Physics 04 (2021), 280. DOI.
F. Chazal, V. de Silva, M. Glisse, and S. Oudot. The Structure and Stability of Persistence Modules. Springer, 2016. DOI.
D. Cohen-Steiner, H. Edelsbrunner, and J. Harer. Stability of persistence diagrams. Discrete & Computational Geometry 37 (2007), 103–120. DOI.
B. Fong, D. I. Spivak, and R. Tuyéras. Backprop as functor: A compositional perspective on supervised learning. In Proceedings of LICS, 2019. Author manuscript.
A. Hatcher. Algebraic Topology. Cambridge University Press, 2002. Author’s copy.
M. W. Hirsch. Differential Topology. Graduate Texts in Mathematics 33. Springer, 1976. Publisher’s edition.
J. Matoušek. Using the Borsuk–Ulam Theorem. Springer, 2003. DOI.
G. Naitzat, A. Zhitnikov, and L.-H. Lim. Topology of deep neural networks. Journal of Machine Learning Research 21(184) (2020), 1–40. Journal article.
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,balance}
\usepackage{xurl}
\usepackage{hyperref}
\usepackage{bookmark}
\definecolor{paperink}{HTML}{202824}
\definecolor{paperaccent}{HTML}{145A51}
\definecolor{papermuted}{HTML}{51655F}
\definecolor{paperrule}{HTML}{B6CBC4}
\hypersetup{
pdftitle={Differential Topology of Parametrized Neural Networks},
pdfsubject={Jet representations, decision boundaries, and reconstruction},
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}Differential topology of parametrized neural networks}
\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{27}{32}\selectfont\bfseries
Differential Topology of\\Parametrized Neural Networks\par}
\vspace{8pt}
{\noindent\large\color{papermuted}
Jet representations, decision boundaries, and reconstruction\par}
\vspace{10pt}
{\noindent\small November 2026\par}
\vspace{9pt}
\noindent\begin{minipage}{\textwidth}\small
\textbf{Abstract.}
A smooth neural architecture represents a space of weights by evaluations
on a data manifold. We give architecture-level conditions for controlling
first jets, decision boundaries, and reconstruction. An affine readout of
immersed features is submersive on first jets, while an arbitrary smooth
feature map with free final biases still gives generic transverse
classification boundaries. We compute a fold discriminant for a
quadratic readout on the circle and a regular family whose decision
boundaries have nontrivial monodromy. For equivariant layers the
stabilizer representation can force an excess intersection even when
the total evaluation is submersive. An explicit torus calculation
exhibits the associated obstruction complex. For autoencoders, exact
ambient recovery requires an embedding, and a spherical example gives
a sharp uniform error bound below the required latent dimension.
Relative mapping objects and descent specify how these results apply
to local parameter and feature presentations.
\par\medskip
\noindent\textbf{Keywords.}\enspace
Parametrized networks, transversality, representation theory, singularities,
monodromy, autoencoders.
\end{minipage}
\vspace{17pt}
\end{@twocolumnfalse}
]
\thispagestyle{plain}
\section{Networks as families of smooth maps}
\label{sec:families}
A parametrized neural network is a family of maps realized by a
specified architecture. Differential topology studies these maps
through their derivatives, their fibers, and their behavior along
paths of parameters. This makes it possible to ask which local
perturbations a layer actually supplies and which global topological
constraints remain even if local perturbations are plentiful.
A basic example is an encoder with a smooth, trainable readout.
The readout can have enough derivatives to move a decision boundary,
while a narrow intermediate space still obstructs reconstruction
on the whole data manifold.
The starting points are Hirsch's treatment of jets and transversality
\cite[Chs.~2--4]{hirsch} and their extension to relative families,
derived pullbacks, and equivariant representations in
\cite[Chs.~18--22, 25--28]{higher-hirsch}.
Earlier categorical models of learning represent layers by parametrized
maps, including their compositions and shared weights
\cite{fong-spivak-tuyeras,previous-networks}. Here the map from a
prescribed architecture to the function object is subjected to
differential-topological tests. Related empirical work investigates
changes in data topology across network layers \cite{naitzat}.
Our statements concern smooth realizations and their included
perturbations.
\subsection{Relative evaluation and representations}
Let $\mathcal E$ be the $\infty$-topos of sheaves of spaces on
smooth test manifolds, and let $B$ be a coefficient or geometry
base. For $X,Y\to B$ write $[X,Y]_B$ for the internal mapping
object in $\mathcal E_{/B}$. A model with weights $P\to B$ is an
evaluation
\[
F:P\times_B X\longrightarrow Y
\]
over $B$, or its transpose $P\to[X,Y]_B$. For example,
if $B=*$ and $X,Y$ are represented by smooth manifolds, an
evaluation on $P$ is a smooth map $P\times X\to Y$ on ordinary
representable parameter charts. The classifying map records
the realizations available to that architecture. Two different
points of $P$ can specify the same evaluation, and a change
of coordinates acts on the presenting parameter space.
A parameter shared by multiple factors of the network is a
pullback along the corresponding diagonal into the product
of their parameter spaces.
Internal exponentials commute with base change in $B$.
For an effective epimorphism $U\to B$ with \v{C}ech nerve
$U_\bullet$, we have the descent equivalence
\[
\E_{/[X,Y]_B}\simeq
\lim_{[n]\in\boldsymbol\Delta}
\E_{/[X|_{U_n},Y|_{U_n}]_{U_n}}
\]
from \cite[Ch.~18]{higher-hirsch}. Thus compatible evaluations and
parameter presentations on geometry charts glue. On charts of
smooth bundles, a change of input and output coordinates transforms
the vertical derivative by the derivative of the output change on
the left and the inverse input change on the right. It preserves rank,
so rank strata and regularity conditions are intrinsic to the
glued model. For an action, the same formula is conjugation by
the linear representations on tangent spaces. Its restriction to
stabilizers becomes a concrete selection rule in
Section~\ref{sec:eq-readout}.
\subsection{Which kind of realization is measured}
A differentiable classifier with real logit $f_p$ has a
decision boundary $f_p^{-1}(0)$ and sublevel sets
$f_p^{-1}((-\infty,c])$. A scalar strictly monotone link, such
as the logistic link, preserves the boundary when the corresponding
threshold is transported. A classifier with several logits admits
a stratification by the indices realizing the maximum. An autoencoder,
by contrast, is a factorization of an approximation to the
identity on a data manifold. Its topology depends on the encoder's
image in the latent space and on the codomain of the decoder.
We use smooth activations for all jet and regular-value assertions.
The uniform reconstruction obstruction requires only continuous
realizations, so it also applies to continuous piecewise-linear
activations.
\par
\section{Readouts and the available jet directions}
\label{sec:readout-jets}
Let $M$ be an $n$-dimensional smooth manifold without boundary.
Fix a smooth feature extractor $\phi:M\to\mathbb R^d$ and
use the affine readout
\begin{equation}
\label{eq:affine-readout}
f_{A,b}(x)=A\phi(x)+b,\qquad
(A,b)\in\mathcal A
=\operatorname{Hom}(\mathbb R^d,\mathbb R^q)\times\mathbb R^q .
\end{equation}
The choice to hold the feature extractor fixed specifies a subfamily
of whatever full architecture realizes $\phi$. This separates the
derivatives supplied by the readout from those provided by
perturbing upstream layers.
\begin{theorem}[First-jet control by an affine readout]
\label{thm:jet-surjectivity}
If $\phi$ is an immersion, then
$J:\mathcal A\times M\to J^1(M,\mathbb R^q)$,
$(A,b,x)\mapsto j^1_x f_{A,b}$, is a submersion.
Without the immersion assumption, the corresponding map to
$J^0(M,\mathbb R^q)$ remains a submersion. In either case, for a
finite or countable family of smooth jet submanifolds at the
indicated order, the parameters for which the corresponding
jet section is transverse to all of them form a conull residual
set.
\end{theorem}
\begin{proof}
At fixed $x$, use jet coordinates $(y,L)$ with
$L:T_xM\to\mathbb R^q$. The differential in the readout
directions is
\[
(\dot A,\dot b)
\longmapsto
\bigl(\dot A\phi(x)+\dot b,\ \dot A D\phi_x\bigr).
\]
Given any $L_0:T_xM\to\mathbb R^q$ and immersive $\phi$, extend
$L_0(D\phi_x)^{-1}$ linearly from $D\phi_x(T_xM)$ to
$\mathbb R^d$ to obtain $\dot A$. Then any value variation
$y_0$ is realized by $\dot b=y_0-\dot A\phi(x)$.
This proves surjectivity in the vertical jet coordinates.
Base coordinates are covered because $J$ lies over the
projection to $M$. For $J^0$ only the value variation is
needed, and free biases supply it at every $x$.
A submersion is transverse to each smooth jet submanifold $S$.
Its preimage is smooth, and regular values of the projection
of the preimage to $\mathcal A$ are exactly the parameters
whose jet extensions are transverse to $S$. This is the
parametric transversality argument
\cite[Ch.~6]{higher-hirsch}. Sard's theorem gives a null
exceptional set. The critical locus has a countable compact
exhaustion because it is closed in a second-countable
manifold. Its images under the projection are compact
null sets and therefore nowhere dense. Taking the countable
intersection of the good parameter sets proves the claim.
\end{proof}
For a binary classifier, Theorem~\ref{thm:jet-surjectivity}
implies that generically the decision boundary is empty or a
codimension-one smooth submanifold, even if $\phi$ has
singular points. The free final bias alone suffices. A
fixed-bias family can have additional constraints.
For example, if a feature identically vanishes on a
region, a bias-free readout may vanish on that entire region.
For vector outputs, immersion of $\phi$ supplies every
first-order normal perturbation at a given input. It makes
no assertion that independent perturbations at multiple
inputs can be realized by the same $\dot A$.
\subsection{Several classes and several points}
Consider $q$ logits and the open part of the tie stratum
where a subset $I$ of $r$ classes is exactly the set of
maximizers. Fix $i_0\in I$. Equalities
$y_i-y_{i_0}=0$ for $i\in I\setminus\{i_0\}$ are $r-1$
independent linear equations, and strict inequalities
separate the other coordinates. This is a smooth
codimension-$(r-1)$ submanifold of $\mathbb R^q$.
Consequently, the inverse image for a generic readout
is empty or has codimension $r-1$. No $r$-way maximum
tie occurs if $r>n+1$. The conclusion also holds when
the bias space is restricted to vectors with coordinate
sum zero. The differences of any bias vector are
unchanged by subtracting a constant vector so the
variation still surjects onto the normal directions to
each tie. Thus the ordinary redundancy of adding the
same constant to every logit can be removed without
losing the relevant perturbations.
There is a complementary global result when $\phi$
is an embedding. For different points $x,x'$ the
map $(A,x,x')\mapsto A(\phi(x)-\phi(x'))$ is a submersion
to $\mathbb R^q$. Indeed, $\phi(x)-\phi(x')$ is nonzero
and a linear map can assign it any value. The source
in the input variables is $M\times M\setminus\Delta$,
of dimension $2n$. When $q>2n$, parametric
transversality forces absence of off-diagonal coincidences
for a conull residual set of $A$. By the first-jet argument,
$r$-rank derivative strata of codimension
$(n-r)(q-r)$ are avoided when that codimension exceeds $n$.
This formula follows by choosing an invertible $r$-minor
and using block elimination: the remaining
$(q-r)\times(n-r)$ block after elimination must be zero.
In particular for $q\ge2n$ all nonimmersive strata
are avoided. For compact $M$ and $q\ge2n+1$ the generic
readout of an embedding is thus an embedding. An affine
bias has no effect on this two-point test. This is a version
of the projection argument for embeddings applied to
the allowed neural readouts.
\subsection{A two-parameter boundary discriminant}
\label{sec:threshold-discriminant}
Use the unit circle as a data manifold with coordinate
$x=(\cos t,\sin t)$. A network with one square unit and
affine readout can realize
\[
F(a,b,t)=\cos t+a\cos(2t)-b
\]
because $\cos(2t)=2(\cos t)^2-1$.
Its zero locus in $(a,b,t)$ is smooth since the last bias
derivative is $-1$. Projection to $(a,b)$ is proper
because the circle is compact. A parameter belongs
to the discriminant when a root also satisfies
\[
\partial_t F=-\sin t(1+4a\cos t)=0.
\]
Thus the discriminant consists of the two lines
$b=a+1$ and $b=a-1$ and the curve
$b=-a-1/(8a)$ for $|a|\ge1/4$.
At $|a|=1/4$ the latter joins a line with a degenerate
critical point. At $a=1/4$ and $t=\pi$ the first nonzero
variation of $\cos t+a\cos 2t$ in $t-\pi$ is quartic,
with coefficient $1/8$.
\begin{figure}[tbp]
\centering
\begin{tikzpicture}[x=2.3cm,y=.86cm,font=\small]
\draw[->,gray!85] (-1.04,0) -- (1.08,0) node[right] {$a$};
\draw[->,gray!85] (0,-2.02) -- (0,2.03) node[above] {$b$};
\draw[paperaccent,thick,domain=-.94:.94] plot (\x,{\x+1});
\draw[paperaccent,thick,domain=-.94:.94] plot (\x,{\x-1});
\draw[orange!62!black,thick,domain=.25:.94,samples=60]
plot (\x,{-\x-1/(8*\x)});
\draw[orange!62!black,thick,domain=-.94:-.25,samples=60]
plot (\x,{-\x-1/(8*\x)});
\draw[dashed,gray!85] (.5,-1.96) -- (.5,1.90);
\node[anchor=west] at (.54,.32) {$a=\tfrac12$};
\fill[paperaccent] (.5,1.5) circle (1.3pt);
\fill[paperaccent] (.5,-.5) circle (1.3pt);
\fill[orange!62!black] (.5,-.75) circle (1.3pt);
\fill[paperink] (.25,-.75) circle (1.25pt);
\fill[paperink] (-.25,.75) circle (1.25pt);
\end{tikzpicture}
\caption{Boundary discriminant of the quadratic circle score.
The lines come from $t=0,\pi$ and the curved branches from
$\cos t=-1/(4a)$. The black points mark degenerate critical
points. The dashed slice gives the regular counts in the text.}
\label{fig:boundary-discriminant}
\end{figure}
For example, fix $a=1/2$. With $u=\cos t$, the logit
equals $u^2+u-1/2-b$. The minimum as a function of
$u\in[-1,1]$ occurs at $u=-1/2$, while the endpoint
values before subtracting $b$ are $-1/2$ and $3/2$.
Every interior root in $u$ yields two roots on the
circle and an endpoint root yields one. The number
of boundary points at regular thresholds is therefore
\begin{center}
\begin{tabular}{@{}lc@{}}
\toprule
Regular threshold $b$ & Boundary points \\
\midrule
$(-\infty,-3/4)$ & $0$\\
$(-3/4,-1/2)$ & $4$\\
$(-1/2,3/2)$ & $2$\\
$(3/2,\infty)$ & $0$\\
\bottomrule
\end{tabular}
\end{center}
At the lower threshold two critical points enter
simultaneously because the network score is even in $t$.
At the other thresholds there is a single nondegenerate
critical point. Each isolated nondegenerate event has
the fold normal form, so it adds or removes an opposite-sign
pair of local roots \cite[Ch.~20]{higher-hirsch}.
The unsigned boundary topology changes, while the signed
count at every regular threshold is zero.
More generally, for a compact $M$ and a scalar
evaluation $F:P\times M\to\mathbb R$ with $0$ a regular
value, the projection of the zero locus is a proper
submersion off its discriminant. It is a smooth
bundle over each regular chamber in its image
\cite[Ch.~20]{higher-hirsch}. Over the other chambers the
fibers are empty. Local triviality concerns a whole chamber. It leaves
open the possibility of monodromy around a loop in that
chamber, even if the point count is fixed. Section~\ref{sec:boundary-monodromy}
computes that effect in a different network.
\par
\section{A stabilizer obstruction to regular readouts}
\label{sec:eq-readout}
A model with moving parameters can be equivariant even when
most of its fixed-parameter maps have few symmetries. Let a
compact Lie group $G$ act smoothly on manifolds $P,X$, and
linearly on a finite-dimensional real space $V$. Suppose a
smooth readout $F:P\times X\to V$ satisfies
\[
F(gp,gx)=gF(p,x).
\]
A single realization $F_p(x)=F(p,x)$ is equivariant for
$K=G_p$. At a zero $x$ of $F_p$, the subgroup governing its
derivative is $H=K_x=G_p\cap G_x$. The general relation between
stabilizers and normal derivatives is developed in
\cite[Chs.~25--26]{higher-hirsch}.
\begin{proposition}[Regularity at a parameter stabilizer]
\label{prop:eq-readout}
If $F_p$ is transverse to zero at $x$, then
$T_xX/T_x(Kx)$ admits a surjective $H$-linear map to $V$.
In particular its $\sigma$-isotypic summand has dimension
at least that of $V_\sigma$ for each real irreducible
$\sigma$ of $H$. Conversely, a surjection of these
representations is realizable as the normal derivative of
a smooth $K$-equivariant map defined near the orbit $Kx$.
\end{proposition}
\begin{proof}
For $h\in H$, differentiate $F_p(hx')=hF_p(x')$ at $x$.
The result is $D_x F_p\circ D_xh=h\circ D_xF_p$.
The derivative vanishes on $T_x(Kx)$ since $F_p$ is zero
along the orbit. Surjectivity is preserved when the derivative
factors through this quotient. Because $H$ is compact, an
averaged inner product gives an $H$-invariant complement of
the kernel of a surjection. Therefore a copy of the target
sits in the source representation, with the asserted
inequality on each isotypic part.
For the converse, identify $T_xX/T_x(Kx)$ with the normal
representation $N_x$ for an invariant metric. The compact
slice theorem gives an invariant neighborhood of $Kx$
modeled on $K\times_H S$, where $S$ is an invariant
neighborhood of $0$ in $N_x$
\cite[Ch.~25]{higher-hirsch}. For an $H$-linear surjection
$L:N_x\to V$, the assignment $[k,s]\mapsto kL(s)$ is
well defined by equivariance of $L$. Its normal derivative
at $[1,0]$ is $L$, and translation by $K$ transports this
surjectivity along the orbit.
\end{proof}
The converse concerns arbitrary equivariant smooth germs.
An architecture restricts which such germs can occur.
In the total space the parameter derivative contributes an
additional $H$-map $T_pP\to V$. Surjectivity at $(p,x)$
tests the sum of its image and the image of the feature
derivative. The two regularity tests correspond to different
domains of the same evaluation.
\subsection{Two sign readouts on a torus}
Take $G=C_2$ and $X=(\mathbb R/2\pi\mathbb Z)^2$, with
the reflection $r(\theta,\phi)=(-\theta,\phi)$. Write
$\mathbb R_+$ and $\mathbb R_-$ for the trivial and sign
representations. Use two real sign readouts, so that
$V=\mathbb R_-^2$. The fixed locus consists of the two
circles $\theta=0,\pi$. A $G$-equivariant readout vanishes
there because $V^G=0$. On each fixed circle the source
tangent representation is $\mathbb R_-\oplus\mathbb R_+$:
the normal angle changes sign and the angle along the circle
is fixed. Its derivative can have rank at most one by
Proposition~\ref{prop:eq-readout}. Every smooth equivariant
map of this feature and output type therefore has a critical
zero at every point of the fixed locus. Increasing the
number of trainable coefficients within this equivariance
constraint leaves the representation bound in force. A
single sign readout can be regular along the circles, as
is demonstrated by the first output below by itself.
The obstruction concerns the simultaneous pair of
constraints. Increasing output multiplicity makes their
joint zero set a different regularity problem.
Consider the specific realization
\[
f(\theta,\phi)=(\sin\theta,\ \sin\theta\cos\phi).
\]
The coordinate features $u=\sin\theta$ and $v=\cos\phi$
are respectively sign and trivial features. A square
activation realizes their product using
$uv=((u+v)^2-(u-v)^2)/4$. Under reflection the two inputs
$u+v,u-v$ are swapped and simultaneously negated, so their
squares are swapped and the difference changes sign.
Both readouts therefore have the required transformation
law. The zero set is exactly the pair of fixed circles.
\subsection{Functions and infinitesimal equations}
Near a fixed circle, $t=\sin\theta$ is a transverse
coordinate, since its derivative in the angular chart is
$1$ at $\theta=0$ and $-1$ at $\theta=\pi$. Write
$c=\cos\phi$ and $A$ for the smooth germ ring on the
ambient chart at a point of the circle. The equations
are $t=0$ and $tc=0$. The ordinary ideal is $(t)$,
so the ordinary fiber has a smooth one-dimensional
function ring on this chart.
For the derived fiber, use the smooth Koszul model
\cite[Chs.~21--22]{higher-hirsch}.
Its cochain algebra is $A\otimes\bigwedge(\eta_1,\eta_2)$
with $|\eta_i|=-1$, $d\eta_1=t$, and $d\eta_2=tc$.
The invertible change $\eta_2'=\eta_2-c\eta_1$ gives
$d\eta_2'=0$. Multiplication by $t$ on smooth germs is
injective because $t\ne0$ on a dense open subset of the
chart. The first generator thus resolves $A/(t)$ and
the second gives cohomology
\[
H^0=A/(t),\qquad
H^{-1}=(A/(t))\eta_2'.
\]
The remaining negative cohomology vanishes. The generator
$\eta_2'$ transforms by sign. The tangent calculation gives
the same excess on the dual side:
\[
T_x^{\rm der}
=\bigl[\,
\mathbb R_-\oplus\mathbb R_+
\xrightarrow{\ (a,b)\mapsto(a,ca)\ }
\mathbb R_-^2
\,\bigr].
\]
The terms have cohomological degrees zero and one.
The kernel is the trivial tangent line of the fixed circle.
The quotient of the target by the image is the sign line,
explicitly identified by $(u_1,u_2)\mapsto u_2-cu_1$.
The virtual dimension is zero. The ordinary dimension
is one, and the derived equation record specifies the excess
direction and its stabilizer action.
The squared readout loss provides another comparison.
With the usual invariant Euclidean metric on the outputs,
its local expression is
\[
\ell(t,\phi)=\tfrac12\|f(t,\phi)\|^2
=\tfrac12 t^2(1+\cos^2\phi).
\]
At $t=0$ the Hessian has normal eigenvalue
$1+\cos^2\phi$ and tangential eigenvalue zero in the
chosen coordinate basis. Thus each fixed circle is a
smooth critical manifold with positive-definite Hessian
on its normal line. This loss detects movement away from
the circle. The excess target line survives even though
the loss controls every normal displacement. At a zero,
the general identity $D^2\ell=(Df)^*Df$ identifies the
null directions of the Hessian with $\ker Df$, whereas
the derived tangent complex encodes the cokernel in a
separate degree.
\subsection{A moving bias and the symmetric fiber}
Take a bias $b\in P=V$ transforming by the same sign action,
and set $F(b,x)=f(x)-b$. The total zero equation is regular
because $D_bF=-I_V$. Its ordinary total incidence space
$Z=\{(b,x):b=f(x)\}$ is the smooth graph of $f$.
Projection $Z\to P$ identifies with $f:X\to V$ and is
singular along its fiber over zero. The parameter $b=0$
has stabilizer $G$, and the readout on this slice is
exactly the original rank-deficient equivariant map.
In contrast, if $0<|b_1|<1$ and $|b_2/b_1|<1$,
the fiber has four points. The equations prescribe
$\sin\theta=b_1$ and $\cos\phi=b_2/b_1$, with two choices
for each angle. The derivative determinant is
$-\cos\theta\sin\theta\sin\phi$, which is nonzero at
all four points. At such a bias the parameter stabilizer
is trivial.
The quotient also retains the type of the parameter.
For this finite group, the smooth action stacks give
a map $[Z/G]\to[P/G]$ as in
\cite[Ch.~17]{higher-hirsch}. The zero-parameter orbit
in the base is $BG=[\{0\}/G]$. In the classical smooth
setting the pullback over this orbit is $[C/G]$,
where $C$ is the pair of fixed circles. It has the
product description $C\times BG$. In derived smooth
stacks the homotopy pullback over that orbit is
$[Z_0^{\rm der}/G]$, with $Z_0^{\rm der}$ the derived
zero fiber computed above. Indeed, pulling back along
the atlas of $BG$ trivializes the torsor and gives
the derived incidence square on the manifolds. Gluing
this calculation under the torsor's changes of frame
recovers its $G$-action. Thus the total evaluation can
be made regular by the bias, while its symmetric
fiber keeps both the stabilizer and the excess
equation of the fixed-parameter model.
\par
\section{Decision boundaries with parameter monodromy}
\label{sec:boundary-monodromy}
The regular fibers of a score can carry a permutation
representation of the parameter space's fundamental group.
An invariant computed separately at each parameter can miss
this information. We apply the signed-fiber, relative mapping,
discriminant, and persistence constructions of
\cite[Chs.~9, 18, 20, and~28]{higher-hirsch} to one exact
network whose decision boundary has this property.
\subsection{An exact rotation-covariant score}
Fix an integer $k\geq2$. Let the input and parameter spaces
both be the unit circle in $\mathbb C$, and consider
\begin{equation}
s(p,z)=\operatorname{Im}(\overline p\,z^k).
\label{eq:circle-network-score}
\end{equation}
The real and imaginary coordinates of $z^k$ can be computed
by a finite network with affine maps and the activation
$x\mapsto x^2$. Indeed,
\[
ab=\tfrac14\bigl((a+b)^2-(a-b)^2\bigr)
\]
implements a real multiplication, hence the real and imaginary
parts of a complex multiplication. Iterating this operation
computes the indicated power. If a scalar needs to pass through
an activation, the identity is
$x=((x+1)^2-(x-1)^2)/4$. With $p=p_1+ip_2$ the final affine
layer reads out $p_1\operatorname{Im}(z^k)-p_2\operatorname{Re}(z^k)$.
Thus $p_1,p_2$ may be treated as the only trainable weights,
constrained to the parameter circle. The classifier thresholds
the score at zero.
For $g\in S^1$, act on inputs by $z\mapsto gz$ and on
parameters by $p\mapsto g^kp$. The feature $z^k$ transforms
by the character $g\mapsto g^k$, with its real two-dimensional
form in the network. The final expression pairs that feature
with a parameter of the same character, giving
$s(g^kp,gz)=s(p,z)$. At a fixed $p$ the rotations stabilizing
the parameter are the $k$th roots of unity, denoted $\mu_k$.
They act on the entire zero fiber of the classifier.
\subsection{Oriented sheets and their transport}
Orient the input circle by increasing angle. In the
coordinates $p=e^{i\theta}$ and $z=e^{i\varphi}$, the score
is $\sin(k\varphi-\theta)$, with input derivative
$k\cos(k\varphi-\theta)$. A boundary point is therefore
regular and has positive or negative derivative according
as $z^k=p$ or $z^k=-p$. Write
\[
B^\pm=\{(p,z):z^k=\pm p\},\qquad
B=B^+\amalg B^-,
\]
with projection $\pi:B\to P=S^1$.
\begin{proposition}[Permutation of boundary points]
\label{prop:circle-boundary-transport}
Each of $B^+$ and $B^-$ is a connected $k$-sheeted
covering of $P$. The positive generator of $\pi_1(P)$
acts on each signed fiber by a $k$-cycle. Neither covering
admits a continuous section.
\end{proposition}
\begin{proof}
Projection onto $z$ identifies each component with a circle.
The projection onto $p$ is then $z\mapsto z^k$ on the positive
component and $z\mapsto -z^k$ on the negative one. Over an
angular interval the roots can be written as
\[
z_j^\pm(\theta)
=\exp\!\left(i\,\frac{\theta+c_\pm+2\pi j}{k}\right),
\qquad c_+=0,\quad c_-=\pi,
\]
for $j$ reduced modulo $k$. Increasing $\theta$ by $2\pi$
sends label $j$ to $j+1$. A continuous section would select
a point fixed by this permutation. Equivalently, its map
into the input circle would have a degree which, when
multiplied by $k$, gives $1$, contradicting integral degree.
\end{proof}
Pullback along the parameter cover $r(\alpha)=\alpha^k$
trivializes the positive component.
\[
\begin{tikzcd}[column sep=2.1em,row sep=1.7em]
S^1\times\mu_k \arrow[r,"h"] \arrow[d,"\operatorname{pr}_1"'] &
B^+ \arrow[d,"\pi"]\\
S^1 \arrow[r,"r"'] & P
\end{tikzcd}
\]
Here $h(\alpha,\zeta)=(\alpha^k,\alpha\zeta)$. Choosing a
fixed $k$th root of $-1$ gives the corresponding
trivialization for $B^-$. More generally, pullback along
$\alpha\mapsto\alpha^r$ for a positive integer $r$ replaces
the permutation by $j\mapsto j+r$. Its cycles have length
$k/\gcd(k,r)$, and there are $\gcd(k,r)$ of them. Thus the
cover with $r=k$ is the least positive winding cover that
permits a global labeling of all the sheets. In each fiber
the signed count is $k-k=0$. The transport still gives two copies of the cyclic
permutation representation. This persists on a range of
perturbed networks. Suppose a smooth perturbed score
$\widetilde s$ satisfies, for some $0<\eta<1$,
\[
\|\widetilde s-s\|_\infty<\eta,\qquad
\|\partial_\varphi(\widetilde s-s)\|_\infty
< k\sqrt{1-\eta^2}.
\]
At a zero of any linear interpolation between the scores
we have $|s|<\eta$, hence
$|\partial_\varphi s|\geq k\sqrt{1-\eta^2}$.
Thus zero remains regular in the input direction and its
derivative sign is preserved at the interpolated zeros.
The relative zero set is a proper local diffeomorphism
over the cylinder of interpolation times and parameters.
Path lifting along the interpolation identifies the two
signed covers at the ends of the cylinder.
\subsection{Persistence over the parameter space}
Use the closed sublevels
$F_p(t)=\{z:s(p,z)\leq t\}$. At $t<-1$ the fiber is empty.
At $t=-1$ it consists of the $k$ minima. For $-1<t<1$
it is a disjoint union of $k$ closed arcs, each containing
one minimum. At $t=1$ it becomes the whole circle.
Over a field $\mathbb K$, the degree-zero barcode therefore
has one copy of $[-1,\infty)$ and $k-1$ copies of $[-1,1)$.
Degree one has one copy of $[1,\infty)$. These statements
follow because inclusions before $1$ preserve the components,
while the inclusion into the circle sends all component
classes to the same class. Each $p$ has the same barcodes.
The total sublevels
$\mathcal F(t)=\{(p,z):s(p,z)\leq t\}\to P$, with their
inclusions for increasing $t$, define a diagram of spaces
over $P$. A local lift $\alpha(p)^k=p$ simultaneously
trivializes the diagram by $z=\alpha(p)u$, since the score
becomes $\operatorname{Im}(u^k)$. The transitions rotate
$u$ by elements of $\mu_k$, and transport around the base
permutes the $k$ components for every $-1\le t<1$. Fiberwise
$H_0$ is therefore a local system with permutation module
$\mathbb K[\mu_k]$ in that range. At $t=1$ the structure map
is the augmentation $\mathbb K[\mu_k]\to\mathbb K$, where
$\mathbb K$ carries trivial transport. Its kernel gives
the local system of reduced degree-zero classes that die.
If the characteristic divides $k$, this augmentation has
no equivariant section. An invariant element is a multiple of
the sum of the $k$ basis vectors, whose augmentation is zero.
This extension is invisible in the barcode of each fiber,
where the augmentation is a surjective map of vector spaces.
Over $\mathbb C$, let $\omega=e^{2\pi i/k}$, with $e_j$ the
component basis of $H_0$. The character vector
$\sum_j\omega^{-\ell j}e_j$ acquires the factor
$\omega^\ell$ under monodromy. The trivial character survives
the merge at $1$ and each nontrivial character dies there.
The $k-1$ numerically equal finite bars consequently carry
distinct transport characters.
For parameters $p,q$, the chord distance $\delta=|p-q|$
bounds the uniform score change, since
\[
|s(p,z)-s(q,z)|
\leq |\overline p-\overline q|\,|z^k|=\delta.
\]
Thus $F_p(t)\subseteq F_q(t+\delta)$ and
$F_q(t)\subseteq F_p(t+\delta)$. The successive inclusion
maps are exactly the shift by $2\delta$, giving a
$\delta$-interleaving both before and after taking homology.
The same proof applies to any uniform score perturbation.
For q-tame persistent homology the stability theorems bound
the bottleneck distance by that uniform change
\cite{chazal-persistence,cohen-steiner-stability}.
In this rotation family the distance between fiber barcodes
is zero. The sublevel diagram over $P$ additionally records
which components are identified by going around a loop.
For example, the constant family of scores
$s_0(p,z)=\operatorname{Im}(z^k)$ has the same barcode at
every parameter. Its component local system is trivial,
whereas the generator of the parameter loop acts on the
components for $s$ by the nonidentity permutation.
The two filtered diagrams therefore cannot be equivalent
over $P$. For two general scores on the same parameter
and input spaces, a uniform bound on their difference
also gives the shifted inclusions of the \emph{total}
sublevels over $P$. This makes the sublevel diagram over
the parameter space a stable comparison object while
still retaining the representation.
\par
\section{Reconstruction through a Euclidean latent space}
\label{sec:reconstruction}
Let \(X\) be a nonempty compact smooth manifold without boundary,
and fix an embedding \(i:X\hookrightarrow\mathbb R^D\).
An encoder \(e:X\to\mathbb R^m\) and ambient decoder
\(d:\mathbb R^m\to\mathbb R^D\) realize the reconstruction
\(x\mapsto d(e(x))\). Here the latent width \(m\) is the dimension
of the intermediate Euclidean space. The encoder and decoder may be
evaluation maps of neural architectures at selected parameters.
We first allow arbitrary smooth maps to isolate the topological
constraints on any such architecture. This applies the tube and
factorization constructions of \cite[Chs.~8, 24, 27]{higher-hirsch}.
For related effects of data topology on autoencoding, see
\cite{batson-autoencoders}.
\subsection{Exact recovery and smooth families}
\begin{theorem}[Smooth ambient reconstruction]
\label{thm:ambient-reconstruction}
A smooth encoder \(e:X\to\mathbb R^m\) admits a smooth
ambient decoder on all of \(\mathbb R^m\) satisfying
\(d\circ e=i\) if and only if \(e\) is a smooth embedding.
Moreover \(e\) extends to a smooth encoder on all of
\(\mathbb R^D\) agreeing with it along \(i(X)\).
\end{theorem}
\begin{proof}
If \(d\circ e=i\), injectivity of \(i\) forces injectivity of \(e\).
Differentiating gives \(D_x i=D_{e(x)}d\,D_x e\), so \(D_xe\)
is injective. The compact-to-Hausdorff criterion makes \(e\) a
homeomorphism onto its image. That and immersion give an embedding.
Conversely let \(U\subset\mathbb R^m\) be a tubular neighborhood
of \(e(X)\) with smooth retraction \(\pi:U\to e(X)\). On \(U\) the
map \(d_0=i\circ e^{-1}\circ\pi\) satisfies \(d_0\circ e=i\).
Choose a smooth bump \(\chi\) equal to one near \(e(X)\) and with
compact support in \(U\). Set \(d=\chi d_0\) on \(U\) and \(d=0\)
off \(U\). The support condition makes this extension smooth,
and the equality holds along the image. To extend \(e\) from
\(i(X)\) to \(\mathbb R^D\), apply the same construction to
\(e\circ i^{-1}\) using a tube about \(i(X)\).
\end{proof}
Consequently the least possible exact latent width in unrestricted
smooth ambient reconstruction is the smooth Euclidean embedding
dimension of \(X\). Derivative rank alone does not determine that
width. For example, the double traversal \(z\mapsto z^2\) from
\(S^1\) to the unit circle in \(\mathbb R^2\) is a smooth immersion,
but a decoder cannot separate the two points at any encoded value.
The tube construction supplies a smooth map in the ambient model
class. Exact realization by a prescribed finite neural architecture
requires a separate representability argument.
\begin{corollary}[A decoder varying smoothly with parameters]
\label{cor:family-decoder}
Let \(P\) be a smooth manifold and
\(e:P\times X\to\mathbb R^m\) a smooth family. There exists a
jointly smooth \(d:P\times\mathbb R^m\to\mathbb R^D\) with
\(d(p,e(p,x))=i(x)\) for all \(p,x\) if and only if each \(e_p\)
is a smooth embedding.
\end{corollary}
\begin{proof}
Necessity follows by restricting to any \(p\). For sufficiency,
the map \(j(p,x)=(p,e(p,x))\) is an injective immersion, since a
vector in its derivative kernel has zero parameter component and
then zero \(X\)-component. The preimage of a compact set \(K\)
is closed in the compact set \(\operatorname{pr}_P(K)\times X\).
Thus \(j\) is proper and is a closed embedding. Take a smooth
tubular retraction \(\pi\) to its image in \(P\times\mathbb R^m\).
On this tube use \(i\circ\operatorname{pr}_X\circ j^{-1}\circ\pi\).
A smooth cutoff equal to one near the closed image and supported
in the tube lets us extend this map by zero. The cutoff can be
chosen by a partition of unity subordinate to the tube and the
complement of the image, with locally finite supports when \(P\)
is noncompact. Its value on \(j(p,x)\) is \(i(x)\), as required.
\end{proof}
\subsection{Uniform approximation and a sharp spherical bound}
The existence of a uniform error gap extends beyond exact recovery.
Suppose \(\dim X=n>0\). Choose a tubular neighborhood
\(U\subset\mathbb R^D\) of \(i(X)\) and a number \(\delta>0\) such
that all points at distance less than \(\delta\) from \(i(X)\)
belong to \(U\). Such a number exists by compactness.
\begin{proposition}[A closed-manifold error gap]
\label{prop:closed-gap}
If \(m\le n\), every continuous encoder and ambient decoder satisfy
\(\sup_{x\in X}\|d(e(x))-i(x)\|\ge\delta\).
\end{proposition}
\begin{proof}
If the supremum were less than \(\delta\), the straight segment
from \(i(x)\) to \(d(e(x))\) would lie in \(U\) for every \(x\).
Put \(V=d^{-1}(U)\), an open subset of \(\mathbb R^m\) containing
\(e(X)\), and let \(r:U\to X\) be the tubular retraction with
\(r\circ i=\mathrm{id}_X\). Applying \(r\) to those segments gives
a homotopy to the identity for the composite
\[
X\xrightarrow{e}V\xrightarrow{\,d|_V\,}U\xrightarrow{r}X.
\]
With coefficients in \(\mathbb F_2\), the identity on \(H_n(X)\)
would then factor through \(H_n(V)\). For \(m>0\), every
component of \(V\) is a noncompact \(m\)-manifold, whose ordinary
homology vanishes in degrees at least \(m\)
\cite[Prop.~3.29]{hatcher-at}. For \(m=0\) its positive-degree
homology also vanishes. The mod-two fundamental class of any
component of \(X\) is nonzero, yielding a contradiction.
\end{proof}
The use of the open latent subset \(V\) is essential because the
decoder may leave the tubular neighborhood elsewhere in latent
space. The argument applies to continuous layers, including
piecewise smooth realizations. It bounds uniform error for an
entire data manifold and therefore also bounds each eligible
member of a parameter family.
\begin{proposition}[Precise bounds on the unit sphere]
\label{prop:sphere-gap}
Take \(X=S^n\) and let \(i:S^n\hookrightarrow\mathbb R^{n+1}\)
be the standard unit sphere, with \(n\ge1\).
For \(m\le n\), the optimal uniform error among continuous
encoders and ambient decoders is \(1\).
For \(m\ge n+1\), exact ambient recovery is possible.
If the decoder is constrained to take values in \(S^n\)
on all of \(\mathbb R^m\), the uniform error is \(2\)
for every width \(m\).
\end{proposition}
\begin{proof}
For \(m\le n\), apply Borsuk--Ulam to the composite of the
encoder and the coordinate inclusion of \(\mathbb R^m\) in
\(\mathbb R^n\). It gives an \(x\) with \(e(x)=e(-x)\)
\cite{matousek-borsuk}.
If \(u\) is their common decoded value,
\(2\le\|x-u\|+\|-x-u\|\). The zero decoder attains error \(1\).
For \(m\ge n+1\), a coordinate inclusion and projection give
exact recovery. Finally a sphere-valued reconstruction through
all of \(\mathbb R^m\) is nullhomotopic. If it avoided
antipodal coincidences, its straight segments to the identity,
normalized to unit length, would give a homotopy to the identity.
This would contradict its zero degree. Hence it reaches error
\(2\), the largest distance between points of the sphere.
\end{proof}
The optimal errors for the unit sphere are summarized below.
\begin{center}
\begin{tabular}{@{}lcc@{}}
\hline
Decoder range & \(m\le n\) & \(m\ge n+1\)\\
\hline
Ambient \(\mathbb R^{n+1}\) & \(1\) & \(0\)\\
Sphere \(S^n\) & \(2\) & \(2\)\\
\hline
\end{tabular}
\end{center}
For a sphere of radius \(R>0\), rescaling replaces the positive
entries by \(R\) and \(2R\), respectively. The bound holds for every
parameter of an appropriate family. Its proof chooses an
antipodal pair separately at each parameter and needs no
continuous choice of witnesses.
Uniform and average losses have different implications. For example,
at width \(n\) choose stereographic coordinates \(\sigma\) from the
complement of the north pole of \(S^n\) to \(\mathbb R^n\).
Multiply \(\sigma\) by a smooth cutoff equal to one outside a small
polar cap \(C\) and equal to zero near the pole, extending the
encoder by zero there. The smooth decoder \(d=\sigma^{-1}\) gives
exact recovery off \(C\). For normalized spherical measure \(\mu\),
the squared error averaged over the sphere is at most \(4\mu(C)\),
which can be made arbitrarily small while the uniform error is \(2\).
Thus the uniform obstruction does not force a positive lower bound
on this average across all smooth maps of width \(n\).
A uniform Lipschitz bound prevents this concentration. If the
reconstruction has Lipschitz constant at most \(L\) with respect to
chord distance, its error function has Lipschitz constant at most
\(1+L\). For ambient width \(m\le n\), some point has error at least
\(1\), so every point within chord distance \(1/(2(1+L))\) of it
has error at least \(1/2\). Its mean squared error is at least
one quarter of the normalized measure of such a cap, a positive
bound at fixed \(L\). This estimate states the additional analytic
hypothesis under which the pointwise obstruction controls
a population loss.
\par
\section{Conclusion}
The analysis separates three conditions on a parametrized network.
The available weight directions determine where the total jet
map is transverse. The symmetry representation determines which
normal derivatives a fixed realization can have at a stabilizer.
The topology of the domain and latent factorization determines
which global reconstructions are possible. A model can satisfy
one of these conditions while exhibiting an obstruction governed
by another.
Examples on the circle and torus make these distinctions explicit.
Ordinary point counts jump at the fold discriminant. They can
also remain constant while boundary sheets are permuted along a
regular parameter loop. The resulting persistence module over
the parameter retains that permutation representation. In the
equivariant torus model the squared residual controls normal
movement, while the tangent complex records an excess target
direction. The autoencoder estimates bound uniform reconstruction
error and give a separate mean-square bound under a Lipschitz
constraint.
Relative evaluation and descent explain how these calculations
pass between local presentations of a model. The assumptions
used for the calculations are geometric: smoothness of the
relevant charts, specified perturbation directions, properness
for topology of fibers, and an explicit choice of decoder range.
They apply to a realized architecture once those hypotheses
have been checked.
\par
\balance
\begin{thebibliography}{99}
\small\raggedright
\bibitem{higher-hirsch}
Anonymous. \emph{Differential Topology in Higher Categories:
From Regular Values and Cobordism to Derived Smooth Families}.
Submitted manuscript, 2026.
\bibitem{previous-networks}
Anonymous. Representation theory of parametrized neural models
on spaces. \emph{Higher Categorical Semantics}, 2026.
\href{https://higher-categorical-semantics.pages.dev/research/representation-theory-of-parametrized-neural-models-on-spaces-ae10d3ed}{Published article}.
\bibitem{batson-autoencoders}
J. Batson, C. G. Haaf, Y. Kahn, and D. A. Roberts.
Topological obstructions to autoencoding.
\emph{Journal of High Energy Physics} 04 (2021), 280.
\href{https://doi.org/10.1007/JHEP04(2021)280}{DOI}.
\bibitem{chazal-persistence}
F. Chazal, V. de Silva, M. Glisse, and S. Oudot.
\emph{The Structure and Stability of Persistence Modules}.
Springer, 2016.
\href{https://doi.org/10.1007/978-3-319-42545-0}{DOI}.
\bibitem{cohen-steiner-stability}
D. Cohen-Steiner, H. Edelsbrunner, and J. Harer.
Stability of persistence diagrams.
\emph{Discrete \& Computational Geometry} 37 (2007), 103--120.
\href{https://doi.org/10.1007/s00454-006-1276-5}{DOI}.
\bibitem{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{hatcher-at}
A. Hatcher. \emph{Algebraic Topology}.
Cambridge University Press, 2002.
\href{https://pi.math.cornell.edu/~hatcher/AT/at1-tc.html}{Author's copy}.
\bibitem{hirsch}
M. W. Hirsch. \emph{Differential Topology}.
Graduate Texts in Mathematics 33. Springer, 1976.
\href{https://link.springer.com/book/10.1007/978-1-4684-9449-5}{Publisher's edition}.
\bibitem{matousek-borsuk}
J. Matou\v{s}ek. \emph{Using the Borsuk--Ulam Theorem}.
Springer, 2003.
\href{https://doi.org/10.1007/978-3-540-76649-0}{DOI}.
\bibitem{naitzat}
G. Naitzat, A. Zhitnikov, and L.-H. Lim.
Topology of deep neural networks.
\emph{Journal of Machine Learning Research} 21(184) (2020), 1--40.
\href{https://www.jmlr.org/papers/v21/20-345.html}{Journal article}.
\end{thebibliography}
\end{document}