betaFeedback

WorksMathematics

Secant dimensions and bilinear decomposition widths for finite Heisenberg groups

Abstract

Let G=UT_3(\F_q) be the finite Heisenberg group. We study the image closures and secant varieties of a family of equivariant bilinear maps between representations induced from central characters. Evaluation at a basepoint identifies the ambient space with M_{q^2}(\C). A single channel is represented by ATdiag⁡(d)BA^T\operatorname{diag}(d)B, with AA and BB in prescribed twisted-convolution algebras. We prove that its affine image closure has dimension 3q2−23q^2-2 for every central-character pair and that its affine second-secant closure has the expected dimension for every prime power qq. For q≥3q\ge3, the proof constructs two differential images with zero intersection. We also prove a subgroup-lifting theorem for channel parameterizations on finite abelian groups with bicharacter coefficients: a parameter point with differential rank r(3∣H∣−2)r(3|H|-2) on a subgroup HH extends by zero to a parameter point with differential rank r(3∣K∣−2)r(3|K|-2) on the ambient group KK, provided that every parameter vector is nonzero. Exact Jacobian certificates determine all secant dimensions for q=2,3,4,5q=2,3,4,5, with generic widths 2,4,6,92,4,6,9. The lifting theorem yields specified higher-secant dimensions on infinite families of extension fields. Attainment of the dimension lower bound for arbitrary qq remains a conjecture.

Introduction

Let qq be a prime power and let G=UT3(Fq)G = UT_3(\mathbb{F}_q) be the group of upper unitriangular 3×33 \times3 matrices over Fq\mathbb{F}_q. Its center ZZ is isomorphic to the additive group of Fq\mathbb{F}_q. Fix a nontrivial additive character ψ\psi of Fq\mathbb{F}_q, and put ψt(z)=ψ(tz)\psi_t(z)=\psi(tz) for t∈Fqt\in\mathbb{F}_q. We consider the induced representations

Ht=Ind⁡ZGψt.\mathcal{H}_t=\operatorname{Ind}^{G}_{Z}\psi_t.

Each Ht\mathcal{H}_t has dimension q2q^2. Pointwise multiplication defines an equivariant bilinear map from Hs×Ht\mathcal{H}_s\times\mathcal{H}_t to Hs+t\mathcal{H}_{s+t}. Thus the central characters determine the possible input and output types of a bilinear map.

Equivariant linear maps between these representations can be described by twisted convolutions. This is a finite version of the twisted-convolution calculus for Heisenberg representations; see Folland [1]. The relevant representation theory is part of finite Stone–von Neumann theory [2, 3]. In the preprint of Bich Van Nguyen and Nguyen Cao Manh Thang (2025; arXiv:2507.08397), explicit decompositions of linear equivariant feature spaces and filters are studied for UT3(F3)UT_3(\mathbb{F}_3). Here we consider sums of bilinear maps of the form

ΓA,B,C(f,g)=C[(Af)(Bg)].\Gamma_{A,B,C}(f,g)=C[(Af)(Bg)].

where AA, BB, and CC are equivariant linear operators on Hs\mathcal{H}_s, Ht\mathcal{H}_t, and Hs+t\mathcal{H}_{s+t}, respectively. We call such a map a fusion channel. The fusion width of an equivariant bilinear map is the least number of channels whose sum is that map.

Evaluation at a basepoint identifies the space of these equivariant bilinear maps with Mq2(C)M_{q^2}(\mathbb{C}). In these coordinates, a channel has the form ATdiag⁡(d)BA^{T}\operatorname{diag}(d)B, where dd is arbitrary and AA, BB belong to the corresponding commutant algebras. Let Xs,tX_{s,t} be the affine closure of the one-channel image. The closure of the sum of rr channels is the affine cone over σr(PXs,t)\sigma_r(\mathbb{P}X_{s,t}). We determine its dimension at order two for every prime power and every pair of central characters. We also prove a lifting theorem for parameterizations defined by bicharacters on finite abelian groups.

The space of admissible equivariant bilinear maps has dimension q4q^4. A single channel has 3q23q^2 parameters, and two independent scalings preserve its value. We prove that these scalings account for the generic fiber dimension: the affine closure Xs,tX_{s,t} of the one-channel image satisfies

dim⁡Xs,t=3q2−2\dim X_{s,t}=3q^2-2

for every q,s,tq,s,t (Theorem 5.7). The proof uses a permutation Jacobian minor at (I,I,eo)(I,I,e_o) and is independent of the input types. Consequently, the generic fusion width is at least ⌈q4/(3q2−2)⌉\lceil q^4/(3q^2-2)\rceil.

We next prove that

dim⁡im⁡Φ2=6q2−4(q≥3)\dim\operatorname{im}\Phi_2=6q^2-4 \qquad(q\ge3)

for every central-character pair (Theorem 6.1). We construct two parameter points whose differential images have dimension 3q2−23q^2-2 and intersect trivially. The argument uses the supports of matrix entries and explicit character factors. A separate final step handles characteristic two.

Theorem 9.1 is stated for finite abelian groups and arbitrary complex-valued bicharacters. If an rr-channel parameter point on a subgroup H≤KH\le K has differential rank r(3∣H∣−2)r(3|H|-2) and every parameter vector is nonzero, then the differential at its zero extension to KK has rank r(3∣K∣−2)r(3|K|-2). In particular, the affine closure of the one-channel image has dimension 3∣K∣−23|K|-2 for any such group and bicharacters (Corollary 9.2). Exact modular certificates determine all secant dimensions for q=2,3,4,5q=2,3,4,5, with generic widths 2, 4, 6, 9 (Theorem 8.3). Lifting parameter points below the filling order proves the expected dimensions through orders 3, 5, 8 for q=3f,4f,5fq=3^f,4^f,5^f, respectively, for every f≥1f\geq1 and every type pair (Corollary 9.5). The characteristic-two conclusion also holds for q=2fq=2^f with every f≥4f\geq4, using an additive subgroup of Fq×{0}\mathbb{F}_q\times\{0\}. These ranges do not reach the first filling order for large extension degree.

Proposition 7.2 shows that permuting the three central weights (s,t,−s−t)(s,t,-s-t) preserves channel decompositions up to a linear change of coordinates. This reduces the number of type pairs in the exact checks. Two rank certificates per type class suffice to determine all secant dimensions in each of these small fields by Lemma 8.2. For arbitrary qq, rank factorization also gives a constructive upper bound of q2q^2 channels for every target.

The untwisted case is related to group-based phylogenetic models. In Fourier coordinates its one-channel variety is the toric tripod for G/Z≅Fq2G/Z\cong\mathbb{F}_q^2. Its dimension 3q2−23q^2-2 already follows from the claw-dimension computation in the proof of Proposition 5.2 of Baños et al. [4]. Theorem 5.7 gives a single permutation minor for all input types, including the twisted cases. Untwisted all-orders nondefectivity is a special case of their Conjecture 5.1, and their Proposition 5.3 proves the untwisted q=2q=2 second secant. Their Section 5.2 also uses finite-field Jacobian ranks to certify dimensions. Our contributions concern the uniform second-secant theorem, subgroup lifting, and the particular small-field certificates and their consequences.

There are also connections with structured matrix decompositions and equivariant bilinear complexity. Couveignes and Ezome [5] study the complexity of multiplication in finite field extensions under equivariance. Huhtanen et al. [6] study sums of diagonal–circulant–diagonal matrices. Their Theorem 2.1 gives the rank-one inclusion, and their Definition 4.3 introduces DCD-rank. After replacing the second Fourier index by its negative, our untwisted parameterization has the same form for the additive group Fq2\mathbb{F}_q^2. This group is not cyclic, so our results do not settle the cyclic DCD-rank question. The rank-factorization upper bound is a standard consequence of rank-one inclusion. Lin et al. [7] approximate SO(3)SO(3)-equivariant tensor products by low-rank decompositions, obtaining approximately equivariant networks with a bound on the equivariance error. Our width concerns exact representation by the specified fusion channels.

In Algebraic Expressivity Certificates for Shallow Polynomial Neural Networks, Akbari and Jamshidi (2026; arXiv:2609.28500, Propositions 2–3 and Theorem 6) study Veronese secant closures, elimination certificates and the quadratic determinantal case for bias-free shallow power-sum networks. Their width counts scalar pure powers; our width counts the prescribed equivariant bilinear channels. Remark 10.2 gives a one-channel example whose scalar basepoint polynomial has quadratic Waring rank 2q22q^2. The distinction between an image and its closure is relevant to both parameterizations.

Sections 2–4 describe the induced representations, their commutants and the admissible bilinear maps. Sections 5–9 contain the dimension and width results, the type symmetries and the exact certificates. Section 10 gives the toric interpretation and distinguishes fusion width from quadratic Waring rank. Sections 11–13 discuss the remaining conjecture, reproducibility and further questions. Appendices D and E give polynomial realization and approximation consequences and explain their relation to earlier results.

Finite Heisenberg groups and induced representations

All representation spaces and algebraic varieties are over C\mathbb{C}. Finite coefficient fields are used only for computational verification.

Let q=pfq=p^f be a prime power and write

g(a,b,c)=(1ab01c001).g(a,b,c)=\begin{pmatrix}1&a&b\\0&1&c\\0&0&1\end{pmatrix}.

Then

g(a,b,c)g(a′,b′,c′)=g(a+a′,b+b′+ac′,c+c′).(1)g(a,b,c)g(a',b',c')=g(a+a',b+b'+ac',c+c'). \tag*{(1)}

The center and quotient are

Z={g(0,b,0):b∈Fq}≅(Fq,+),G/Z≅(Fq2,+).Z=\{g(0,b,0):b\in\mathbb{F}_q\}\cong(\mathbb{F}_q,+),\qquad G/Z\cong(\mathbb{F}_q^2,+).

Fix a nontrivial additive character ψ:Fq→C×\psi:\mathbb{F}_q\to\mathbb{C}^{\times}. The nondegenerate trace pairing implies that

ψt(b):=ψ(tb),t∈Fq,\psi_t(b):=\psi(tb),\qquad t\in\mathbb{F}_q,

exhausts all additive characters of Fq\mathbb{F}_q.

Definition 2.1 (Induced representation) For t∈Fqt\in\mathbb{F}_q, define

Ht=Ind⁡ZGψt={F:G→C:F(gz)=ψt(z)−1F(g)}.\mathcal{H}_t=\operatorname{Ind}_{Z}^{G}\psi_t=\{F:G\to\mathbb{C}:F(gz)=\psi_t(z)^{-1}F(g)\}.

The left action is [LhF](g)=F(h−1g)[L_hF](g)=F(h^{-1}g).

Choose the section

r(x,y)=g(x,0,y),(x,y)∈Fq2.r(x,y)=g(x,0,y),\qquad(x,y)\in\mathbb{F}_q^2.

Restriction to the section identifies Ht\mathcal{H}_t with CFq2\mathbb{C}^{\mathbb{F}_q^2}, hence

dim⁡CHt=q2.(2)\dim_{\mathbb{C}}\mathcal{H}_t=q^2. \tag*{(2)}

A direct calculation from (1) gives, for h=g(α,β,γ)h=g(\alpha,\beta,\gamma),

[Lhtf](x,y)=ψt(β−αγ+αy)f(x−α,y−γ).(3)[L_h^t f](x,y)=\psi_t(\beta-\alpha\gamma+\alpha y)f(x-\alpha,y-\gamma). \tag*{(3)}

In particular, the center acts on all of Ht\mathcal{H}_t by the scalar character ψt\psi_t. For later use we record the irreducible representations explicitly.

Proposition 2.2 (Elementary finite Stone–von Neumann decomposition) For each t≠0t \ne0, define πt\pi_t on CFq\mathbb{C}^{\mathbb{F}_q} by

[πt(g(a,b,c))φ](x)=ψt(b+cx)φ(x+a).(4)[\pi_t(g(a,b,c))\varphi](x)=\psi_t(b+cx)\varphi(x+a). \tag*{(4)}

Then πt\pi_t is irreducible of dimension qq, has central character ψt\psi_t, and the representations πt\pi_t, t≠0t \ne0, are pairwise inequivalent. Together with the q2q^2 one-dimensional characters factoring through G/ZG/Z, they exhaust the irreducible representations of GG.

Proof. (1) shows directly that (4) is a representation. Its character is

χt(g(a,b,c))={qψt(b),a=c=0,0,otherwise.\chi_t(g(a,b,c))= \begin{cases} q\psi_t(b), & a=c=0,\\ 0, & \text{otherwise}. \end{cases}

Indeed, a nonzero translation has zero trace, while for a=0a=0, c≠0c\ne0, the trace is ψt(b)∑xψt(cx)=0\psi_t(b)\sum_{x}\psi_t(cx)=0. Therefore

⟨χt,χt⟩G=1q3∑b∈Fqq2=1,\langle\chi_t,\chi_t\rangle_G=\frac{1}{q^3}\sum_{b\in\mathbb{F}_q}q^2=1,

so πt\pi_t is irreducible. Distinct tt give distinct central characters and hence inequivalent representations. Since [G,G]=Z[G,G]=Z, there are q2q^2 one-dimensional representations. Their squared dimensions contribute q2q^2; the q−1q-1 representations πt\pi_t contribute (q−1)q2(q-1)q^2. The sum is q3=∣G∣q^3=|G|, so the list is complete. □

Corollary 2.3 For t≠0t\ne0,

Ht≅q πt,\mathcal{H}_t\cong q\,\pi_t,

while H0\mathcal{H}_0 is the regular representation of G/ZG/Z.

Proof. For t≠0t\ne0, Frobenius reciprocity gives

[Ht:πt]=dim⁡Hom⁡Z(πt∣Z,ψt)=q,[\mathcal{H}_t:\pi_t]=\dim\operatorname{Hom}_Z(\pi_t|_Z,\psi_t)=q,

because ZZ acts on the qq-dimensional space of πt\pi_t by ψt\psi_t. No other irreducible has this central character. The statement for t=0t=0 follows because Ind⁡ZG1\operatorname{Ind}_Z^G 1 is the regular representation of the quotient. □

Twisted convolutions and commutants

For (u,v)∈Fq2(u,v)\in\mathbb{F}_q^2, define the twisted shift

[Ru,vtf](x,y)=ψt(−xv)f(x+u,y+v).(5)[R^t_{u,v}f](x,y)=\psi_t(-xv)f(x+u,y+v). \tag*{(5)}

These are right translations expressed in the chosen section.

Lemma 3.1 (Commutation and cocycle law) For all h∈Gh\in G and (u,v)∈Fq2(u,v)\in\mathbb{F}_q^2,

LhtRu,vt=Ru,vtLht.L_h^tR_{u,v}^t=R_{u,v}^tL_h^t.

Moreover,

Ru,vtRu′,v′t=ψt(−uv′)Ru+u′,v+v′t.(6)R_{u,v}^tR_{u',v'}^t=\psi_t(-uv')R_{u+u',v+v'}^t. \tag*{(6)}

Proof. Substitution into (3) and (5) proves commutation. For the product,

ψt(−xv′)ψt(−(x+u)v′)=ψt(−uv′)ψt(−x(v+v′)),\psi_t(-xv')\psi_t(-(x+u)v')=\psi_t(-uv')\psi_t(-x(v+v')),

which yields (6).

Theorem 3.2 (Commutants and twisted convolutions) For every t∈Fqt \in\mathbb{F}_q,

End⁡G(Ht)=span⁡C{Ru,vt:(u,v)∈Fq2},dim⁡CEnd⁡G(Ht)=q2.(7)\operatorname{End}_G(\mathcal{H}_t)=\operatorname{span}_{\mathbb{C}}\{R^t_{u,v}:(u,v)\in\mathbb{F}_q^2\},\qquad\dim_{\mathbb{C}}\operatorname{End}_G(\mathcal{H}_t)=q^2. \tag*{(7)}

Consequently every GG-equivariant linear map T:Ht→HtT:\mathcal{H}_t\to\mathcal{H}_t has a unique twisted-convolution representation

[Tkf](x,y)=∑u,v∈Fqk(u,v)ψt(−xv)f(x+u,y+v).(8)[T_kf](x,y)=\sum_{u,v\in\mathbb{F}_q}k(u,v)\psi_t(-xv)f(x+u,y+v). \tag*{(8)}

Proof. If t=0t=0, then H0\mathcal{H}_0 is the regular representation of the abelian group G/ZG/Z, hence the multiplicity-free sum of its q2q^2 characters. Therefore dim⁡End⁡G(H0)=q2\dim\operatorname{End}_G(\mathcal{H}_0)=q^2. If t≠0t\ne0, then Ht≅qπt\mathcal{H}_t\cong q\pi_t, so Schur’s lemma gives

End⁡G(Ht)≅Mq(C),dim⁡End⁡G(Ht)=q2.\operatorname{End}_G(\mathcal{H}_t)\cong M_q(\mathbb{C}),\qquad\dim\operatorname{End}_G(\mathcal{H}_t)=q^2.

The twisted shifts lie in the commutant by the previous lemma. They are linearly independent: applying a linear relation to the delta function at (0,0)(0,0) produces functions supported at distinct points (−u,−v)(-u,-v). Since there are q2q^2 shifts, they form a basis.

Remark 3.3 For t=0t=0, (8) is ordinary convolution on the abelian quotient Fq2\mathbb{F}_q^2. For t≠0t\ne0, the phase in (8) records the Heisenberg cocycle. Thus the same coordinate space Fq2\mathbb{F}_q^2 carries different convolution algebras according to the central character.

Equivariant bilinear maps

Throughout, “bilinear” means complex-bilinear. For types s,ts,t, pointwise multiplication

ms,t(f,g)=fgm_{s,t}(f,g)=fg

is a GG-equivariant bilinear map Hs×Ht→Hs+t\mathcal{H}_s\times\mathcal{H}_t\to\mathcal{H}_{s+t}, because the phases in (3) multiply.

Theorem 4.1 (Selection rule and exact dimension) For s,t,u∈Fqs,t,u\in\mathbb{F}_q,

Bil⁡G(Hs,Ht;Hu)=0if u≠s+t.(9)\operatorname{Bil}_G(\mathcal{H}_s,\mathcal{H}_t;\mathcal{H}_u)=0\qquad\text{if }u\ne s+t. \tag*{(9)}

If u=s+tu=s+t, then

dim⁡CBil⁡G(Hs,Ht;Hs+t)=q4.(10)\dim_{\mathbb{C}}\operatorname{Bil}_G(\mathcal{H}_s,\mathcal{H}_t;\mathcal{H}_{s+t})=q^4. \tag*{(10)}

Proof. Complex-bilinear equivariant maps are naturally

Hom⁡G(Hs⊗Ht,Hu).\operatorname{Hom}_G(\mathcal{H}_s\otimes\mathcal{H}_t,\mathcal{H}_u).

For finite groups induction agrees with coinduction. Since Hu=Ind⁡ZGψu\mathcal{H}_u = \operatorname{Ind}_{Z}^{G}\psi_u, Frobenius reciprocity in the target variable gives

Hom⁡G(Hs⊗Ht,Hu)≅Hom⁡Z((Hs⊗Ht)∣Z,ψu).(11)\operatorname{Hom}_{G}(\mathcal{H}_s \otimes\mathcal{H}_t,\mathcal{H}_u) \cong\operatorname{Hom}_{Z}((\mathcal{H}_s \otimes\mathcal{H}_t)|_{Z},\psi_u). \tag*{(11)}

The center acts on the tensor product by ψsψt=ψs+t\psi_s\psi_t=\psi_{s+t}. Since dim⁡(Hs⊗Ht)=q4\dim(\mathcal{H}_s\otimes\mathcal{H}_t)=q^4, as a ZZ-module it is ψs+t⊕q4\psi_{s+t}^{\oplus q^4}. Equation (11) therefore has dimension zero unless u=s+tu=s+t, and dimension q4q^4 in the admissible case. □\square

Let N=q2N=q^2 and fix the basepoint o=(0,0)∈G/Zo=(0,0)\in G/Z.

Lemma 4.2 (Basepoint functional model) For every admissible pair (s,t)(s,t), evaluation at oo induces a linear isomorphism

ev⁡o:Bil⁡G(Hs,Ht;Hs+t)⟶Bil⁡C(Hs,Ht;C).\operatorname{ev}_{o}:\operatorname{Bil}_{G}(\mathcal{H}_s,\mathcal{H}_t;\mathcal{H}_{s+t})\longrightarrow\operatorname{Bil}_{\mathbb{C}}(\mathcal{H}_s,\mathcal{H}_t;\mathbb{C}).

Equivalently, every scalar complex-bilinear functional at the basepoint extends uniquely to a GG-equivariant Hs+t\mathcal{H}_{s+t}-valued bilinear map.

Proof. Apply Frobenius reciprocity to W=Hs⊗HtW=\mathcal{H}_s\otimes\mathcal{H}_t:

Hom⁡G(W,Ind⁡ZGψs+t)≅Hom⁡Z(W,ψs+t).\operatorname{Hom}_{G}(W,\operatorname{Ind}_{Z}^{G}\psi_{s+t})\cong\operatorname{Hom}_{Z}(W,\psi_{s+t}).

Under the induced-function realization, this isomorphism is evaluation of the output at the identity coset o=Zo=Z. On WW, the center acts by the scalar character ψsψt=ψs+t\psi_s\psi_t=\psi_{s+t}. Consequently every complex-linear functional W→CW\to\mathbb{C} is automatically ZZ-equivariant when the target C\mathbb{C} is given the character ψs+t\psi_{s+t}. Hence

Hom⁡Z(W,ψs+t)=Hom⁡C(W,C)≅Bil⁡C(Hs,Ht;C),\operatorname{Hom}_{Z}(W,\psi_{s+t})=\operatorname{Hom}_{\mathbb{C}}(W,\mathbb{C})\cong\operatorname{Bil}_{\mathbb{C}}(\mathcal{H}_s,\mathcal{H}_t;\mathbb{C}),

which proves both existence and uniqueness of the equivariant extension. □\square

Choosing the standard coordinate basis of Cq2\mathbb{C}^{q^2}, Lemma 4.2 yields

Bil⁡G(Hs,Ht;Hs+t)≅MN(C),(12)\operatorname{Bil}_{G}(\mathcal{H}_s,\mathcal{H}_t;\mathcal{H}_{s+t})\cong M_N(\mathbb{C}), \tag*{(12)}

where M∈MN(C)M\in M_N(\mathbb{C}) represents the bilinear functional B(f,g)(o)=fTMgB(f,g)(o)=f^{T}Mg. We use this identification in the Jacobian computations below.

Fusion width and a general lower bound

Definition 5.1 (Fusion channel) A fusion channel from Hs×Ht\mathcal{H}_s\times\mathcal{H}_t to Hs+t\mathcal{H}_{s+t} is

ΓA,B,C(f,g)=C[(Af)(Bg)],(13)\Gamma_{A,B,C}(f,g)=C[(Af)(Bg)], \tag*{(13)}

where

A∈End⁡G(Hs),B∈End⁡G(Ht),C∈End⁡G(Hs+t).A\in\operatorname{End}_{G}(\mathcal{H}_s),\quad B\in\operatorname{End}_{G}(\mathcal{H}_t),\quad C\in\operatorname{End}_{G}(\mathcal{H}_{s+t}).

For r≥1r\geq1, define the polynomial parameterization

Φr((Aℓ,Bℓ,Cℓ)ℓ=1r)=∑ℓ=1rΓAℓ,Bℓ,Cℓ.(14)\Phi_r((A_\ell,B_\ell,C_\ell)_{\ell=1}^{r})=\sum_{\ell=1}^{r}\Gamma_{A_\ell,B_\ell,C_\ell}. \tag*{(14)}

The operators AA and BB act on the two inputs, and CC acts on their pointwise product. If dd is the basepoint row of CC, the coefficient matrix of one channel is

M=ATdiag⁡(d)B,A∈End⁡G(Hs),B∈End⁡G(Ht),d∈CN.M = A^{T}\operatorname{diag}(d)B,\qquad A \in\operatorname{End}_{G}(\mathcal{H}_{s}),\qquad B \in\operatorname{End}_{G}(\mathcal{H}_{t}),\qquad d \in\mathbb{C}^{N}.

The vector dd is arbitrary because the basepoint rows of the twisted shifts form the coordinate basis. Thus fusion width is the number of summands in this structured matrix decomposition. It is not ordinary matrix rank: taking A=B=IA = B = I and d=1d = 1 gives M=INM = I_{N} in one channel. The same example is considered in Remark 10.2.

Definition 5.2 (Generic fusion width) Let

Vs,t=Bil⁡G(Hs,Ht;Hs+t).\mathcal{V}_{s,t} = \operatorname{Bil}_{G}(\mathcal{H}_{s},\mathcal{H}_{t};\mathcal{H}_{s+t}).

The width of an individual target is the least nonnegative number of channels representing it. We set width⁡(0)=0\operatorname{width}(0) = 0 and Φ0=0\Phi_{0} = 0. The generic fusion width rgen(q;s,t)r_{\mathrm{gen}}(q; s,t) is the least rr for which Φr\Phi_{r} is dominant onto Vs,t\mathcal{V}_{s,t}.

Here and throughout, generic is meant in the complex-algebraic sense: the asserted property holds on a nonempty Zariski-open subset of the affine target space. Consequently its complement lies in a proper algebraic subset and has Lebesgue measure zero after identifying the target with a real Euclidean space. Because the image of a morphism is constructible, dominance here has an exact representability meaning, not merely a border-rank meaning.

Proposition 5.3 (Dominance implies exact generic width) If Φr\Phi_{r} is dominant and Φr−1\Phi_{r-1} is not, then there exists a nonempty Zariski-open subset U⊂Vs,tU \subset\mathcal{V}_{s,t} such that every element of UU has exact fusion width rr.

Proof. By Chevalley’s theorem, the dense constructible set im⁡Φr\operatorname{im}\Phi_{r} contains a nonempty open subset of the irreducible affine space Vs,t\mathcal{V}_{s,t}. Since im⁡Φr−1\operatorname{im}\Phi_{r-1} is a proper closed subset, removing it leaves a nonempty open subset on which width is at least rr and at most rr. □

Proposition 5.4 (Constructive exact upper bound) Let MM be the basepoint coefficient matrix of an admissible equivariant bilinear map BB. Then

width⁡(B)≤rank⁡(M)≤N=q2.\operatorname{width}(B) \le\operatorname{rank}(M) \le N = q^{2}.

Given a rank factorization of MM, the corresponding channel kernels are obtained explicitly. A factorization and the resulting kernels can be computed using O(N3)O(N^{3}) arithmetic operations.

Proof. Write M=∑ℓ=1kaℓbℓTM = \sum_{\ell=1}^{k} a_{\ell}b_{\ell}^{T}, where k=rank⁡(M)k = \operatorname{rank}(M). The basepoint row of the twisted convolution with coefficient kernel aℓa_{\ell} is aℓTa_{\ell}^{T}, independently of its type; similarly for bℓb_{\ell}. Choose the output kernel c=eoc = e_{o}, so that its convolution operator is C=IC = I and its basepoint row is eoTe_{o}^{T}. Consequently

A(aℓ)Tdiag⁡(eo)B(bℓ)=aℓbℓT.(15)A(a_{\ell})^{T}\operatorname{diag}(e_{o})B(b_{\ell}) = a_{\ell}b_{\ell}^{T}. \tag*{(15)}

Each summand is therefore one channel. Evaluation at the basepoint is an isomorphism on admissible equivariant bilinear maps, so equality of these matrices gives equality of the maps. Gaussian rank factorization costs O(N3)O(N^3) arithmetic operations, and writing the kernels costs O(kN)O(kN) operations. This is an arithmetic, rather than bit, complexity statement.

Corollary 5.5 (A worst-case consequence of dominance) If Φr\Phi_r is dominant, then every target has fusion width at most min⁡{N,2r}\min\{N,2r\}. In particular, the worst-case widths for q=2,3,4,5q=2,3,4,5 are at most 4,8,12,184,8,12,18, respectively.

Proof. The image contains a nonempty Zariski-open subset UU of Vs,tV_{s,t}. For any M∈Vs,tM \in V_{s,t}, the two nonempty open subsets UU and M−UM-U intersect because the affine space Vs,tV_{s,t} is irreducible. Thus MM is a sum of two elements in im⁡Φr\operatorname{im}\Phi_r. Apply Proposition 5.4 for the other bound. □

Proposition 5.4 gives a decomposition of every target, but need not attain the generic minimum. The factor-two bound follows from the standard dense-open-set argument; see Blekherman and Teitler [8]. The rank-one construction also occurs in the DCD-matrix setting of Huhtanen et al. [6].

Lemma 5.6 (Scaling symmetries) The one-channel parameterization is invariant under

(A,B,C)⟼(λA,μB,(λμ)−1C),λ,μ∈C×.(16)(A,B,C) \longmapsto(\lambda A,\mu B,(\lambda\mu)^{-1}C), \qquad\lambda,\mu\in\mathbb{C}^{\times}. \tag*{(16)}

Hence the Zariski closure of the one-channel image has dimension at most

3q2−2.3q^2-2.

More generally,

dim⁡im⁡Φr‾≤r(3q2−2).(17)\dim\overline{\operatorname{im}\Phi_r} \le r(3q^2-2). \tag*{(17)}

Proof. Each commutant has dimension q2q^2. On the dense open subset where A,B,C≠0A,B,C \ne0, the torus (C×)2(\mathbb{C}^{\times})^2 in (15) acts freely and preserves every channel. For rr channels the product torus (C×)2r(\mathbb{C}^{\times})^{2r} acts freely on a dense open subset of the parameter space and is contained in each fiber. The fiber-dimension theorem gives (16). □

We now prove that the preceding dimension bound is attained for every prime power and every type pair.

Theorem 5.7 (Dimension of the one-channel variety) Let Xs,t⊂Vs,tX_{s,t} \subset V_{s,t} be the affine Zariski closure of the one-channel image. For every prime power qq and every s,t∈Fqs,t \in\mathbb{F}_q,

dim⁡Xs,t=3q2−2.\dim X_{s,t}=3q^2-2.

The projectivized variety has dimension 3q2−33q^2-3, and the generic fiber has dimension two. At (A,B,d)=(I,I,eo)(A,B,d)=(I,I,e_o), the Jacobian has an explicitly selected (3q2−2)(3q^2-2)-square permutation submatrix. Its determinant is (−1)q2−1(-1)^{q^2-1} in the ordering below, and all its singular values equal one.

Proof. Put N=q2N=q^2 and K=Fq2K=\mathbb{F}_q^2. In the basepoint matrix model, the one-channel map is

F(A,B,d)=ATdiag⁡(d)B.F(A,B,d)=A^{T}\operatorname{diag}(d)B.

The vector dd is arbitrary, since the basepoint rows of the twisted shifts are the coordinate rows. At (I,I,d)(I,I,d) the differential is

dF(I,I,d)(X,Y,z)=XTdiag⁡(d)+diag⁡(z)+diag⁡(d)Y.(18)dF_{(I,I,d)}(X,Y,z)=X^{T}\operatorname{diag}(d)+\operatorname{diag}(z)+\operatorname{diag}(d)Y. \tag*{(18)}

Write X=∑hxhRhsX=\sum_h x_hR_h^s and Y=∑hyhRhtY=\sum_h y_hR_h^t. For h=(u,v)≠oh=(u,v)\ne o and i=(ix,iy)i=(i_x,i_y), the (i,i−h)(i,i-h) entry is

xhψs(−(ix−u)v)di−h+y−hψt(ixv)di.(19)x_h\psi_s\bigl(-(i_x-u)v\bigr)d_{i-h}+y_{-h}\psi_t(i_xv)d_i. \tag*{(19)}

Now set d=eod=e_o. Select the NN columns ziz_i and the NN diagonal rows. For each h≠oh\ne o, select columns xh,y−hx_h,y_{-h} and rows (o,−h),(h,o)(o,-h),(h,o), in that order. All selected rows and columns are distinct, including in characteristic two. Since the basepoint row of RhR_h is ehTe_h^T, the selected matrix is

IN⊕⨁h≠o(0110).(20)I_N\oplus\bigoplus_{h\ne o} \begin{pmatrix} 0&1\\ 1&0 \end{pmatrix}. \tag*{(20)}

This proves the determinant and singular-value assertions, and gives differential rank at least 3N−23N-2. The dimension bound in Lemma 5.6 gives the reverse inequality. Since Xs,tX_{s,t} is a nonzero affine cone, its projectivization has dimension 3N−33N-3. The generic-fiber dimension is 3N−(3N−2)=23N-(3N-2)=2 by the fiber-dimension theorem. □

Remark 5.8 (Scope of the explicit minor) The selected minor has spectral condition number one in the stated coordinates. This condition number concerns the differential at the specified point. A generic fiber of dimension two need not be a single scaling orbit: common twisted shifts also give discrete parameter symmetries.

Theorem 5.9 (Lower bound for generic width) For every prime power qq and every s,t∈Fqs,t\in\mathbb{F}_q,

rgen(q;s,t)≥⌈q43q2−2⌉.(21)r_{\mathrm{gen}}(q;s,t)\ge\left\lceil\frac{q^4}{3q^2-2}\right\rceil. \tag*{(21)}

Proof. The target dimension is q4q^4 by (10). By (17), the closure of the rr-channel image has dimension at most r(3q2−2)r(3q^2-2). If r(3q2−2)<q4r(3q^2-2)<q^4, dominance is impossible. Solving for rr proves the claim. □

Remark 5.10 By Theorems 5.7 and 6.1, any failure of the expected secant dimension must occur at order at least three.

Nondefectivity of the second secant

Theorem 6.1 (Dimension of the second secant) Let q≥3q\ge3 be a prime power and let s,t∈Fqs,t\in\mathbb{F}_q be arbitrary. Then

dim⁡im⁡Φ2‾=6q2−4.\dim\overline{\operatorname{im}\Phi_2}=6q^2-4.

Thus the second projective secant of the one-channel variety has its expected dimension 6q2−56q^2-5. Together with the q=2q=2 case of Theorem 8.3, this gives

dim⁡im⁡Φ2=min⁡{6q2−4,q4}\dim\operatorname{im}\Phi_2=\min\{6q^2-4,q^4\}

for every prime power qq and every pair of central types. The argument below applies to every q≥3q\geq3; only the case q=2q=2 uses a rank certificate.

Proof. Write K=Fq2K=\mathbb{F}_q^2, N=q2N=q^2, o=(0,0)o=(0,0), a=(1,0)a=(1,0) and b=(0,1)b=(0,1). Put

H={o,a,−b,a−b},J={K∖{o},char⁡Fq≠2,K∖H,char⁡Fq=2.H=\{o,a,-b,a-b\},\qquad J=\begin{cases} K\setminus\{o\}, & \operatorname{char}\mathbb{F}_q\ne2,\\ K\setminus H, & \operatorname{char}\mathbb{F}_q=2. \end{cases}

The set HH has four elements. It is a subgroup only in the characteristic-two case. A stripe of difference hh of a matrix consists of its entries Mi,i−hM_{i,i-h}, i∈Ki\in K. Distinct stripes have disjoint supports.

Choose d∈CKd\in\mathbb{C}^K satisfying these two conditions:

  1. For every h=(u,v)∉Hh=(u,v)\notin H, the following two vectors, restricted to K∖EhK\setminus E_h with Eh={o,a,h,h+b}E_h=\{o,a,h,h+b\}, are linearly independent:

(ψs(−(ix−u)v)di−h)i,(ψt(ixv)di)i.\left(\psi_s\left(-(i_x-u)v\right)d_{i-h}\right)_i,\qquad \left(\psi_t(i_xv)d_i\right)_i.
  1. The three vectors on JJ given by

D=(dk)k,Da=(dk+a)k,W=(ψs+t(kx)dk+b)kD=(d_k)_k,\qquad D_a=(d_{k+a})_k,\qquad W=\left(\psi_{s+t}(k_x)d_{k+b}\right)_k

are linearly independent.

We first prove that such a vector dd exists. Since N≥9N\geq9, the complement of EhE_h has at least N−4≥5N-4\geq5 elements. For any two distinct indices i,ji,j in that complement, the determinant of the corresponding two rows in (i) is a difference of nonzero scalar multiples of the monomials di−hdjd_{i-h}d_j and didj−hd_id_{j-h}. These monomials are distinct: equality of the multisets of their indices would force either h=oh=o or i=ji=j. Hence the selected determinant is a nonzero polynomial.

For (ii) in odd characteristic, choose w∉{o,a,b}w\notin\{o,a,b\}; in characteristic two choose w∉Hw\notin H, which is possible since q≥4q\geq4 and N≥16N\geq16. Specializing d=ewd=e_w, the three distinct indices ww, w−aw-a, w−bw-b all belong to JJ. The columns D,Da,WD,D_a,W on these rows form a diagonal matrix with nonzero diagonal (up to their ordering). Thus this three-square determinant is also a nonzero polynomial. The simultaneous nonvanishing conditions are finitely many nonempty principal Zariski-open subsets of the irreducible space CK\mathbb{C}^K, and their intersection is nonempty.

Choose complex scalars

(α,β)={(1,1),char⁡Fq≠2,(2,3),char⁡Fq=2,(\alpha,\beta)= \begin{cases} (1,1), & \operatorname{char}\mathbb{F}_q\ne2,\\ (2,3), & \operatorname{char}\mathbb{F}_q=2, \end{cases}

and consider the two parameter points

P=(I,I,d),Q=(A,B,eo),A=I+αRas,B=I+βRbt.P=(I,I,d),\qquad Q=(A,B,e_o),\qquad A=I+\alpha R_a^s,\qquad B=I+\beta R_b^t.

Both AA and BB are invertible. The shifts Ras,RbtR_a^s,R_b^t have order pp, where p=char⁡Fqp=\operatorname{char}\mathbb{F}_q. For odd pp, −1-1 is not an eigenvalue. For p=2p=2 their squares are the identity and α2,β2≠1\alpha^2,\beta^2\ne1. Let TP,TQT_P,T_Q denote the two differential images. At PP the diagonal stripe has dimension NN, and every stripe outside HH has dimension two by (i). For the stripes a,−b,a−ba,-b,a-b, evaluate formula (18) at (k+a,k)(k+a,k), (k,k+b)(k,k+b) and (k+a,k+b)(k+a,k+b), respectively, with k∈Jk\in J. This gives the pairs

(D,Da),(W,D),(W,ψt(−1)Da),(D,D_a),\qquad(W,D),\qquad\left(W,\psi_t(-1)D_a\right),

after multiplication of rows by nonzero scalars. In the last two pairs the row indexed by kk is divided by ψt(−kx)\psi_t(-kx). Condition (ii) makes each pair independent. Since the stripes have disjoint supports,

dim⁡TP=N+2(N−1)=3N−2.\dim T_P = N + 2(N-1) = 3N - 2.

Let ToT_o denote the differential image at (I,I,eo)(I,I,e_o); it consists exactly of the matrices supported on the diagonal together with row oo and column oo. The changes X↦XAX \mapsto XA, Y↦YBY \mapsto YB are bijections of the respective commutants. Differentiation therefore gives

TQ=ATToB,dim⁡TQ=3N−2.T_Q = A^{T}T_oB,\qquad\dim T_Q = 3N - 2.

We show that TP∩TQ={0}T_P \cap T_Q = \{0\}. For an element MM in this intersection, write

M=xvT+uyT+ATdiag⁡(z)B,u=eo+αea,v=eo+βeb,M = xv^{T} + uy^{T} + A^{T}\operatorname{diag}(z)B,\qquad u = e_o + \alpha e_a,\quad v = e_o + \beta e_b,

where x,y,z∈CKx,y,z \in\mathbb{C}^{K}. The vectors x,yx,y are arbitrary because the basepoint row maps of the two commutants are isomorphisms. The term indexed by kk in ATdiag⁡(z)BA^{T}\operatorname{diag}(z)B has positions

(k,k),(k+a,k),(k,k+b),(k+a,k+b),(k,k),\quad(k+a,k),\quad(k,k+b),\quad(k+a,k+b),

so this term is supported only on stripes HH. For h∉Hh \notin H, the first two terms may be nonzero only at rows i∈Ehi \in E_h. On the other hand, M∈TPM \in T_P restricts on that stripe to a linear combination of the two vectors in (i). Vanishing on K∖EhK \setminus E_h forces its two coefficients to be zero. Thus all stripes outside HH vanish.

For i∉{o,a}i \notin\{o,a\}, the entries (i,o)(i,o) and (i,b)(i,b) of the first two terms are xix_i and βxi\beta x_i. Unless both differences i,i−bi,i-b belong to HH, a vanishing stripe forces xi=0x_i=0. Similarly, entries (o,j),(a,j)(o,j),(a,j) for j∉{o,b}j \notin\{o,b\} are yj,αyjy_j,\alpha y_j. Consequently,

supp⁡x⊂H∩(H+b),supp⁡y⊂(−H)∩(a−H).\operatorname{supp} x \subset H \cap(H+b),\qquad\operatorname{supp} y \subset(-H) \cap(a-H).

In odd characteristic these two intersections are {o,a}\{o,a\} and {o,b}\{o,b\}, respectively. In characteristic two both intersections equal HH. These identities follow by comparing the independent coordinates along a,ba,b; they also hold in the smallest odd characteristic p=3p=3.

For k∈Jk \in J, the first two terms therefore contribute nothing at the three positions (k+a,k),(k,k+b),(k+a,k+b)(k+a,k),(k,k+b),(k+a,k+b). Formula (19) for TPT_P and division by the nonzero phase ψt(−kx)\psi_t(-kx) in the last two positions give constants λj,μj\lambda_j,\mu_j such that

αzk=λ1dk+μ1dk+a,\alpha z_k = \lambda_1 d_k + \mu_1 d_{k+a},
βzk=λ2ψs+t(kx)dk+b+μ2dk,\beta z_k = \lambda_2\psi_{s+t}(kx)d_{k+b} + \mu_2d_k,
αβzk=λ3ψs+t(kx)dk+b+μ3ψt(−1)dk+a.\alpha\beta z_k = \lambda_3\psi_{s+t}(kx)d_{k+b} + \mu_3\psi_t(-1)d_{k+a}.

Thus (zk)k∈J(z_k)_{k\in J} lies in the intersection of the three planes

span⁡(D,Da),span⁡(W,D),span⁡(W,Da).\operatorname{span}(D,D_a),\qquad\operatorname{span}(W,D),\qquad\operatorname{span}(W,D_a).

Their intersection is zero by (ii). Hence zk=0z_k=0 for k∈Jk\in J and each pair of coefficients λj,μj\lambda_j,\mu_j is zero. These are the coefficients of the full stripes of M∈TPM \in T_P, so the stripes a,−b,a−ba,-b,a-b vanish at every index, not just on JJ. We conclude that all off-diagonal entries of MM are zero.

It remains to treat the possible diagonal entries. In odd characteristic, only zoz_o remains. The matrix ATdiag⁡(z)B=zouvTA^{T}\operatorname{diag}(z)B=z_ouv^{T} can be absorbed into xvT+uyTxv^{T}+uy^{T}. This matrix is supported on {o,a}×{o,b}\{o,a\}\times\{o,b\}; its only possible diagonal entry is MooM_{oo}. Because α=β=1\alpha=\beta=1, its entries obey

Moo+Mab−Mob−Mao=0.M_{oo}+M_{ab}-M_{ob}-M_{ao}=0.

The other three entries are off-diagonal and vanish. Hence Moo=0M_{oo}=0.

In characteristic two, both xvT+uyTxv^{T}+uy^{T} and the remaining terms indexed by k∈Hk\in H have support in H×HH\times H. Thus M=diag⁡(m)M=\operatorname{diag}(m) with supp⁡m⊆H\operatorname{supp}m\subseteq H. Since TQ=ATToBT_Q=A^{T}T_oB and both shifts square to the identity,

L=(I−α(Ras)T)M(I−βRbt)∈To;L=(I-\alpha(R_a^s)^T)M(I-\beta R_b^t)\in T_o;

the scalar denominators omitted from A−TA^{-T} and B−1B^{-1} are nonzero. The four terms in this expression have the four distinct differences o,a,b,a+bo,a,b,a+b. Put c=a+bc=a+b and θk=ψt(−kx)\theta_k=\psi_t(-kx). Direct evaluation at four off-diagonal positions gives

La,b=αβmo,Lb,a=αβθcmc,Lc,b=−αmb,La,c=−βθama.L_{a,b}=\alpha\beta m_o,\qquad L_{b,a}=\alpha\beta\theta_c m_c,\qquad L_{c,b}=-\alpha m_b,\qquad L_{a,c}=-\beta\theta_a m_a.

These entries avoid row oo, column oo, and the diagonal, so L∈ToL\in T_o forces each to vanish. All four scalar multipliers are nonzero; therefore mo=ma=mb=mc=0m_o=m_a=m_b=m_c=0.

In both cases M=0M=0, proving the required trivial intersection. The differential of Φ2\Phi_2 at (P,Q)(P,Q) has rank 6N−46N-4, which equals the universal upper bound. For N≥9N\ge9 this is less than N2N^2, so the asserted non-filling second-secant dimension follows. For q=2q=2 the exact certificate in Section 8 supplies the filling dimension 16. □

Remark 6.2. The admissible vectors dd form a nonempty Zariski-open set defined by the two-square and three-square minors used in the proof. The higher-secant question is considered in Section 9 and in Conjecture 11.1.

Symmetries of central-character pairs

In this section, we describe the changes of type that preserve fusion width and secant dimensions.

Proposition 7.1 (Scaling and exchange symmetries) Fusion width is invariant under

(s,t)↦(λs,λt),λ∈Fq×,(s,t)\mapsto(\lambda s,\lambda t),\qquad\lambda\in\mathbb{F}_q^\times,

and under exchange (s,t)↔(t,s)(s,t)\leftrightarrow(t,s). These transformations induce linear isomorphisms of the ambient bilinear spaces that preserve every channel image and every secant closure.

Proof. For λ≠0\lambda\ne0, the map

σλ:g(a,b,c)↦g(λa,λb,c)\sigma_\lambda:g(a,b,c)\mapsto g(\lambda a,\lambda b,c)

is a group automorphism. On the center it sends b↦λbb\mapsto\lambda b, and therefore transports central character ψt\psi_t to ψλt\psi_{\lambda t}. Precomposition with σλ\sigma_\lambda transports the induced representation spaces, their commutants, pointwise products, and hence every channel decomposition. Thus simultaneous nonzero scaling preserves fusion width. Exchange invariance follows from commutativity of pointwise multiplication after interchanging the two preprocessing operators. □

Proposition 7.2 (Triality of the three central weights) The one-channel varieties and all their secant closures are preserved, up to linear isomorphism of their ambient spaces, by permutations of the triple

(s,t,−s−t).(s,t,-s-t).

In particular, besides simultaneous scaling and input exchange, the transformation (s,t)↦(−s−t,t)(s,t)\mapsto(-s-t,t) preserves the entire secant-dimension profile and every exact fusion-width stratum.

Proof. For every type uu, the pairing

βu(f,h)=∑x∈Fq2f(x)h(x),f∈Hu,h∈H−u,\beta_u(f,h)=\sum_{x\in\mathbb{F}_q^2} f(x)h(x), \qquad f\in\mathcal{H}_u,\quad h\in\mathcal{H}_{-u},

is nondegenerate and GG-invariant: the opposite central phases cancel, and the remaining action permutes the summands. It identifies H−u\mathcal{H}_{-u} with the dual of Hu\mathcal{H}_u. Consequently matrix transpose induces a linear isomorphism End⁡G(Hu)→End⁡G(H−u)\operatorname{End}_G(\mathcal{H}_u)\to\operatorname{End}_G(\mathcal{H}_{-u}). Pairing an output of type u=s+tu=s+t with a third input hh of type −u-u identifies each equivariant bilinear map with a GG-invariant trilinear form. A channel becomes

βu(C[(Af)(Bg)],h)=∑x(Af)(x)(Bg)(x)(CTh)(x).\beta_u(C[(Af)(Bg)],h)=\sum_x (Af)(x)(Bg)(x)(C^{T}h)(x).

This expression is symmetric in the three factors. Permuting the factors and their preprocessing operators, with transpose when a factor is designated as the output, carries channels bijectively to channels. These transformations are linear on the ambient spaces of multilinear forms and preserve sums of any number of channels, proving all assertions. □

For pairs with both entries nonzero, simultaneous scaling reduces to (1,r)(1,r), and exchange sends r↦r−1r\mapsto r^{-1}. Table 1 gives the orbit representatives retained in the exact certificates: These representatives cover all ordered pairs in the corresponding fields.

fieldrepresentatives of ordered type pairs
F2\mathbb{F}_2(0,0),(0,1),(1,1)(0,0), (0,1), (1,1)
F3\mathbb{F}_3(0,0),(0,1),(1,1),(1,2)(0,0), (0,1), (1,1), (1,2)
F4\mathbb{F}_4(0,0),(0,1),(1,1),(1,α)(0,0), (0,1), (1,1), (1,\alpha), α2+α+1=0\alpha^2+\alpha+1=0
F5\mathbb{F}_5(0,0),(0,1),(1,1),(1,2),(1,4)(0,0), (0,1), (1,1), (1,2), (1,4)

Table 1. Archived type representatives under scaling and input exchange.

They retain the original scaling/exchange convention for compatibility with the frozen certificates. Triality gives the smaller lists

qqrepresentatives under scaling and triality
2(0,0),(0,1)(0,0), (0,1)
3(0,0),(0,1),(1,1)(0,0), (0,1), (1,1)
4(0,0),(0,1),(1,α)(0,0), (0,1), (1,\alpha)
5(0,0),(0,1),(1,1)(0,0), (0,1), (1,1)

Table 4.

Indeed, a triple with a zero entry and not all entries zero is equivalent to (0,1,−1)(0,1,-1). In F3\mathbb{F}_3 all nonzero zero-sum triples are proportional to (1,1,1)(1,1,1); in F4\mathbb{F}_4 they are permutations of (1,α,α2)(1,\alpha,\alpha^2) up to scale; in F5\mathbb{F}_5 they are equivalent to (1,1,3)(1,1,3). Thus only 2, 3, 3, 3 classes are needed, respectively; the additional archived classes are redundant checks.

Exact computer-assisted certification for q=2,3,4,5q=2,3,4,5

For each type representative, we give exact Jacobian certificates at the width immediately before filling and at the first filling width. The supplement also checks every smaller width.

Jacobian coordinates

Choose the twisted-shift bases of the three commutants. Write a channel as kernels a,b,c∈CNa,b,c \in\mathbb{C}^{N}. Let A(a),B(b),C(c)A(a),B(b),C(c) be the corresponding N×NN \times N matrices. If coc_o denotes the row of C(c)C(c) producing the output at the basepoint, then under the identification (12) the channel coefficient matrix is

M(a,b,c)=A(a)Tdiag⁡(co)B(b).(22)M(a,b,c)=A(a)^{T}\operatorname{diag}(c_o)B(b). \tag*{(22)}

Thus Φr\Phi_r is a polynomial map from A3Nr\mathbb{A}^{3Nr} to AN2\mathbb{A}^{N^2}, and its Jacobian has N2N^2 rows and 3Nr3Nr columns.

Lifting modular rank to characteristic zero

Lemma 8.1 (Modular maximal-minor certificate) Let RR be the cyclotomic integer ring containing the character values occurring in a fixed qq-model, and let JJ be a Jacobian matrix with entries in RR evaluated at a parameter point in RR. Suppose a ring homomorphism R→kR \to k to a finite field sends some m×mm \times m minor of JJ to a nonzero element of kk. Then that minor is nonzero over C\mathbb{C}, so the complex Jacobian rank is at least mm.

Proof. If the determinant were zero in RR, every ring homomorphism would send it to zero. A nonzero reduction therefore proves that the algebraic integer determinant is nonzero. Embedding the cyclotomic ring into C\mathbb{C} preserves nonzeroness. □\square

If m=N2m=N^2, the differential is surjective. Since source and target are smooth affine spaces, the parameterization is dominant. Combined with the lower bound, this proves exact generic width.

Lemma 8.2 (Maximal rank passes to subcollections) Let PP and VV be complex affine spaces, let F:P→VF:P \to V be a polynomial one-channel parameterization over C\mathbb{C}, let d=dim⁡F(P)‾d=\dim\overline{F(P)}, and let Φr\Phi_r be the sum of rr copies of FF. Suppose that at a point (p1,…,pm)(p_1,\ldots,p_m),

rank⁡dΦm=md.\operatorname{rank} d\Phi_m=md.

Then, for every subset S⊆{1,…,m}S \subseteq\{1,\ldots,m\}, the sum parameterization on the points indexed by SS has differential rank ∣S∣d|S|d. Hence dim⁡im⁡Φr‾=rd\dim\overline{\operatorname{im}\Phi_r}=rd for 1≤r≤m1 \le r \le m.

Proof. Write Ti=im⁡dFpiT_i=\operatorname{im}dF_{p_i}. Each TiT_i has dimension at most dd, since the rank of the differential is bounded above by the dimension of the image closure in characteristic zero. The hypothesis states

md=dim⁡(T1+⋯+Tm)≤∑idim⁡Ti≤md.md=\dim(T_1+\cdots+T_m)\leq\sum_i\dim T_i\leq md.

All inequalities are equalities. Thus each TiT_i has dimension dd and their sum is direct. Every subcollection is therefore a direct sum of dimension ∣S∣d|S|d. The general sum-dimension upper bound proves equality for the corresponding image closures.

Applied after modular lifting with m=r∗−1m=r_\ast-1, Lemma 8.2 shows that a penultimate rank (r∗−1)(3N−2)(r_\ast-1)(3N-2), together with a full-rank certificate at r∗r_\ast, already proves the entire profile. This reduces the necessary number of rank witnesses to two per type class. For q=4,5q=4,5, the archive additionally retains and verifies every sub-filling minor as a redundant check; the q=2,3q=2,3 scripts verify each prefix rank directly.

Certified cases

Table 2 summarizes the exact arithmetic choices. For q=3q=3 and q=5q=5, the auxiliary prime is chosen so that the finite field contains the needed roots of unity. For q=2,4q=2,4, additive characters take values {±1}\{\pm1\}, so reduction modulo 101 is immediate. The specialization maps used by the certificates are explicit:

qqN=q2N=q^2target N2N^2thresholdcoefficient fieldranks
24162F101\mathbb{F}_{101}, ω2=−1\omega_2=-110→1610\to16
39814F109\mathbb{F}_{109}, ω3=45\omega_3=4575→8175\to81
4162566F101\mathbb{F}_{101}, ψt(z)=(−1)Tr⁡(tz)\psi_t(z)=(-1)^{\operatorname{Tr}(tz)}230→256230\to256
5256259F101\mathbb{F}_{101}, ω5=36\omega_5=36584→625584\to625

Table 2. Exact coefficient fields and ranks at the penultimate and filling widths.

Z[ζ3]⟶F109,ζ3⟼45,\mathbb{Z}[\zeta_3]\longrightarrow\mathbb{F}_{109},\qquad\zeta_3\longmapsto45,

where 45 has exact order 3 modulo 109, and

Z[ζ5]⟶F101,ζ5⟼36,\mathbb{Z}[\zeta_5]\longrightarrow\mathbb{F}_{101},\qquad\zeta_5\longmapsto36,

where 36 has exact order 5 modulo 101. For q=2q=2 the nontrivial additive character is (−1)z(-1)^z, and for q=4q=4 the additive character is (−1)Tr⁡(tz)(-1)^{\operatorname{Tr}(tz)}, so all character values lie in Z\mathbb{Z} and we use the ordinary reduction Z→F101\mathbb{Z}\to\mathbb{F}_{101}. The verifier checks the required root orders before constructing the Jacobians.

Here the two ranks displayed are for widths r∗−1r_\ast-1 and r∗r_\ast, with

r∗=⌈q43q2−2⌉.r_\ast=\left\lceil\frac{q^4}{3q^2-2}\right\rceil.

For q=2q=2, a fixed width-two witness gives rank 10 at width one and full rank 16 at width two for each representative (0,0),(0,1),(1,1)(0,0),(0,1),(1,1). The corresponding certified 16×1616 \times16 maximal-minor determinants modulo 101 are

87,61,37,87,\quad61,\quad37,

respectively. Thus the smallest field already attains the parameter-count threshold ⌈16/10⌉=2\lceil16/10\rceil=2.

For q=3q=3, the same fixed parameter point gives the complete rank sequence

25, 50, 75, 8125,\ 50,\ 75,\ 81

for widths 1,2,3,41,2,3,4, respectively, in every type-pair orbit. The certified 81×8181 \times81 minor determinants modulo 109 for the representatives (0,0),(0,1),(1,1),(1,2)(0,0),(0,1),(1,1),(1,2) are, respectively,

16,21,71,34.16,\quad21,\quad71,\quad34.

For q=4,5q=4,5, the supplement stores the parameter arrays as fixed JSON data, records the pivot columns of an explicit maximal minor, and verifies its determinant without using a pseudorandom number generator. For q=4q=4, the certified 256×256256 \times256 determinants modulo 101 for (0,0),(0,1),(1,1),(1,α)(0,0),(0,1),(1,1),(1,\alpha) (with the implementation encoding α\alpha by the integer 2) are

91,49,12,47.91,\quad49,\quad12,\quad47.

For q=5q=5, the certified 625×625625 \times625 determinants modulo 101 for (0,0),(0,1),(1,1),(1,2),(1,4)(0,0),(0,1),(1,1),(1,2),(1,4) are

49,78,43,24,85.49,\quad78,\quad43,\quad24,\quad85.

A redundant exact recheck at every sub-filling width gives, for every representative type pair,

qqJacobian ranks at widths 1,2,…,r∗1,2,\ldots,r_\ast
210, 16
325, 50, 75, 81
446, 92, 138, 184, 230, 256
573, 146, 219, 292, 365, 438, 511, 584, 625.

Table 5.

Before filling, each rank equals r(3q2−2)r(3q^2-2). At the first filling width it equals q4q^4. All ranks and determinants are computed by exact modular arithmetic.

Theorem 8.3 (Secant dimensions for q=2,3,4,5q=2,3,4,5) For q∈{2,3,4,5}q \in\{2,3,4,5\}, every central-character pair (s,t)(s,t), and every r≥1r \ge1,

dim⁡im⁡Φr‾=min⁡{r(3q2−2),q4}.(23)\dim\overline{\operatorname{im}\Phi_r}=\min\{r(3q^2-2),q^4\}. \tag*{(23)}

Equivalently, every projective secant through the first filling order has its expected dimension. In particular,

rgen(2;s,t)=2,rgen(3;s,t)=4,rgen(4;s,t)=6,rgen(5;s,t)=9.(24)r_{\mathrm{gen}}(2;s,t)=2,\quad r_{\mathrm{gen}}(3;s,t)=4,\quad r_{\mathrm{gen}}(4;s,t)=6,\quad r_{\mathrm{gen}}(5;s,t)=9. \tag*{(24)}

Proof. Theorem 5.7 gives the upper bound dim⁡im⁡Φr≤min⁡{r(3q2−2),q4}\dim\operatorname{im}\Phi_r \le\min\{r(3q^2-2),q^4\}. For each certified type-pair representative, the exact modular Jacobian ranks displayed above attain this bound at every width through r∗r_\ast. By Lemma 8.1, the same ranks hold as lower bounds in characteristic zero, proving equality in (23) up to the filling order. At r=r∗r=r_\ast the map is dominant; hence it remains dominant for every larger rr by setting the extra channels to zero. § 7 transports the result from the representatives to all ordered type pairs. The width values then follow from Proposition 5.3 and the general lower bound. □\square

Remark 8.4 (Scope of the computation) The structure of the induced representations, the twisted-convolution characterization, the bilinear selection rule, the target dimension q4q^4, the exact one-channel dimension 3q2−23q^2-2, the general lower bound, and the type-pair symmetry reductions are proved analytically. Theorem 6.1 also proves order-two nondefectivity analytically for every q≥3q \ge3. Computation supplies the complete small-field profiles and hence the base witnesses used in Corollary 9.5; the lifting theorem itself is analytic.

Lifting witnesses from additive subgroups

Theorem 9.1 (Lifting from a subgroup) Let KK be a finite abelian group, let H≤KH \le K, and write N=∣K∣N=|K| and n=∣H∣n=|H|. Let βs,βt:K×K→C×\beta_s,\beta_t:K\times K\to\mathbb{C}^{\times} be bicharacters. Consider the polynomial channel map

FK(a,b,d)i,j=∑z∈Kai−zbj−zdzβs(z,i−z)βt(z,j−z),i,j∈K,F_K(a,b,d)_{i,j}=\sum_{z\in K}a_{i-z}b_{j-z}d_z\beta_s(z,i-z)\beta_t(z,j-z),\qquad i,j\in K,

and define FHF_H by restricting the indices and bicharacters to HH. Suppose that, at rr parameter triples pℓ=(aℓ,bℓ,dℓ)p_\ell=(a_\ell,b_\ell,d_\ell) on HH, all three vectors in each triple are nonzero and

rank⁡d(∑ℓ=1rFH)(p1,…,pr)=r(3n−2).\operatorname{rank} d\left(\sum_{\ell=1}^{r}F_H\right)_{(p_1,\ldots,p_r)}=r(3n-2).

Extend each parameter vector by zero from HH to KK. At the resulting parameter point on $K,

rank⁡d(∑ℓ=1rFK)=r(3N−2).\operatorname{rank} d\left(\sum_{\ell=1}^{r}F_K\right)=r(3N-2).

Consequently the rr-channel image closure on KK has dimension r(3N−2)r(3N-2).

Proof. The two scaling directions

(aℓ,0,−dℓ),(0,bℓ,−dℓ)(a_\ell,0,-d_\ell),\qquad(0,b_\ell,-d_\ell)

lie in the kernel of the restricted Jacobian. They are independent, including across channels. The rank hypothesis says that these 2r2r vectors span its entire kernel. It follows that each of the three partial Jacobians, obtained by varying only all aa-vectors, only all bb-vectors, or only all dd-vectors, is injective and has rank rnrn. Indeed, a vector in such a partial kernel would have to be a linear combination of the displayed scaling directions; vanishing of the other two components and nonzeroness of the parameter vectors force all its coefficients to vanish. Let UU be representatives for K/HK/H, including 00. At the zero-extended point, the columns indexed by parameters in HH have support in H×HH\times H and reproduce the restricted Jacobian. For every u∈U∖{0}u \in U \setminus\{0\} the remaining columns split into three groups with the following mutually disjoint supports:

varied parametersoutput block
au+ha_{u+h}, h∈Hh \in H(u+H)×H(u+H) \times H
bu+hb_{u+h}, h∈Hh \in HH×(u+H)H \times(u+H)
du+hd_{u+h}, h∈Hh \in H(u+H)×(u+H)(u+H) \times(u+H)

Table 1 (PDF p. 19).

No other output blocks occur in the differential. The same block supports hold simultaneously for every channel.

Each outside block is equivalent, by invertible row and column diagonal scalings, to the corresponding partial Jacobian on HH. Here are those scalings explicitly. In the aa-block write i=u+i′i=u+i', j=j′j=j', with i′,j′∈Hi',j' \in H, and denote the outside parameter by au+ha_{u+h}. The extra phase is

βs(z,u)=βs(i′,u)βs(h,u)−1,z=i′−h.\beta_s(z,u)=\beta_s(i',u)\beta_s(h,u)^{-1}, \qquad z=i'-h.

Thus the row factor is βs(i′,u)\beta_s(i',u) and the column factor is βs(h,u)−1\beta_s(h,u)^{-1}. In the bb-block the factors are βt(j′,u)\beta_t(j',u) and βt(h,u)−1\beta_t(h,u)^{-1}. In the dd-block write i=u+i′i=u+i', j=u+j′j=u+j', z=u+hz=u+h. The extra phase is

βs(u,i′−h)βt(u,j′−h)=βs(u,i′)βt(u,j′)[βs(u,h)βt(u,h)]−1.\beta_s(u,i'-h)\beta_t(u,j'-h)=\beta_s(u,i')\beta_t(u,j')[\beta_s(u,h)\beta_t(u,h)]^{-1}.

These factors again separate into a row factor and a column factor, and all are nonzero. Therefore every outside partial block has rank rnrn.

All displayed blocks have disjoint row and column supports, so their ranks add. The complete rank is

r(3n−2)+3(∣K/H∣−1)rn=r(3N−2).r(3n-2)+3(|K/H|-1)rn=r(3N-2).

The scaling argument of Lemma 5.6 also applies to FKF_K and gives image dimension at most r(3N−2)r(3N-2). The computed Jacobian rank gives the reverse inequality. □\square

Corollary 9.2 (One-channel dimension for finite abelian groups) Let KK be a finite abelian group and let βs,βt:K×K→C×\beta_s,\beta_t:K \times K \to\mathbb{C}^{\times} be bicharacters. For the parameterization FKF_K of Theorem 9.1,

dim⁡im⁡FK‾=3∣K∣−2.\dim\overline{\operatorname{im} F_K}=3|K|-2.

Proof. Take H={0}H=\{0\} and r=1r=1. Since βs(0,0)=βt(0,0)=1\beta_s(0,0)=\beta_t(0,0)=1, the restricted map is FH(a,b,d)=abdF_H(a,b,d)=abd. Its differential at (1,1,1)(1,1,1) has rank 1=3∣H∣−21=3|H|-2. Theorem 9.1 gives the assertion, including when KK is trivial. □\square

Remark 9.3 (Necessity of the nonvanishing hypothesis) The rank hypothesis in Theorem 9.1 does not suffice if a parameter vector is zero. Indeed, take H={0}H=\{0\}, r=1r=1, and (a,b,d)=(0,1,1)(a,b,d)=(0,1,1). The restricted differential has rank 1=3∣H∣−21=3|H|-2. After zero extension to a nontrivial group KK, it is

dFK(x,y,z)i,j=xiδj,0.dF_K(x,y,z)_{i,j}=x_i\delta_{j,0}.

Its rank is ∣K∣<3∣K∣−2|K|<3|K|-2.

Remark 9.4 (Application to the twisted Heisenberg channel) For K=(Fq2,+)K=(\mathbb{F}_q^2,+), take

βt((x,y),(u,v))=ψt(−xv).\beta_t((x,y),(u,v))=\psi_t(-xv).

These are bicharacters, and the displayed polynomial channel is exactly the basepoint matrix parameterization A(a)Tdiag⁡(d)B(b)A(a)^{T}\operatorname{diag}(d)B(b). Hence Theorem 9.1 applies to every central-character

pair without an untwistedness assumption. The hypothesis requires rank r(3∣H∣−2)r(3|H|-2). A filling witness of smaller rank does not satisfy this hypothesis. Equality r(3∣H∣−2)=∣H∣2r(3|H|-2)=|H|^2 is allowed; the witnesses used below have rank strictly less than ∣H∣2|H|^2.

Corollary 9.5 (Dimensions over extension fields) For every central-character pair, the expected affine secant dimension

dim⁡im⁡Φr‾=r(3q2−2)\dim\overline{\operatorname{im}\Phi_r}=r(3q^2-2)

holds in each of the following ranges:

  1. q=3fq=3^f, f≥1f\ge1, and 1≤r≤31\le r\le3;

  2. q=5fq=5^f, f≥1f\ge1, and 1≤r≤81\le r\le8;

  3. q=2fq=2^f, with either f=2f=2 or f≥4f\ge4, and 1≤r≤51\le r\le5.

For q=3,4,5q=3,4,5 these are the certified cases of Theorem 8.3; for the remaining extension fields they are consequences of subgroup lifting.

Proof For the first two assertions take H=Fp2⊆Fq2H=\mathbb{F}_p^2\subseteq\mathbb{F}_q^2, where p=3p=3 or 55. Write the fixed nontrivial additive character as ψ(z)=exp⁡(2πiTr⁡Fq/Fp(κz)/p)\psi(z)=\exp(2\pi i\operatorname{Tr}_{\mathbb{F}_q/\mathbb{F}_p}(\kappa z)/p) for some κ≠0\kappa\ne0. The restricted bicharacters are precisely the prime-field models with effective central types

s0=Tr⁡Fq/Fp(κs),t0=Tr⁡Fq/Fp(κt).s_0=\operatorname{Tr}_{\mathbb{F}_q/\mathbb{F}_p}(\kappa s),\qquad t_0=\operatorname{Tr}_{\mathbb{F}_q/\mathbb{F}_p}(\kappa t).

Every such pair was certified at all sub-filling orders up to 33 for p=3p=3 and up to 88 for p=5p=5. A corresponding prefix of the frozen witness supplies the hypothesis of Theorem 9.1, by Lemma 8.2. Lemma 8.2 also shows that each individual channel differential has rank 3∣H∣−23|H|-2. Each triple therefore has nonzero components: a zero component would reduce its differential rank to at most ∣H∣<3∣H∣−2|H|<3|H|-2.

For characteristic two and f≥4f\ge4, choose a four-dimensional F2\mathbb{F}_2-linear subspace U⊆FqU\subseteq\mathbb{F}_q and let H=U×{0}H=U\times\{0\}. Then H≅(F2)4≅(F24,+)H\cong(\mathbb{F}_2)^4\cong(\mathbb{F}_2^4,+) and both restricted bicharacters are trivial, for every ambient pair (s,t)(s,t). The untwisted q=4q=4 certificates supply rank r(3⋅16−2)r(3\cdot16-2) for r≤5r\le5. Apply the lifting theorem. The case f=2f=2 is already part of the small-field theorem. □

Remark 9.6 Corollary 9.5 uses the exact certificates on the base groups and the subgroup-lifting theorem. Its ranges concern the stated secant orders and do not determine the first filling width for arbitrary qq.

Secant geometry and comparison of decomposition widths

Let Xs,t⊂Vs,tX_{s,t}\subset V_{s,t} be the affine cone obtained as the Zariski closure of one-channel maps. We use the convention σr(PX)\sigma_r(\mathbb{P}X) for the Zariski closure of projective linear spans of rr points of PX\mathbb{P}X, so σ1(PX)=PX\sigma_1(\mathbb{P}X)=\mathbb{P}X. With this convention, the closure of the width-rr image is the affine cone over σr(PXs,t)\sigma_r(\mathbb{P}X_{s,t}). Thus the generic fusion width is the first secant order whose affine cone fills the ambient bilinear space.

By Theorem 5.7,

dim⁡PXs,t=3q2−3\dim\mathbb{P}X_{s,t}=3q^2-3

@page 20/30 starts_mid=1 ends_mid=1 for every type pair. Hence the expected projective dimension of the rr-th secant is

expdim⁡σr(PXs,t)=min⁡{r(3q2−2)−1,q4−1}.(25)\operatorname{expdim} \sigma_r(\mathbb{P}X_{s,t})=\min\{r(3q^2-2)-1,q^4-1\}. \tag*{(25)}

Thus Conjecture 11.1 concerns nondefectivity at the first expected filling order. For the untwisted case, we obtain a toric description. Let

A=G/Z≅(Fq2,+),A=G/Z\cong(\mathbb{F}_q^2,+),

and take Fourier transform on the finite abelian group AA. Ordinary convolution operators diagonalize. If their Fourier multipliers are denoted by Aa,Bb,CkA_a,B_b,C_k, then pointwise multiplication in the spatial domain becomes convolution in frequency. Equivariance forces output frequency k=a+bk=a+b, and one channel has coefficient array

Da,b=AaBbCa+b,a,b∈A^.(26)D_{a,b}=A_aB_bC_{a+b},\qquad a,b\in\widehat{A}. \tag*{(26)}

After replacing the third index by its negative, (26) is the standard monomial parameterization of a group-based three-leaf claw (tripod) model for the finite abelian group AA. Consequently:

Proposition 10.1 (Untwisted toric correspondence) The projective one-channel variety for (s,t)=(0,0)(s,t)=(0,0) is the group-based tripod toric variety for A≅Fq2A\cong\mathbb{F}_q^2, up to a relabeling of the third Fourier index. The projectivizations of the width-rr image closures are its rr-th secant varieties. The displayed Fourier formula depends on the normalization convention for the finite Fourier transform: with a normalized inverse transform a nonzero global scalar factor may appear in Da,bD_{a,b}. Such a factor is absorbed into the third kernel CC, and therefore does not alter the affine/projective one-channel variety or any fusion-width statement.

This identifies the untwisted fusion problem with a familiar nondefectivity question from algebraic statistics [4]. More precisely, its all-orders nondefectivity is a special case of their Conjecture 5.1. Their Proposition 5.3 proves the q=2q=2 untwisted second-secant case, since A=F22A=\mathbb{F}_2^2 gives the Kimura three-parameter tripod. We include that case for completeness and reproducibility, not as a new theorem. If s≠0s\ne0 or t≠0t\ne0, at least one input commutant is isomorphic to the noncommutative algebra Mq(C)M_q(\mathbb{C}). Thus the simultaneous scalar Fourier diagonalization used for (s,t)=(0,0)(s,t)=(0,0) does not apply to these pairs. Theorem 6.1 proves order-two nondefectivity uniformly in these twisted sectors. The exact certificates reach the first filling order for every prime power q≤5q\le5, and Theorem 9.1 propagates their sub-filling ranks to the infinite families above.

For the polynomial parameterization used here, filling of the secant closure is equivalent to dominance. In either formulation, constructibility supplies a nonempty Zariski-open subset of the actual image, so the generic statement is about exact width. Exact width and border width may still differ at special targets. The full-rank Jacobian certificates provide an explicit sufficient test for this dominance.

Remark 10.2 (Comparison with quadratic Waring rank) The two notions of width differ even for one fusion channel. For any type pair, take A=B=IA=B=I and choose the output kernel so that its basepoint row is d=1d=1. Equation (22) then gives M=INM=I_N, where N=q2N=q^2. This nonzero equivariant bilinear map has fusion width one. Its scalar basepoint polynomial on Hs⊕Ht≅C2N\mathcal{H}_s\oplus\mathcal{H}_t\cong\mathbb{C}^{2N} is

p(f,g)=fTg=(fg)T(0IN/2IN/20)(fg).p(f,g)=f^{T}g=\begin{pmatrix}f\\g\end{pmatrix}^{T}\begin{pmatrix}0&I_{N/2}\\I_{N/2}&0\end{pmatrix}\begin{pmatrix}f\\g\end{pmatrix}.

The symmetric coefficient matrix has rank 2N2N, so at least 2N2N squares of linear forms are required. The identity

p(f,g)=14∑i=1N((fi+gi)2−(fi−gi)2)p(f,g)=\frac{1}{4}\sum_{i=1}^{N}\left((f_i+g_i)^2-(f_i-g_i)^2\right)

attains this bound over C\mathbb{C}. Thus the quadratic Waring rank is 2N2N, as in the symmetric-rank characterization of Akbari and Jamshidi (2026), Theorem 6. Here the Waring rank refers to the scalar basepoint polynomial, whereas fusion width refers to its equivariant vector-valued extension.

A conjecture on generic width

Theorem 8.3 leads to the following conjecture.

Conjecture 11.1 (Generic fusion width) For every prime power qq and every s,t∈Fqs,t\in\mathbb{F}_q,

rgen(q;s,t)=⌈q43q2−2⌉.(27)r_{\mathrm{gen}}(q;s,t)=\left\lceil\frac{q^4}{3q^2-2}\right\rceil. \tag*{(27)}

For q=2,3,4,5q=2,3,4,5, Theorem 8.3 proves the conjecture and also gives the expected dimensions at every smaller secant order. These cases include the extension field F4\mathbb{F}_4. They do not exclude the possibility of a defect at a larger prime power.

By Theorem 5.7, Conjecture 11.1 is equivalent to nondefectivity at the first expected filling order. It makes no assertion about every smaller secant order. In the untwisted case, the question belongs to the theory of group-based toric mixtures. For nonzero central characters, the commutants are matrix algebras, so the same Fourier reduction is unavailable. Theorem 6.1 settles order two uniformly, and Corollary 9.5 settles further orders on infinite families. Reaching the first filling order for arbitrary qq requires additional control of sums of more differential images.

The lower bound satisfies

q43q2−2=q23+O(1).\frac{q^4}{3q^2-2}=\frac{q^2}{3}+O(1).

Thus generic representation requires a number of channels that grows quadratically with qq. The width is measured with respect to the prescribed channel map. The certificates do not give an algorithm attaining the generic minimum for an arbitrary input, nor do they establish accuracy or efficiency on a learning task.

Reproducibility

Online Resource 1 contains stored parameter points, specified nonzero minors and certificates for all secant dimensions, and read-only verification scripts. Table 3 summarizes the entry points. The verifiers compare the computed values with the stored indices and determinants and do not modify the certificate files.

Entry pointPurpose
verify_all.pyCheck hashes, run the certificate suite and confirm unchanged file bytes
verify_checksums.pyValidate the SHA-256 manifest before mathematical verification
verify_q2_all_twisted_sectors.py, verify_q3_all_twisted_sectors.pyVerify all prefix ranks and fixed terminal determinants
verify_q45_frozen_certificates.pyVerify stored hashes, penultimate ranks and stored terminal minors
verify_full_secant_profiles_q45.pyRecompute every stored sub-filling minor for q=4,5q = 4, 5
verify_independent_coordinates.pyReconstruct the Jacobians directly from coefficient sums, independently of the matrix-product implementation
decompose_basepoint_matrix.pyExact rational rank-factorization example for Proposition 5.4
verify_theory_upgrades.pyExact checks of the second-secant parameter points and a subgroup lift

Table 3. Verification and constructive-reconstruction entry points in Online Resource 1.

The tested environment was Python 3.12.14 and NumPy 2.3.5 on Windows 10. The q=2,3q = 2, 3 verifiers use Python integer arithmetic. The q=4,5q = 4, 5 verifiers use integer arrays reduced modulo 101, with modular inverses; no floating-point rank routines enter the certificates. For these fixed dimensions, unreduced dot products have magnitude at most 25⋅100325 \cdot100^{3}, well within the signed 64-bit range. Integer parameter residues are lifted to ordinary integers and character coefficients via the displayed cyclotomic specialization; no embedding of a finite field into C\mathbb{C} is assumed.

To reproduce the computations, run pythonverify_all.py. It validates checksums before and after the verification runs, refuses optimized Python execution that would disable legacy assertions, and reports nonzero exit status on a failed check. The independent coordinate verifier is an optional additional audit. The script verify_theory_upgrades.py checks 32 stored two-channel parameter points for q=2,3,4,5,7,8,9,11q = 2, 3, 4, 5, 7, 8, 9, 11, compares two Jacobian constructions, and verifies four width-three lifts from q=3q = 3 to q=9q = 9, each of rank 723. The proofs of Theorems 6.1 and 9.1 do not depend on these checks. They use Python-integer modular elimination over F2311\mathbb{F}_{2311}; small matrix-assembly comparisons at q=3,4q = 3, 4 use NumPy integer arrays within the signed 64-bit range. The witness SHA-256 values are

q4: affb0a1a77348bb33f7ac797529fa152958f8cdbe19616740e8dbf59811e3c15
q5: 3a6d802a5b097dd5d8081a17c2dba3303ba7792156b06236e87f2c74557d4e19.

All certificate data are self-contained and require no external datasets. The full profile can be proved using only the penultimate and terminal rank checks by Lemma 8.2; the additional stored minors make each order separately inspectable.

Conclusion and future work

We have determined the one-channel dimension and the second-secant dimension for the bilinear channel varieties associated with representations of UT3(Fq)UT_3(\mathbb{F}_q) induced from central characters. The affine one-channel variety has dimension 3q2−23q^2 - 2 for every type pair. The affine second-secant closure has dimension 6q2−46q^2 - 4 for q≥3q \ge3, while it fills the ambient space for q=2q = 2. The subgroup-lifting theorem transfers differential rank r(3∣H∣−2)r(3|H| - 2) on HH to rank r(3∣K∣−2)r(3|K| - 2) on KK, provided that every parameter vector is nonzero. Together with the exact small-field certificates, it gives higher-secant dimensions on the extension-field families in Corollary 9.5. For q=2,3,4,5q = 2,3,4,5, all secant dimensions are determined and the generic widths are 2,4,6,92,4,6,9, respectively.

The main unresolved question is Conjecture 11.1, which asserts that the dimension lower bound ⌈q4/(3q2−2)⌉\lceil q^4/(3q^2-2)\rceil is attained for every prime power and every central-character pair. A proof requires control of the differential rank at the first expected filling order. The present subgroup construction preserves a fixed channel order and therefore does not reach that order when the extension degree grows. One possible approach is to combine several translated subgroup supports and determine the intersections of their differential images. This would require an argument beyond the zero-extension construction used here. Determining the dimensions at every smaller secant order is a separate, stronger problem.

A further question concerns reconstruction at small width. The rank-factorization construction gives at most q2q^2 channels for any target, whereas the generic-width results establish existence without giving a procedure that attains the generic minimum. An algorithm for finding such decompositions would require analysis of solvability and numerical stability. The nonsingular minor used in the one-channel proof gives local information at the selected parameter point; it does not provide a conditioning estimate for general decompositions. The polynomial realization and approximation results in Appendices D and E do not supply such quantitative decomposition bounds.

Supplementary information

Online Resource 1 (ESM_1.zip): exact computational certificates and verification code for the secant dimensions, with stored parameter points, nonzero-minor data, reproduction logs and a SHA-256 checksum manifest.

Details of the basepoint matrix formula

Let eie_i, 1≤i≤N1 \le i \le N, be the delta basis of CFq2\mathbb{C}^{\mathbb{F}_q^2}. Write

Af=Af,Bg=BgAf = Af,\qquad Bg = Bg

in matrix notation, and let coc_o be the basepoint row of CC. Then

ΓA,B,C(f,g)(o)=∑z=1N(co)z(Af)z(Bg)z.\Gamma_{A,B,C}(f,g)(o) = \sum_{z=1}^{N}(c_o)_z(Af)_z(Bg)_z.

Expanding gives

∑i,jfigj(∑zAzi(co)zBzj),\sum_{i,j} f_i g_j\left(\sum_z A_{zi}(c_o)_zB_{zj}\right),

so the coefficient matrix is

M=ATdiag⁡(co)B,M = A^T\operatorname{diag}(c_o)B,

which is (22). Differentiating this expression with respect to the twisted-shift kernel coefficients gives the Jacobian implemented in the supplementary scripts.

Dominance and exact representation

For completeness, let X=AmX = \mathbb{A}^m, Y=AnY = \mathbb{A}^n, and Φ:X→Y\Phi: X \to Y be a polynomial map over C\mathbb{C}. If dΦxd\Phi_x has rank nn at one point, then the image has dimension nn, hence Φ\Phi is dominant. Chevalley’s theorem says Φ(X)\Phi(X) is constructible. A dense constructible subset of irreducible YY contains a nonempty open subset. Thus dominance obtained from a Jacobian certificate implies that a generic target is represented by an actual finite parameter point; no limiting argument is needed.

Certificate orbit counts

For reference, simultaneous scaling and exchange classify nonzero ratios by

r∼r−1,r=t/s∈Fq×.r \sim r^{-1},\qquad r=t/s\in\mathbb{F}_q^\times.

Pairs with exactly one zero form one orbit, and (0,0)(0,0) forms another. Thus the representative lists used in § 7 are exhaustive for q=2,3,4,5q=2,3,4,5.

Equivariant polynomial maps

The neural-network interpretation uses group-equivariant convolutions as in Cohen and Welling [9] and Kondor and Trivedi [10]. General equivariant universality results are available in Ravanbakhsh [11] and Nyholm et al. [12]. In Graded Representation Theory of Equivariant Neural Networks, Shayestehfar (2026), arXiv:2609.25776, Theorem 3.6 and Proposition 5.1) describes the degree components and their span for fixed equivariant lifts and readouts. His Proposition 5.3 proves that regular-representation branches span every finite-group real equivariant homogeneous polynomial space. Appendix D gives a direct implementation of this existence principle by central-type linear maps, products and conjugation. The width results in Sections 5–9 concern the dimensions obtained when all channel parameters vary.

We now apply the preceding description of the commutants to equivariant polynomial maps. In this section, depth and width are unrestricted.

Let

E=⨁j=1mHτj⊕dj,F=⨁k=1nHυk⊕ek.E = \bigoplus_{j=1}^{m} \mathcal{H}_{\tau_j}^{\oplus d_j}, \qquad F = \bigoplus_{k=1}^{n} \mathcal{H}_{\upsilon_k}^{\oplus e_k}.

A *-polynomial is a polynomial in the complex coordinates and their conjugates. A typed polynomial network is built from the following operations:

  1. arbitrary GG-equivariant complex-linear maps between channels of the same type;

  2. pointwise products Hs×Ht→Hs+t\mathcal{H}_s \times\mathcal{H}_t \to\mathcal{H}_{s+t};

  3. complex conjugation Ht→H−t\mathcal{H}_t \to\mathcal{H}_{-t};

  4. additions, scalar multiplication, and spatially constant fields of type 0.

A constant node is a scalar multiple of the all-ones field in H0\mathcal{H}_0; an arbitrary fixed nonconstant field would not be equivariant. Conjugation is included only for *-polynomial and real-continuous approximation; purely holomorphic polynomial maps do not require it. Formally, such a network is a finite directed acyclic computation graph. Every edge carries a finite direct sum of typed feature channels; linear nodes are intertwiners, product nodes have two incoming typed channels and one outgoing channel of summed type, and conjugation nodes reverse the type. Fan-out is allowed, so a previously computed feature may feed several later nodes. No limiting operation, division, or data-dependent branching is permitted. Exact representation below means equality with the map computed by such a graph.

Lemma D.1 (Equivariant maps are determined at one basepoint) Let Q:E→HuQ : E \to\mathcal{H}_u be GG-equivariant. If Q(v)(o)=0Q(v)(o) = 0 for every v∈Ev \in E, then Q=0Q = 0.

Proof. Every point of G/ZG/Z is g⋅og \cdot o for some g∈Gg \in G. Equivariance expresses Q(v)(g⋅o)Q(v)(g \cdot o) in terms of the basepoint value of Q(g−1v)Q(g^{-1}v), up to the known fiber action. Hence all coordinates vanish. □\square

The general spanning result of Shayestehfar (2026, Proposition 5.3) also yields existence in the present network class. Apply it degree by degree to the underlying real representations and then complexify. The complexified input is E⊕E‾E \oplus\overline{E}, with the original input represented by (v,v‾)(v,\overline{v}). The complexified output is F⊕F‾F \oplus\overline{F}, and projection onto FF recovers the original map. The regular hidden space decomposes as C[G]≅⨁t∈G^Ht\mathbb{C}[G] \cong\bigoplus_{t \in\widehat{G}} \mathcal{H}_t. Its equivariant linear maps split by central type, and its pointwise powers expand into products HsHt⊂Hs+t\mathcal{H}_s\mathcal{H}_t \subset\mathcal{H}_{s+t}. Conjugation supplies the conjugate input factors, and constants have type zero. Thus these branches can be implemented by the allowed operations. The proof below gives a direct basepoint construction. Neither argument gives a minimal-width decomposition.

Theorem D.2 (Exact realization of equivariant ∗\ast-polynomials) Every GG-equivariant ∗\ast-polynomial map Q:E→FQ : E \to F is represented exactly by a finite typed polynomial network.

Proof. It is enough to consider one output copy Hu\mathcal{H}_u. Set

p(v)=Q(v)(o).p(v) = Q(v)(o).

This is a scalar ∗\ast-polynomial in the input coordinates. Because z∈Zz \in Z fixes the basepoint and acts on the output by ψu(z)\psi_u(z), equivariance gives

p(z⋅v)=ψu(z)p(v).p(z \cdot v) = \psi_u(z)p(v).

Thus only central-weight-uu monomials occur in the character decomposition of pp.

A constant term can occur only when u=0u = 0 and is represented by a constant node. Consider a monomial of positive degree. Each holomorphic factor is an input coordinate f(x,y)f(x,y) of some type tt, and each antiholomorphic factor is its conjugate. By (5),

(Rx,ytf)(o)=f(x,y).(R_{x,y}^{t}f)(o) = f(x,y).

Therefore every holomorphic coordinate factor is obtained at the basepoint by an allowed typed linear operator; an antiholomorphic factor is obtained by following the shift with conjugation, which changes type tt to −t-t. Repeated pointwise multiplication realizes the entire monomial as an Hu\mathcal{H}_u-valued equivariant feature because its total central weight is uu. Summing the monomial subnetworks realizes a map whose basepoint value is pp. By the previous lemma it equals QQ everywhere. Repeating this for each output channel proves the theorem. □\square

Remark D.3 In this construction, the type of each monomial is the sum of the types of its holomorphic factors minus the sum of the types of its antiholomorphic factors.

Universal approximation

The polynomial-realization theorem gives the following approximation result by Stone–Weierstrass approximation and Reynolds averaging.

Equip all finite-dimensional representations with GG-invariant Hermitian norms. Let K⊂EK \subset E be compact and GG-invariant.

Theorem E.1 (Universal approximation) For every continuous GG-equivariant map F0:K→FF_0 : K \to F and every ε>0\varepsilon> 0, there exists a typed polynomial network N:E→FN : E \to F such that

sup⁡v∈K∥N(v)−F0(v)∥<ε.\sup_{v \in K} \lVert N(v) - F_0(v) \rVert< \varepsilon.

Proof. The algebra generated by complex coordinate functions and their conjugates is self-adjoint, contains constants, and separates points of the compact subset KK. By the complex

Stone–Weierstrass theorem, vector-valued *-polynomials approximate F0F_0 uniformly coordinatewise. Choose such a polynomial map P:E→FP:E\to F with uniform error less than ε\varepsilon on KK.

Apply Reynolds averaging:

RP(v)=1∣G∣∑g∈GρF(g)P(ρE(g−1)v).(28)\mathcal{R}P(v)=\frac{1}{\lvert G\rvert}\sum_{g\in G}\rho_F(g)P\left(\rho_E\left(g^{-1}\right)v\right). \tag*{(28)}

This is again a *-polynomial and is exactly GG-equivariant. Since KK is invariant and the representations are unitary,

∥RP(v)−F0(v)∥≤1∣G∣∑g∥P(ρE(g−1)v)−F0(ρE(g−1)v)∥<ε.\left\lVert\mathcal{R}P(v)-F_0(v)\right\rVert\leq\frac{1}{\lvert G\rvert}\sum_g\left\lVert P\left(\rho_E\left(g^{-1}\right)v\right)-F_0\left(\rho_E\left(g^{-1}\right)v\right)\right\rVert<\varepsilon.

Theorem D.2 realizes RP\mathcal{R}P exactly by a typed polynomial network. □

The argument combines Theorem D.2 with polynomial approximation and averaging. General equivariant approximation results are given in Ravanbakhsh [11] and Nyholm et al. [12].

Statements and Declarations

Funding

No funding was received for conducting this study.

Author contributions

Nguyen Bich Van developed the theoretical foundations and proofs and wrote the verification code. Nguyen Trung Kien carried out computations for specific examples and checked the code.

Competing interests

The authors declare that they have no competing financial or non-financial interests that are directly or indirectly related to this work.

Data and code availability

The computational data and code supporting the results are supplied as Online Resource 1. They include the exact computer-assisted certificates and frozen witnesses used for Theorem 8.3, verification scripts, reproduction logs, and a SHA-256 checksum manifest. All computations use exact arithmetic; no external datasets were used.

References

  1. [1]Folland, G.B.: Harmonic Analysis in Phase Space. Annals of Mathematics Studies, vol. 122. Princeton University Press, Princeton, NJ (1989). https://doi.org/10.1515/9781400882427
  2. [2]Gurevich, S., Hadani, R.: Quantization of symplectic vector spaces over finite fields. Journal of Symplectic Geometry 7(4), 475–502 (2009) https://doi.org/10.4310/JSG.2009.v7.n4.a4
  3. [3]Gurevich, S., Hadani, R.: The Weil representation in characteristic two. Advances in Mathematics 230(3), 894–926 (2012) https://doi.org/10.1016/j.aim.2012.03.008
  4. [4]Baños, H., Bushek, N., Davidson, R., Gross, E., Harris, P.E., Krone, R., Long, C., Stewart, A., Walker, R.: Dimensions of group-based phylogenetic mixtures. Bulletin of Mathematical Biology 81, 316–336 (2019) https://doi.org/10.1007/s11538-018-0489-0
  5. [5]Couveignes, J.-M., Ezome, T.: The equivariant complexity of multiplication in finite field extensions. Journal of Algebra 622, 694–720 (2023) https://doi.org/10.1016/j.jalgebra.2023.01.022
  6. [6]Huhtanen, M., Noponen, K., Pingamage Don, C.: Best low rank approximations are a special case of best DCD-matrix approximations. Journal of Fourier Analysis and Applications 32, 49 (2026) https://doi.org/10.1007/s00041-026-10259-x
  7. [7]Lin, Y., Fu, C., Krueger, Z., Yu, H., Nakata, M., Xie, J., Kucukbenli, E., Qian, X., Ji, S.: Tensor decomposition networks for fast machine learning interatomic potential computations. In: Advances in Neural Information Processing Systems, vol. 38, pp. 98313–98337 (2025). https://doi.org/10.52202/085713-2958
  8. [8]Blekherman, G., Teitler, Z.: On maximum, typical and generic ranks. Mathematische Annalen 362, 1021–1031 (2015) https://doi.org/10.1007/s00208-014-1150-3
  9. [9]Cohen, T., Welling, M.: Group equivariant convolutional networks. In: Proceedings of the 33rd International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 48, pp. 2990–2999 (2016). https://proceedings.mlr.press/v48/cohenc16.htmlarxiv.org/abs/1602.07576
  10. [10]Kondor, R., Trivedi, S.: On the generalization of equivariance and convolution in neural networks to the action of compact groups. In: Proceedings of the 35th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 80, pp. 2747–2755 (2018). https://proceedings.mlr.press/v80/kondor18a.htmlarxiv.org/abs/1802.03690
  11. [11]Ravanbakhsh, S.: Universal equivariant multilayer perceptrons. In: Proceedings of the 37th International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 119, pp. 7996–8006 (2020). https://proceedings.mlr.press/v119/ravanbakhsh20a.htmlarxiv.org/abs/2002.02912
  12. [12]Nyholm, E., Carlsson, O., Weiler, M., Persson, D.: Equivariant non-linear maps for neural networks on homogeneous spaces. Mathematical Foundations of Machine Learning 2, 1 (2026) https://doi.org/10.1007/s44439-026-00004-z

Paper details

vlnc:W_bqv4b4wu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents