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 , with and in prescribed twisted-convolution algebras. We prove that its affine image closure has dimension for every central-character pair and that its affine second-secant closure has the expected dimension for every prime power . For , 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 on a subgroup extends by zero to a parameter point with differential rank on the ambient group , provided that every parameter vector is nonzero. Exact Jacobian certificates determine all secant dimensions for , with generic widths . The lifting theorem yields specified higher-secant dimensions on infinite families of extension fields. Attainment of the dimension lower bound for arbitrary remains a conjecture.
Introduction
Let be a prime power and let be the group of upper unitriangular matrices over . Its center is isomorphic to the additive group of . Fix a nontrivial additive character of , and put for . We consider the induced representations
Each has dimension . Pointwise multiplication defines an equivariant bilinear map from to . 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 . Here we consider sums of bilinear maps of the form
where , , and are equivariant linear operators on , , and , 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 . In these coordinates, a channel has the form , where is arbitrary and , belong to the corresponding commutant algebras. Let be the affine closure of the one-channel image. The closure of the sum of channels is the affine cone over . 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 . A single channel has parameters, and two independent scalings preserve its value. We prove that these scalings account for the generic fiber dimension: the affine closure of the one-channel image satisfies
for every (Theorem 5.7). The proof uses a permutation Jacobian minor at and is independent of the input types. Consequently, the generic fusion width is at least .
We next prove that
for every central-character pair (Theorem 6.1). We construct two parameter points whose differential images have dimension 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 -channel parameter point on a subgroup has differential rank and every parameter vector is nonzero, then the differential at its zero extension to has rank . In particular, the affine closure of the one-channel image has dimension for any such group and bicharacters (Corollary 9.2). Exact modular certificates determine all secant dimensions for , 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 , respectively, for every and every type pair (Corollary 9.5). The characteristic-two conclusion also holds for with every , using an additive subgroup of . These ranges do not reach the first filling order for large extension degree.
Proposition 7.2 shows that permuting the three central weights 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 , rank factorization also gives a constructive upper bound of 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 . Its dimension 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 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 . 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 -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 . 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 . Finite coefficient fields are used only for computational verification.
Let be a prime power and write
Then
The center and quotient are
Fix a nontrivial additive character . The nondegenerate trace pairing implies that
exhausts all additive characters of .
Definition 2.1 (Induced representation) For , define
The left action is .
Choose the section
Restriction to the section identifies with , hence
A direct calculation from (1) gives, for ,
In particular, the center acts on all of by the scalar character . For later use we record the irreducible representations explicitly.
Proposition 2.2 (Elementary finite Stone–von Neumann decomposition) For each , define on by
Then is irreducible of dimension , has central character , and the representations , , are pairwise inequivalent. Together with the one-dimensional characters factoring through , they exhaust the irreducible representations of .
Proof. (1) shows directly that (4) is a representation. Its character is
Indeed, a nonzero translation has zero trace, while for , , the trace is . Therefore
so is irreducible. Distinct give distinct central characters and hence inequivalent representations. Since , there are one-dimensional representations. Their squared dimensions contribute ; the representations contribute . The sum is , so the list is complete. □
Corollary 2.3 For ,
while is the regular representation of .
Proof. For , Frobenius reciprocity gives
because acts on the -dimensional space of by . No other irreducible has this central character. The statement for follows because is the regular representation of the quotient. □
Twisted convolutions and commutants
For , define the twisted shift
These are right translations expressed in the chosen section.
Lemma 3.1 (Commutation and cocycle law) For all and ,
Moreover,
Proof. Substitution into (3) and (5) proves commutation. For the product,
which yields (6).
Theorem 3.2 (Commutants and twisted convolutions) For every ,
Consequently every -equivariant linear map has a unique twisted-convolution representation
Proof. If , then is the regular representation of the abelian group , hence the multiplicity-free sum of its characters. Therefore . If , then , so Schur’s lemma gives
The twisted shifts lie in the commutant by the previous lemma. They are linearly independent: applying a linear relation to the delta function at produces functions supported at distinct points . Since there are shifts, they form a basis.
Remark 3.3 For , (8) is ordinary convolution on the abelian quotient . For , the phase in (8) records the Heisenberg cocycle. Thus the same coordinate space carries different convolution algebras according to the central character.
Equivariant bilinear maps
Throughout, “bilinear” means complex-bilinear. For types , pointwise multiplication
is a -equivariant bilinear map , because the phases in (3) multiply.
Theorem 4.1 (Selection rule and exact dimension) For ,
If , then
Proof. Complex-bilinear equivariant maps are naturally
For finite groups induction agrees with coinduction. Since , Frobenius reciprocity in the target variable gives
The center acts on the tensor product by . Since , as a -module it is . Equation (11) therefore has dimension zero unless , and dimension in the admissible case.
Let and fix the basepoint .
Lemma 4.2 (Basepoint functional model) For every admissible pair , evaluation at induces a linear isomorphism
Equivalently, every scalar complex-bilinear functional at the basepoint extends uniquely to a -equivariant -valued bilinear map.
Proof. Apply Frobenius reciprocity to :
Under the induced-function realization, this isomorphism is evaluation of the output at the identity coset . On , the center acts by the scalar character . Consequently every complex-linear functional is automatically -equivariant when the target is given the character . Hence
which proves both existence and uniqueness of the equivariant extension.
Choosing the standard coordinate basis of , Lemma 4.2 yields
where represents the bilinear functional . 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 to is
where
For , define the polynomial parameterization
The operators and act on the two inputs, and acts on their pointwise product. If is the basepoint row of , the coefficient matrix of one channel is
The vector 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 and gives in one channel. The same example is considered in Remark 10.2.
Definition 5.2 (Generic fusion width) Let
The width of an individual target is the least nonnegative number of channels representing it. We set and . The generic fusion width is the least for which is dominant onto .
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 is dominant and is not, then there exists a nonempty Zariski-open subset such that every element of has exact fusion width .
Proof. By Chevalley’s theorem, the dense constructible set contains a nonempty open subset of the irreducible affine space . Since is a proper closed subset, removing it leaves a nonempty open subset on which width is at least and at most . □
Proposition 5.4 (Constructive exact upper bound) Let be the basepoint coefficient matrix of an admissible equivariant bilinear map . Then
Given a rank factorization of , the corresponding channel kernels are obtained explicitly. A factorization and the resulting kernels can be computed using arithmetic operations.
Proof. Write , where . The basepoint row of the twisted convolution with coefficient kernel is , independently of its type; similarly for . Choose the output kernel , so that its convolution operator is and its basepoint row is . Consequently
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 arithmetic operations, and writing the kernels costs operations. This is an arithmetic, rather than bit, complexity statement.
Corollary 5.5 (A worst-case consequence of dominance) If is dominant, then every target has fusion width at most . In particular, the worst-case widths for are at most , respectively.
Proof. The image contains a nonempty Zariski-open subset of . For any , the two nonempty open subsets and intersect because the affine space is irreducible. Thus is a sum of two elements in . 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
Hence the Zariski closure of the one-channel image has dimension at most
More generally,
Proof. Each commutant has dimension . On the dense open subset where , the torus in (15) acts freely and preserves every channel. For channels the product torus 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 be the affine Zariski closure of the one-channel image. For every prime power and every ,
The projectivized variety has dimension , and the generic fiber has dimension two. At , the Jacobian has an explicitly selected -square permutation submatrix. Its determinant is in the ordering below, and all its singular values equal one.
Proof. Put and . In the basepoint matrix model, the one-channel map is
The vector is arbitrary, since the basepoint rows of the twisted shifts are the coordinate rows. At the differential is
Write and . For and , the entry is
Now set . Select the columns and the diagonal rows. For each , select columns and rows , in that order. All selected rows and columns are distinct, including in characteristic two. Since the basepoint row of is , the selected matrix is
This proves the determinant and singular-value assertions, and gives differential rank at least . The dimension bound in Lemma 5.6 gives the reverse inequality. Since is a nonzero affine cone, its projectivization has dimension . The generic-fiber dimension is 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 and every ,
Proof. The target dimension is by (10). By (17), the closure of the -channel image has dimension at most . If , dominance is impossible. Solving for 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 be a prime power and let be arbitrary. Then
Thus the second projective secant of the one-channel variety has its expected dimension . Together with the case of Theorem 8.3, this gives
for every prime power and every pair of central types. The argument below applies to every ; only the case uses a rank certificate.
Proof. Write , , , and . Put
The set has four elements. It is a subgroup only in the characteristic-two case. A stripe of difference of a matrix consists of its entries , . Distinct stripes have disjoint supports.
Choose satisfying these two conditions:
For every , the following two vectors, restricted to with , are linearly independent:
The three vectors on given by
are linearly independent.
We first prove that such a vector exists. Since , the complement of has at least elements. For any two distinct indices in that complement, the determinant of the corresponding two rows in (i) is a difference of nonzero scalar multiples of the monomials and . These monomials are distinct: equality of the multisets of their indices would force either or . Hence the selected determinant is a nonzero polynomial.
For (ii) in odd characteristic, choose ; in characteristic two choose , which is possible since and . Specializing , the three distinct indices , , all belong to . The columns 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 , and their intersection is nonempty.
Choose complex scalars
and consider the two parameter points
Both and are invertible. The shifts have order , where . For odd , is not an eigenvalue. For their squares are the identity and . Let denote the two differential images. At the diagonal stripe has dimension , and every stripe outside has dimension two by (i). For the stripes , evaluate formula (18) at , and , respectively, with . This gives the pairs
after multiplication of rows by nonzero scalars. In the last two pairs the row indexed by is divided by . Condition (ii) makes each pair independent. Since the stripes have disjoint supports,
Let denote the differential image at ; it consists exactly of the matrices supported on the diagonal together with row and column . The changes , are bijections of the respective commutants. Differentiation therefore gives
We show that . For an element in this intersection, write
where . The vectors are arbitrary because the basepoint row maps of the two commutants are isomorphisms. The term indexed by in has positions
so this term is supported only on stripes . For , the first two terms may be nonzero only at rows . On the other hand, restricts on that stripe to a linear combination of the two vectors in (i). Vanishing on forces its two coefficients to be zero. Thus all stripes outside vanish.
For , the entries and of the first two terms are and . Unless both differences belong to , a vanishing stripe forces . Similarly, entries for are . Consequently,
In odd characteristic these two intersections are and , respectively. In characteristic two both intersections equal . These identities follow by comparing the independent coordinates along ; they also hold in the smallest odd characteristic .
For , the first two terms therefore contribute nothing at the three positions . Formula (19) for and division by the nonzero phase in the last two positions give constants such that
Thus lies in the intersection of the three planes
Their intersection is zero by (ii). Hence for and each pair of coefficients is zero. These are the coefficients of the full stripes of , so the stripes vanish at every index, not just on . We conclude that all off-diagonal entries of are zero.
It remains to treat the possible diagonal entries. In odd characteristic, only remains. The matrix can be absorbed into . This matrix is supported on ; its only possible diagonal entry is . Because , its entries obey
The other three entries are off-diagonal and vanish. Hence .
In characteristic two, both and the remaining terms indexed by have support in . Thus with . Since and both shifts square to the identity,
the scalar denominators omitted from and are nonzero. The four terms in this expression have the four distinct differences . Put and . Direct evaluation at four off-diagonal positions gives
These entries avoid row , column , and the diagonal, so forces each to vanish. All four scalar multipliers are nonzero; therefore .
In both cases , proving the required trivial intersection. The differential of at has rank , which equals the universal upper bound. For this is less than , so the asserted non-filling second-secant dimension follows. For the exact certificate in Section 8 supplies the filling dimension 16. □
Remark 6.2. The admissible vectors 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
and under exchange . These transformations induce linear isomorphisms of the ambient bilinear spaces that preserve every channel image and every secant closure.
Proof. For , the map
is a group automorphism. On the center it sends , and therefore transports central character to . Precomposition with 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
In particular, besides simultaneous scaling and input exchange, the transformation preserves the entire secant-dimension profile and every exact fusion-width stratum.
Proof. For every type , the pairing
is nondegenerate and -invariant: the opposite central phases cancel, and the remaining action permutes the summands. It identifies with the dual of . Consequently matrix transpose induces a linear isomorphism . Pairing an output of type with a third input of type identifies each equivariant bilinear map with a -invariant trilinear form. A channel becomes
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 , and exchange sends . Table 1 gives the orbit representatives retained in the exact certificates: These representatives cover all ordered pairs in the corresponding fields.
| field | representatives of ordered type pairs |
| , | |
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
| representatives under scaling and triality | |
| 2 | |
| 3 | |
| 4 | |
| 5 |
Table 4.
Indeed, a triple with a zero entry and not all entries zero is equivalent to . In all nonzero zero-sum triples are proportional to ; in they are permutations of up to scale; in they are equivalent to . Thus only 2, 3, 3, 3 classes are needed, respectively; the additional archived classes are redundant checks.
Exact computer-assisted certification for
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 . Let be the corresponding matrices. If denotes the row of producing the output at the basepoint, then under the identification (12) the channel coefficient matrix is
Thus is a polynomial map from to , and its Jacobian has rows and columns.
Lifting modular rank to characteristic zero
Lemma 8.1 (Modular maximal-minor certificate) Let be the cyclotomic integer ring containing the character values occurring in a fixed -model, and let be a Jacobian matrix with entries in evaluated at a parameter point in . Suppose a ring homomorphism to a finite field sends some minor of to a nonzero element of . Then that minor is nonzero over , so the complex Jacobian rank is at least .
Proof. If the determinant were zero in , 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 preserves nonzeroness.
If , 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 and be complex affine spaces, let be a polynomial one-channel parameterization over , let , and let be the sum of copies of . Suppose that at a point ,
Then, for every subset , the sum parameterization on the points indexed by has differential rank . Hence for .
Proof. Write . Each has dimension at most , since the rank of the differential is bounded above by the dimension of the image closure in characteristic zero. The hypothesis states
All inequalities are equalities. Thus each has dimension and their sum is direct. Every subcollection is therefore a direct sum of dimension . The general sum-dimension upper bound proves equality for the corresponding image closures.
Applied after modular lifting with , Lemma 8.2 shows that a penultimate rank , together with a full-rank certificate at , already proves the entire profile. This reduces the necessary number of rank witnesses to two per type class. For , the archive additionally retains and verifies every sub-filling minor as a redundant check; the scripts verify each prefix rank directly.
Certified cases
Table 2 summarizes the exact arithmetic choices. For and , the auxiliary prime is chosen so that the finite field contains the needed roots of unity. For , additive characters take values , so reduction modulo 101 is immediate. The specialization maps used by the certificates are explicit:
| target | threshold | coefficient field | ranks | ||
| 2 | 4 | 16 | 2 | , | |
| 3 | 9 | 81 | 4 | , | |
| 4 | 16 | 256 | 6 | , | |
| 5 | 25 | 625 | 9 | , |
Table 2. Exact coefficient fields and ranks at the penultimate and filling widths.
where 45 has exact order 3 modulo 109, and
where 36 has exact order 5 modulo 101. For the nontrivial additive character is , and for the additive character is , so all character values lie in and we use the ordinary reduction . The verifier checks the required root orders before constructing the Jacobians.
Here the two ranks displayed are for widths and , with
For , a fixed width-two witness gives rank 10 at width one and full rank 16 at width two for each representative . The corresponding certified maximal-minor determinants modulo 101 are
respectively. Thus the smallest field already attains the parameter-count threshold .
For , the same fixed parameter point gives the complete rank sequence
for widths , respectively, in every type-pair orbit. The certified minor determinants modulo 109 for the representatives are, respectively,
For , 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 , the certified determinants modulo 101 for (with the implementation encoding by the integer 2) are
For , the certified determinants modulo 101 for are
A redundant exact recheck at every sub-filling width gives, for every representative type pair,
| Jacobian ranks at widths | |
| 2 | 10, 16 |
| 3 | 25, 50, 75, 81 |
| 4 | 46, 92, 138, 184, 230, 256 |
| 5 | 73, 146, 219, 292, 365, 438, 511, 584, 625. |
Table 5.
Before filling, each rank equals . At the first filling width it equals . All ranks and determinants are computed by exact modular arithmetic.
Theorem 8.3 (Secant dimensions for ) For , every central-character pair , and every ,
Equivalently, every projective secant through the first filling order has its expected dimension. In particular,
Proof. Theorem 5.7 gives the upper bound . For each certified type-pair representative, the exact modular Jacobian ranks displayed above attain this bound at every width through . By Lemma 8.1, the same ranks hold as lower bounds in characteristic zero, proving equality in (23) up to the filling order. At the map is dominant; hence it remains dominant for every larger 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.
Remark 8.4 (Scope of the computation) The structure of the induced representations, the twisted-convolution characterization, the bilinear selection rule, the target dimension , the exact one-channel dimension , the general lower bound, and the type-pair symmetry reductions are proved analytically. Theorem 6.1 also proves order-two nondefectivity analytically for every . 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 be a finite abelian group, let , and write and . Let be bicharacters. Consider the polynomial channel map
and define by restricting the indices and bicharacters to . Suppose that, at parameter triples on , all three vectors in each triple are nonzero and
Extend each parameter vector by zero from to . At the resulting parameter point on $K,
Consequently the -channel image closure on has dimension .
Proof. The two scaling directions
lie in the kernel of the restricted Jacobian. They are independent, including across channels. The rank hypothesis says that these vectors span its entire kernel. It follows that each of the three partial Jacobians, obtained by varying only all -vectors, only all -vectors, or only all -vectors, is injective and has rank . 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 be representatives for , including . At the zero-extended point, the columns indexed by parameters in have support in and reproduce the restricted Jacobian. For every the remaining columns split into three groups with the following mutually disjoint supports:
| varied parameters | output block |
| , | |
| , | |
| , |
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 . Here are those scalings explicitly. In the -block write , , with , and denote the outside parameter by . The extra phase is
Thus the row factor is and the column factor is . In the -block the factors are and . In the -block write , , . The extra phase is
These factors again separate into a row factor and a column factor, and all are nonzero. Therefore every outside partial block has rank .
All displayed blocks have disjoint row and column supports, so their ranks add. The complete rank is
The scaling argument of Lemma 5.6 also applies to and gives image dimension at most . The computed Jacobian rank gives the reverse inequality.
Corollary 9.2 (One-channel dimension for finite abelian groups) Let be a finite abelian group and let be bicharacters. For the parameterization of Theorem 9.1,
Proof. Take and . Since , the restricted map is . Its differential at has rank . Theorem 9.1 gives the assertion, including when is trivial.
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 , , and . The restricted differential has rank . After zero extension to a nontrivial group , it is
Its rank is .
Remark 9.4 (Application to the twisted Heisenberg channel) For , take
These are bicharacters, and the displayed polynomial channel is exactly the basepoint matrix parameterization . Hence Theorem 9.1 applies to every central-character
pair without an untwistedness assumption. The hypothesis requires rank . A filling witness of smaller rank does not satisfy this hypothesis. Equality is allowed; the witnesses used below have rank strictly less than .
Corollary 9.5 (Dimensions over extension fields) For every central-character pair, the expected affine secant dimension
holds in each of the following ranges:
, , and ;
, , and ;
, with either or , and .
For 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 , where or . Write the fixed nontrivial additive character as for some . The restricted bicharacters are precisely the prime-field models with effective central types
Every such pair was certified at all sub-filling orders up to for and up to for . 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 . Each triple therefore has nonzero components: a zero component would reduce its differential rank to at most .
For characteristic two and , choose a four-dimensional -linear subspace and let . Then and both restricted bicharacters are trivial, for every ambient pair . The untwisted certificates supply rank for . Apply the lifting theorem. The case 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 .
Secant geometry and comparison of decomposition widths
Let be the affine cone obtained as the Zariski closure of one-channel maps. We use the convention for the Zariski closure of projective linear spans of points of , so . With this convention, the closure of the width- image is the affine cone over . Thus the generic fusion width is the first secant order whose affine cone fills the ambient bilinear space.
By Theorem 5.7,
@page 20/30 starts_mid=1 ends_mid=1 for every type pair. Hence the expected projective dimension of the -th secant is
Thus Conjecture 11.1 concerns nondefectivity at the first expected filling order. For the untwisted case, we obtain a toric description. Let
and take Fourier transform on the finite abelian group . Ordinary convolution operators diagonalize. If their Fourier multipliers are denoted by , then pointwise multiplication in the spatial domain becomes convolution in frequency. Equivariance forces output frequency , and one channel has coefficient array
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 . Consequently:
Proposition 10.1 (Untwisted toric correspondence) The projective one-channel variety for is the group-based tripod toric variety for , up to a relabeling of the third Fourier index. The projectivizations of the width- image closures are its -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 . Such a factor is absorbed into the third kernel , 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 untwisted second-secant case, since gives the Kimura three-parameter tripod. We include that case for completeness and reproducibility, not as a new theorem. If or , at least one input commutant is isomorphic to the noncommutative algebra . Thus the simultaneous scalar Fourier diagonalization used for 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 , 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 and choose the output kernel so that its basepoint row is . Equation (22) then gives , where . This nonzero equivariant bilinear map has fusion width one. Its scalar basepoint polynomial on is
The symmetric coefficient matrix has rank , so at least squares of linear forms are required. The identity
attains this bound over . Thus the quadratic Waring rank is , 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 and every ,
For , Theorem 8.3 proves the conjecture and also gives the expected dimensions at every smaller secant order. These cases include the extension field . 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 requires additional control of sums of more differential images.
The lower bound satisfies
Thus generic representation requires a number of channels that grows quadratically with . 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 point | Purpose |
verify_all.py | Check hashes, run the certificate suite and confirm unchanged file bytes |
verify_checksums.py | Validate the SHA-256 manifest before mathematical verification |
verify_q2_all_twisted_sectors.py, verify_q3_all_twisted_sectors.py | Verify all prefix ranks and fixed terminal determinants |
verify_q45_frozen_certificates.py | Verify stored hashes, penultimate ranks and stored terminal minors |
verify_full_secant_profiles_q45.py | Recompute every stored sub-filling minor for |
verify_independent_coordinates.py | Reconstruct the Jacobians directly from coefficient sums, independently of the matrix-product implementation |
decompose_basepoint_matrix.py | Exact rational rank-factorization example for Proposition 5.4 |
verify_theory_upgrades.py | Exact 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 verifiers use Python integer arithmetic. The 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 , 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 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 , compares two Jacobian constructions, and verifies four width-three lifts from to , 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 ; small matrix-assembly comparisons at 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 induced from central characters. The affine one-channel variety has dimension for every type pair. The affine second-secant closure has dimension for , while it fills the ambient space for . The subgroup-lifting theorem transfers differential rank on to rank on , 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 , all secant dimensions are determined and the generic widths are , respectively.
The main unresolved question is Conjecture 11.1, which asserts that the dimension lower bound 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 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 , , be the delta basis of . Write
in matrix notation, and let be the basepoint row of . Then
Expanding gives
so the coefficient matrix is
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 , , and be a polynomial map over . If has rank at one point, then the image has dimension , hence is dominant. Chevalley’s theorem says is constructible. A dense constructible subset of irreducible 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
Pairs with exactly one zero form one orbit, and forms another. Thus the representative lists used in § 7 are exhaustive for .
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
A *-polynomial is a polynomial in the complex coordinates and their conjugates. A typed polynomial network is built from the following operations:
arbitrary -equivariant complex-linear maps between channels of the same type;
pointwise products ;
complex conjugation ;
additions, scalar multiplication, and spatially constant fields of type 0.
A constant node is a scalar multiple of the all-ones field in ; 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 be -equivariant. If for every , then .
Proof. Every point of is for some . Equivariance expresses in terms of the basepoint value of , up to the known fiber action. Hence all coordinates vanish.
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 , with the original input represented by . The complexified output is , and projection onto recovers the original map. The regular hidden space decomposes as . Its equivariant linear maps split by central type, and its pointwise powers expand into products . 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 -polynomials) Every -equivariant -polynomial map is represented exactly by a finite typed polynomial network.
Proof. It is enough to consider one output copy . Set
This is a scalar -polynomial in the input coordinates. Because fixes the basepoint and acts on the output by , equivariance gives
Thus only central-weight- monomials occur in the character decomposition of .
A constant term can occur only when and is represented by a constant node. Consider a monomial of positive degree. Each holomorphic factor is an input coordinate of some type , and each antiholomorphic factor is its conjugate. By (5),
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 to . Repeated pointwise multiplication realizes the entire monomial as an -valued equivariant feature because its total central weight is . Summing the monomial subnetworks realizes a map whose basepoint value is . By the previous lemma it equals everywhere. Repeating this for each output channel proves the theorem.
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 -invariant Hermitian norms. Let be compact and -invariant.
Theorem E.1 (Universal approximation) For every continuous -equivariant map and every , there exists a typed polynomial network such that
Proof. The algebra generated by complex coordinate functions and their conjugates is self-adjoint, contains constants, and separates points of the compact subset . By the complex
Stone–Weierstrass theorem, vector-valued *-polynomials approximate uniformly coordinatewise. Choose such a polynomial map with uniform error less than on .
Apply Reynolds averaging:
This is again a *-polynomial and is exactly -equivariant. Since is invariant and the representations are unitary,
Theorem D.2 realizes 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]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]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]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]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]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]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]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]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]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]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]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]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