Introduction
The permanent of an n × n n \times n n × n matrix Y Y Y is the polynomial
per n ( Y ) = ∑ σ ∈ S n ∏ i = 1 n Y i , σ ( i ) , \operatorname{per}_n(Y) = \sum_{\sigma\in S_n} \prod_{i=1}^{n} Y_{i,\sigma(i)}, per n ( Y ) = σ ∈ S n ∑ i = 1 ∏ n Y i , σ ( i ) ,
where S n S_n S n is the set of permutations of { 1 , … , n } \{1, \ldots, n\} { 1 , … , n } . The determinant is given by the same sum with the sign of σ \sigma σ attached to each term. The determinant can be computed by a polynomial number of arithmetic operations. The arithmetic complexity of the permanent is a central open problem in algebraic complexity theory.
Valiant [22 ] formulated this problem in terms of determinantal representations. A determinantal representation of a polynomial F F F is a square matrix whose determinant is F F F and whose entries are affine-linear forms in the variables of F F F , that is, polynomials of degree at most one. The determinantal complexity (DC) of F F F , denoted by dc ( F ) \operatorname{dc}(F) dc ( F ) , is the smallest size of such a matrix. Valiant showed that a polynomial computed by an arithmetic formula of size s s s has a determinantal representation of size linear in s s s . He also proved that the permanent is complete for VNP, an algebraic analogue of NP, in characteristic different from two. Valiant’s determinantal complexity conjecture asks whether dc ( per n ) \operatorname{dc}(\operatorname{per}_{n}) dc ( per n ) grows faster than every polynomial in n n n .
The border determinantal complexity dc ‾ ( F ) \underline{\operatorname{dc}}(F) dc ( F ) permits limits. It is the least m m m such that F F F is a coefficientwise limit of determinants of m × m m \times m m × m matrices with affine-linear entries. The size m m m remains fixed during the limit. Every exact representation is also such a limit, and neither a size-m m m determinant nor a limit of size-m m m determinants can have degree above m m m . Consequently, for every nonzero polynomial F F F ,
deg F ≤ dc ‾ ( F ) ≤ dc ( F ) . (1) \deg F \le\underline{\operatorname{dc}}(F) \le\operatorname{dc}(F). \tag*{(1)} deg F ≤ dc ( F ) ≤ dc ( F ) . (1)
In particular, a lower bound on border DC also gives the same lower bound on exact DC. Throughout the paper the ground field is C \mathbb{C} C , and a form means a homogeneous polynomial: all of its monomials have the same total degree.
Mignon and Ressayre [17 ] proved the quadratic lower bound dc ( per n ) ≥ n 2 / 2 \operatorname{dc}(\operatorname{per}_{n}) \ge n^{2}/2 dc ( per n ) ≥ n 2 /2 in characteristic zero. Landsberg, Manivel, and Ressayre [14 ] proved that the same bound holds for border DC. More recently, OpenAI’s September 2026 preprint [19 ] reports a cubic lower bound, dc ‾ ( per n ) = Ω ( n 3 ) \underline{\operatorname{dc}}(\operatorname{per}_{n}) = \Omega(n^{3}) dc ( per n ) = Ω ( n 3 ) over C \mathbb{C} C . Related lower bounds from smooth restrictions and polar counts appear in the author’s earlier work [20 , 21 ] . In addition to their lower bound for the permanent, Landsberg, Manivel, and Ressayre exhibit forms for which the second inequality in (1) is strict. The present paper concerns this distinction between exact and border complexity.
Geometric complexity theory (GCT), introduced by Mulmuley and Sohoni [18 ] , studies orbit closures associated with polynomials such as the determinant. Let
O n = G L n 2 ⋅ det n \mathcal{O}_{n} = \mathrm{GL}_{n^{2}} \cdot\det_{n} O n = GL n 2 ⋅ n det
be the orbit of the determinant under invertible linear changes of its n 2 n^{2} n 2 variables. Its closure O n ‾ \overline{\mathcal{O}_{n}} O n also contains their limits, and the determinant boundary is
∂ O n ‾ = O n ‾ ∖ O n . \partial\overline{\mathcal{O}_{n}} = \overline{\mathcal{O}_{n}} \setminus\mathcal{O}_{n}. ∂ O n = O n ∖ O n .
Closures are taken in the Zariski topology; over C \mathbb{C} C the orbit closure also agrees with the closure in the usual complex topology. A polynomial is concise in its ambient variables if no invertible linear change of coordinates reduces their number. Equivalently, its first partial derivatives are linearly independent.
Every nonzero degree-n n n form in O n ‾ \overline{\mathcal{O}_{n}} O n has border DC n n n . If it is concise and lies on the boundary, its exact DC is larger than n n n . Indeed, a size-n n n affine representation would have homogeneous degree-n n n part det L \det L det L , where L L L is its linear part. Conciseness in n 2 n^{2} n 2 variables would force this linear substitution to be invertible, placing the form in O n \mathcal{O}_{n} O n itself. Thus the orbit-closure formulation distinguishes border complexity from the size required by an exact representation.
Landsberg, Manivel, and Ressayre [14 Proposition 3.5.1] constructed such a concise boundary component for every odd n ≥ 3 n \ge3 n ≥ 3 . Their construction starts with odd skew-symmetric matrices. Chen [6 Theorem 5.5.3] proved that adding one row and column gives another component. Hüttenhain [10 ] treats the traceless-determinant component and a further concise component at size four. The leading determinant coefficient of a singular matrix space gives a more general source of boundary points [12 Section 6.7.1] . The bounded-rank spaces of Huang and Landsberg [9 ] and the exterior-multiplication spaces studied by Landsberg and Manivel [13 ] provide further starting spaces for this construction.
These examples lead to two questions. How large can the difference between exact and border DC be on an explicit boundary component? How many distinct components can be obtained by varying the starting matrix space and adding rows and columns? We prove a linear lower bound for the first question and a uniform component count for the second.
Our contributions
Our main result is the following.
Theorem 1.1. For every n ≥ 10 n \ge10 n ≥ 10 , there is an explicit homogeneous form f n f_n f n of degree n n n , concise in n 2 n^2 n 2 variables, whose orbit closure is a determinant-boundary component and which satisfies
dc ( f n ) = n , dc ( f n ) ≥ 2 n − 1. \operatorname{dc}(f_n)=n,\qquad\operatorname{dc}(f_n)\ge2n-1. dc ( f n ) = n , dc ( f n ) ≥ 2 n − 1. One may take f n = K 2 , n − 10 f_n=K_{2,n-10} f n = K 2 , n − 10 , defined in § 3.5 . For every odd n ≥ 3 n\ge3 n ≥ 3 , the unpadded odd-skew form Q n , 0 Q_{n,0} Q n , 0 also has these properties.
The defining determinant limits give dc ( f n ) ≤ n \operatorname{dc}(f_n)\le n dc ( f n ) ≤ n , and the degree gives the reverse inequality. The exact lower bound comes from a property of the whole hypersurface f n = 0 f_n=0 f n = 0 : every polynomial matrix with determinant f n f_n f n has a free rank-one cokernel over its coordinate ring. For the unpadded odd-skew forms, this follows from unique factorization in that ring. For the Koszul forms, it follows from a calculation of the divisor class group and the depth of the two possible nonfree cokernels.
The passage from freeness to a size bound is a general degree argument. Suppose an affine matrix M M M represents an irreducible homogeneous polynomial of degree d d d and has a free rank-one cokernel on its zero set. Its adjugate factors there as u v T uv^T u v T , where u u u and v v v are nonzero right- and left-kernel vectors. Each vector has a coordinate of degree at least d − 1 d-1 d − 1 . Since a cofactor of a size-m m m affine matrix has degree at most m − 1 m-1 m − 1 , it follows that m ≥ 2 d − 1 m\ge2d-1 m ≥ 2 d − 1 . This is Theorem 4.1 . The hypothesis is proved for every exact representation of the forms in Theorem 1.1 ; the resulting bound therefore applies to unrestricted affine determinantal representations.
In particular, the exact-minus-border gap is at least n − 1 n-1 n − 1 and
dc ( f n ) dc ( f n ) ≥ 2 − 1 n . \frac{\operatorname{dc}(f_n)}{\operatorname{dc}(f_n)}\ge2-\frac{1}{n}. dc ( f n ) dc ( f n ) ≥ 2 − n 1 .
For example, the degree-ten construction has border DC ten and exact DC at least nineteen. These are lower bounds on the exact complexity; the theorem does not determine its value.
We next construct three families of boundary components. They arise from odd skew-symmetric matrices, the block spaces of Huang and Landsberg, and middle Koszul maps. Each starting space can be padded by any number of rows and columns.
Theorem 1.2. Every form in the three families defined in § 3 is concise in n 2 n^2 n 2 variables, and its orbit closure is a determinant-boundary component. For fixed n n n , the components obtained from different listed parameters are pairwise distinct. Including the traceless-determinant component, the boundary has at least
⌊ n − 1 2 ⌋ + ⌊ n 3 ⌋ + κ ( n ) , κ ( n ) = # { p ≥ 2 : ( 2 p + 1 p ) ≤ n } , (2) \left\lfloor\frac{n-1}{2}\right\rfloor+\left\lfloor\frac{n}{3}\right\rfloor+\kappa(n),\qquad
\kappa(n)=\#\left\{p\ge2:\binom{2p+1}{p}\le n\right\}, \tag*{(2)} ⌊ 2 n − 1 ⌋ + ⌊ 3 n ⌋ + κ ( n ) , κ ( n ) = # { p ≥ 2 : ( p 2 p + 1 ) ≤ n } , (2) irreducible components for every $n\ge 3.
The odd-skew family includes every odd starting size r r r from three to n n n , with n − r n-r n − r added rows and columns. It therefore unifies the unpadded construction of [14 ] , Chen’s one-row extension, and arbitrary padding of the smallest skew block. The block family has starting sizes 6, 9, 12, … \ldots … , and the Koszul family has starting sizes 10, 35, 126, … \ldots … . The count in (2) is a lower bound; it does not classify the entire boundary. The proofs identify singular or higher-multiplicity loci determined by each polynomial and then compute its stabilizer, the group of linear changes of variables preserving that polynomial. The same loci distinguish the components. We also obtain general criteria for componenthood and for preservation of componenthood under padding.
Our final general lower bound excludes exact complexity n + 1 n+1 n + 1 for every concise boundary point.
Theorem 1.3. Let n ≥ 3 n \ge3 n ≥ 3 , and let f f f be a concise degree-n n n form on the determinant boundary in n 2 n^2 n 2 variables. Then
dc ( f ) = n , dc ( f ) ≥ n + 2. \operatorname{dc}(f)=n,\qquad\operatorname{dc}(f)\ge n+2. dc ( f ) = n , dc ( f ) ≥ n + 2.
Theorem 7.1 proves the stronger statement that a homogeneous degree-n n n form with at least n 2 − n + 2 n^2-n+2 n 2 − n + 2 essential variables cannot have exact DC equal to n + 1 n+1 n + 1 . For the padded odd-skew forms with starting size three, an explicit upper bound gives
n + 2 ≤ dc ( Q 3 , n − 3 ) ≤ n + 4. n+2\le\operatorname{dc}(Q_{3,n-3})\le n+4. n + 2 ≤ dc ( Q 3 , n − 3 ) ≤ n + 4.
Thus the bound 2 n − 1 2n-1 2 n − 1 in Theorem 1.1 applies to the unpadded odd-skew forms and all padded Koszul forms, rather than to every padded odd-skew form.
Organization of the paper
Section 2 gives a proof overview using the smallest odd-skew example. Section 3 defines the three families, and Section 4 proves the general degree bound. Sections 5 and 6 establish the cokernel property needed for Theorem 1.1. Section 7 treats general concise boundary points. The component proofs occupy Sections 8, 9, and 10, followed by the component count and general padding criteria. The remaining sections develop bounds from affine level sets, a reduction to regular representations of the determinant, and a criterion for rational matrix realizations. They also explain the limitations of these methods for obtaining a superlinear gap. Section 20 records the open questions.
Proof overview
The smallest odd-skew form illustrates both the determinant limit and the degree argument for exact complexity. Let x = ( x 1 , x 2 , x 3 ) T x=(x_1,x_2,x_3)^{T} x = ( x 1 , x 2 , x 3 ) T , let S S S be a symmetric 3 × 3 3\times3 3 × 3 matrix with six independent entries, and put
A ( x ) = ( 0 − x 3 x 2 x 3 0 − x 1 − x 2 x 1 0 ) . A(x)=\begin{pmatrix}
0 & -x_3 & x_2\\
x_3 & 0 & -x_1\\
-x_2 & x_1 & 0
\end{pmatrix}. A ( x ) = 0 x 3 − x 2 − x 3 0 x 1 x 2 − x 1 0 .
Then A ( x ) x = 0 A(x)x=0 A ( x ) x = 0 and adj A ( x ) = x x T \operatorname{adj} A(x)=xx^{T} adj A ( x ) = x x T . Expanding the determinant gives
det ( A ( x ) + t S ) = t x T S x + t 3 det S . (3) \det(A(x)+tS)=t\,x^{T}Sx+t^{3}\det S. \tag*{(3)} det ( A ( x ) + tS ) = t x T S x + t 3 det S . (3)
For nonzero t t t , the entries of A ( x ) + t S A(x)+tS A ( x ) + tS are an invertible linear function of the nine variables, since a matrix decomposes uniquely into its skew-symmetric and symmetric parts. Multiplying one row by t − 1 t^{-1} t − 1 therefore gives a polynomial in the determinant orbit O 3 \mathcal{O}_3 O 3 , and
f = x T S x = lim t → 0 t − 1 det ( A ( x ) + t S ) . f=x^{T}Sx=\lim_{t\to0}t^{-1}\det(A(x)+tS). f = x T S x = t → 0 lim t − 1 det ( A ( x ) + tS ) .
Thus dc ( f ) = 3 \operatorname{dc}(f)=3 dc ( f ) = 3 : the limit proves the upper bound, and the degree proves the lower bound.
We next outline why every exact representation of f f f has size at least five. Let M M M be an m × m m \times m m × m affine matrix with det M = f \det M=f det M = f , and put R = C [ x , S ] / ( f ) R=\mathbb{C}[x,S]/(f) R = C [ x , S ] / ( f ) . The matrix M M M is singular over R R R . The main algebraic input is that
coker R M = R m / M R m ≃ R . \operatorname{coker}_{R} M = R^{m}/MR^{m} \simeq R. coker R M = R m / M R m ≃ R .
This means that the cokernel has one global generator, with no nonzero relation over R R R . It implies that every row annihilating M M M is a multiple of one row v T v^{T} v T . Since the rows of the adjugate annihilate M M M modulo f f f , we obtain
adj M = u v T over R , M u = 0 , v T M = 0 , \operatorname{adj} M = uv^{T}\quad\text{over }R,\qquad Mu=0,\qquad v^{T}M=0, adj M = u v T over R , M u = 0 , v T M = 0 ,
with u , v u,v u , v nonzero.
Suppose all coordinates of u u u had degree at most one. Lift them to polynomials of degree at most one before taking the quotient by f f f . The entries of M u Mu M u then have degree at most two and are divisible by the cubic f f f . Hence they vanish identically. This is impossible, since a matrix with nonzero determinant has no nonzero kernel over the field of rational functions. Therefore u u u has a coordinate of degree at least two. The same holds for v v v .
Degrees add in the graded domain R R R . Consequently some entry of u v T uv^{T} u v T , and hence of adj M \operatorname{adj} M adj M , has degree at least four. A cofactor of an affine matrix of size m m m has degree at most m − 1 m-1 m − 1 , so m ≥ 5 m\geq5 m ≥ 5 . For an irreducible homogeneous polynomial of degree d d d , the same argument gives a coordinate of degree at least d − 1 d-1 d − 1 in each vector and yields m ≥ 2 d − 1 m\geq2d-1 m ≥ 2 d − 1 . § 4 proves this statement in full.
It remains to justify the cokernel hypothesis for every exact representation. The polynomial x T S x x^{T}Sx x T S x is linear in the entries of S S S , and its coefficient polynomials are the products x i x j x_i x_j x i x j . They vanish simultaneously only at x = 0 x=0 x = 0 , a set of codimension three. § 5 proves that, more generally, a polynomial linear in one group of variables defines a factorial coordinate ring when its coefficients have a common zero set of codimension at least three. Every determinant cokernel on such a hypersurface is free. This proves the hypothesis for x T S x x^{T}Sx x T S x and for all unpadded odd-skew forms.
The Koszul forms require a different freeness argument. Away from the zero value of their starting variables, they become ordinary determinants in the coefficient variables. A classification for the generic determinant leaves three possible rank-one cokernels. One is free. The other two extend to modules whose depths are too small on the omitted set, so neither can be the cokernel of a determinant representation. § 6 proves this assertion and establishes the lower bound at every n ≥ 10 n\geq10 n ≥ 10 .
Finally, the component statements require an independent geometric argument. The defining limits give points of the determinant orbit closure. To show that their orbit closures are entire boundary components, we compute their stabilizer dimensions. We distinguish the components using singular and multiplicity loci that are preserved under every invertible linear change of variables.
Three families of boundary forms
We define the forms in Theorems 1.1 and 1.2 . Each family is obtained from a singular matrix space by taking a leading determinant coefficient. The construction of this coefficient is standard [12 Section 6.7.1] . The component and distinctness statements are proved in § 8 , § 10 , and § 11 .
Leading coefficients and padding
Let L ⊂ Mat r L \subset\operatorname{Mat}_r L ⊂ Mat r be a linear space whose generic rank is r − c r-c r − c , where c ≥ 1 c \ge1 c ≥ 1 . Choose a complementary space S L S_L S L , so Mat r = L ⊕ S L \operatorname{Mat}_r = L \oplus S_L Mat r = L ⊕ S L , and let A A A and S S S be the generic matrices of these two spaces. We call L L L the starting space, or seed space. For k ≥ 0 k \ge0 k ≥ 0 , put n = r + k n = r+k n = r + k and introduce independent matrices
U ∈ Mat r , k , V ∈ Mat k , r , Z ∈ Mat k U \in\operatorname{Mat}_{r,k}, \qquad V \in\operatorname{Mat}_{k,r}, \qquad Z \in\operatorname{Mat}_k U ∈ Mat r , k , V ∈ Mat k , r , Z ∈ Mat k .
Define
M t = ( A + t S U t V Z ) , F L , k = [ t c ] det M t . (4) M_t = \begin{pmatrix} A+tS & U \\ tV & Z \end{pmatrix}, \qquad F_{L,k} = [t^c]\det M_t. \tag*{(4)} M t = ( A + tS t V U Z ) , F L , k = [ t c ] det M t . (4)
Here [ t c ] [t^c] [ t c ] denotes the coefficient of t c t^c t c , and the empty blocks are omitted when k = 0 k=0 k = 0 . There are r 2 + 2 r k + k 2 = n 2 r^2+2rk+k^2=n^2 r 2 + 2 r k + k 2 = n 2 independent variables. The addition of these blocks is called padding.
All coefficients of t j t^j t j for j < c j<c j < c vanish. On det Z ≠ 0 \det Z \ne0 det Z = 0 , Schur elimination gives
det M t = det Z det ( A + t ( S − U Z − 1 V ) ) . \det M_t = \det Z \det\left(A+t(S-UZ^{-1}V)\right). det M t = det Z det ( A + t ( S − U Z − 1 V ) ) .
The vanishing follows from rank A ≤ r − c A \le r-c A ≤ r − c , and extends to every value of the variables because it is a polynomial identity. The leading coefficient is nonzero. Indeed, at a rank-r − c r-c r − c point of L L L , every direction in L L L induces the zero map from the kernel of A A A to its cokernel. Since S L S_L S L complements L L L in the full matrix space, its induced maps between those c c c -dimensional spaces are arbitrary. A perturbation invertible between them has a nonzero coefficient of t c t^c t c .
For t ≠ 0 t \ne0 t = 0 , the entries of M t M_t M t are an invertible linear parametrization of Mat n \operatorname{Mat}_n Mat n . The factor t − c t^{-c} t − c can be absorbed into one row, so
F L , k = lim t → 0 t − c det M t ∈ O ‾ n . F_{L,k}=\lim_{t\to0}t^{-c}\det M_t\in\overline{\mathcal O}_n. F L , k = t → 0 lim t − c det M t ∈ O n .
Every nonzero term of F L , k F_{L,k} F L , k has total degree n n n . Hence dc ( F L , k ) = n \operatorname{dc}(F_{L,k})=n dc ( F L , k ) = n . The three choices of L L L below give the families used in this paper.
Odd skew-symmetric matrices
Let r = 2 s + 1 ≥ 3 r=2s+1\ge3 r = 2 s + 1 ≥ 3 , take L = Skew r L=\operatorname{Skew}_r L = Skew r , and take the symmetric matrices as its complement. An odd skew-symmetric matrix is singular, since det A = det ( − A ) = − det A \det A=\det(-A)=-\det A det A = det ( − A ) = − det A , and its generic rank is r − 1 r-1 r − 1 . Thus c = 1 c=1 c = 1 .
The signed maximal Pfaffians form a polynomial kernel vector:
p i ( A ) = ( − 1 ) i + 1 Pf ( A i ^ ) , p ( A ) = ( p 1 ( A ) , … , p r ( A ) ) T . p_i(A)=(-1)^{i+1}\operatorname{Pf}(A_{\hat{i}}), \qquad p(A)=(p_1(A),\ldots,p_r(A))^T. p i ( A ) = ( − 1 ) i + 1 Pf ( A i ^ ) , p ( A ) = ( p 1 ( A ) , … , p r ( A ) ) T .
Here A i ^ A_{\hat{i}} A i ^ denotes the matrix obtained by deleting row and column i i i . The Pfaffian of an even-size skew matrix is the signed sum over pairings of its indices; its square is the determinant. Each p i p_i p i has degree s s s , and the Pfaffian identities give
A p = 0 , adj A = p p T . Ap=0, \qquad\operatorname{adj} A=pp^T. A p = 0 , adj A = p p T .
Consequently the unpadded form is
Q r , 0 ( A , S ) = p ( A ) T S p ( A ) = [ t ] det ( A + t S ) . (5) Q_{r,0}(A,S)=p(A)^TSp(A)=[t]\det(A+tS). \tag*{(5)} Q r , 0 ( A , S ) = p ( A ) T S p ( A ) = [ t ] det ( A + tS ) . (5)
For k ≥ 1 k\ge1 k ≥ 1 , the padded form is
Q r , k = det Z p T S p − p T U adj Z V p . (6) Q_{r,k}=\det Z\,p^TSp-p^TU\operatorname{adj}Z\,Vp. \tag*{(6)} Q r , k = det Z p T S p − p T U adj Z V p . (6)
This identity follows from Schur elimination of Z Z Z and adj A = p p T \operatorname{adj} A = pp^{T} adj A = p p T . Although elimination initially requires Z Z Z to be invertible, the resulting polynomial identity holds everywhere. For r = 3 r = 3 r = 3 and k = 0 k = 0 k = 0 , this is the cubic x T S x x^{T}Sx x T S x from § 2 .
Theorem 8.1 proves that every Q r , k Q_{r,k} Q r , k has an orbit closure that is a boundary component, and that distinct values of r r r give distinct components at a fixed final size n = r + k n = r + k n = r + k . § 5 proves the bound dc ( Q r , 0 ) ≥ 2 r − 1 \operatorname{dc}(Q_{r,0}) \ge2r - 1 dc ( Q r , 0 ) ≥ 2 r − 1 for the unpadded forms. No such bound is asserted for every padding: § 7 gives an upper bound n + 4 n + 4 n + 4 when r = 3 r = 3 r = 3 .
Block matrix spaces
For h ≥ 2 h \ge2 h ≥ 2 , let a , b a,b a , b be scalar variables and let F F F be a generic h × h h \times h h × h matrix. Set
A h ( a , b , F ) = ( 0 a I h b I h − a I h 0 F − b I h − F 0 ) . A_h(a,b,F) =
\begin{pmatrix}
0 & aI_h & bI_h \\
-aI_h & 0 & F \\
-bI_h & -F & 0
\end{pmatrix}. A h ( a , b , F ) = 0 − a I h − b I h a I h 0 − F b I h F 0 .
These spaces were constructed by Huang and Landsberg [9 ] . Their dimension is h 2 + 2 h^2 + 2 h 2 + 2 , their matrix size is r = 3 h r = 3h r = 3 h , and their generic rank is 2 h 2h 2 h . The matrices
K R = ( F − b I h a I h ) , K L = ( F , − b I h , a I h ) K_R =
\begin{pmatrix}
F \\
-bI_h \\
aI_h
\end{pmatrix},
\qquad
K_L = (F,-bI_h,aI_h) K R = F − b I h a I h , K L = ( F , − b I h , a I h )
satisfy A h K R = 0 A_hK_R = 0 A h K R = 0 and K L A h = 0 K_LA_h = 0 K L A h = 0 . For a complementary matrix S S S and padding blocks U , V , Z U,V,Z U , V , Z , define
H h , k = [ t h ] det ( A h + t S U t V Z ) . H_{h,k} = [t^h]\det
\begin{pmatrix}
A_h + tS & U \\
tV & Z
\end{pmatrix}. H h , k = [ t h ] det ( A h + tS t V U Z ) .
Equivalently, the coefficient calculation in § 10 gives
H h , k = det ( K L S K R K L U V K R Z ) . H_{h,k} = \det
\begin{pmatrix}
K_LSK_R & K_LU \\
VK_R & Z
\end{pmatrix}. H h , k = det ( K L S K R V K R K L U Z ) .
This is a polynomial of total degree n = 3 h + k n = 3h + k n = 3 h + k . The last display is not an affine representation: the degrees of its four blocks are 3, 2, 2, 1. Theorem 9.1 proves componenthood and pairwise distinctness for every h ≥ 2 h \ge2 h ≥ 2 and k ≥ 0 k \ge0 k ≥ 0 . The complexity bounds for this family are given in § 17 .
Middle Koszul maps
Let p ≥ 2 p \ge2 p ≥ 2 , put V = C 2 p + 1 V = \mathbb{C}^{2p+1} V = C 2 p + 1 , and let E = Λ p V E = \Lambda^p V E = Λ p V . Choose a volume form on V V V . Exterior multiplication by v ∈ V v \in V v ∈ V defines a square matrix A ( v ) : E → E ∗ A(v) : E \to E^* A ( v ) : E → E ∗ through the pairing
A ( v ) ( α , β ) = vol ( v ∧ α ∧ β ) , α , β ∈ E . A(v)(\alpha,\beta) = \operatorname{vol}(v \wedge\alpha\wedge\beta), \qquad\alpha,\beta\in E. A ( v ) ( α , β ) = vol ( v ∧ α ∧ β ) , α , β ∈ E .
In the standard basis, rows and columns are indexed by increasing p p p -element subsets of { 1 , … , 2 p + 1 } \{1,\ldots,2p+1\} { 1 , … , 2 p + 1 } . An entry is zero if the two subsets intersect. Otherwise it is the missing coordinate v a v_a v a , with the sign of the permutation obtained by listing a a a , the row subset, and the column subset. This gives an explicit matrix-entry definition of the exterior-multiplication map.
Its size, rank at every nonzero v v v , and corank are
r = ( 2 p + 1 p ) , R 0 = ( 2 p p ) , c = ( 2 p p − 1 ) , r = R 0 + c . (7) r = \binom{2p+1}{p}, \qquad R_0 = \binom{2p}{p}, \qquad c = \binom{2p}{p-1}, \qquad r = R_0 + c. \tag*{(7)} r = ( p 2 p + 1 ) , R 0 = ( p 2 p ) , c = ( p − 1 2 p ) , r = R 0 + c . (7)
For the rank calculation, take v = e 1 v=e_1 v = e 1 . The basis vectors indexed by subsets containing 1 lie in the kernel. The remaining basis vectors are paired invertibly with their complementary subsets in { 2 , … , 2 p + 1 } \{2,\ldots,2p+1\} { 2 , … , 2 p + 1 } . A change of coordinates in V V V induces invertible changes in the exterior powers, so the rank is the same for every nonzero v v v .
Choose a complement S p S_p S p to A ( V ) A(V) A ( V ) in Mat r \operatorname{Mat}_r Mat r , let S S S range over this complement, and define
K p , k = [ t c ] det ( A ( v ) + t S U t V 0 Z ) , n = r + k . \mathcal{K}_{p,k}=[t^c]\det\begin{pmatrix} A(v)+tS & U \\ tV_0 & Z \end{pmatrix}, \qquad n=r+k. K p , k = [ t c ] det ( A ( v ) + tS t V 0 U Z ) , n = r + k .
The notation V 0 V_0 V 0 distinguishes the lower-left block from the vector space V V V . The polynomial has degree R 0 R_0 R 0 in v v v and degree c + k c+k c + k in the coefficient variables S , U , V 0 , Z S,U,V_0,Z S , U , V 0 , Z , hence total degree n n n . Theorem 10.1 proves that its orbit closure is a boundary component. § 6 proves
dc ( K p , k ) = n , dc ( K p , k ) ≥ 2 n − 1 \operatorname{dc}(\mathcal{K}_{p,k})=n,\qquad\operatorname{dc}(\mathcal{K}_{p,k})\ge2n-1 dc ( K p , k ) = n , dc ( K p , k ) ≥ 2 n − 1
for every p ≥ 2 p\ge2 p ≥ 2 and $k\ge 0.
The size-ten construction
For p = 2 p=2 p = 2 , the preceding matrix has size ten and depends on v 1 , … , v 5 v_1,\ldots,v_5 v 1 , … , v 5 . Its rows and columns are indexed by
12 , 13 , 14 , 15 , 23 , 24 , 25 , 34 , 35 , 45. 12,13,14,15,23,24,25,34,35,45. 12 , 13 , 14 , 15 , 23 , 24 , 25 , 34 , 35 , 45.
For example, the entry rule gives
A ( v ) 12 , 34 = v 5 , A ( v ) 12 , 35 = − v 4 , A ( v ) 12 , 13 = 0. A(v)_{12,34}=v_5,\qquad A(v)_{12,35}=-v_4,\qquad A(v)_{12,13}=0. A ( v ) 12 , 34 = v 5 , A ( v ) 12 , 35 = − v 4 , A ( v ) 12 , 13 = 0.
The matrix is symmetric and has rank six at every nonzero v v v . Its four-dimensional kernel at v = e 1 v=e_1 v = e 1 is spanned by the pair coordinates containing 1. An invertible change of the five coordinates induces an invertible change of the ten pair coordinates, with coefficients given by the corresponding 2 × 2 2\times2 2 × 2 minors. This verifies rank invariance directly in this case.
An explicit complement is
S = { S ∈ Mat 10 : tr ( A ( e i ) T S ) = 0 for i = 1 , … , 5 } . (8) S=\left\{S\in\operatorname{Mat}_{10}:\operatorname{tr}(A(e_i)^T S)=0\text{ for }i=1,\ldots,5\right\}. \tag*{(8)} S = { S ∈ Mat 10 : tr ( A ( e i ) T S ) = 0 for i = 1 , … , 5 } . (8)
The five matrices A ( e i ) A(e_i) A ( e i ) have disjoint supports and six entries in { 1 , − 1 } \{1,-1\} { 1 , − 1 } each, so tr ( A ( e i ) T A ( e j ) ) = 6 δ i j \operatorname{tr}(A(e_i)^T A(e_j))=6\delta_{ij} tr ( A ( e i ) T A ( e j )) = 6 δ ij . Thus S S S is indeed a complement and has dimension 95. One can choose coordinates on it by solving each of the five equations for one entry with nonzero coefficient and taking the other 95 entries freely. The disjoint supports make these choices independent. The matrix S S S is not required to be symmetric.
For any n ≥ 10 n\ge10 n ≥ 10 , put k = n − 10 k=n-10 k = n − 10 and set
f n = K 2 , k = [ t 4 ] det ( A ( v ) + t S U t V 0 Z ) . (9) f_n=\mathcal{K}_{2,k}=[t^4]\det\begin{pmatrix} A(v)+tS & U \\ tV_0 & Z \end{pmatrix}. \tag*{(9)} f n = K 2 , k = [ t 4 ] det ( A ( v ) + tS t V 0 U Z ) . (9)
The number of variables is
5 + 95 + 20 k + k 2 = ( 10 + k ) 2 = n 2 . 5+95+20k+k^2=(10+k)^2=n^2. 5 + 95 + 20 k + k 2 = ( 10 + k ) 2 = n 2 .
This is an explicit polynomial definition by coefficient extraction from a finite determinant. It has degree six in v v v and degree k + 4 k+4 k + 4 in the other variables. The general construction gives
f n = lim t → 0 t − 4 det ( A ( v ) + t S U t V 0 Z ) , dc ( f n ) = n . f_n=\lim_{t\to0}t^{-4}\det\begin{pmatrix} A(v)+tS & U \\ tV_0 & Z \end{pmatrix},\qquad\operatorname{dc}(f_n)=n. f n = t → 0 lim t − 4 det ( A ( v ) + tS t V 0 U Z ) , dc ( f n ) = n .
The global cokernel theorem and the component calculation complete the proof of Theorem 1.1 .
Component statements
The parameters of the three families are summarized below. In each case the final size is the starting size plus k k k .
Table 3.
For every member of these families, the component proofs establish
dim Stab ( F ) = 2 n 2 − 1 , dim ( G L n 2 ⋅ F ) = n 4 − 2 n 2 + 1. \dim\operatorname{Stab}(F)=2n^2-1,\qquad\dim(\mathrm{GL}_{n^2}\cdot F)=n^4-2n^2+1. dim Stab ( F ) = 2 n 2 − 1 , dim ( GL n 2 ⋅ F ) = n 4 − 2 n 2 + 1.
The stabilizer consists of the invertible linear changes of variables preserving F F F . Since the determinant stabilizer has dimension 2 n 2 − 2 2n^2-2 2 n 2 − 2 , its orbit has dimension n 4 − 2 n 2 + 2 n^4-2n^2+2 n 4 − 2 n 2 + 2 . Each constructed orbit closure is therefore a proper irreducible closed subset of codimension one in O n ‾ \overline{O_n} O n , and hence a boundary component.
The component distinction is obtained from loci preserved by linear equivalence. For the odd-skew family, the translation dimension of the largest singular component recovers the starting size r r r . For the block family, the span of the corresponding singular component has a different codimension. The Koszul family has a singular locus of different codimension, and a higher-multiplicity locus recovers p p p . The detailed calculations also prove conciseness in all n 2 n^2 n 2 variables.
A lower bound from kernel degrees
We prove the general degree bound used in Theorem 1.1 . Let f ∈ T = C [ x 1 , … , x N ] f \in T=\mathbb{C}[x_1,\ldots,x_N] f ∈ T = C [ x 1 , … , x N ] be irreducible and homogeneous, and let R = T / ( f ) R=T/(f) R = T / ( f ) be the coordinate ring of its zero hypersurface. Elements of R R R are polynomial classes modulo multiples of f f f . Since f f f is irreducible, R R R is a domain.
For a square matrix M M M with det M = f \det M=f det M = f , its cokernel over R R R is
coker R M = R m / M R m . \operatorname{coker}_{R} M=R^m/MR^m. coker R M = R m / M R m .
It is free of rank one if every class can be written uniquely as a b ab ab for a fixed class b b b and some a ∈ R a \in R a ∈ R . This is a global condition. In particular it is stronger than requiring the specialized matrix to have corank one at every point of f = 0 f=0 f = 0 . The following theorem concerns one representation. Sections 5 and 6 show that its hypothesis holds for every representation of the forms in the main result.
Theorem 4.1 (Kernel-degree bound). Let f ∈ C [ x 1 , … , x N ] f \in\mathbb{C}[x_1,\ldots,x_N] f ∈ C [ x 1 , … , x N ] be homogeneous and irreducible of degree d ≥ 2 d \ge2 d ≥ 2 . Let M M M be an m × m m \times m m × m matrix of affine-linear forms with det M = f \det M=f det M = f . If its cokernel over R = C [ x 1 , … , x N ] / ( f ) R=\mathbb{C}[x_1,\ldots,x_N]/(f) R = C [ x 1 , … , x N ] / ( f ) is free of rank one, then
Proof. We first factor the adjugate over R R R . The adjugate is the transpose of the matrix of cofactors and satisfies
M adj M = ( adj M ) M = f I m . M\operatorname{adj}M=(\operatorname{adj}M)M=fI_m. M adj M = ( adj M ) M = f I m .
Thus its columns are right-kernel vectors, and its rows are left-kernel vectors, when we work modulo f f f . The assumed isomorphism coker R M ≅ R \operatorname{coker}_{R} M \cong R coker R M ≅ R gives the quotient map
π : R m ⟶ R , ker π = M R m . \pi:R^{m}\longrightarrow R,\qquad\ker\pi=MR^{m}. π : R m ⟶ R , ker π = M R m .
There is a column vector v ∈ R m v\in R^{m} v ∈ R m with π ( z ) = v T z \pi(z)=v^{T}z π ( z ) = v T z . A row vector w T w^{T} w T satisfies w T M = 0 w^{T}M=0 w T M = 0 exactly when its map R m → R R^{m}\to R R m → R vanishes on M R m MR^{m} M R m . It then factors uniquely through π \pi π . Every R R R -linear map R → R R\to R R → R is multiplication by one element of R R R . Consequently every such row has the form a v T av^{T} a v T . In particular the rows of the adjugate have this form, so
adj M = u v T over R (10) \operatorname{adj} M=uv^{T}\quad\text{over }R \tag*{(10)} adj M = u v T over R (10)
for a column vector u ∈ R m u\in R^{m} u ∈ R m .
The adjugate is nonzero modulo f f f . Indeed, differentiating det M = f \det M=f det M = f expresses each partial derivative of f f f as a linear combination of cofactors with constant coefficients. If every cofactor were divisible by f f f , every derivative of f f f would also be divisible by f f f . This is impossible: at least one partial derivative is nonzero, and its degree is d − 1 d-1 d − 1 . Thus both u u u and v v v in (10) are nonzero. Since R R R is a domain, the identities
( M u ) v T = 0 , v T M = 0 (Mu)v^{T}=0,\qquad v^{T}M=0 ( M u ) v T = 0 , v T M = 0
give M u = 0 Mu=0 M u = 0 and v T M = 0 v^{T}M=0 v T M = 0 .
We next bound the degrees of u u u and v v v . Because f f f is homogeneous, R R R is graded by total degree. For a nonzero element of R R R , use its largest nonzero homogeneous degree. This is also the smallest possible degree of a polynomial representing that element. Set
α = max u i ≠ 0 deg R u i , β = max v j ≠ 0 deg R v j . \alpha=\max_{u_i\ne0}\deg_R u_i,\qquad\beta=\max_{v_j\ne0}\deg_R v_j. α = u i = 0 max deg R u i , β = v j = 0 max deg R v j .
Suppose, for a contradiction, that α ≤ d − 2 \alpha\le d-2 α ≤ d − 2 . Choose polynomial representatives u ~ i \widetilde{u}_i u i of degree at most d − 2 d-2 d − 2 . Every coordinate of M u ~ M\widetilde{u} M u then has degree at most d − 1 d-1 d − 1 . It vanishes modulo f f f , so it is a multiple of f f f . A polynomial of degree below d d d cannot be a nonzero multiple of f f f . Therefore
M u ~ = 0 already over T . M\widetilde{u}=0\quad\text{already over }T. M u = 0 already over T .
But M M M has nonzero determinant over T T T , so it has no nonzero polynomial kernel vector. The same argument for the left kernel gives
α ≥ d − 1 , β ≥ d − 1. \alpha\ge d-1,\qquad\beta\ge d-1. α ≥ d − 1 , β ≥ d − 1.
This argument applies to any nonzero kernel vectors over R R R .
Finally, choose coordinates u i u_i u i and v j v_j v j whose degrees are α \alpha α and β \beta β . Degrees add for nonzero products in a graded domain. Thus (10) gives
deg R ( adj M ) i j = α + β ≥ 2 d − 2. \deg_R(\operatorname{adj} M)_{ij}=\alpha+\beta\ge2d-2. deg R ( adj M ) ij = α + β ≥ 2 d − 2.
A cofactor of an affine-linear m × m m\times m m × m matrix has degree at most m − 1 m-1 m − 1 . Passing to R R R cannot increase that degree. Hence
m − 1 ≥ 2 d − 2 , m-1\ge2d-2, m − 1 ≥ 2 d − 2 ,
which is the claimed inequality.
Normality of R R R is not a hypothesis of Theorem 4.1 . It will be used later to establish the required freeness for the particular forms under consideration.
Corollary 4.2. Keep the hypotheses on f f f and the cokernel, but allow entries of M M M to have degree at most e ≥ 1 e \ge1 e ≥ 1 . Then
( m + 1 ) e ≥ 2 d . (m+1)e \ge2d. ( m + 1 ) e ≥ 2 d . Proof. If d > e d > e d > e , the same lifting argument gives α , β ≥ d − e \alpha,\beta\ge d-e α , β ≥ d − e . A cofactor has degree at most e ( m − 1 ) e(m-1) e ( m − 1 ) , so e ( m − 1 ) ≥ 2 ( d − e ) e(m-1) \ge2(d-e) e ( m − 1 ) ≥ 2 ( d − e ) . If d ≤ e d \le e d ≤ e , the claimed inequality follows from m ≥ 1 m \ge1 m ≥ 1 . The adjugate is again nonzero modulo f f f : differentiating det M = f \det M=f det M = f gives the same divisibility contradiction, even though the derivatives of the entries are no longer constant.
Regularity and global freeness
An affine representation is regular at the origin if M ( 0 ) M(0) M ( 0 ) has rank m − 1 m-1 m − 1 . This condition alone does not imply the conclusion of Theorem 4.1 . The matrix
( 0 x y − y 1 + a b x c 1 + d ) \begin{pmatrix}
0 & x & y \\
-y & 1+a & b \\
x & c & 1+d
\end{pmatrix} 0 − y x x 1 + a c y b 1 + d
has rank two at the origin and determinant
b x 2 + ( d − a ) x y − c y 2 . bx^2+(d-a)xy-cy^2. b x 2 + ( d − a ) x y − c y 2 .
This is an irreducible homogeneous cubic represented at size three. Its size is below 2 deg f − 1 2\deg f-1 2 deg f − 1 , and its cokernel is not globally free. Thus regularity at one point cannot replace the global hypothesis in the theorem.
One useful sufficient condition is that R R R be a normal unique factorization domain. Standard facts about hypersurface modules then force the cokernel of every polynomial representation of f f f to be free. § 5 explains those facts and proves the condition for the unpadded odd-skew forms. The Koszul forms require a different argument, given in § 6 .
Sharpness
The bound is sharp for individual representations with free cokernel. The following explicit matrix is the standard construction from a two-path algebraic branching program (ABP) [11 Section 3] .
For d ≥ 2 d \ge2 d ≥ 2 , consider
g d = x 1 ⋯ x d + y 1 ⋯ y d . g_d=x_1\cdots x_d+y_1\cdots y_d. g d = x 1 ⋯ x d + y 1 ⋯ y d .
This polynomial is irreducible. Indeed, viewed as a polynomial in x d x_d x d , it has degree one and its two coefficients x 1 ⋯ x d − 1 x_1\cdots x_{d-1} x 1 ⋯ x d − 1 and y 1 ⋯ y d − 1 y_1\cdots y_{d-1} y 1 ⋯ y d − 1 are relatively prime. Gauss’s lemma reduces irreducibility to the degree-one polynomial over the fraction field of the other variables.
Let H x H_x H x be the ( d − 1 ) × ( d − 1 ) (d-1)\times(d-1) ( d − 1 ) × ( d − 1 ) matrix with ones on the diagonal, entries − x 2 , … , − x d − 1 -x_2,\ldots,-x_{d-1} − x 2 , … , − x d − 1 immediately above the diagonal, and zero elsewhere. Define H y H_y H y in the same way, and put
H = ( H x 0 0 H y ) , a T = ( x 1 e 1 T , y 1 e 1 T ) , b = ( x d e d − 1 y d e d − 1 ) . H=\begin{pmatrix}
H_x & 0 \\
0 & H_y
\end{pmatrix},
\qquad
a^T=(x_1e_1^T,y_1e_1^T),
\qquad
b=\begin{pmatrix}
x_de_{d-1} \\
y_de_{d-1}
\end{pmatrix}. H = ( H x 0 0 H y ) , a T = ( x 1 e 1 T , y 1 e 1 T ) , b = ( x d e d − 1 y d e d − 1 ) .
Here e i e_i e i is the i i i th coordinate vector of C d − 1 \mathbb{C}^{d-1} C d − 1 . The matrix H H H has determinant one, and its inverse sums products along the two chains. In particular
a T H − 1 b = x 1 ⋯ x d + y 1 ⋯ y d = g d . a^T H^{-1}b=x_1\cdots x_d+y_1\cdots y_d=g_d. a T H − 1 b = x 1 ⋯ x d + y 1 ⋯ y d = g d .
Therefore
M d = ( 0 a T − b H ) M_d=\begin{pmatrix} 0 & a^T \\ -b & H \end{pmatrix} M d = ( 0 − b a T H )
satisfies det M d = g d \det M_d=g_d det M d = g d . {#eq:11}
It has size 2 d − 1 2d-1 2 d − 1 . Since H H H is invertible over the polynomial ring, polynomial row and column operations reduce M d M_d M d to the block matrix diag ( g d , H ) \operatorname{diag}(g_d,H) diag ( g d , H ) . Its cokernel over g d = 0 g_d=0 g d = 0 is consequently free of rank one. Equality holds in Theorem 4.1 .
The corresponding ABP consists of two directed paths of length d d d , with common starting and ending vertices and no shared internal vertices. A path contributes the product of its edge labels; the program adds the two products. The entries of the two kernel vectors are products along initial and final portions of these paths. Their degrees are therefore at most d − 1 d-1 d − 1 .
The polynomial g d g_d g d also has a homogeneous determinant of size d d d : use diagonal entries x i x_i x i , entries y 1 , … , y d − 1 y_1,\ldots,y_{d-1} y 1 , … , y d − 1 immediately above the diagonal, and ( − 1 ) d − 1 y d (-1)^{d-1}y_d ( − 1 ) d − 1 y d in the bottom-left corner. Only the diagonal and the full cycle contribute to its determinant, giving g d g_d g d . Thus dc ( g d ) = d \operatorname{dc}(g_d)=d dc ( g d ) = d , and this smaller representation has nonfree cokernel. The example proves sharpness for representations satisfying the lemma; it does not determine the exact complexity of our boundary families.
A superlinear lower bound therefore requires information beyond the hypotheses of Theorem 4.1 . Possible additional data include the number of coordinates required by compatible kernel vectors and the relations among them for a specific family of forms.
Odd-skew forms and factorial hypersurfaces
We prove that every polynomial representation of an unpadded odd-skew form has free rank-one cokernel. The kernel-degree theorem will then give a lower bound on its determinantal complexity. The proof uses a general criterion for a polynomial that is linear in one group of variables. Write
F ( x , y ) = y 1 g 1 ( x ) + ⋯ + y s g s ( x ) . F(x,y)=y_1g_1(x)+\cdots+y_sg_s(x). F ( x , y ) = y 1 g 1 ( x ) + ⋯ + y s g s ( x ) .
For each x x x at which some g i ( x ) g_i(x) g i ( x ) is nonzero, the equation F ( x , y ) = 0 F(x,y)=0 F ( x , y ) = 0 can be solved for one coordinate of y y y . Its fiber is therefore an ( s − 1 ) (s-1) ( s − 1 ) -dimensional vector space. The exceptional fibers lie over the common zero set of the g i g_i g i . We will show that codimension at least three for this set implies factoriality of the hypersurface ring, and hence freeness of every determinant cokernel.
Preliminaries on determinant cokernels
We recall the standard facts about modules over a hypersurface ring that will be used in this section and the next.
The coordinate ring of the hypersurface F = 0 F=0 F = 0 is R = C [ x , y ] / ( F ) R=\mathbb{C}[x,y]/(F) R = C [ x , y ] / ( F ) . In this ring, two polynomials are identified when their difference is divisible by F F F . For a matrix M M M of size m m m , its cokernel over this ring is
coker R M = R m / M R m . \operatorname{coker}_{R} M=R^m/MR^m. coker R M = R m / M R m .
it consists of vectors of functions, with vectors in the image of M M M identified with zero. It is free of rank one if a single element generates it with no relation, so that it is isomorphic to R R R . Throughout, freeness means global freeness over R R R . A one-dimensional cokernel at each individual point does not by itself imply this condition.
A domain is normal if it is integrally closed in its fraction field. It is factorial if it is a unique factorization domain (UFD). Over a normal factorial ring, every rank-one reflexive module is free.
Determinant cokernels on an irreducible normal hypersurface are rank-one reflexive modules, so this fact applies to them. A finite module C C C is reflexive if the natural map
C ⟶ C ∗ ∗ , C ∗ = Hom R ( C , R ) , C \longrightarrow C^{**}, \qquad C^* = \operatorname{Hom}_R(C,R), C ⟶ C ∗∗ , C ∗ = Hom R ( C , R ) ,
is an isomorphism. Over a normal domain, rank-one reflexive modules are recorded by the divisor class group Cl ( R ) \operatorname{Cl}(R) Cl ( R ) ; the zero class is represented by R R R itself. Thus Cl ( R ) = 0 \operatorname{Cl}(R)=0 Cl ( R ) = 0 says precisely that every such module is free. Normality allows us to compute this group after removing a set of codimension at least two. Here codimension is the ambient dimension minus the dimension of the set removed.
At a local ring, the depth of a module is the largest length of a sequence from the maximal ideal that acts successively without zero divisors on the module and its quotients. Depth is at most the dimension of the local ring. A module is maximal Cohen–Macaulay , abbreviated MCM, when equality holds at every prime in its support. For a normal hypersurface ring, a rank-one MCM module is reflexive. A polynomial matrix whose determinant is an irreducible polynomial supplies a rank-one MCM module through its cokernel. We verify that last assertion below. These standard facts, including the depth facts used in the next section, are treated in [3 ] .
A factoriality criterion
Theorem 5.1. Let g 1 , … , g s ∈ C [ x 1 , … , x d ] g_1,\ldots,g_s \in\mathbb{C}[x_1,\ldots,x_d] g 1 , … , g s ∈ C [ x 1 , … , x d ] . Suppose their common zero set has codimension at least three. Set
F ( x , y ) = ∑ i = 1 s y i g i ( x ) , R = C [ x , y ] / ( F ) . F(x,y)=\sum_{i=1}^{s} y_i g_i(x), \qquad R=\mathbb{C}[x,y]/(F). F ( x , y ) = i = 1 ∑ s y i g i ( x ) , R = C [ x , y ] / ( F ) . Then R R R is a normal factorial domain. Every square polynomial matrix M ( x , y ) M(x,y) M ( x , y ) with det M = F \det M=F det M = F satisfies
coker R M ≅ R . \operatorname{coker}_R M \cong R. coker R M ≅ R . Consequently, at every point of F = 0 F=0 F = 0 , the matrix M M M has corank one. The entries of M M M need not be linear, and the g i g_i g i need not be homogeneous.
Proof. We first prove normality and factoriality, and then apply these properties to the cokernel of an arbitrary representation.
Normality. Let
Z = V ( g 1 , … , g s ) , B = A d ∖ Z , X = V ( F ) . Z=V(g_1,\ldots,g_s), \qquad B=\mathbb{A}^d\setminus Z, \qquad X=V(F). Z = V ( g 1 , … , g s ) , B = A d ∖ Z , X = V ( F ) .
There is no common nonconstant factor of the g i g_i g i : such a factor would give a codimension-one subset of Z Z Z . Over the fraction field C ( x ) \mathbb{C}(x) C ( x ) , the polynomial F F F is a nonzero linear polynomial in the y i y_i y i , hence irreducible. Gauss’s lemma now shows that F F F is irreducible in C [ x , y ] \mathbb{C}[x,y] C [ x , y ] as well.
For x ∈ B x\in B x ∈ B , the equation F ( x , y ) = 0 F(x,y)=0 F ( x , y ) = 0 cuts out an ( s − 1 ) (s-1) ( s − 1 ) -dimensional vector space. These vector spaces vary as the kernel of the surjective map
O B s ⟶ O B , ( y i ) ⟼ ∑ i g i y i . \mathcal{O}_B^s \longrightarrow\mathcal{O}_B,\qquad(y_i)\longmapsto\sum_i g_i y_i. O B s ⟶ O B , ( y i ) ⟼ i ∑ g i y i .
Its kernel is a vector bundle: on the open set D ( g i ) D(g_i) D ( g i ) , where g i g_i g i is invertible, solve for y i y_i y i to obtain a product with affine ( s − 1 ) (s-1) ( s − 1 ) -space. Thus the part X B X_B X B of X X X over B B B is smooth. The part we have omitted is Z × A s Z\times\mathbb{A}^s Z × A s . If Z Z Z has codimension c ≥ 3 c\geq3 c ≥ 3 in A d \mathbb{A}^d A d , this part has codimension c − 1 ≥ 2 c-1\geq2 c − 1 ≥ 2 in X X X . Hence X X X is regular in codimension one. A hypersurface ring is Cohen–Macaulay, so it also satisfies the depth condition S 2 S_2 S 2 : at each prime its depth is at least the smaller of two and the local dimension. Serre’s normality criterion combines these two properties to show that R R R is normal.
The divisor class group. Because the omitted part has codimension at least two, it does not change the divisor class group:
Cl ( X ) = Cl ( X B ) . \operatorname{Cl}(X)=\operatorname{Cl}(X_B). Cl ( X ) = Cl ( X B ) .
Likewise, removing Z Z Z from affine space leaves Cl ( B ) = 0 \operatorname{Cl}(B)=0 Cl ( B ) = 0 . On a smooth space, the divisor class group is the Picard group, which records line bundles. The total space of a vector bundle over this smooth base has the same Picard group as the base. In the present case this fact is also visible on the cover D ( g i ) D(g_i) D ( g i ) . There the coordinate ring of X B X_B X B is a polynomial extension of the UFD C [ x ] g i \mathbb{C}[x]_{g_i} C [ x ] g i , and every invertible function on an overlap comes from the base. The transition functions for a line bundle on X B X_B X B therefore give exactly a line bundle on B B B . It follows that
Cl ( X B ) = Pic ( X B ) = Pic ( B ) = Cl ( B ) = 0. \operatorname{Cl}(X_B)=\operatorname{Pic}(X_B)=\operatorname{Pic}(B)=\operatorname{Cl}(B)=0. Cl ( X B ) = Pic ( X B ) = Pic ( B ) = Cl ( B ) = 0.
A normal domain with zero divisor class group is factorial. This proves the first assertion.
Determinant cokernels. Let T = C [ x , y ] T=\mathbb{C}[x,y] T = C [ x , y ] , let M M M have size m m m , and set C = coker T M C=\operatorname{coker}_T M C = coker T M . Since det M = F ≠ 0 \det M=F\ne0 det M = F = 0 , the matrix is injective over T T T . We have the exact sequence
0 ⟶ T m → M T m ⟶ C ⟶ 0. 0\longrightarrow T^m\xrightarrow{M}T^m\longrightarrow C\longrightarrow0. 0 ⟶ T m M T m ⟶ C ⟶ 0.
The identity M adj ( M ) = F I m M\operatorname{adj}(M)=F I_m M adj ( M ) = F I m shows that F F F annihilates C C C , so C C C is an R R R -module and equals the cokernel computed over R R R . At the generic point of F = 0 F=0 F = 0 , the determinant vanishes to order one. The matrix therefore has corank one there, and C C C has rank one over R R R .
The displayed two-term free resolution, localized at any prime in its support, gives C C C depth equal to the dimension of the local hypersurface ring. Thus C C C is MCM. Over the normal ring R R R , it is reflexive; factoriality then gives C ≃ R C\simeq R C ≃ R . Finally, specialize at any point of X X X . The fiber of C C C has dimension one, and this fiber is the cokernel of the specialized matrix. Its dimension is exactly the corank of that matrix. □
The same theorem gives a lower bound in terms of the number of essential variables.
Corollary 5.2. If F F F in Theorem 5.1 depends essentially on all N N N of its variables, then every affine determinantal representation of size m m m satisfies
m ( m + 1 ) ≥ 2 N . m(m+1)\ge2N. m ( m + 1 ) ≥ 2 N .
Proof. Essential dependence makes the linear part of the representing affine map injective: a vector in its kernel would give a direction along which the polynomial never changes. The image is therefore an N N N -dimensional affine space of matrices. On F = 0 F=0 F = 0 every matrix has rank m − 1 m-1 m − 1 , by the theorem; elsewhere it has rank m m m . Meshulam’s bound for an affine space with this minimum rank [16 Theorem 2] gives
N ≤ m 2 − ( m 2 ) = m ( m + 1 ) 2 . N\le m^2-\binom{m}{2}=\frac{m(m+1)}{2}. N ≤ m 2 − ( 2 m ) = 2 m ( m + 1 ) .
Application to unpadded odd-skew forms
Let r ≥ 3 r \ge3 r ≥ 3 be odd. Take a generic skew-symmetric matrix A A A of size r r r , its signed maximal-Pfaffian vector p ( A ) p(A) p ( A ) , and an independent symmetric matrix S S S . The polynomial is
Q r , 0 ( A , S ) = p ( A ) T S p ( A ) . Q_{r,0}(A,S)=p(A)^{T}Sp(A). Q r , 0 ( A , S ) = p ( A ) T S p ( A ) .
It is linear in the entries of S S S . The coefficients of those entries are the products p i p j p_i p_j p i p j , with a factor of two on off-diagonal entries. Thus their common zero set is exactly p ( A ) = 0 p(A)=0 p ( A ) = 0 . Equivalently, A A A has rank at most r − 3 r-3 r − 3 . This set has codimension three in the space of skew-symmetric matrices. In the language of ideals, the coefficient ideal is ( p 1 , … , p r ) 2 (p_1,\ldots,p_r)^2 ( p 1 , … , p r ) 2 . Theorem 5.1 now applies.
Corollary 5.3. Every square polynomial matrix with determinant Q r , 0 Q_{r,0} Q r , 0 has free rank-one cokernel over C [ A , S ] / ( Q r , 0 ) \mathbb{C}[A,S]/(Q_{r,0}) C [ A , S ] / ( Q r , 0 ) , and therefore has corank one at every zero of Q r , 0 Q_{r,0} Q r , 0 . In particular, every affine representation is regular at the origin, so dc ( Q r , 0 ) = r dc ( Q r , 0 ) \operatorname{dc}(Q_{r,0})=r\operatorname{dc}(Q_{r,0}) dc ( Q r , 0 ) = r dc ( Q r , 0 ) . Since Q r , 0 Q_{r,0} Q r , 0 is homogeneous of degree r r r , the kernel-degree lemma gives
dc ( Q r , 0 ) ≥ 2 r − 1. \operatorname{dc}(Q_{r,0}) \ge2r-1. dc ( Q r , 0 ) ≥ 2 r − 1. The freeness assertion holds for every polynomial representation. Thus its application to affine representations imposes no restriction on their size or matrix form.
The degree-one hypothesis. The hypothesis concerns the degree of the whole polynomial in S S S . For example, the polynomial det ( A + S ) \det(A+S) det ( A + S ) has Q r , 0 Q_{r,0} Q r , 0 as its first S S S -term, and hence the same coefficient ideal for that term. It is itself an exact determinant of size r r r . Its higher S S S -terms change the hypersurface: the equation is no longer the linear equation in S S S used above. The proof relies on the degree-one dependence of the entire polynomial on S S S .
The conclusion applies to the unpadded forms Q r , 0 Q_{r,0} Q r , 0 . No corresponding freeness assertion for every padded odd-skew form is used here.
Cokernels of the padded Koszul forms
We prove that every polynomial representation of a padded Koszul form has free rank-one cokernel. Unlike the factorial rings of the preceding section, the hypersurface rings here have nontrivial rank-one reflexive modules. We will show that none of the nonfree modules is MCM, which excludes them as determinant cokernels.
Away from v = 0 v=0 v = 0 , the polynomial is locally a generic determinant in a smaller set of variables. The classification of rank-one MCM modules over a generic determinant gives three possibilities: the free module and two nonfree modules. We first prove that a global reflexive class is determined by its restriction to this open set. We then construct the two nonfree candidates across v = 0 v=0 v = 0 and compute their depths. Both fail the MCM condition there. The only remaining possibility is the free module. We use the definitions and standard module-theoretic facts recalled in § 5 .
Local determinant form
Fix p ≥ 2 p \ge2 p ≥ 2 and set
d = 2 p + 1 , r = ( 2 p + 1 p ) , R 0 = ( 2 p p ) , c = ( 2 p p − 1 ) . d=2p+1,\qquad r=\binom{2p+1}{p},\qquad R_0=\binom{2p}{p},\qquad c=\binom{2p}{p-1}. d = 2 p + 1 , r = ( p 2 p + 1 ) , R 0 = ( p 2 p ) , c = ( p − 1 2 p ) .
Here d d d is the number of v v v -variables, r r r is the size of the seed matrix, R 0 R_0 R 0 is its rank at nonzero v v v , and c = r − R 0 c=r-R_0 c = r − R 0 is its corank there. For the smallest case p = 2 p=2 p = 2 , these numbers are d = 5 d=5 d = 5 , r = 10 r=10 r = 10 , R 0 = 6 R_0=6 R 0 = 6 , and c = 4 c=4 c = 4 . Padding this case gives every final size n ≥ 10 n\ge10 n ≥ 10 .
Let V = C d V=\mathbb{C}^d V = C d and E = ⋀ p V E=\bigwedge^p V E = ⋀ p V . Fix a volume form on V V V . The seed matrix records the alternating-product pairing
A ( v ) : E ⟶ E ∗ , A ( v ) ( α , β ) = vol ( v ∧ α ∧ β ) . A(v):E\longrightarrow E^*,\qquad A(v)(\alpha,\beta)=\operatorname{vol}(v\wedge\alpha\wedge\beta). A ( v ) : E ⟶ E ∗ , A ( v ) ( α , β ) = vol ( v ∧ α ∧ β ) .
For v ≠ 0 v\ne0 v = 0 , its left and right kernels are both
U v = v ∧ ⋀ p − 1 V . U_v=v\wedge\bigwedge^{p-1}V. U v = v ∧ ⋀ p − 1 V .
a space of dimension c c c . The first nonzero determinant coefficient is the restriction of the perturbation to these left and right kernel spaces.
To add k k k rows and columns, put n = r + k n=r+k n = r + k and define
f = K p , k = [ t c ] det ( A ( v ) + t S U t V 0 Z ) . f=\mathcal{K}_{p,k}=[t^c]\det\left(\begin{array}{cc} A(v)+tS & U \\ tV_0 & Z \end{array}\right). f = K p , k = [ t c ] det ( A ( v ) + tS t V 0 U Z ) .
The notation [ t c ] [t^c] [ t c ] means the coefficient of t c t^c t c . The matrix S S S ranges over a linear complement to A ( V ) A(V) A ( V ) in Mat r \operatorname{Mat}_r Mat r ; its role is to provide the matrix directions that the seed does not already use. The independent matrices U , V 0 , Z U,V_0,Z U , V 0 , Z have sizes r × k r\times k r × k , k × r k\times r k × r , and k × k k\times k k × k . Write Y Y Y for all these coefficient variables and put q = c + k q=c+k q = c + k . There are n 2 n^2 n 2 variables altogether.
On any sufficiently small open set with v ≠ 0 v\ne0 v = 0 , choose bases in the kernel spaces. The determinant coefficient then has the form
f ( v , Y ) = u ( v ) det q ( κ v ( Y ) ) , u ( v ) ≠ 0. (11) f(v,Y)=u(v)\det\nolimits_q\bigl(\kappa_v(Y)\bigr),\qquad u(v)\ne0. \tag*{(11)} f ( v , Y ) = u ( v ) det q ( κ v ( Y ) ) , u ( v ) = 0. (11)
Here u u u is invertible on the chosen open set, and κ v \kappa_v κ v is the surjective map that restricts the coefficient bilinear form to U v ⊕ C k U_v\oplus\mathbb{C}^k U v ⊕ C k on both sides. Thus, after a linear change of the Y Y Y -coordinates, f f f is an ordinary determinant of size q q q , times an invertible factor. The remaining Y Y Y -coordinates are independent variables that do not enter that determinant. We can therefore apply the classification of rank-one MCM modules over a generic determinant.
Rank-one MCM modules over the generic determinant
Let F F F be a field, let X X X be a generic q × q q\times q q × q matrix with q ≥ 2 q\ge2 q ≥ 2 , and let z z z be any list of independent extra variables. Set
A F = F [ z , X ] / ( det X ) . A_F=F[z,X]/(\det X). A F = F [ z , X ] / ( det X ) .
The ring A F A_F A F is normal. Its class group is Z \mathbb{Z} Z , with opposite generators represented by coker X \operatorname{coker}X coker X and coker X T \operatorname{coker}X^T coker X T . Although the class group has infinitely many elements, only three of them are represented by rank-one MCM modules. Up to isomorphism, those modules are
A F , coker X , coker X T . A_F,\qquad\operatorname{coker}X,\qquad\operatorname{coker}X^T. A F , coker X , coker X T .
This is the generic-determinant classification of Bruns and Bruns–Vetter, in the form recorded by Buchweitz–Leuschke [4 Section 4.4] . The cokernels can be computed over F [ z , X ] F[z,X] F [ z , X ] or over A F A_F A F , since det X \det X det X annihilates them. We use this classification only over a field with polynomial extra variables; no version over an arbitrary coefficient ring is being assumed.
The classification has the following consequence for polynomial matrix representations.
Lemma 6.1. Let F F F be a field, q ≥ 2 q \ge2 q ≥ 2 , and let z z z denote any list of additional variables. Suppose a square polynomial matrix M ( X , z ) M(X,z) M ( X , z ) satisfies
det M = a det q X , a ∈ F ∗ . \det M = a\det\nolimits_q X,\qquad a \in F^{*}. det M = a det q X , a ∈ F ∗ . Then, along X = 0 X = 0 X = 0 , either M ( 0 , z ) M(0,z) M ( 0 , z ) has corank one for every z z z , or it has corank q q q for every z z z .
Proof. The matrix gives an injective map between free F [ z , X ] F[z,X] F [ z , X ] -modules, whose cokernel is annihilated by det X \det X det X . Its two-term free resolution makes this cokernel MCM over A F A_F A F . It has rank one there, because the determinant vanishes to order one at the generic point of det X = 0 \det X = 0 det X = 0 . The classification therefore applies.
Now set X = 0 X = 0 X = 0 . The free module has a one-dimensional fiber. Each of coker X X X and coker X T X^T X T has a q q q -dimensional fiber, since the specialized matrix is zero. These dimensions do not depend on z z z . The dimension of the cokernel fiber is exactly the corank of the specialized matrix M ( 0 , z ) M(0,z) M ( 0 , z ) .
Applying the lemma over F = C ( v ) F = \mathbb{C}(v) F = C ( v ) gives corank one or q q q along Y = 0 Y = 0 Y = 0 at generic v v v . Here the seed variables are treated as rational scalars, and the coefficient variables remain polynomial variables. The behavior at v = 0 v = 0 v = 0 will exclude the two nonfree possibilities.
Global freeness
Theorem 6.2. For p ≥ 2 p \ge2 p ≥ 2 and k ≥ 0 k \ge0 k ≥ 0 , set
R = C [ v , Y ] / ( K p , k ) . \mathcal{R} = \mathbb{C}[v,Y]/(K_{p,k}). R = C [ v , Y ] / ( K p , k ) . Every rank-one MCM module over R \mathcal{R} R is free. Consequently, every square polynomial matrix M M M with det M = K p , k \det M = K_{p,k} det M = K p , k satisfies
coker R M ≅ R . \operatorname{coker}_{\mathcal{R}} M \cong\mathcal{R}. coker R M ≅ R . At every zero of the polynomial, the matrix has corank one. In particular, every affine representation is regular at the origin.
Proof. Write B = C [ v 1 , … , v d ] B = \mathbb{C}[v_1,\ldots,v_d] B = C [ v 1 , … , v d ] and T = B [ Y ] T = B[Y] T = B [ Y ] .
The divisor class group. We first prove that R \mathcal{R} R is a normal domain. Equation (11) shows that f f f is irreducible over C ( v ) \mathbb{C}(v) C ( v ) . There is no nonconstant factor of f f f that lies in B B B . Indeed, such a factor would vanish at some nonzero v v v , forcing f ( v , Y ) f(v,Y) f ( v , Y ) to be zero for every Y Y Y , contrary to the local determinant description. Gauss’s lemma shows that f f f is irreducible over B B B .
On v ≠ 0 v \ne0 v = 0 , the singular locus has codimension four in the ambient space, just as it does for a generic determinant. The remaining set v = 0 v = 0 v = 0 has ambient codimension d ≥ 5 d \ge5 d ≥ 5 . Thus the hypersurface is regular in codimension one. As a hypersurface it satisfies S 2 S_2 S 2 , so Serre’s criterion makes R \mathcal{R} R normal.
Let X = Spec R \mathcal{X} = \operatorname{Spec}\mathcal{R} X = Spec R , and remove the exceptional set:
X ∘ = X ∖ { v = 0 } . \mathcal{X}^{\circ} = \mathcal{X} \setminus\{v = 0\}. X ∘ = X ∖ { v = 0 } . The polynomial vanishes identically when v = 0 v = 0 v = 0 , so the set removed has codimension d − 1 ≥ 4 d - 1 \ge4 d − 1 ≥ 4 in X \mathcal{X} X . It does not change the class group:
Cl ( X ) = Cl ( X ∘ ) . \operatorname{Cl}(\mathcal{X}) = \operatorname{Cl}(\mathcal{X}^{\circ}). Cl ( X ) = Cl ( X ∘ ) .
Projection onto the v v v -coordinates makes X ∘ \mathcal{X}^{\circ} X ∘ a family over A d ∖ { 0 } \mathbb{A}^{d} \setminus\{0\} A d ∖ { 0 } . Locally this family is a generic determinant hypersurface, with independent affine coordinates. In particular, the family is flat and its fibers are integral. Passing from the base to its fraction field gives a map
Cl ( R ) ⟶ Cl ( C ( v ) [ Y ] / ( f ) ) = Z . (12) \operatorname{Cl}(\mathcal{R}) \longrightarrow\operatorname{Cl}(\mathbb{C}(v)[Y]/(f)) = \mathbb{Z}. \tag*{(12)} Cl ( R ) ⟶ Cl ( C ( v ) [ Y ] / ( f )) = Z . (12)
We claim that this map is injective.
The only divisor classes lost on passing to the fraction field are generated by vertical prime divisors: divisors lying over a proper subset of the v v v -space. Flatness and the dimension formula show that each such divisor lies over a divisor of the base, rather than over a subset of higher codimension. A base divisor is defined by an irreducible polynomial h ( v ) h(v) h ( v ) . In a local product chart, its inverse image is a generic determinant hypersurface over the domain B / ( h ) B/(h) B / ( h ) , together with extra affine coordinates. It is therefore integral and generically reduced. Its multiplicity in the divisor of h h h is one. Thus the whole vertical divisor is principal: it is already the divisor of the single function h h h . Every class in the kernel is consequently zero, proving injectivity.
Consequently, the class of a rank-one reflexive module is determined by its class after passage to C ( v ) \mathbb{C}(v) C ( v ) .
The nonfree reflexive classes. Let K = ker A ( v ) K = \ker A(v) K = ker A ( v ) as a B B B -module. Its restriction away from the origin is the kernel bundle U v U_v U v . The exact Koszul complex gives a free resolution of K K K over the whole base:
0 ⟶ B ( − p ) ⟶ B ( − ( p − 1 ) ) ⊗ V ⟶ ⋯ ⟶ B ( − 1 ) ⊗ ⋀ p − 1 V ⟶ K ⟶ 0. (13) 0 \longrightarrow B(-p) \longrightarrow B(-(p-1)) \otimes V \longrightarrow\cdots\longrightarrow B(-1) \otimes\bigwedge^{p-1} V \longrightarrow K \longrightarrow0. \tag*{(13)} 0 ⟶ B ( − p ) ⟶ B ( − ( p − 1 )) ⊗ V ⟶ ⋯ ⟶ B ( − 1 ) ⊗ ⋀ p − 1 V ⟶ K ⟶ 0. (13)
Here B ( − j ) B(-j) B ( − j ) denotes the grading shift with generator in degree j j j . We will use the length and ranks of this resolution and the fact that its differentials are linear in v v v .
The pairing on the ambient vector spaces gives a map coker A → K ∗ \operatorname{coker} A \to K^{*} coker A → K ∗ , where K ∗ = Hom B ( K , B ) K^{*} = \operatorname{Hom}_{B}(K,B) K ∗ = Hom B ( K , B ) . It is an isomorphism away from the origin. We need it at the origin as well. A truncation of the same Koszul complex shows that coker A A A has projective dimension p + 1 p+1 p + 1 and depth p p p there. Here projective dimension is the shortest possible length of a projective resolution; the resolutions in use are free. The module is locally free elsewhere. Its positive depth excludes torsion supported at the origin. It is therefore torsion-free, and it satisfies S 2 S_{2} S 2 because p ≥ 2 p \ge2 p ≥ 2 . It follows that coker A A A is reflexive. The dual K ∗ K^{*} K ∗ is reflexive too. Isomorphisms between reflexive modules extend across sets of codimension at least two, so the map extends to an isomorphism on all of Spec B \operatorname{Spec} B Spec B :
K ∗ ≃ coker A , K^{*} \simeq\operatorname{coker} A, K ∗ ≃ coker A ,
up to the grading shift fixed by the convention for A A A .
We record the numerical information supplied by the two Koszul resolutions. The depths here are at the origin of the v v v -space:
pd B K = p − 1 , depth B K = p + 2 , pd B K ∗ = p + 1 , depth B K ∗ = p , Ext B i ( K ∗ , B ) = 0 ( 1 ≤ i ≤ p ) . (14) \begin{aligned}
\operatorname{pd}_{B} K &= p-1, & \operatorname{depth}_{B} K &= p+2, \\
\operatorname{pd}_{B} K^{*} &= p+1, & \operatorname{depth}_{B} K^{*} &= p, \\
\operatorname{Ext}^{i}_{B}(K^{*},B) &= 0 && (1 \le i \le p). \tag*{(14)}
\end{aligned} pd B K pd B K ∗ Ext B i ( K ∗ , B ) = p − 1 , = p + 1 , = 0 depth B K depth B K ∗ = p + 2 , = p , ( 1 ≤ i ≤ p ) . (14)
The depth formulas use the equality between depth plus projective dimension and the dimension 2 p + 1 2p+1 2 p + 1 of the regular local ring. The last line follows by dualizing the truncated Koszul complex: its positive-degree cohomology occurs only at the last term.
Add the padding directions by setting
P = K ⊕ B k , P ∗ = K ∗ ⊕ B k , P = K \oplus B^{k}, \qquad P^{*} = K^{*} \oplus B^{k}, P = K ⊕ B k , P ∗ = K ∗ ⊕ B k ,
and extend both modules to T = B [ Y ] T = B[Y] T = B [ Y ] . The coefficient variables induce a map
ϕ Y : P ⟶ P ∗ . \phi_{Y}: P \longrightarrow P^{*}. ϕ Y : P ⟶ P ∗ .
When v ≠ 0 v \ne0 v = 0 , this is precisely the restriction of the coefficient bilinear form to U v ⊕ C k U_v \oplus\mathbb{C}^{k} U v ⊕ C k . It is invertible over the full fraction field, hence injective. Define
C = coker ϕ Y . C = \operatorname{coker}\phi_{Y}. C = coker ϕ Y .
Away from v = 0 v = 0 v = 0 , this is the cokernel module of the smaller generic determinant. We next verify that it extends to a reflexive module on the entire hypersurface.
Let q \mathfrak{q} q be any prime of T T T containing ( v ) (v) ( v ) , and put h = dim T q h = \dim T_{\mathfrak{q}} h = dim T q . The Koszul differentials have entries in ( v ) (v) ( v ) , so their resolutions remain minimal after this localization. The displayed projective dimensions therefore give
depth T q P q = h − p + 1 , depth T q P q ∗ = h − p − 1. \operatorname{depth}_{T_{\mathfrak{q}}} P_{\mathfrak{q}} = h - p + 1, \qquad\operatorname{depth}_{T_{\mathfrak{q}}} P^{*}_{\mathfrak{q}} = h - p - 1. depth T q P q = h − p + 1 , depth T q P q ∗ = h − p − 1.
Apply the depth lemma to 0 ⟶ P ⟶ P ∗ ⟶ C ⟶ 0 0 \longrightarrow P \longrightarrow P^{*} \longrightarrow C \longrightarrow0 0 ⟶ P ⟶ P ∗ ⟶ C ⟶ 0 . Since the first module has strictly larger depth than the second, the cokernel has the second depth:
depth T q C q = h − p − 1 ≥ p ≥ 2. (15) \operatorname{depth}_{T_{\mathfrak{q}}} C_{\mathfrak{q}} = h - p - 1 \ge p \ge2. \tag*{(15)} depth T q C q = h − p − 1 ≥ p ≥ 2. (15)
The inequality uses h ≥ 2 p + 1 h \ge2p + 1 h ≥ 2 p + 1 , since q \mathfrak{q} q contains all 2 p + 1 2p + 1 2 p + 1 seed variables.
This lower bound has two immediate uses. First, C C C has no nonzero submodule supported on v = 0 v = 0 v = 0 . Such a submodule would have an associated prime containing ( v ) (v) ( v ) , and at an associated prime the module has depth zero. Equation (15) excludes that possibility. Off v = 0 v = 0 v = 0 , the local determinant calculation gives f C = 0 fC = 0 f C = 0 . The submodule f C fC f C is therefore supported on v = 0 v = 0 v = 0 , so it must be zero everywhere. We can now regard C C C as an R \mathcal{R} R -module, not just a T T T -module. Second, any torsion of C C C would be supported on v = 0 v = 0 v = 0 , since the generic determinant cokernel is torsion-free off that set. The same argument excludes this torsion.
Off v = 0 v = 0 v = 0 , the module C C C is rank-one MCM. At primes over v = 0 v = 0 v = 0 , Equation (15) gives S 2 S_{2} S 2 . The depth is unchanged when calculated over T q T_{\mathfrak{q}} T q or its quotient by f f f , because f f f annihilates the module. Thus C C C is torsion-free and satisfies S 2 S_{2} S 2 over the normal ring R \mathcal{R} R : it is a rank-one reflexive module.
Its class over C ( v ) \mathbb{C}(v) C ( v ) is the generator represented by the cokernel of the generic determinant. Hence the injection (12) is also surjective. We have now identified Cl ( R ) \operatorname{Cl}(\mathcal{R}) Cl ( R ) with Z \mathbb{Z} Z , with C C C representing one generator and its dual representing the opposite generator.
The depth of the nonfree classes. Localize at the prime ( v ) (v) ( v ) . At this prime the coefficient variables are generic, and dim T ( v ) = 2 p + 1 \dim T_{(v)} = 2p + 1 dim T ( v ) = 2 p + 1 . Formula (15) becomes
depth C ( v ) = p < 2 p = dim R ( v ) . \operatorname{depth} C_{(v)} = p < 2p = \dim\mathcal{R}_{(v)}. depth C ( v ) = p < 2 p = dim R ( v ) .
The module C C C therefore fails the MCM condition. For p = 2 p = 2 p = 2 , this is the concrete inequality 2 < 4 2 < 4 2 < 4 . We need the same conclusion for the opposite class. Its reflexive representative is the dual
C ∗ = Hom R ( C , R ) . C^{*} = \operatorname{Hom}_{\mathcal{R}}(C,\mathcal{R}). C ∗ = Hom R ( C , R ) .
The hypersurface change-of-rings isomorphism identifies it with Ext T 1 ( C , T ) \operatorname{Ext}^{1}_{T}(C,T) Ext T 1 ( C , T ) , with a grading shift if we retain gradings. To compute this module, dualize the exact sequence 0 ⟶ P → ϕ Y P ∗ ⟶ C ⟶ 0 0 \longrightarrow P \xrightarrow{\phi_Y} P^{*} \longrightarrow C \longrightarrow0 0 ⟶ P ϕ Y P ∗ ⟶ C ⟶ 0 over T T T . Here Hom T ( C , T ) = 0 \operatorname{Hom}_{T}(C,T)=0 Hom T ( C , T ) = 0 , since C C C is annihilated by the nonzero polynomial f f f . Reflexivity gives Hom T ( P ∗ , T ) = P \operatorname{Hom}_{T}(P^{*},T)=P Hom T ( P ∗ , T ) = P , and (17) gives Ext T 1 ( P ∗ , T ) = 0 \operatorname{Ext}^{1}_{T}(P^{*},T)=0 Ext T 1 ( P ∗ , T ) = 0 . The resulting sequence is
0 ⟶ P → ϕ Y T P ∗ ⟶ Ext T 1 ( C , T ) ⟶ 0. 0 \longrightarrow P \xrightarrow{\phi_Y^{T}} P^{*} \longrightarrow\operatorname{Ext}^{1}_{T}(C,T) \longrightarrow0. 0 ⟶ P ϕ Y T P ∗ ⟶ Ext T 1 ( C , T ) ⟶ 0.
Thus C ∗ ≃ coker ϕ Y T C^{*}\simeq\operatorname{coker}\phi_Y^{T} C ∗ ≃ coker ϕ Y T . The same depth calculation applies to this transposed map and gives
depth R ( v ) C ( v ) ∗ = p < 2 p = dim R ( v ) . \operatorname{depth}_{\mathcal{R}_{(v)}} C^{*}_{(v)}=p<2p=\dim\mathcal{R}_{(v)}. depth R ( v ) C ( v ) ∗ = p < 2 p = dim R ( v ) .
The opposite class is not MCM either.
Now let D D D be any rank-one MCM R \mathcal{R} R -module. It is reflexive, and it remains MCM after passage to C ( v ) \mathbb{C}(v) C ( v ) . The generic-determinant classification permits only the classes 0 0 0 , 1 1 1 , and − 1 -1 − 1 . Our class-group isomorphism says that globally these are represented by R \mathcal{R} R , C C C , and C ∗ C^{*} C ∗ , respectively. Rank-one reflexive modules in the same class are isomorphic. Since C C C and C ∗ C^{*} C ∗ are not MCM, the only possibility for D D D is D ≃ R D\simeq\mathcal{R} D ≃ R .
Finally, a matrix with determinant f f f has a two-term free resolution over T T T , so its cokernel is rank-one MCM over R \mathcal{R} R , exactly as in § 5 . It is therefore free. At any zero of f f f , its one-dimensional fiber is the cokernel of the specialized matrix, proving the corank assertion. □ \square □
The determinantal complexity bound
The freeness theorem verifies the hypothesis of the kernel-degree bound for every exact representation of a padded Koszul form.
Corollary 6.3. For p ≥ 2 p\ge2 p ≥ 2 , k ≥ 0 k\ge0 k ≥ 0 , and n = ( 2 p + 1 p ) + k n=\binom{2p+1}{p}+k n = ( p 2 p + 1 ) + k ,
dc ( K p , k ) = n , dc ( K p , k ) ≥ 2 n − 1. \operatorname{dc}(K_{p,k})=n,\qquad\operatorname{dc}(K_{p,k})\ge2n-1. dc ( K p , k ) = n , dc ( K p , k ) ≥ 2 n − 1. In particular, f n = K 2 , n − 10 f_n=K_{2,n-10} f n = K 2 , n − 10 has exact-minus-border gap at least n − 1 n-1 n − 1 for every n ≥ 10 n\ge10 n ≥ 10 .
Proof. The polynomial has degree n n n , so its border complexity is at least n n n . Its defining determinant limit gives the matching upper bound. It is homogeneous and irreducible, and Theorem 6.2 shows that every exact representation has free cokernel. The kernel-degree lemma, Theorem 4.1 , then gives size at least 2 n − 1 2n-1 2 n − 1 .
Betti numbers of the natural cokernel
The Betti numbers of the natural nonfree module C C C grow faster than the seed size. Since C C C has been excluded as a determinant cokernel, this growth does not directly give a larger lower bound for exact representations, whose cokernels belong to the free class. The calculation is as follows. At generic coefficient values, C C C has a minimal linear mapping-cone resolution. Its Betti numbers, the numbers of free generators in successive terms of a minimal resolution, are
β 0 ( C ) = r , β i ( C ) = ( 2 p + 2 p + 1 − i ) ( 1 ≤ i ≤ p + 1 ) . \beta_{0}(C)=r,\qquad\beta_i(C)=\binom{2p+2}{p+1-i}\qquad(1\le i\le p+1). β 0 ( C ) = r , β i ( C ) = ( p + 1 − i 2 p + 2 ) ( 1 ≤ i ≤ p + 1 ) .
Their sum is
∑ i β i ( C ) = 2 2 p + 1 = Θ ( r log r ) . \sum_i \beta_i(C)=2^{2p+1}=\Theta(r\sqrt{\log r}). i ∑ β i ( C ) = 2 2 p + 1 = Θ ( r log r ) .
To obtain these numbers, combine the free resolutions of K K K and K ∗ K^* K ∗ using the map φ Y \varphi_Y φ Y . This is the mapping-cone construction: the terms of the two resolutions appear in adjacent degrees. Over the fraction field of the coefficient variables, the invertible padding block cancels. Every remaining comparison map has positive v v v -degree, so the resulting resolution is minimal at ( v ) (v) ( v ) . Adjacent binomial ranks add by Pascal’s identity, giving the formula for β i \beta_i β i . Symmetry of the binomial coefficients of order 2 p + 2 2p+2 2 p + 2 gives the sum.
The projective dimension p + 1 p+1 p + 1 is the depth defect that excludes C C C and its dual from being determinant cokernels. It does not transfer their growing Betti sum to the size of a representation using the free class. In fact, the seed module ker A ( v ) \operatorname{ker} A(v) ker A ( v ) already has total Betti number
4 p − r = Θ ( r log r ) , 4^p-r=\Theta\left(r\sqrt{\log r}\right), 4 p − r = Θ ( r log r ) ,
even though the matrix A ( v ) A(v) A ( v ) has size only r r r . A superlinear lower bound would need an additional quantitative restriction on the representations that remain possible.
Exact complexity of general boundary points
We now prove that every concise degree-n n n point of the determinant boundary has exact complexity at least n + 2 n+2 n + 2 . The argument excludes size n + 1 n+1 n + 1 under the weaker hypothesis of at least n 2 − n + 2 n^2-n+2 n 2 − n + 2 essential variables.
A size-( n + 1 ) (n+1) ( n + 1 ) representation of a degree-n n n form has a homogeneous linear part consisting entirely of singular matrices. A theorem of de Seguins Pazzis classifies such matrix spaces when their dimension is sufficiently large. The resulting block form allows the original polynomial to be written as a homogeneous determinant of size n n n .
Write ess ( f ) \operatorname{ess}(f) ess ( f ) for the number of essential variables of a homogeneous form f f f . This is the smallest number of variables left after an invertible linear change of coordinates, or, equivalently, the dimension of the span of the first partial derivatives. Throughout this section, matrix entries may be affine-linear, as required by the usual definition of dc.
Theorem 7.1. Let f f f be a nonzero homogeneous complex form of degree n ≥ 3 n\ge3 n ≥ 3 . If ess ( f ) ≥ n 2 − n + 2 \operatorname{ess}(f)\ge n^2-n+2 ess ( f ) ≥ n 2 − n + 2 , then
dc ( f ) ≤ n + 1 ⟹ f = det n ( L ) \operatorname{dc}(f)\le n+1\quad\Longrightarrow\quad f=\det_n(L) dc ( f ) ≤ n + 1 ⟹ f = n det ( L ) for a matrix L L L of homogeneous linear forms. Thus exact complexity n + 1 n+1 n + 1 is impossible in this range of essential dimensions.
Proof. Work in N = ess ( f ) N=\operatorname{ess}(f) N = ess ( f ) essential variables. If a representation has size at most n n n , its size must be exactly n n n , because its determinant has degree n n n . Taking the degree-n n n homogeneous part then gives a homogeneous size-n n n representation. It remains to consider size n + 1 n+1 n + 1 .
The homogeneous linear part. Write
f ( x ) = det ( A 0 + L ( x ) ) , f(x)=\det(A_0+L(x)), f ( x ) = det ( A 0 + L ( x )) ,
where A 0 A_0 A 0 is constant and L L L is homogeneous linear, both of size n + 1 n+1 n + 1 . The linear map L L L is injective. Indeed, if L ( w ) = 0 L(w)=0 L ( w ) = 0 for a nonzero direction w w w , then the determinant and hence f f f would be unchanged on every line x + λ w x+\lambda w x + λ w . That would contradict our choice of essential variables.
Homogenizing the determinant identity gives
det ( L ( x ) + t A 0 ) = t f ( x ) . \det(L(x) + tA_0) = tf(x). det ( L ( x ) + t A 0 ) = t f ( x ) .
Setting t = 0 t = 0 t = 0 shows that every matrix in L ( C N ) L(\mathbb{C}^N) L ( C N ) is singular. This is an N N N -dimensional linear space in Mat n + 1 \operatorname{Mat}_{n+1} Mat n + 1 . Its maximum rank is n n n : if all ranks were at most n − 1 n - 1 n − 1 , the coefficient of t t t would vanish, contradicting f ≠ 0 f \ne0 f = 0 .
The compression-space classification. Theorem 1.7 of de Seguins Pazzis [7 ] classifies sufficiently large spaces of this kind. For upper rank n n n in square size n + 1 n + 1 n + 1 , its dimension threshold is
( n + 1 ) n − 2 n + 2 = n 2 − n + 2. (n+1)n - 2n + 2 = n^2 - n + 2. ( n + 1 ) n − 2 n + 2 = n 2 − n + 2.
Our hypothesis meets that threshold, and its exceptional finite-field case does not occur over C \mathbb{C} C . After constant row and column changes, the space lies in a compression space , meaning a space with a prescribed rectangular zero block:
L = ( A C B 0 ) , A ∈ Mat s , q , C ∈ Mat s , s + 1 , B ∈ Mat q + 1 , q , s + q = n . L = \begin{pmatrix} A & C \\ B & 0 \end{pmatrix}, \qquad A \in\operatorname{Mat}_{s,q}, \quad C \in\operatorname{Mat}_{s,s+1}, \quad B \in\operatorname{Mat}_{q+1,q}, \quad s+q=n. L = ( A B C 0 ) , A ∈ Mat s , q , C ∈ Mat s , s + 1 , B ∈ Mat q + 1 , q , s + q = n .
The row and column changes may multiply the determinant by a nonzero constant. We absorb that constant into one row of the final representation.
Consider the coefficient of t t t in det ( L + t A 0 ) \det(L+tA_0) det ( L + t A 0 ) . A determinant term contributing to that coefficient uses exactly one entry from A 0 A_0 A 0 . It must use that entry in the zero block. Of the remaining entries, s s s come from C C C and q q q from B B B . No entry comes from A A A . Thus the entire polynomial f f f depends only on the entries of B B B and C C C .
If 2 ≤ s ≤ n − 2 2 \le s \le n-2 2 ≤ s ≤ n − 2 , the number of those entries is at most
s ( s + 1 ) + q ( q + 1 ) = n 2 + n − 2 s ( n − s ) ≤ n 2 − 3 n + 8 < n 2 − n + 2 ≤ N . \begin{aligned}
s(s+1)+q(q+1)=n^2+n-2s(n-s) \\
&\le n^2-3n+8<n^2-n+2\le N.
\end{aligned} s ( s + 1 ) + q ( q + 1 ) = n 2 + n − 2 s ( n − s ) ≤ n 2 − 3 n + 8 < n 2 − n + 2 ≤ N .
This contradicts essential dependence on N N N variables. The interior range occurs only for n ≥ 4 n \ge4 n ≥ 4 , which is exactly the range in which the displayed inequality is strict. We are left with s = 0 , 1 , n − 1 , n s=0,1,n-1,n s = 0 , 1 , n − 1 , n .
Reduction to size $n. If s = 0 s=0 s = 0 , the linear matrix has the form L = [ B , 0 ] L=[B,0] L = [ B , 0 ] . Its coefficient of t t t is det [ B , d ] \det[B,d] det [ B , d ] , where d d d is a fixed column from A 0 A_0 A 0 . This column is nonzero since f ≠ 0 f \ne0 f = 0 . A constant row change makes it a coordinate vector. Expansion along it leaves a homogeneous determinant of size n n n . The case s = n s=n s = n is the transpose of this argument.
If s = 1 s=1 s = 1 , write C = ( c 1 , c 2 ) C=(c_1,c_2) C = ( c 1 , c 2 ) , and denote the two constant columns in the lower-right block of A 0 A_0 A 0 by d 1 , d 2 d_1,d_2 d 1 , d 2 . Expanding along the first row gives
f = ± ( c 1 det [ B , d 2 ] − c 2 det [ B , d 1 ] ) = ± det [ B , c 1 d 2 − c 2 d 1 ] . \begin{aligned}
f=\pm(c_1\det[B,d_2]-c_2\det[B,d_1]) \\
&=\pm\det[B,c_1d_2-c_2d_1].
\end{aligned} f = ± ( c 1 det [ B , d 2 ] − c 2 det [ B , d 1 ]) = ± det [ B , c 1 d 2 − c 2 d 1 ] .
This is again a homogeneous determinant of size n n n . Transposing handles s = n − 1 s=n-1 s = n − 1 . These four cases include all possibilities, also when n = 3 n=3 n = 3 .
Corollary 7.2. For n ≥ 3 n \ge3 n ≥ 3 , every degree-n n n point of the determinant boundary that is concise in its n 2 n^2 n 2 variables satisfies
dc ( f ) = n , dc ( f ) ≥ n + 2. \operatorname{dc}(f)=n, \qquad\operatorname{dc}(f)\ge n+2. dc ( f ) = n , dc ( f ) ≥ n + 2. Its orbit closure need not itself be a boundary component.
Proof. The degree gives dc ( f ) ≥ n \operatorname{dc}(f) \ge n dc ( f ) ≥ n , and membership in the determinant orbit closure gives the reverse inequality. Suppose that f f f had an exact representation of size at most n + 1 n+1 n + 1 . The theorem would give a homogeneous size-n n n representation. Conciseness makes its linear map from C n 2 \mathbb{C}^{n^2} C n 2 to Mat n \operatorname{Mat}_n Mat n injective, hence an isomorphism. The polynomial would then be an invertible linear change of variables of det n \det_n det n , putting it in the orbit itself. That contradicts its being a boundary point. □ \square □
Bounds for padded odd-skew and block forms
The theorem applies to the padded odd-skew forms Q r , k Q_{r,k} Q r , k and the block forms H h , k H_{h,k} H h , k . For these families, explicit affine matrices also give the following upper bounds.
For r = 3 r=3 r = 3 , the Pfaffian vector is linear. Write it as x ∈ C 3 x \in\mathbb{C}^3 x ∈ C 3 . Then
Q 3 , k = det ( 0 x T 0 0 0 I 3 S U x 0 I 3 0 0 0 V Z ) , block sizes 1 , 3 , 3 , k . Q_{3,k}=\det\begin{pmatrix}
0 & x^{T} & 0 & 0\\
0 & I_3 & S & U\\
x & 0 & I_3 & 0\\
0 & 0 & V & Z
\end{pmatrix},
\qquad\text{block sizes }1,3,3,k. Q 3 , k = det 0 0 x 0 x T I 3 0 0 0 S I 3 V 0 U 0 Z , block sizes 1 , 3 , 3 , k .
The two identity blocks allow ordinary row and column elimination; the remaining block determinant is det Z x T S x − x T U adj ( Z ) V x \det Z\,x^{T}Sx-x^{T}U\operatorname{adj}(Z)Vx det Z x T S x − x T U adj ( Z ) V x , as required. The total matrix size is k + 7 = n + 4 k+7=n+4 k + 7 = n + 4 . Therefore
n + 2 ≤ dc ( Q 3 , n − 3 ) ≤ n + 4. n+2 \le\operatorname{dc}(Q_{3,n-3}) \le n+4. n + 2 ≤ dc ( Q 3 , n − 3 ) ≤ n + 4.
The same construction works for the block family because its kernel matrices K L K_L K L and K R K_R K R are linear; these matrices are defined in § 10 . With r = 3 h r=3h r = 3 h , it gives
H h , k = det ( 0 K L 0 0 0 I r S U K R 0 I r 0 0 0 V Z ) , block sizes h , r , r , k . H_{h,k}=\det\begin{pmatrix}
0 & K_L & 0 & 0\\
0 & I_r & S & U\\
K_R & 0 & I_r & 0\\
0 & 0 & V & Z
\end{pmatrix},
\qquad\text{block sizes }h,r,r,k. H h , k = det 0 0 K R 0 K L I r 0 0 0 S I r V 0 U 0 Z , block sizes h , r , r , k .
After eliminating the identity blocks, the two off-diagonal signs can be removed simultaneously. The result is exactly (22) . The matrix size is h + 2 r + k = n + 4 h h+2r+k=n+4h h + 2 r + k = n + 4 h , so
n + 2 ≤ dc ( H h , k ) ≤ n + 4 h . n+2 \le\operatorname{dc}(H_{h,k}) \le n+4h. n + 2 ≤ dc ( H h , k ) ≤ n + 4 h .
These intervals do not determine the exact complexities. For odd r > 3 r>3 r > 3 , the entries of the Pfaffian vector have degree greater than one. Substituting them into the first block matrix would no longer give an affine representation, so it would not prove an n + 4 n+4 n + 4 upper bound.
The dimension threshold. The threshold n 2 − n + 2 n^2-n+2 n 2 − n + 2 comes from the matrix-space classification. We do not claim that it is optimal for determinantal complexity. For a hypothetical representation of size n + 2 n+2 n + 2 , the same classification would require at least n 2 + n + 2 n^2+n+2 n 2 + n + 2 essential variables. The boundary polynomials have n 2 n^2 n 2 variables, so this argument does not exclude exact size n + 2 n+2 n + 2 .
Regularity of the smallest odd-skew form
For the smallest odd-skew form, regularity at the origin also follows from an elementary argument about linear spaces in its zero set.
Proposition 7.3. Every affine determinantal representation of the nine-variable cubic Q 3 , 0 = x T S x Q_{3,0}=x^{T}Sx Q 3 , 0 = x T S x has constant matrix of corank one.
Proof. First observe that Q 3 , 0 = 0 Q_{3,0}=0 Q 3 , 0 = 0 contains no linear subspace of codimension at most two. To see this, let a linear zero subspace project to a t t t -dimensional space T T T in the three x x x -coordinates. Its kernel in the six-dimensional space Sym 3 \operatorname{Sym}_3 Sym 3 must consist of quadratic forms vanishing on T T T . Restriction to T T T imposes t ( t + 1 ) / 2 t(t+1)/2 t ( t + 1 ) /2 independent conditions, so that kernel has dimension at most 6 − t ( t + 1 ) / 2 6-t(t+1)/2 6 − t ( t + 1 ) /2 . The whole subspace consequently has dimension at most
t + 6 − t ( t + 1 ) 2 ≤ 6. t+6-\frac{t(t+1)}{2}\le6. t + 6 − 2 t ( t + 1 ) ≤ 6.
A subspace of codimension at most two in nine variables would have dimension at least seven, which is impossible.
Now make constant row and column changes so that the constant matrix of a representation has an identity block and a zero block of size a a a . Since the polynomial vanishes to order three at the origin, 1 ≤ a ≤ 3 1\le a\le3 1 ≤ a ≤ 3 .
If a = 3 a=3 a = 3 , the degree-three term is the determinant of the corresponding 3 × 3 3\times3 3 × 3 linear block. Conciseness in nine variables would make this an invertible linear change of variables of det 3 \det_3 det 3 . But Q 3 , 0 Q_{3,0} Q 3 , 0 lies on the boundary and is not in that orbit, a contradiction.
If a = 2 a=2 a = 2 , the 2 × 2 2\times2 2 × 2 linear block is identically singular: otherwise its determinant would give a nonzero quadratic term. A linear space of singular 2 × 2 2\times2 2 × 2 matrices has a common kernel or a common cokernel. Indeed, two rank-one matrices with independent column factors and independent row factors have an invertible sum; excluding that possibility throughout the space forces one of the two common factors. After transposing if necessary, the block has the form
( ℓ 1 ℓ 2 0 0 ) . \begin{pmatrix}
\ell_1 & \ell_2\\
0 & 0
\end{pmatrix}. ( ℓ 1 0 ℓ 2 0 ) .
Eliminate the invertible constant block by a Schur complement. The cubic term must then lie in the ideal ( ℓ 1 , ℓ 2 ) (\ell_1,\ell_2) ( ℓ 1 , ℓ 2 ) , so it vanishes on the linear subspace ℓ 1 = ℓ 2 = 0 \ell_1=\ell_2=0 ℓ 1 = ℓ 2 = 0 . This subspace has codimension at most two, contrary to the first paragraph.
Both a = 2 a=2 a = 2 and a = 3 a=3 a = 3 have been excluded. Hence a = 1 a=1 a = 1 . □
The factoriality argument in § 5 strengthens this conclusion: every representation has corank one at {ital}every{/ital} zero, and its cokernel is globally free of rank one. The exact complexity of Q 3 , 0 Q_{3,0} Q 3 , 0 remains undetermined; the bounds above leave the values 5, 6, and 7.
Odd-skew boundary components
We prove that every odd-skew form remains a boundary component after arbitrary padding, and that distinct skew sizes give distinct components at a fixed final size. The degeneration follows from a determinant coefficient identity. The component and distinctness statements follow from the singular locus and the dimension of the linear stabilizer.
Let r = 2 m + 1 ≥ 3 r=2m+1\ge3 r = 2 m + 1 ≥ 3 , let k ≥ 0 k\ge0 k ≥ 0 , and put n = r + k n=r+k n = r + k . The padded form uses the independent entries of the following five blocks:
A ∈ Skew r , S ∈ Sym r , U ∈ Mat r , k , V ∈ Mat k , r , Z ∈ Mat k . A\in\operatorname{Skew}_r,\quad S\in\operatorname{Sym}_r,\quad U\in\operatorname{Mat}_{r,k},\quad V\in\operatorname{Mat}_{k,r},\quad Z\in\operatorname{Mat}_k. A ∈ Skew r , S ∈ Sym r , U ∈ Mat r , k , V ∈ Mat k , r , Z ∈ Mat k .
Here A A A is skew-symmetric, S S S is symmetric, and U , V , Z U,V,Z U , V , Z are unrestricted. The skew and symmetric entries together give r 2 r^2 r 2 variables. Including the padding blocks, there are r 2 + 2 r k + k 2 = n 2 r^2+2rk+k^2=n^2 r 2 + 2 r k + k 2 = n 2 variables. The signed maximal Pfaffians of A A A give a polynomial kernel vector. For an even-size skew matrix, the Pfaffian is a polynomial whose square is its determinant. Define
p i ( A ) = ( − 1 ) i + 1 Pf ( A i ^ ) , 1 ≤ i ≤ r . p_i(A)=(-1)^{i+1}\operatorname{Pf}(A_{\hat{i}}),\qquad1\leq i\leq r. p i ( A ) = ( − 1 ) i + 1 Pf ( A i ^ ) , 1 ≤ i ≤ r .
Write p ( A ) p(A) p ( A ) for the column of these r r r polynomials. Two standard Pfaffian identities give both a kernel vector and a factorization of the adjugate:
A p ( A ) = 0 , adj A = p ( A ) p ( A ) T . (16) Ap(A)=0,\qquad\operatorname{adj}A=p(A)p(A)^{T}. \tag*{(16)} A p ( A ) = 0 , adj A = p ( A ) p ( A ) T . (16)
For k ≥ 1 k\geq1 k ≥ 1 , define the padded odd-skew form by
Q r , k = det Z p ( A ) T S p ( A ) − p ( A ) T U adj Z V p ( A ) . (17) Q_{r,k}=\det Z\,p(A)^{T}Sp(A)-p(A)^{T}U\operatorname{adj}ZVp(A). \tag*{(17)} Q r , k = det Z p ( A ) T S p ( A ) − p ( A ) T U adj Z V p ( A ) . (17)
For k = 0 k=0 k = 0 , set Q r , 0 = P r = p ( A ) T S p ( A ) Q_{r,0}=P_r=p(A)^{T}Sp(A) Q r , 0 = P r = p ( A ) T S p ( A ) and omit the empty blocks. Each p i p_i p i has degree m m m , so Q r , k Q_{r,k} Q r , k is homogeneous of degree n n n . For r = 3 r=3 r = 3 , the entries of p p p are the three skew variables, up to signs. Thus Q 3 , 0 = p T S p Q_{3,0}=p^{T}Sp Q 3 , 0 = p T S p is the cubic consisting of a quadratic form in p p p with independent coefficients.
Write O n = GL n 2 ⋅ det n \mathcal{O}_n=\operatorname{GL}_{n^2}\cdot\det_n O n = GL n 2 ⋅ det n for the affine orbit under invertible linear changes of variables. Its boundary is ∂ O n = O n ‾ ∖ O n \partial\mathcal{O}_n=\overline{\mathcal{O}_n}\setminus\mathcal{O}_n ∂ O n = O n ∖ O n . A form is concise if no invertible linear change of variables can make one variable disappear. Equivalently, its first partial derivatives are linearly independent. Its stabilizer is the group of invertible linear changes of all n 2 n^2 n 2 variables that leave the polynomial unchanged. The orbit–stabilizer dimension formula gives
dim ( GL n 2 ⋅ Q ) = n 4 − dim Stab ( Q ) . \dim(\operatorname{GL}_{n^2}\cdot Q)=n^4-\dim\operatorname{Stab}(Q). dim ( GL n 2 ⋅ Q ) = n 4 − dim Stab ( Q ) .
Since dim Stab ( det n ) = 2 n 2 − 2 \dim\operatorname{Stab}(\det_n)=2n^2-2 dim Stab ( det n ) = 2 n 2 − 2 , a boundary orbit with stabilizer dimension 2 n 2 − 1 2n^2-1 2 n 2 − 1 has codimension one in O n ‾ \overline{\mathcal{O}_n} O n . Its closure is therefore a boundary component.
Theorem 8.1. For every odd r ≥ 3 r\geq3 r ≥ 3 and every k ≥ 0 k\geq0 k ≥ 0 , with n = r + k n=r+k n = r + k , the form Q r , k Q_{r,k} Q r , k is concise and
dim Stab GL n 2 ( Q r , k ) = 2 n 2 − 1 , dim ( GL n 2 ⋅ Q r , k ) = n 4 − 2 n 2 + 1. \dim\operatorname{Stab}_{\operatorname{GL}_{n^2}}(Q_{r,k})=2n^2-1,\qquad\dim(\operatorname{GL}_{n^2}\cdot Q_{r,k})=n^4-2n^2+1. dim Stab GL n 2 ( Q r , k ) = 2 n 2 − 1 , dim ( GL n 2 ⋅ Q r , k ) = n 4 − 2 n 2 + 1. Its orbit closure is an irreducible component of ∂ O n \partial\mathcal{O}_n ∂ O n . For fixed n n n , these components are pairwise distinct as odd r r r ranges from 3 3 3 to n n n .
Corollary 8.2. For every n ≥ 3 n\geq3 n ≥ 3 , the determinant boundary has at least ⌊ ( n − 1 ) / 2 ⌋ \lfloor(n-1)/2\rfloor ⌊( n − 1 ) /2 ⌋ pairwise distinct concise components. Together with the traceless-determinant component, it has at least ⌊ ( n − 1 ) / 2 ⌋ + 1 \lfloor(n-1)/2\rfloor+1 ⌊( n − 1 ) /2 ⌋ + 1 irreducible components.
This is a lower bound on the number of components, not a classification of the whole boundary. The odd-skew theorem includes the following cases:
Table 1.
Pfaffian identities and degeneration
The map A ↦ p ( A ) A \mapsto p(A) A ↦ p ( A ) attains every nonzero vector of C r \mathbb{C}^{r} C r . This permits the kernel vector to be varied freely in the calculations below. Indeed, for x ≠ 0 x \ne0 x = 0 , choose a skew form of rank r − 1 r-1 r − 1 with kernel C x \mathbb{C}x C x . Its Pfaffian vector is a nonzero multiple of x x x . Rescaling A A A rescales p ( A ) p(A) p ( A ) by an m m m -th power, so any nonzero multiple can be obtained.
For later use, put
L x = { A ∈ Skew r : A x = 0 } , x ≠ 0. L_x = \{A \in\operatorname{Skew}_r : Ax = 0\}, \qquad x \ne0. L x = { A ∈ Skew r : A x = 0 } , x = 0.
This is a linear space. A dense open subset has rank r − 1 r-1 r − 1 , since the induced skew form on the quotient by C x \mathbb{C}x C x has even dimension r − 1 r-1 r − 1 . On that open subset, p ( A ) p(A) p ( A ) is a nonzero multiple of x x x . Polynomial identities proved there therefore extend to every A ∈ L x A \in L_x A ∈ L x .
For the degeneration, introduce a parameter t t t and consider
M t = ( A + t S U t V Z ) . M_t = \begin{pmatrix} A+tS & U \\ tV & Z \end{pmatrix}. M t = ( A + tS t V U Z ) .
For t ≠ 0 t \ne0 t = 0 , its entries are an invertible linear parametrization of Mat n \operatorname{Mat}_n Mat n : the upper-left block separates into arbitrary skew and symmetric parts, and the other blocks are independent. On det Z ≠ 0 \det Z \ne0 det Z = 0 ,
det M t = det Z det ( A + t ( S − U Z − 1 V ) ) = t det Z tr ( adj A ( S − U Z − 1 V ) ) + O ( t 2 ) = t Q r , k + O ( t 2 ) . \begin{aligned}
\det M_t = \det Z \det\left(A+t(S-UZ^{-1}V)\right) \\
&= t\det Z\operatorname{tr}\left(\operatorname{adj} A(S-UZ^{-1}V)\right)+O(t^2) \\
&= tQ_{r,k}+O(t^2).
\end{aligned} det M t = det Z det ( A + t ( S − U Z − 1 V ) ) = t det Z tr ( adj A ( S − U Z − 1 V ) ) + O ( t 2 ) = t Q r , k + O ( t 2 ) .
The calculation used Z − 1 Z^{-1} Z − 1 only to simplify the algebra. Both sides of the resulting coefficient identity are polynomials, and the identity holds on the dense open set det Z ≠ 0 \det Z \ne0 det Z = 0 , so it holds everywhere. Multiplying one row by t − 1 t^{-1} t − 1 gives a determinant in O ‾ n \overline{\mathcal{O}}_n O n whose limit is Q r , k Q_{r,k} Q r , k . For k = 0 k=0 k = 0 , use t − 1 det ( A + t S ) t^{-1}\det(A+tS) t − 1 det ( A + tS ) instead. Thus every Q r , k Q_{r,k} Q r , k lies in O ‾ n \overline{\mathcal{O}}_n O n . The singular-locus calculation below excludes membership in O n \mathcal{O}_n O n , and the stabilizer calculation then proves componenthood.
Singular locus and translation invariants
The singular locus of Q = 0 Q=0 Q = 0 is the common zero set of its first partial derivatives. Every linear stabilizer of Q Q Q preserves this locus. For Q r , k Q_{r,k} Q r , k , its unique largest component determines the coefficient directions and recovers the skew size r r r .
Put a = ( r 2 ) a=\binom{r}{2} a = ( 2 r ) , the number of skew variables, and N = n 2 N=n^2 N = n 2 , the total number of variables. Define
Y r = { A ∈ Skew r : rank A ≤ r − 3 } , M r , k = Y r × C N − a . Y_r=\{A\in\operatorname{Skew}_r:\operatorname{rank} A\le r-3\}, \qquad M_{r,k}=Y_r\times\mathbb{C}^{N-a}. Y r = { A ∈ Skew r : rank A ≤ r − 3 } , M r , k = Y r × C N − a .
The set Y r Y_r Y r is the zero set of the maximal Pfaffians. The skew normal form shows that it is irreducible of codimension three in Skew r \operatorname{Skew}_r Skew r . In particular, dim M r , k = N − 3 \dim M_{r,k}=N-3 dim M r , k = N − 3 .
Proposition 8.3. For k = 0 k=0 k = 0 , the singular locus of Q r , 0 Q_{r,0} Q r , 0 is exactly M r , 0 M_{r,0} M r , 0 . For k ≥ 1 k\ge1 k ≥ 1 , it has exactly two irreducible components:
Sing V ( Q r , k ) = M r , k ∪ C r , k , dim M r , k = N − 3 , dim C r , k = N − 4. \operatorname{Sing} V(Q_{r,k})=M_{r,k}\cup C_{r,k}, \qquad\dim M_{r,k}=N-3, \qquad\dim C_{r,k}=N-4. Sing V ( Q r , k ) = M r , k ∪ C r , k , dim M r , k = N − 3 , dim C r , k = N − 4. Thus M r , k M_{r,k} M r , k is always the unique largest singular component.
Proof. If p ( A ) = 0 p(A)=0 p ( A ) = 0 , all first derivatives vanish, since every term of Q r , k Q_{r,k} Q r , k contains two entries of p ( A ) p(A) p ( A ) . When k = 0 k=0 k = 0 , the S i i S_{ii} S ii derivatives are p i ( A ) 2 p_i(A)^2 p i ( A ) 2 , so no other points are singular.
On p ( A ) ≠ 0 p(A)\ne0 p ( A ) = 0 , the form is a determinant in a smaller set of coefficient coordinates. For k ≥ 1 k\ge1 k ≥ 1 , put
W = ( p T S p p T U V p Z ) , Q r , k = det W . W=\begin{pmatrix}p^{T}Sp & p^{T}U\\ Vp & Z\end{pmatrix},\qquad Q_{r,k}=\det W. W = ( p T S p V p p T U Z ) , Q r , k = det W .
Fix A A A with p ( A ) ≠ 0 p(A)\ne0 p ( A ) = 0 . The map ( S , U , V , Z ) ↦ W (S,U,V,Z)\mapsto W ( S , U , V , Z ) ↦ W is a surjective linear map onto Mat k + 1 \operatorname{Mat}_{k+1} Mat k + 1 . Consequently the derivatives in these blocks vanish precisely when rank W ≤ k − 1 \operatorname{rank}W\le k-1 rank W ≤ k − 1 . At such a point the entire differential of det W \det W det W is zero, so the A A A -derivatives vanish too.
These singular points form an irreducible set of dimension N − 4 N-4 N − 4 . Indeed, on a chart p i ( A ) ≠ 0 p_i(A)\ne0 p i ( A ) = 0 use one symmetric entry, one entry in each column of U U U , and one entry in each row of V V V to replace coordinates by the entries of W W W . This is a regular invertible coordinate change on the chart. The fibers are an affine space times the irreducible rank-≤ k − 1 \le k-1 ≤ k − 1 matrix variety. The base is the irreducible open set of rank-r − 1 r-1 r − 1 skew matrices. The dimension is
a + ( N − a − ( k + 1 ) 2 ) + ( ( k + 1 ) 2 − 4 ) = N − 4. a+\left(N-a-(k+1)^2\right)+\left((k+1)^2-4\right)=N-4. a + ( N − a − ( k + 1 ) 2 ) + ( ( k + 1 ) 2 − 4 ) = N − 4.
Its closure is C r , k C_{r,k} C r , k . All other singular points are in M r , k M_{r,k} M r , k . The two varieties do not contain one another: C r , k C_{r,k} C r , k has points outside M r , k M_{r,k} M r , k and has smaller dimension. This proves the claim.
The determinant singular locus has dimension N − 4 N-4 N − 4 , whereas Sing V ( Q r , k ) \operatorname{Sing}V(Q_{r,k}) Sing V ( Q r , k ) contains M r , k M_{r,k} M r , k of dimension N − 3 N-3 N − 3 . Therefore Q r , k Q_{r,k} Q r , k is not linearly equivalent to det n \det_n det n . Together with the degeneration, this proves boundary membership.
The coefficient directions in M r , k M_{r,k} M r , k are translation directions. The skew directions contribute no translations, as the next lemma shows. Define the translation space of a variety X X X by
Trans ( X ) = { v : X + t v = X for every t ∈ C } . \operatorname{Trans}(X)=\{v:X+tv=X\text{ for every }t\in\mathbb{C}\}. Trans ( X ) = { v : X + t v = X for every t ∈ C } .
It is a vector subspace, and its dimension is preserved by linear equivalence.
Lemma 8.4. The translation space of Y r Y_r Y r is zero. Consequently
Trans ( M r , k ) = { 0 } × C N − a , dim Trans ( M r , k ) = n 2 − ( r 2 ) . \operatorname{Trans}(M_{r,k})=\{0\}\times\mathbb{C}^{N-a},\qquad\dim\operatorname{Trans}(M_{r,k})=n^2-\binom{r}{2}. Trans ( M r , k ) = { 0 } × C N − a , dim Trans ( M r , k ) = n 2 − ( 2 r ) .
Proof. For r = 3 r=3 r = 3 , Y r = { 0 } Y_r=\{0\} Y r = { 0 } . Otherwise suppose B ≠ 0 B\ne0 B = 0 is a translation direction. Since 0 ∈ Y r 0\in Y_r 0 ∈ Y r , we have B ∈ Y r B\in Y_r B ∈ Y r . Write rank B = 2 s \operatorname{rank}B=2s rank B = 2 s . After skew congruence, its support is a 2 s 2s 2 s -dimensional block. On the complement choose a skew matrix A A A of maximal rank r − 2 s − 1 r-2s-1 r − 2 s − 1 , and extend it by zero on the support of B B B . Then rank A ≤ r − 3 \operatorname{rank}A\le r-3 rank A ≤ r − 3 , but
rank ( A + B ) = ( r − 2 s − 1 ) + 2 s = r − 1. \operatorname{rank}(A+B)=(r-2s-1)+2s=r-1. rank ( A + B ) = ( r − 2 s − 1 ) + 2 s = r − 1.
This contradicts translation invariance. The product assertion follows.
These facts already prove pairwise inequivalence of the polynomials for fixed n n n . A linear equivalence must preserve the unique largest singular component and the dimension of its translation space. That dimension n 2 − ( r 2 ) n^2-\binom{r}{2} n 2 − ( 2 r ) determines r r r .
They also simplify the stabilizer argument. Any linear stabilizer preserves M r , k M_{r,k} M r , k , hence its translation space, which is exactly the space of all ( S , U , V , Z ) (S,U,V,Z) ( S , U , V , Z ) variables. Therefore every infinitesimal stabilizer has
δ A = T ( A ) . (18) \delta A=T(A). \tag*{(18)} δ A = T ( A ) . (18)
Here an infinitesimal stabilizer is a linear vector field δ \delta δ satisfying δ Q = 0 \delta Q = 0 δ Q = 0 . Equation (18) therefore follows from the polynomial’s singular locus, without assuming that the stabilizer preserves the original block decomposition.
Linear Pfaffian syzygies
The stabilizer calculation uses two spaces of linear polynomial relations, or syzygies . The first consists of vector-valued changes annihilated by p T p^{T} p T . The second consists of symmetric matrix-valued changes annihilated by B ↦ p T B p B \mapsto p^{T}Bp B ↦ p T B p .
Lemma 8.5. Let u : Skew r → C r u : \operatorname{Skew}_{r} \to\mathbb{C}^{r} u : Skew r → C r and F : Skew r → Sym r F : \operatorname{Skew}_{r} \to\operatorname{Sym}_{r} F : Skew r → Sym r be linear. Then
p ( A ) T u ( A ) = 0 for all A ⟺ u ( A ) = A b for a constant b , p(A)^{T}u(A) = 0 \text{ for all } A \quad\Longleftrightarrow\quad u(A) = Ab \text{ for a constant } b, p ( A ) T u ( A ) = 0 for all A ⟺ u ( A ) = A b for a constant b , p ( A ) T F ( A ) p ( A ) = 0 for all A ⟺ F ( A ) = A B − B T A for a constant B . p(A)^{T}F(A)p(A) = 0 \text{ for all } A \quad\Longleftrightarrow\quad F(A) = AB - B^{T}A \text{ for a constant } B. p ( A ) T F ( A ) p ( A ) = 0 for all A ⟺ F ( A ) = A B − B T A for a constant B . The two spaces have dimensions r r r and r 2 − 1 r^{2} - 1 r 2 − 1 , respectively.
Proof. Fix x ≠ 0 x \ne0 x = 0 . On the dense rank-r − 1 r-1 r − 1 subset of L x L_x L x , the hypotheses imply respectively x T u ( A ) = 0 x^{T}u(A) = 0 x T u ( A ) = 0 and x T F ( A ) x = 0 x^{T}F(A)x = 0 x T F ( A ) x = 0 . They extend to all of L x L_x L x . For fixed x x x , either expression is a linear functional of the skew matrix A A A . The standard nondegenerate pairing identifies such a functional with a two-form, which we call K ( x ) K(x) K ( x ) . Its polynomial degree in x x x is one or two, respectively. The annihilator of L x L_x L x is x ∧ C r x \wedge\mathbb{C}^{r} x ∧ C r , so x ∧ K ( x ) = 0 x \wedge K(x) = 0 x ∧ K ( x ) = 0 .
The identity x ∧ K ( x ) = 0 x \wedge K(x) = 0 x ∧ K ( x ) = 0 gives a polynomial factorization with a factor x x x . The following contraction identity produces the other factor with the required degree. On polynomial-valued exterior forms define
D = ∑ i = 1 r ∂ x i ι e i ∗ . D = \sum_{i=1}^{r} \partial_{x_i}\iota_{e_i^{*}}. D = i = 1 ∑ r ∂ x i ι e i ∗ . For an exterior q q q -form ω \omega ω whose coefficients have degree d d d , the contraction rule and Euler’s identity give
D ( x ∧ ω ) + x ∧ D ω = ( r + d − q ) ω . D(x \wedge\omega) + x \wedge D\omega= (r + d - q)\omega. D ( x ∧ ω ) + x ∧ D ω = ( r + d − q ) ω . Apply this with q = 2 q = 2 q = 2 to obtain
K ( x ) = x ∧ D K ( x ) r + d − 2 . K(x) = x \wedge\frac{DK(x)}{r + d - 2}. K ( x ) = x ∧ r + d − 2 D K ( x ) . For d = 1 d = 1 d = 1 the second factor is a constant vector. For d = 2 d = 2 d = 2 it is B x Bx B x for a constant matrix. Translating the pairing back, and absorbing a nonzero scalar into the parameters, gives the displayed formulas. Their converse follows from A p ( A ) = 0 Ap(A) = 0 A p ( A ) = 0 .
The vector parameter has zero kernel. The kernel of B ↦ [ A ↦ A B − B T A ] B \mapsto[A \mapsto AB - B^{T}A] B ↦ [ A ↦ A B − B T A ] consists exactly of scalar matrices: diagonal entries first force all off-diagonal entries of B B B to vanish, and off-diagonal entries then force its diagonal entries to agree. This proves the dimensions.
The unpadded stabilizer
The unpadded stabilizer is one of the quotients in the padded stabilizer calculation. We prove the required upper bound here; boundary membership will give the reverse inequality.
Proposition 8.6. The Lie algebra of the full linear stabilizer of P r ( A , S ) = p ( A ) T S p ( A ) P_r(A,S) = p(A)^{T}Sp(A) P r ( A , S ) = p ( A ) T S p ( A ) has dimension at most $2r^{2} - 1.
Proof. By (18) , a stabilizer vector field has δ A = T ( A ) \delta A=T(A) δ A = T ( A ) and δ S = F ( A ) + R ( S ) \delta S=F(A)+R(S) δ S = F ( A ) + R ( S ) . The change F ( A ) F(A) F ( A ) replaces one S S S by an A A A , so its contribution has one more A A A -degree than the contributions from T ( A ) T(A) T ( A ) and R ( S ) R(S) R ( S ) . Terms of different degrees cannot cancel. Lemma 8.5 therefore treats F F F separately and gives at most r 2 − 1 r^2-1 r 2 − 1 parameters. For the other part put h ( A ) = d p A [ T ( A ) ] h(A)=d p_A[T(A)] h ( A ) = d p A [ T ( A )] . Its equation is
2 p T S h + p T R ( S ) p = 0. 2p^TSh+p^TR(S)p=0. 2 p T S h + p T R ( S ) p = 0.
Let R ∗ R^* R ∗ be the transpose map for the trace pairing on symmetric matrices. Then
R ∗ ( p p T ) = − ( p h T + h p T ) . (19) R^*(pp^T)=-(ph^T+hp^T). \tag*{(19)} R ∗ ( p p T ) = − ( p h T + h p T ) . (19)
The diagonal entry ( R ∗ ( x x T ) ) i i (R^*(xx^T))_{ii} ( R ∗ ( x x T ) ) ii is a quadratic polynomial in x x x . It vanishes when x i = 0 x_i=0 x i = 0 , since every nonzero x x x can be realized as p ( A ) p(A) p ( A ) . Thus it is divisible by x i x_i x i , and we can write
( R ∗ ( x x T ) ) i i = − 2 x i ( C x ) i . (R^*(xx^T))_{ii}=-2x_i(Cx)_i. ( R ∗ ( x x T ) ) ii = − 2 x i ( C x ) i .
for one constant matrix C C C . Substitute x = p ( A ) x=p(A) x = p ( A ) in (19) and cancel the nonzero polynomial p i p_i p i to obtain h ( A ) = C p ( A ) h(A)=Cp(A) h ( A ) = C p ( A ) . The matrix identity determines R R R from C C C , since rank-one symmetric matrices span Sym r \operatorname{Sym}_r Sym r . Projection ( T , R ) ↦ C (T,R)\mapsto C ( T , R ) ↦ C therefore has image dimension at most r 2 r^2 r 2 .
It remains to prove that the projection ( T , R ) ↦ C (T,R)\mapsto C ( T , R ) ↦ C is injective. Suppose C = 0 C=0 C = 0 . Then h = 0 h=0 h = 0 and R = 0 R=0 R = 0 . Differentiating A p ( A ) = 0 Ap(A)=0 A p ( A ) = 0 gives T ( A ) p ( A ) = 0 T(A)p(A)=0 T ( A ) p ( A ) = 0 . For fixed x ≠ 0 x\ne0 x = 0 , density in L x L_x L x implies T ( A ) x = 0 T(A)x=0 T ( A ) x = 0 whenever A x = 0 Ax=0 A x = 0 . In particular, if A A A has rank two, the skew form T ( A ) T(A) T ( A ) kills ker A \ker A ker A . Skew forms on the two-dimensional quotient by that kernel form a one-dimensional space, so T ( A ) T(A) T ( A ) is a scalar multiple of A A A . Every decomposable two-form is therefore an eigenvector of T T T .
Elementary wedges form a basis. Sums of two elementary wedges sharing a vector are decomposable, and force the corresponding eigenvalues to agree. The graph of pairs joined when they share a vertex is connected for r ≥ 3 r\ge3 r ≥ 3 . Hence T = c Id T=c\operatorname{Id} T = c Id . Euler’s identity gives h ( A ) = d p A [ c A ] = c m p ( A ) h(A)=d p_A[cA]=cm\,p(A) h ( A ) = d p A [ c A ] = c m p ( A ) , so h = 0 h=0 h = 0 and m ≥ 1 m\ge1 m ≥ 1 force c = 0 c=0 c = 0 . The projection is injective. Adding the F F F -space proves the bound r 2 + ( r 2 − 1 ) = 2 r 2 − 1 r^2+(r^2-1)=2r^2-1 r 2 + ( r 2 − 1 ) = 2 r 2 − 1 .
Infinitesimal determinant preservers
The linear changes of a generic matrix that preserve its determinant up to scalar are infinitesimal row and column operations. We recall an elementary proof for use in the padding calculation.
Lemma 8.7. For k ≥ 1 k\ge1 k ≥ 1 , the space of linear maps J : Mat k → Mat k J:\operatorname{Mat}_k\to\operatorname{Mat}_k J : Mat k → Mat k satisfying D ( det ) Z [ J ( Z ) ] = c det Z D(\det)_Z[J(Z)]=c\det Z D ( det ) Z [ J ( Z )] = c det Z for some constant c c c has dimension 2 k 2 − 1 2k^2-1 2 k 2 − 1 . Every such map is J ( Z ) = L Z + Z R J(Z)=LZ+ZR J ( Z ) = L Z + Z R , with c = tr L + tr R c=\operatorname{tr}L+\operatorname{tr}R c = tr L + tr R . The subspace with c = 0 c=0 c = 0 has dimension 2 k 2 − 2 2k^2-2 2 k 2 − 2 .
Proof. The case k = 1 k=1 k = 1 is immediate. Write J i j ( Z ) = ∑ a , b c i j a b z a b J_{ij}(Z)=\sum_{a,b}c_{ij}^{ab}z_{ab} J ij ( Z ) = ∑ a , b c ij ab z ab and expand ∑ i , j J i j ( Z ) cof i j ( Z ) \sum_{i,j}J_{ij}(Z)\operatorname{cof}_{ij}(Z) ∑ i , j J ij ( Z ) cof ij ( Z ) . If a ≠ i a\ne i a = i and b ≠ j b\ne j b = j , choose a cofactor monomial containing z a b z_{ab} z ab . Multiplication by z a b z_{ab} z ab gives a square, with row i i i missing, row a a a doubled, column j j j missing and column b b b doubled. These degrees uniquely identify its coefficient, which must vanish. Thus J i j J_{ij} J ij uses only entries in row i i i or column j j j .
Monomials with column j j j missing and column b b b doubled have exactly two possible cofactor contributions, with opposite signs. They force the coefficient of z i b z_{ib} z ib in J i j J_{ij} J ij to be independent of i i i . The row version gives the corresponding independence of j j j . These supply the off-diagonal parts of L Z + Z R LZ+ZR L Z + Z R . After subtracting them, write J i j = d i j z i j J_{ij}=d_{ij}z_{ij} J ij = d ij z ij . Permutation monomials give ∑ i d i , σ ( i ) = c \sum_i d_{i,\sigma(i)}=c ∑ i d i , σ ( i ) = c for every σ \sigma σ . Permutations differing on two rows imply the rectangle relations, hence d i j = a i + b j d_{ij}=a_i+b_j d ij = a i + b j . These supply the diagonal parts. The kernel of ( L , R ) ↦ J (L,R)\mapsto J ( L , R ) ↦ J is ( a I , − a I ) (aI,-aI) ( a I , − a I ) , proving dimension 2 k 2 − 1 2k^2-1 2 k 2 − 1 . The trace functional is nonzero, so c = 0 c=0 c = 0 removes one dimension.
We also use linear independence of the entries of adj Z \operatorname{adj} Z adj Z . For k ≥ 2 k \ge2 k ≥ 2 their distinct missing row and column degrees prove this. For k = 1 k = 1 k = 1 , the sole entry is 1. Thus an identity linear in adj Z \operatorname{adj} Z adj Z can be tested by replacing adj Z \operatorname{adj} Z adj Z with an arbitrary constant matrix.
The stabilizer of a padded odd-skew form
For k ≥ 1 k \ge1 k ≥ 1 , we filter the stabilizer by its action on Z Z Z , then by its induced action on the unpadded form, and then by the A A A -dependent parts of U , V U,V U , V . The residual action involves only the coefficient variables. The dimensions of these successive images and kernels add to the full stabilizer dimension.
Let δ \delta δ be an arbitrary linear vector field annihilating Q r , k Q_{r,k} Q r , k . Put Y = ( S , U , V , Z ) Y=(S,U,V,Z) Y = ( S , U , V , Z ) . Equation (18) gives δ A = T ( A ) \delta A=T(A) δ A = T ( A ) . Terms from δ Y ( A ) \delta Y(A) δ Y ( A ) have degree 2 m + 1 2m+1 2 m + 1 in A A A ; terms from δ A ( A ) \delta A(A) δ A ( A ) and δ Y ( Y ) \delta Y(Y) δ Y ( Y ) have degree 2 m 2m 2 m . These groups satisfy separate identities.
The action on the padding block
Throughout this calculation, δ Z ( U ) \delta Z(U) δ Z ( U ) means the part of δ Z \delta Z δ Z that is linear in the variables U U U ; the same notation is used for every pair of blocks. Comparing the different degrees in the independent blocks lets us isolate these parts.
The degree-2 m + 1 2m+1 2 m + 1 terms containing S S S force δ Z ( A ) = 0 \delta Z(A)=0 δ Z ( A ) = 0 , by cofactor independence. Terms quadratic in S S S force δ Z ( S ) = 0 \delta Z(S)=0 δ Z ( S ) = 0 . The terms containing one S S S and one U U U give
( p T S p ) tr ( adj Z δ Z ( U ) ) = p T U adj Z δ V ( S ) p . (20) (p^{T}Sp)\operatorname{tr}(\operatorname{adj} Z\,\delta Z(U))=p^{T}U\operatorname{adj} Z\,\delta V(S)p. \tag*{(20)} ( p T S p ) tr ( adj Z δ Z ( U )) = p T U adj Z δ V ( S ) p . (20)
Fix x = p ( A ) ≠ 0 x=p(A)\ne0 x = p ( A ) = 0 and restrict to U U U with x T U = 0 x^{T}U=0 x T U = 0 . Choosing S S S with x T S x ≠ 0 x^{T}Sx\ne0 x T S x = 0 shows that the linear functional U ↦ tr ( adj Z δ Z ( U ) ) U\mapsto\operatorname{tr}(\operatorname{adj} Z\,\delta Z(U)) U ↦ tr ( adj Z δ Z ( U )) vanishes on this kernel for every x x x . Every elementary matrix U U U belongs to one of these kernels, by choosing x x x orthogonal to its nonzero column. Hence the functional is zero. Cofactor independence gives δ Z ( U ) = 0 \delta Z(U)=0 δ Z ( U ) = 0 . Then (20) , using arbitrary U U U and invertible Z Z Z , forces δ V ( S ) = 0 \delta V(S)=0 δ V ( S ) = 0 . The S , V S,V S , V terms similarly give δ Z ( V ) = 0 \delta Z(V)=0 δ Z ( V ) = 0 and δ U ( S ) = 0 \delta U(S)=0 δ U ( S ) = 0 .
Thus δ Z \delta Z δ Z depends only on Z Z Z . The terms involving S S S and no U , V U,V U , V give
det Z ( 2 p T S d p A [ T ( A ) ] + p T δ S ( S ) p ) + D ( det ) Z [ δ Z ( Z ) ] p T S p = 0. \det Z\left(2p^{T}S\,\mathrm{d}p_{A}[T(A)]+p^{T}\delta S(S)p\right)+D(\det)_{Z}[\delta Z(Z)]p^{T}Sp=0. det Z ( 2 p T S d p A [ T ( A )] + p T δ S ( S ) p ) + D ( det ) Z [ δ Z ( Z )] p T S p = 0.
The first bracket depends only on A , S A,S A , S . Choosing them with p T S p ≠ 0 p^{T}Sp\ne0 p T S p = 0 shows that D ( det ) Z [ δ Z ( Z ) ] = c det Z D(\det)_{Z}[\delta Z(Z)]=c\det Z D ( det ) Z [ δ Z ( Z )] = c det Z for a constant c c c . By Lemma 8.7 , the image of the Z Z Z -projection has dimension at most 2 k 2 − 1 2k^{2}-1 2 k 2 − 1 .
Projection to the unpadded stabilizer
Take the kernel of this projection, so δ Z = 0 \delta Z=0 δ Z = 0 . The induced changes
δ A = T ( A ) , δ S = δ S ( A ) + δ S ( S ) \delta A=T(A),\qquad\delta S=\delta S(A)+\delta S(S) δ A = T ( A ) , δ S = δ S ( A ) + δ S ( S )
annihilate P r P_r P r . Indeed, the degree-2 m 2m 2 m equation above and the separate degree-2 m + 1 2m+1 2 m + 1 equation p T δ S ( A ) p = 0 \ p^{T}\delta S(A)p=0 p T δ S ( A ) p = 0 are precisely its stabilizer equations. This second projection has image dimension at most 2 r 2 − 1 2r^{2}-1 2 r 2 − 1 by Proposition 8.6 .
In the next kernel,
δ A = δ Z = 0 , δ S ( A ) = δ S ( S ) = 0 , δ U ( S ) = δ V ( S ) = 0. \delta A=\delta Z=0,\qquad\delta S(A)=\delta S(S)=0,\qquad\delta U(S)=\delta V(S)=0. δ A = δ Z = 0 , δ S ( A ) = δ S ( S ) = 0 , δ U ( S ) = δ V ( S ) = 0.
The parts δ U ( A ) , δ V ( A ) \delta U(A),\delta V(A) δ U ( A ) , δ V ( A ) satisfy p T δ U ( A ) = 0 p^{T}\delta U(A)=0 p T δ U ( A ) = 0 and δ V ( A ) p = 0 \delta V(A)p=0 δ V ( A ) p = 0 . The vector syzygy lemma gives r r r dimensions per column or row, for at most 2 r k 2rk 2 r k dimensions in total.
The residual coefficient action
In the kernel of the preceding projections, all A A A -dependent parts vanish. Since p ( A ) p(A) p ( A ) attains every nonzero vector, the remaining identity in U , V , Z U,V,Z U , V , Z is
det Z δ S = sym ( δ U adj Z V + U adj Z δ V ) , sym B = ( B + B T ) / 2. (21) \det Z\,\delta S = \operatorname{sym}(\delta U\operatorname{adj} ZV + U\operatorname{adj} Z\delta V), \qquad\operatorname{sym} B = (B+B^{T})/2. \tag*{(21)} det Z δ S = sym ( δ U adj Z V + U adj Z δ V ) , sym B = ( B + B T ) /2. (21)
Terms involving Z Z Z alone give δ S ( Z ) = 0 \delta S(Z)=0 δ S ( Z ) = 0 . Terms linear in V V V give
det Z δ S ( V ) = sym ( δ U ( Z ) adj Z V ) . \det Z\,\delta S(V) = \operatorname{sym}(\delta U(Z)\operatorname{adj} ZV). det Z δ S ( V ) = sym ( δ U ( Z ) adj Z V ) .
On det Z ≠ 0 \det Z\ne0 det Z = 0 , let F ( Z ) = δ U ( Z ) Z − 1 F(Z)=\delta U(Z)Z^{-1} F ( Z ) = δ U ( Z ) Z − 1 . Then δ S ( V ) = sym ( F ( Z ) V ) \delta S(V)=\operatorname{sym}(F(Z)V) δ S ( V ) = sym ( F ( Z ) V ) . The map F ↦ [ V ↦ sym ( F V ) ] F\mapsto[V\mapsto\operatorname{sym}(FV)] F ↦ [ V ↦ sym ( F V )] is injective: take V V V with one nonzero column and inspect the corresponding diagonal entry. The left side does not depend on Z Z Z , so F ( Z ) = B F(Z)=B F ( Z ) = B is constant. Hence
δ U ( Z ) = B Z , δ S ( V ) = sym ( B V ) , \delta U(Z)=BZ,\qquad\delta S(V)=\operatorname{sym}(BV), δ U ( Z ) = B Z , δ S ( V ) = sym ( B V ) ,
with at most r k rk r k parameters. Similarly,
δ V ( Z ) = Z C , δ S ( U ) = sym ( U C ) , \delta V(Z)=ZC,\qquad\delta S(U)=\operatorname{sym}(UC), δ V ( Z ) = Z C , δ S ( U ) = sym ( U C ) ,
with at most another r k rk r k parameters.
Terms quadratic in V V V give x T δ U ( V ) adj Z V x = 0 x^{T}\delta U(V)\operatorname{adj} ZVx=0 x T δ U ( V ) adj Z V x = 0 for all x x x . Cofactor independence implies
( x T δ U ( V ) ) a ( V x ) b = 0 for every a , b . (x^{T}\delta U(V))_{a}(Vx)_{b}=0\qquad\text{for every }a,b. ( x T δ U ( V ) ) a ( V x ) b = 0 for every a , b .
The polynomial ring is a domain, so δ U ( V ) = 0 \delta U(V)=0 δ U ( V ) = 0 . Similarly δ V ( U ) = 0 \delta V(U)=0 δ V ( U ) = 0 .
Write P ( U ) = δ U ( U ) P(U)=\delta U(U) P ( U ) = δ U ( U ) and R ( V ) = δ V ( V ) R(V)=\delta V(V) R ( V ) = δ V ( V ) . The remaining equation, for every constant T ∈ Mat k T\in\operatorname{Mat}_{k} T ∈ Mat k , is
( x T P ( U ) ) T ( V x ) + ( x T U ) T ( R ( V ) x ) = 0. (x^{T}P(U))T(Vx)+(x^{T}U)T(R(V)x)=0. ( x T P ( U )) T ( V x ) + ( x T U ) T ( R ( V ) x ) = 0.
It forces preservation of the kernels of U ↦ x T U U\mapsto x^{T}U U ↦ x T U and V ↦ V x V\mapsto Vx V ↦ V x . Taking x = e i x=e_i x = e i shows that each output row of P P P depends only on the corresponding input row. Taking x = e i + e j x=e_i+e_j x = e i + e j makes these row maps agree. Thus P ( U ) = U K P(U)=UK P ( U ) = U K and similarly R ( V ) = J V R(V)=JV R ( V ) = J V for constant matrices K , J K,J K , J . The equation becomes K T + T J = 0 KT+TJ=0 K T + T J = 0 for every T T T . Therefore K = λ I K=\lambda I K = λ I and J = − λ I J=-\lambda I J = − λ I , leaving one dimension.
Stabilizer dimension and componenthood
Adding the dimensions of the four successive images and kernels gives
dim LieStab ( Q r , k ) ≤ ( 2 k 2 − 1 ) + ( 2 r 2 − 1 ) + 2 r k + ( 2 r k + 1 ) = 2 ( r + k ) 2 − 1 = 2 n 2 − 1. \begin{aligned}
\dim\operatorname{LieStab}(Q_{r,k}) &\le(2k^{2}-1)+(2r^{2}-1)+2rk+(2rk+1)\\
&=2(r+k)^{2}-1=2n^{2}-1.
\end{aligned} dim LieStab ( Q r , k ) ≤ ( 2 k 2 − 1 ) + ( 2 r 2 − 1 ) + 2 r k + ( 2 r k + 1 ) = 2 ( r + k ) 2 − 1 = 2 n 2 − 1.
For k = 0 k=0 k = 0 , Proposition 8.6 gives the same bound.
By Lemma 8.7 , the determinant orbit has dimension n 4 − ( 2 n 2 − 2 ) n^{4}-(2n^{2}-2) n 4 − ( 2 n 2 − 2 ) . An orbit is open in its closure, so its boundary has dimension at most one less. We already proved that Q r , k Q_{r,k} Q r , k is in this boundary. In characteristic zero the stabilizer dimension equals its Lie-algebra dimension. Boundary membership gives the reverse bound dim Stab ( Q r , k ) ≥ 2 n 2 − 1 \dim\operatorname{Stab}(Q_{r,k})\ge2n^{2}-1 dim Stab ( Q r , k ) ≥ 2 n 2 − 1 . Equality follows, and the orbit dimension is n 4 − 2 n 2 + 1 n^{4}-2n^{2}+1 n 4 − 2 n 2 + 1 . Its irreducible closure must be a boundary component: any component containing it has at most that dimension and hence must equal it.
Conciseness and distinctness
We complete the proof by establishing conciseness and distinctness of the orbit closures.
A polynomial loses a variable after a linear coordinate change exactly when some nonzero constant directional derivative annihilates it. Let such a derivative annihilate Q r , k Q_{r,k} Q r , k . Its A A A -part has degree 2 m − 1 2m-1 2 m − 1 in A A A , while all other parts have degree 2 m 2m 2 m . The coefficient involving the arbitrary S S S forces d p A [ B ] = 0 d p_A[B]=0 d p A [ B ] = 0 for its constant skew direction B B B . Differentiating A p ( A ) = 0 A p(A)=0 A p ( A ) = 0 gives B p ( A ) = 0 B p(A)=0 B p ( A ) = 0 ; surjectivity of p p p gives B = 0 B=0 B = 0 . For k ≥ 1 k \ge1 k ≥ 1 , the S S S coefficient then forces the constant Z Z Z direction to be zero by cofactor independence. The V V V and U U U coefficients force the constant U , V U,V U , V directions to vanish. Finally, the symmetric direction C C C in S S S satisfies p T C p = 0 p^{T} C p=0 p T C p = 0 for every p p p , so C = 0 C=0 C = 0 . For k = 0 k=0 k = 0 only the first and last steps are needed. This proves conciseness.
Section 8.2 proved that different odd r r r give inequivalent representatives. Their orbit closures are distinct as well. Each orbit has the full dimension of a boundary component. If two closures were equal, both orbits would be dense open subsets of one irreducible variety and would intersect, which is impossible for distinct orbits. This proves Theorem 8.1 .
There are ⌊ ( n − 1 ) / 2 ⌋ \lfloor(n-1)/2\rfloor ⌊( n − 1 ) /2 ⌋ odd integers in [ 3 , n ] [3,n] [ 3 , n ] . The traceless component is represented by a form in n 2 − 1 n^2-1 n 2 − 1 essential variables [10 Section 8.3] . The locus of nonconcise forms is closed, being a rank condition on first derivatives. Thus it is different from every component above. This proves Corollary 8.2.
For example, at n = 7 n=7 n = 7 the invariant is
Table 1 (PDF p. 32).
It distinguishes the three concise components; the traceless component supplies a fourth.
Boundary components from block matrix spaces
We next consider singular matrix spaces of higher generic corank. Their leading determinant coefficient occurs at a higher power of t t t . As in the odd-skew case, the singular locus determines the seed directions, after which the stabilizer dimension can be computed.
The bounded-rank spaces in this section are due to Huang and Landsberg [9 Section 9] ; their 6 × 6 6 \times6 6 × 6 member is case (III) of their Theorem 3.1. The leading-coefficient construction is established in [12 Section 6.7.1] . We prove that these leading coefficients give boundary components after arbitrary padding, and that the resulting components are pairwise distinct and different from the odd-skew components.
Choose an integer h ≥ 2 h \ge2 h ≥ 2 . The original matrix has size r = 3 h r=3h r = 3 h , arranged in three blocks of size h h h . After adding k ≥ 0 k \ge0 k ≥ 0 rows and columns, its size is n = r + k n=r+k n = r + k . Let F ∈ Mat h F \in\operatorname{Mat}_h F ∈ Mat h be a matrix of independent variables, let a , b a,b a , b be scalar variables, and set
A = A h ( a , b , F ) = ( 0 a I h b I h − a I h 0 F − b I h − F 0 ) , K R = ( F − b I h a I h ) , K L = ( F − b I h a I h ) . A=A_h(a,b,F)=
\begin{pmatrix}
0 & aI_h & bI_h\\
-aI_h & 0 & F\\
-bI_h & -F & 0
\end{pmatrix},
\qquad
K_R=
\begin{pmatrix}
F\\
-bI_h\\
aI_h
\end{pmatrix},
\qquad
K_L=\left(F\ \ -bI_h\ \ aI_h\right). A = A h ( a , b , F ) = 0 − a I h − b I h a I h 0 − F b I h F 0 , K R = F − b I h a I h , K L = ( F − b I h a I h ) .
Direct multiplication gives A K R = 0 AK_R=0 A K R = 0 and K L A = 0 K_LA=0 K L A = 0 . Thus K R K_R K R and K L K_L K L provide h h h right and left kernel vectors, respectively. They define the compression of a coefficient matrix to an h × h h \times h h × h matrix.
The space L h \mathcal{L}_h L h has dimension d = h 2 + 2 d = h^2 + 2 d = h 2 + 2 . Choose a linear complement S h \mathcal{S}_h S h to L h \mathcal{L}_h L h in Mat r \operatorname{Mat}_r Mat r . Thus every r × r r \times r r × r matrix is uniquely a sum of a matrix in L h \mathcal{L}_h L h and one in S h \mathcal{S}_h S h . Write S S S for a matrix whose independent coefficients run through this complement. The remaining variables are U ∈ Mat r , k U \in\operatorname{Mat}_{r,k} U ∈ Mat r , k , V ∈ Mat k , r V \in\operatorname{Mat}_{k,r} V ∈ Mat k , r , and Z ∈ Mat k Z \in\operatorname{Mat}_k Z ∈ Mat k . Together these give n 2 n^2 n 2 variables. Define
H h , k = det h + k ( K L S K R K L U V K R Z ) . (22) H_{h,k}=\det\nolimits_{h+k}\begin{pmatrix}K_LSK_R & K_LU \\ VK_R & Z\end{pmatrix}. \tag*{(22)} H h , k = det h + k ( K L S K R V K R K L U Z ) . (22)
The empty blocks are omitted when k = 0 k=0 k = 0 . This formula is useful for studying the polynomial, but it is not an affine determinantal representation of size h + k h+k h + k . Its entries are not all linear: the degrees of the four blocks are 3, 2, 2, 1. Every determinant term has degree 3 h + k = n 3h+k=n 3 h + k = n and degree 2 h 2h 2 h in the variables x = ( a , b , F ) x=(a,b,F) x = ( a , b , F ) .
Theorem 9.1. For every h ≥ 2 h \ge2 h ≥ 2 and k ≥ 0 k \ge0 k ≥ 0 , H h , k H_{h,k} H h , k is concise in n 2 n^2 n 2 variables and
dim Stab ( H h , k ) = 2 n 2 − 1. \dim\operatorname{Stab}(H_{h,k})=2n^2-1. dim Stab ( H h , k ) = 2 n 2 − 1. Its orbit closure is a determinant-boundary component. For fixed n n n , these components are pairwise distinct as h h h ranges from 2 to ⌊ n / 3 ⌋ \lfloor n/3\rfloor ⌊ n /3 ⌋ , and are different from every component in Theorem 8.1 .
Degeneration and singular locus
The generic rank of A A A is 2 h 2h 2 h , so the leading determinant coefficient occurs at t h t^h t h . On a ≠ 0 a \ne0 a = 0 , eliminate the leading 2 h × 2 h 2h \times2h 2 h × 2 h block. That block has determinant a 2 h a^{2h} a 2 h ; the normalized left and right kernel matrices are K L / a K_L/a K L / a and K R / a K_R/a K R / a . Consequently
[ t h ] det ( A + t S ) = det ( K L S K R ) . [t^h]\det(A+tS)=\det(K_LSK_R). [ t h ] det ( A + tS ) = det ( K L S K R ) .
Schur elimination of Z Z Z then proves, first on a det Z ≠ 0 a\det Z \ne0 a det Z = 0 and hence as a polynomial identity,
[ t h ] det ( A + t S U t V Z ) = H h , k . (23) [t^h]\det\begin{pmatrix}A+tS & U \\ tV & Z\end{pmatrix}=H_{h,k}. \tag*{(23)} [ t h ] det ( A + tS t V U Z ) = H h , k . (23)
All lower coefficients vanish. For t ≠ 0 t \ne0 t = 0 , the entries of the displayed matrix give an invertible linear parametrization of Mat n \operatorname{Mat}_n Mat n . Divide its determinant by t h t^h t h , which can be done by scaling one row. Letting t t t tend to zero proves that H h , k H_{h,k} H h , k belongs to O n \mathcal{O}_n O n .
If a ≠ 0 a \ne0 a = 0 or b ≠ 0 b \ne0 b = 0 , the rank of A A A is 2 h 2h 2 h . If a = b = 0 a=b=0 a = b = 0 , it is 2 rank F 2\operatorname{rank}F 2 rank F . Thus its rank-drop locus is the irreducible codimension-three variety
Y h = { a = b = 0 , det F = 0 } ⊂ C d . Y_h=\{a=b=0,\ \det F=0\}\subset\mathbb{C}^d. Y h = { a = b = 0 , det F = 0 } ⊂ C d .
The set Y h Y_h Y h determines the largest singular component of H h , k H_{h,k} H h , k . At a point of Y h Y_h Y h , both K L K_L K L and K R K_R K R lose rank. The matrix in (22) then has a left and a right kernel. If its rank is at most h + k − 2 h+k-2 h + k − 2 , all first derivatives vanish. If its rank is h + k − 1 h+k-1 h + k − 1 , its adjugate pairs the indicated left and right kernels; differentiating either kernel factor leaves the other factor zero. Thus all first derivatives again vanish. This compression argument covers every U , V , Z U,V,Z U , V , Z , including singular Z Z Z .
Explicitly, at a = b = 0 a=b=0 a = b = 0 choose nonzero ℓ , q \ell,q ℓ , q with ℓ F = 0 \ell F=0 ℓ F = 0 and F q = 0 Fq=0 F q = 0 . Then ℓ K L = 0 \ell K_L=0 ℓ K L = 0 and K R q = 0 K_Rq=0 K R q = 0 . These give a left kernel ( ℓ , 0 ) (\ell,0) ( ℓ , 0 ) and a right kernel ( q , 0 ) (q,0) ( q , 0 ) of the compressed matrix. Its derivative paired against these vectors is
ℓ D ( K L S K R ) [ δ x , δ S ] q = 0 , \ell D(K_LSK_R)[\delta x,\delta S]q=0, ℓ D ( K L S K R ) [ δ x , δ S ] q = 0 ,
because each of the three product-rule terms retains a zero outer factor. This proves the asserted vanishing in the rank-h + k − 1 h+k-1 h + k − 1 case.
For fixed x = ( a , b , F ) x=(a,b,F) x = ( a , b , F ) outside Y h Y_h Y h , the map
( S , U , V , Z ) ⟼ W x = ( K L S K R K L U V K R Z ) (S,U,V,Z)\longmapsto W_x=\begin{pmatrix} K_LSK_R & K_LU \\ VK_R & Z \end{pmatrix} ( S , U , V , Z ) ⟼ W x = ( K L S K R V K R K L U Z )
is onto Mat h + k \operatorname{Mat}_{h+k} Mat h + k . Indeed K L , K R K_L,K_R K L , K R have full rank, and K L L h K R = 0 K_L\mathcal{L}_hK_R=0 K L L h K R = 0 allows the full matrix space to be replaced by its complement S h \mathcal{S}_h S h . The S , U , V , Z S,U,V,Z S , U , V , Z derivatives vanish exactly when rank W x ≤ h + k − 2 \operatorname{rank}W_x\le h+k-2 rank W x ≤ h + k − 2 , which also kills the x x x derivatives. This locus is locally a product of the irreducible rank-≤ h + k − 2 \le h+k-2 ≤ h + k − 2 determinantal variety with an affine space, over the irreducible base C d ∖ Y h \mathbb{C}^d\setminus Y_h C d ∖ Y h . It has dimension n 2 − 4 n^2-4 n 2 − 4 . Therefore the unique largest singular component is
M h , k = Y h × C n 2 − d , dim M h , k = n 2 − 3. (24) M_{h,k}=Y_h\times\mathbb{C}^{n^2-d},\qquad\dim M_{h,k}=n^2-3. \tag*{(24)} M h , k = Y h × C n 2 − d , dim M h , k = n 2 − 3. (24)
The generic determinant has singular locus of codimension four. Hence H h , k H_{h,k} H h , k is outside its orbit and is a boundary point.
The determinant hypersurface in Mat h \operatorname{Mat}_h Mat h has no nonzero translation direction, since its first derivatives are independent. Hence the translation space of M h , k M_{h,k} M h , k consists exactly of the coefficient directions. Its linear span is cut out by a = b = 0 a=b=0 a = b = 0 . Therefore
Trans ( M h , k ) = { x = 0 } , dim Trans ( M h , k ) = n 2 − ( h 2 + 2 ) , codim ⟨ M h , k ⟩ = 2. (25) \operatorname{Trans}(M_{h,k})=\{x=0\},\qquad\dim\operatorname{Trans}(M_{h,k})=n^2-(h^2+2),\qquad\operatorname{codim}\langle M_{h,k}\rangle=2. \tag*{(25)} Trans ( M h , k ) = { x = 0 } , dim Trans ( M h , k ) = n 2 − ( h 2 + 2 ) , codim ⟨ M h , k ⟩ = 2. (25)
The last equality uses the fact that singular matrices span Mat h \operatorname{Mat}_h Mat h for h ≥ 2 h\ge2 h ≥ 2 . Any linear stabilizer therefore preserves the S , U , V , Z S,U,V,Z S , U , V , Z directions: its x x x -part has the form δ x = T ( x ) \delta x=T(x) δ x = T ( x ) .
The same calculation proves conciseness. A constant direction annihilating H h , k H_{h,k} H h , k must preserve its unique largest singular component, and thus has zero x x x -part. The surjective compression above and conciseness of a generic determinant force its compressed S , U , V , Z S,U,V,Z S , U , V , Z direction to vanish for every x x x . This kills Z , U , V Z,U,V Z , U , V , since the kernel matrices have no common linear annihilators. Finally the common constant kernel of S ↦ K L S K R S\mapsto K_LSK_R S ↦ K L S K R inside Mat r \operatorname{Mat}_r Mat r is exactly L h \mathcal{L}_h L h . To see this, expand a constant block matrix B = ( B i j ) B=(B_{ij}) B = ( B ij ) :
K L B K R = F B 11 F − b ( F B 12 + B 21 F ) + a ( F B 13 + B 31 F ) + b 2 B 22 − a b ( B 23 + B 32 ) + a 2 B 33 . K_LBK_R=FB_{11}F-b(FB_{12}+B_{21}F)+a(FB_{13}+B_{31}F)+b^2B_{22}-ab(B_{23}+B_{32})+a^2B_{33}. K L B K R = F B 11 F − b ( F B 12 + B 21 F ) + a ( F B 13 + B 31 F ) + b 2 B 22 − ab ( B 23 + B 32 ) + a 2 B 33 .
The coefficients give B 11 = B 22 = B 33 = 0 B_{11}=B_{22}=B_{33}=0 B 11 = B 22 = B 33 = 0 , B 23 = − B 32 B_{23}=-B_{32} B 23 = − B 32 , B 12 = c I = − B 21 B_{12}=cI=-B_{21} B 12 = c I = − B 21 , and B 13 = e I = − B 31 B_{13}=eI=-B_{31} B 13 = e I = − B 31 . These describe precisely L h \mathcal{L}_h L h . The chosen complement has zero intersection with it, proving conciseness.
Linear syzygies and the action on the seed space
The stabilizer splits into the action on the seed variables x = ( a , b , F ) x=(a,b,F) x = ( a , b , F ) , the maps from seed variables to coefficient variables, and the action among coefficient variables. We first compute the first two contributions. The linear syzygies are matrix-valued because the kernels have h h h columns. They follow from the next identity.
Lemma 9.2. For h ≥ 2 h\ge2 h ≥ 2 , linear maps X , Y : Mat h → Mat h X,Y:\operatorname{Mat}_h\to\operatorname{Mat}_h X , Y : Mat h → Mat h satisfy F X ( F ) + Y ( F ) F = 0 FX(F)+Y(F)F=0 F X ( F ) + Y ( F ) F = 0 identically if and only if
X ( F ) = B F + λ ( F ) I , Y ( F ) = − F B − λ ( F ) I X(F)=BF+\lambda(F)I,\qquad Y(F)=-FB-\lambda(F)I X ( F ) = B F + λ ( F ) I , Y ( F ) = − F B − λ ( F ) I for a unique constant matrix B B B and a unique scalar linear form λ \lambda λ .
Proof. Put C = X ( I ) C=X(I) C = X ( I ) ; evaluation at I I I gives Y ( I ) = − C Y(I)=-C Y ( I ) = − C . The linear coefficient after substituting F = I + t M F=I+tM F = I + tM yields Y ( M ) = − X ( M ) − M C + C M Y(M)=-X(M)-MC+CM Y ( M ) = − X ( M ) − M C + C M . The quadratic coefficient then gives [ M , X ( M ) − C M ] = 0 [M,X(M)-CM]=0 [ M , X ( M ) − C M ] = 0 for all M M M . A linear map commuting with its argument has the form M ↦ c M + μ ( M ) I M\mapsto cM+\mu(M)I M ↦ c M + μ ( M ) I . One elementary verification is as follows. For a diagonal matrix D D D , its image is diagonal, first for distinct diagonal entries and then for all D D D by linearity. Polarization with each matrix unit E i j E_{ij} E ij shows that the off-diagonal part of its image is c i j E i j c_{ij}E_{ij} c ij E ij and that the differences between diagonal entries of the image of D D D are c i j ( d i − d j ) c_{ij}(d_i-d_j) c ij ( d i − d j ) . Comparing three indices forces one common c c c ; for h = 2 h=2 h = 2 the two opposite pairs give the same conclusion. After subtracting c c c times the identity operator, polarization between distinct matrix units forces every remaining diagonal image to be scalar. Commutation with E i j E_{ij} E ij itself supplies the same conclusion in the 2 × 2 2 \times2 2 × 2 case. This proves the stated form of a commuting map. Substitution now gives the displayed formulas with B = C + c I B=C+cI B = C + c I and λ = μ \lambda=\mu λ = μ . Uniqueness follows since B F BF B F cannot be a scalar matrix for every F F F unless B = 0 B=0 B = 0 .
Lemma 9.3. The space of linear maps B : x ↦ S h B:x\mapsto\mathcal{S}_h B : x ↦ S h satisfying K L B ( x ) K R = 0 K_LB(x)K_R=0 K L B ( x ) K R = 0 has dimension 14 h 2 − 4 14h^2-4 14 h 2 − 4 . The left and right vector syzygy spaces K L u ( x ) = 0 K_Lu(x)=0 K L u ( x ) = 0 and v ( x ) K R = 0 v(x)K_R=0 v ( x ) K R = 0 each have dimension r = 3 h r=3h r = 3 h .
Proof. Combine B 23 + B 32 B_{23}+B_{32} B 23 + B 32 into one block, since the opposite pair belongs to L h \mathcal{L}_h L h . There are eight block-valued linear functions remaining. Decompose each into its a a a , b b b , and F F F parts. We compare the independent groups of monomials separately. For example, the a F 2 aF^2 a F 2 terms are linear in a a a and quadratic in the entries of F F F . The F 3 F^3 F 3 coefficient forces B 11 ( F ) = 0 B_{11}(F)=0 B 11 ( F ) = 0 . The a F 2 aF^2 a F 2 equation is
F C F + F X ( F ) + Y ( F ) F = 0. FCF+FX(F)+Y(F)F=0. F C F + F X ( F ) + Y ( F ) F = 0.
By Lemma 9.2, its solutions are X = ( B − C ) F + λ ( F ) I X=(B-C)F+\lambda(F)I X = ( B − C ) F + λ ( F ) I , Y = − F B − λ ( F ) I Y=-FB-\lambda(F)I Y = − F B − λ ( F ) I : there are 3 h 2 3h^2 3 h 2 parameters, including C C C . The b F 2 bF^2 b F 2 equation contributes another 3 h 2 3h^2 3 h 2 . The a 2 F a^2F a 2 F , b 2 F b^2F b 2 F , a b F abF ab F equations determine respectively the F F F parts of B 33 B_{33} B 33 , B 22 B_{22} B 22 , B 23 + B 32 B_{23}+B_{32} B 23 + B 32 from the a , b a,b a , b parts of B 12 B_{12} B 12 , B 21 B_{21} B 21 , B 13 B_{13} B 13 , B 31 B_{31} B 31 ; these supply 8 h 2 8h^2 8 h 2 free parameters. The a 3 a^3 a 3 , b 3 b^3 b 3 coefficients set B 33 ( a ) = B 22 ( b ) = 0 B_{33}(a)=B_{22}(b)=0 B 33 ( a ) = B 22 ( b ) = 0 , and the a 2 b a^2b a 2 b , a b 2 ab^2 a b 2 equations leave 2 h 2 2h^2 2 h 2 more parameters. Thus the count before the remaining quotient is 16 h 2 16h^2 16 h 2 . The two scalar opposite block pairs in L h \mathcal{L}_h L h each permit an arbitrary linear form in d = h 2 + 2 d=h^2+2 d = h 2 + 2 variables. Quotienting subtracts 2 d 2d 2 d , leaving 14 h 2 − 4 14h^2-4 14 h 2 − 4 .
For the left vector equation, write u = ( u 0 , u 1 , u 2 ) u=(u_0,u_1,u_2) u = ( u 0 , u 1 , u 2 ) , so that F u 0 − b u 1 + a u 2 = 0 Fu_0-bu_1+au_2=0 F u 0 − b u 1 + a u 2 = 0 . The part quadratic in F F F forces u 0 ( F ) = 0 u_0(F)=0 u 0 ( F ) = 0 by invertibility of generic F F F . Coefficient comparison then gives
u 0 = a p + b q , u 1 = F q + a w , u 2 = − F p + b w u_0=ap+bq,\qquad u_1=Fq+aw,\qquad u_2=-Fp+bw u 0 = a p + b q , u 1 = F q + a w , u 2 = − F p + b w
for arbitrary constant p , q , w ∈ C h p,q,w\in\mathbb{C}^h p , q , w ∈ C h . These are exactly the columns A c Ac A c with constant c c c . They have dimension 3 h 3h 3 h . Transposing the calculation proves the right-vector statement.
The induced action on x x x must preserve Y h Y_h Y h and its linear span. Consequently ( δ a , δ b ) (\delta a,\delta b) ( δ a , δ b ) is any linear function of ( a , b ) (a,b) ( a , b ) , and
δ F = C F + F D + a E + b G . (26) \delta F=CF+FD+aE+bG. \tag*{(26)} δ F = C F + F D + a E + b G . (26)
The familiar determinant-similitude calculation gives dimension 2 h 2 − 1 2h^2-1 2 h 2 − 1 for F ↦ C F + F D F\mapsto CF+FD F ↦ C F + F D ; it holds also for h = 2 h=2 h = 2 . Including the other terms, the space of possible T T T has dimension
4 + ( 2 h 2 − 1 ) + 2 h 2 = 4 h 2 + 3. (27) 4+(2h^2-1)+2h^2=4h^2+3. \tag*{(27)} 4 + ( 2 h 2 − 1 ) + 2 h 2 = 4 h 2 + 3. (27)
Every action in (26) extends to a stabilizer of the whole polynomial by a compensating action on the coefficient blocks. Indeed, changes in ( a , b ) (a,b) ( a , b ) and the translations F ↦ F + a E + b G F\mapsto F+aE+bG F ↦ F + a E + b G induce constant block actions on K L , K R K_L,K_R K L , K R , which are canceled by opposite actions on S , U , V S,U,V S , U , V . For the remaining term C F + F D CF+FD C F + F D , explicitly
δ K L = C K L + K L diag ( D , − C , − C ) , δ K R = diag ( C , − D , − D ) K R + K R D . \delta K_L=CK_L+K_L\operatorname{diag}(D,-C,-C),\qquad\delta K_R=\operatorname{diag}(C,-D,-D)K_R+K_RD. δ K L = C K L + K L diag ( D , − C , − C ) , δ K R = diag ( C , − D , − D ) K R + K R D .
Cancel the internal block actions on S , U , V S,U,V S , U , V . The compressed matrix then changes by left multiplication by diag ( C , 0 k ) \operatorname{diag}(C,0_k) diag ( C , 0 k ) and right multiplication by diag ( D , 0 k ) \operatorname{diag}(D,0_k) diag ( D , 0 k ) . Subtracting τ S \tau S τ S and τ U \tau U τ U from δ S , δ U \delta S,\delta U δ S , δ U , with τ = ( tr C + tr D ) / h \tau=(\operatorname{tr} C+\operatorname{tr} D)/h τ = ( tr C + tr D ) / h , cancels the resulting determinant character. Whenever necessary project δ S \delta S δ S to the chosen complement; this does not change its compression. Hence the upper bound in (27) is attained.
The coefficient stabilizer
For fixed x x x , the polynomial is an ordinary determinant after compression. A linear coefficient symmetry must induce a determinant symmetry for every x x x . The following lemma determines the actions satisfying these simultaneous conditions.
We retain the block notation of the construction. Fix h ≥ 2 h\ge2 h ≥ 2 and k ≥ 0 k\ge0 k ≥ 0 , and put r = 3 h r=3h r = 3 h and n = r + k n=r+k n = r + k . For F ∈ Mat h F\in\operatorname{Mat}_h F ∈ Mat h and a , b ∈ C a,b\in\mathbb{C} a , b ∈ C , write
A ( x ) = ( 0 a I h b I h − a I h 0 F − b I h − F 0 ) , K R ( x ) = ( F − b I h a I h ) , K L ( x ) = ( F − b I h a I h ) . A(x)=\begin{pmatrix}0&aI_h&bI_h\\-aI_h&0&F\\-bI_h&-F&0\end{pmatrix},\qquad K_R(x)=\begin{pmatrix}F\\-bI_h\\aI_h\end{pmatrix},\qquad K_L(x)=\begin{pmatrix}F&-bI_h&aI_h\end{pmatrix}. A ( x ) = 0 − a I h − b I h a I h 0 − F b I h F 0 , K R ( x ) = F − b I h a I h , K L ( x ) = ( F − b I h a I h ) .
Then A ( x ) K R ( x ) = 0 A(x)K_R(x)=0 A ( x ) K R ( x ) = 0 and K L ( x ) A ( x ) = 0 K_L(x)A(x)=0 K L ( x ) A ( x ) = 0 . Let L ⊂ Mat r L\subset\operatorname{Mat}_r L ⊂ Mat r be the space of the matrices A ( x ) A(x) A ( x ) , and let L t o p ⊂ Mat n L_{\mathrm{top}}\subset\operatorname{Mat}_n L top ⊂ Mat n be its embedding in the upper-left block. Adding an element of L t o p L_{\mathrm{top}} L top does not change the compressed coefficient matrix. We therefore identify the coefficient space with the quotient
Y = Mat n / L t o p . \mathcal{Y}=\operatorname{Mat}_n/L_{\mathrm{top}}. Y = Mat n / L top .
After choosing a complement to L L L , its coordinates are Y = ( S , U , V , Z ) Y=(S,U,V,Z) Y = ( S , U , V , Z ) , with block sizes r × r r\times r r × r , r × k r\times k r × k , k × r k\times r k × r , and k × k k\times k k × k , respectively. Set
B R ( x ) = diag ( K R ( x ) , I k ) , B L ( x ) = diag ( K L ( x ) , I k ) , B_R(x)=\operatorname{diag}(K_R(x),I_k),\qquad B_L(x)=\operatorname{diag}(K_L(x),I_k), B R ( x ) = diag ( K R ( x ) , I k ) , B L ( x ) = diag ( K L ( x ) , I k ) ,
and
M x ( Y ) = B L ( x ) Y B R ( x ) = ( K L ( x ) S K R ( x ) K L ( x ) U V K R ( x ) Z ) , H h , k ( x , Y ) = det M x ( Y ) . (28) M_x(Y)=B_L(x)YB_R(x)=\begin{pmatrix}K_L(x)SK_R(x)&K_L(x)U\\VK_R(x)&Z\end{pmatrix},\qquad H_{h,k}(x,Y)=\det M_x(Y). \tag*{(28)} M x ( Y ) = B L ( x ) Y B R ( x ) = ( K L ( x ) S K R ( x ) V K R ( x ) K L ( x ) U Z ) , H h , k ( x , Y ) = det M x ( Y ) . (28)
These expressions are well defined on the quotient Y \mathcal{Y} Y . The matrices B L B_L B L and B R B_R B R combine the seed kernel vectors with the padding coordinates. We call Y ↦ B L Y B R Y\mapsto B_LYB_R Y ↦ B L Y B R the compression map.
Lemma 9.4. Let p h , k \mathfrak{p}_{h,k} p h , k be the space of constant linear operators R : Y → Y R:\mathcal{Y}\to\mathcal{Y} R : Y → Y such that
D Y H h , k ( x , Y ) [ R ( Y ) ] = 0 for every x , Y . D_YH_{h,k}(x,Y)[R(Y)]=0\qquad\text{for every }x,Y. D Y H h , k ( x , Y ) [ R ( Y )] = 0 for every x , Y . Then
dim p h , k = 2 k 2 + 2 r k . \dim\mathfrak{p}_{h,k}=2k^2+2rk. dim p h , k = 2 k 2 + 2 r k . In particular , p h , 0 = 0 \mathfrak{p}_{h,0}=0 p h , 0 = 0 . More explicitly , after representing the quotient by a complement and writing π : Mat r → Mat r / L \pi:\operatorname{Mat}_r\to\operatorname{Mat}_r/L π : Mat r → Mat r / L for the quotient map, all its elements are uniquely parametrized by
B 0 ∈ Mat r , k , C 0 ∈ Mat k , r , E 0 , G 0 ∈ Mat k , α ∈ C , h α + tr E 0 + tr G 0 = 0 , B_0\in\operatorname{Mat}_{r,k},\quad C_0\in\operatorname{Mat}_{k,r},\quad E_0,G_0\in\operatorname{Mat}_k,\quad\alpha\in\mathbb{C},\quad h\alpha+\operatorname{tr}E_0+\operatorname{tr}G_0=0, B 0 ∈ Mat r , k , C 0 ∈ Mat k , r , E 0 , G 0 ∈ Mat k , α ∈ C , h α + tr E 0 + tr G 0 = 0 ,
through the vector fields
δ S = π ( α S + B 0 V + U C 0 ) , δ U = α U + B 0 Z + U E 0 , δ V = G 0 V + Z C 0 , δ Z = G 0 Z + Z E 0 . \begin{aligned}
\delta S &= \pi(\alpha S + B_0V + UC_0), & \delta U &= \alpha U + B_0Z + UE_0,\\
\delta V &= G_0V + ZC_0, & \delta Z &= G_0Z + ZE_0.
\end{aligned} δ S δ V = π ( α S + B 0 V + U C 0 ) , = G 0 V + Z C 0 , δ U δ Z = α U + B 0 Z + U E 0 , = G 0 Z + Z E 0 .
For k = 0 k = 0 k = 0 , the empty blocks are omitted and the trace condition forces α = 0 \alpha= 0 α = 0 .
Proof. We prove that a coefficient symmetry preserves the kernel of each compression and hence induces an infinitesimal determinant preserver. Comparing these induced actions as x x x varies determines their parameters.
We use the trace pairing to identify dual matrix spaces. First, the common constant kernel of Y ↦ M x ( Y ) Y \mapsto M_x(Y) Y ↦ M x ( Y ) is precisely L t o p L_{\mathrm{top}} L top . Indeed, the seed compression expands as
K L S K R = F S 11 F − b ( F S 12 + S 21 F ) + a ( F S 13 + S 31 F ) + b 2 S 22 − a b ( S 23 + S 32 ) + a 2 S 33 . K_LSK_R = FS_{11}F - b(FS_{12} + S_{21}F) + a(FS_{13} + S_{31}F) + b^2S_{22} - ab(S_{23} + S_{32}) + a^2S_{33}. K L S K R = F S 11 F − b ( F S 12 + S 21 F ) + a ( F S 13 + S 31 F ) + b 2 S 22 − ab ( S 23 + S 32 ) + a 2 S 33 .
Its quadratic part in F F F gives S 11 = 0 S_{11} = 0 S 11 = 0 . The coefficient of b b b gives F S 12 + S 21 F = 0 FS_{12} + S_{21}F = 0 F S 12 + S 21 F = 0 for every F F F , hence S 12 = c I h S_{12} = cI_h S 12 = c I h and S 21 = − c I h S_{21} = -cI_h S 21 = − c I h . The coefficient of a a a gives the analogous scalar pair S 13 = d I h S_{13} = dI_h S 13 = d I h , S 31 = − d I h S_{31} = -dI_h S 31 = − d I h . The other coefficients give S 22 = S 33 = 0 S_{22} = S_{33} = 0 S 22 = S 33 = 0 and S 23 = − S 32 S_{23} = -S_{32} S 23 = − S 32 , exactly the matrices in L L L . Furthermore, K L U = 0 K_LU = 0 K L U = 0 for every x x x forces the three block rows of U U U to vanish, and similarly V K R = 0 VK_R = 0 V K R = 0 forces V = 0 V = 0 V = 0 ; the Z Z Z block is unchanged. It follows, in particular, that the dual compression spaces introduced below span Y ∗ \mathcal{Y}^* Y ∗ .
For the comparison it is enough to use values with a ≠ 0 a \ne0 a = 0 . Rescale them to a = 1 a = 1 a = 1 , and write t = b t = b t = b . Polynomial identities obtained on this dense set extend to all seed values. For every F , t F,t F , t , the map M x : Y → Mat h + k M_x : \mathcal{Y} \to\operatorname{Mat}_{h+k} M x : Y → Mat h + k is surjective: the blocks S 33 S_{33} S 33 , U 3 U_3 U 3 , V 3 V_3 V 3 , Z Z Z already realize every compressed matrix. We claim that R ∈ p h , k R \in\mathfrak{p}_{h,k} R ∈ p h , k preserves ker M x \ker M_x ker M x . For Y 0 ∈ ker M x Y_0 \in\ker M_x Y 0 ∈ ker M x , substitute Y + z Y 0 Y + zY_0 Y + z Y 0 in its stabilizer identity and take the coefficient of z z z . This gives
D ( det ) h + k [ M x ( R ( Y 0 ) ) ] = 0 for every T ∈ Mat h + k . D(\det)_{h+k}[M_x(R(Y_0))] = 0 \qquad\text{for every } T \in\operatorname{Mat}_{h+k}. D ( det ) h + k [ M x ( R ( Y 0 ))] = 0 for every T ∈ Mat h + k .
The generic determinant is concise in its entries, so M x ( R ( Y 0 ) ) = 0 M_x(R(Y_0)) = 0 M x ( R ( Y 0 )) = 0 , proving the claim.
Dualizing the compression maps, the transpose R ∗ R^* R ∗ preserves each subspace
W x = { B R ( x ) T B L ( x ) : T ∈ Mat h + k } ⊂ Y ∗ . W_x = \{B_R(x)TB_L(x) : T \in\operatorname{Mat}_{h+k}\} \subset\mathcal{Y}^*. W x = { B R ( x ) T B L ( x ) : T ∈ Mat h + k } ⊂ Y ∗ .
Write T = ( H B C D ) T = \begin{pmatrix} H & B \\ C & D \end{pmatrix} T = ( H C B D ) in block sizes h , k h,k h , k . On the chosen chart its image is
W F , t ( T ) = ( F H F − t F H F H F B − t H F t 2 H − t H − t B H F − t H H B C F − t C C D ) . (29) W_{F,t}(T) =
\begin{pmatrix}
FHF & -tFH & FH & FB\\
-tHF & t^2H & -tH & -tB\\
HF & -tH & H & B\\
CF & -tC & C & D
\end{pmatrix}. \tag*{(29)} W F , t ( T ) = F H F − t H F H F C F − tF H t 2 H − t H − tC F H − t H H C F B − tB B D . (29)
The submatrix on its third and fourth block rows and columns is T T T itself. Thus the uniquely induced operator Ψ F , t \Psi_{F,t} Ψ F , t , defined by
R ∗ ( W F , t ( T ) ) = W F , t ( Ψ F , t ( T ) ) , R^*(W_{F,t}(T)) = W_{F,t}(\Psi_{F,t}(T)), R ∗ ( W F , t ( T )) = W F , t ( Ψ F , t ( T )) ,
has coefficients polynomial in F , t F,t F , t , of total degree at most two. The induced primal operator preserves the determinant, and the dual of a left–right determinant Lie action is again such an action. The classical determinant stabilizer description therefore gives
Ψ F , t ( T ) = E F , t T + T G F , t , tr E F , t + tr G F , t = 0. (30) \Psi_{F,t}(T) = E_{F,t}T + TG_{F,t}, \qquad\operatorname{tr} E_{F,t} + \operatorname{tr} G_{F,t} = 0. \tag*{(30)} Ψ F , t ( T ) = E F , t T + T G F , t , tr E F , t + tr G F , t = 0. (30)
The left and right matrices are unique up to adding a scalar identity to one and subtracting it from the other. We fix this ambiguity before comparing degrees. The kernel of ( E , G ) ↦ [ T ↦ E T + T G ] (E,G) \mapsto[T \mapsto ET+TG] ( E , G ) ↦ [ T ↦ E T + T G ] consists of ( c I , − c I ) (cI,-cI) ( c I , − c I ) . The condition tr E 11 = 0 \operatorname{tr} E_{11}=0 tr E 11 = 0 , where the block E 11 E_{11} E 11 has size h h h , removes this kernel and gives a fixed linear splitting of this finite-dimensional map. Applying that splitting to Ψ F , t \Psi_{F,t} Ψ F , t makes all entries of E F , t , G F , t E_{F,t},G_{F,t} E F , t , G F , t polynomial in F , t F,t F , t . This also fixes the scalar ambiguity throughout the proof.
The repeated blocks in (29) impose degree constraints on this action. In particular, a block occurring both without a factor and with a factor t 2 t^2 t 2 must satisfy both degree bounds.
Let H ′ H' H ′ be the upper-left block of Ψ F , t ( T ) \Psi_{F,t}(T) Ψ F , t ( T ) . Its appearance as the core (3, 3) block of (29) shows that it has total degree at most two in F , t F,t F , t . Its appearance as t 2 H ′ t^2H' t 2 H ′ in the core (2, 2) block shows that it is independent of F , t F,t F , t . Since
H ′ = E 11 H + E 12 C + H G 11 + B G 21 . H'=E_{11}H+E_{12}C+HG_{11}+BG_{21}. H ′ = E 11 H + E 12 C + H G 11 + B G 21 .
the matrices E 12 , G 21 E_{12},G_{21} E 12 , G 21 and the operator H ↦ E 11 H + H G 11 H\mapsto E_{11}H+HG_{11} H ↦ E 11 H + H G 11 are constant. The fixed gauge implies that E 11 , G 11 E_{11},G_{11} E 11 , G 11 themselves are constant. For input C C C alone, the left side of the defining equation for Ψ F , t \Psi_{F,t} Ψ F , t has degree at most one in F , t F,t F , t , whereas the output core (1, 1) block is F E 12 C F FE_{12}CF F E 12 C F . Its quadratic degree in F F F forces E 12 = 0 E_{12}=0 E 12 = 0 . The same argument with input B B B gives G 21 = 0 G_{21}=0 G 21 = 0 .
The upper-left action is constant. To prove that it is scalar, put Θ ( H ) = E 11 H + H G 11 \Theta(H)=E_{11}H+HG_{11} Θ ( H ) = E 11 H + H G 11 . For input H H H and t = 0 t=0 t = 0 , the only input block of degree two in F F F is F H F FHF F H F . Comparing degree two in the core (1, 1) output therefore gives a fixed linear map J : Mat h → Mat h J:\operatorname{Mat}_h\to\operatorname{Mat}_h J : Mat h → Mat h with
J ( F H F ) = F Θ ( H ) F . J(FHF)=F\Theta(H)F. J ( F H F ) = F Θ ( H ) F .
At F = I h F=I_h F = I h , this says J = Θ J=\Theta J = Θ , so
Θ ( F H F ) = F Θ ( H ) F . (31) \Theta(FHF)=F\Theta(H)F. \tag*{(31)} Θ ( F H F ) = F Θ ( H ) F . (31)
Replacing F F F by I h + s X I_h+sX I h + s X and taking the coefficient of s s s gives
Θ ( X H + H X ) = X Θ ( H ) + Θ ( H ) X . \Theta(XH+HX)=X\Theta(H)+\Theta(H)X. Θ ( X H + H X ) = X Θ ( H ) + Θ ( H ) X .
If B ∗ = Θ ( I h ) B_*=\Theta(I_h) B ∗ = Θ ( I h ) , setting H = I h H=I_h H = I h yields Θ ( X ) = ( X B ∗ + B ∗ X ) / 2 \Theta(X)=(XB_*+B_*X)/2 Θ ( X ) = ( X B ∗ + B ∗ X ) /2 . Substitution back in the preceding identity gives [ H , [ X , B ∗ ] ] = 0 [H,[X,B_*]]=0 [ H , [ X , B ∗ ]] = 0 for every H , X H,X H , X . Thus [ X , B ∗ ] [X,B_*] [ X , B ∗ ] is central; its trace is zero, so it is zero. Therefore B ∗ B_* B ∗ is scalar and Θ ( H ) = α H \Theta(H)=\alpha H Θ ( H ) = α H . The gauge tr E 11 = 0 \operatorname{tr}E_{11}=0 tr E 11 = 0 now gives
E 11 = 0 , G 11 = α I h , E 12 = G 21 = 0. (32) E_{11}=0,\qquad G_{11}=\alpha I_h,\qquad E_{12}=G_{21}=0. \tag*{(32)} E 11 = 0 , G 11 = α I h , E 12 = G 21 = 0. (32)
It remains to determine the interactions with the padding. These are governed by the blocks of degree one in F , t F,t F , t in (29) . Suppose k > 0 k>0 k > 0 . The upper-right block of the smaller matrix is now
B ′ = H G 12 + B G 22 . B'=HG_{12}+BG_{22}. B ′ = H G 12 + B G 22 .
For input B B B alone, the top extra block of its image is F B G 22 ( F , t ) FBG_{22}(F,t) F B G 22 ( F , t ) , whereas all input entries have total degree at most one. Any positive-degree homogeneous part of G 22 G_{22} G 22 would produce degree at least two after multiplication by generic F F F and B B B , and must vanish. Hence G 22 G_{22} G 22 is constant; similarly E 22 E_{22} E 22 is constant. Denote them by G 0 , E 0 G_0,E_0 G 0 , E 0 .
For input H H H , the second extra block is − t H G 12 ( F , t ) -tHG_{12}(F,t) − t H G 12 ( F , t ) . It has degree at most two, so G 12 G_{12} G 12 has degree at most one. Write
G 12 ( F , t ) = B 3 + t B 2 + Γ ( F ) G_{12}(F,t)=B_3+tB_2+\Gamma(F) G 12 ( F , t ) = B 3 + t B 2 + Γ ( F )
with Γ \Gamma Γ linear. Comparing the terms quadratic in F F F in the top extra block, the only input of that degree is again F H F FHF F H F . Hence a fixed linear map J 0 : Mat h → Mat h , k J_0 : \operatorname{Mat}_h \to\operatorname{Mat}_{h,k} J 0 : Mat h → Mat h , k satisfies
J 0 ( F H F ) = F H Γ ( F ) . J_0(FHF)=FH\Gamma(F). J 0 ( F H F ) = F H Γ ( F ) .
At F = I h F=I_h F = I h this gives J 0 ( H ) = H B 1 J_0(H)=HB_1 J 0 ( H ) = H B 1 , where B 1 = Γ ( I h ) B_1=\Gamma(I_h) B 1 = Γ ( I h ) . Consequently F H Γ ( F ) = F H F B 1 FH\Gamma(F)=FHFB_1 F H Γ ( F ) = F H F B 1 for all F , H F,H F , H , so Γ ( F ) = F B 1 \Gamma(F)=FB_1 Γ ( F ) = F B 1 on the invertible locus, and hence everywhere. Therefore
G 12 = F B 1 + t B 2 + B 3 = K L B 0 , B 0 = ( B 1 − B 2 B 3 ) ∈ Mat r , k . G_{12}=FB_1+tB_2+B_3=K_LB_0,\qquad B_0=\begin{pmatrix}B_1\\-B_2\\B_3\end{pmatrix}\in\operatorname{Mat}_{r,k}. G 12 = F B 1 + t B 2 + B 3 = K L B 0 , B 0 = B 1 − B 2 B 3 ∈ Mat r , k .
The row version of the same argument gives E 21 = C 0 K R E_{21}=C_0K_R E 21 = C 0 K R for a constant C 0 ∈ Mat k , r C_0\in\operatorname{Mat}_{k,r} C 0 ∈ Mat k , r . We have proved that the fiber pair is necessarily
E F , t = ( 0 0 C 0 K R E 0 ) , G F , t = ( α I h K L B 0 0 G 0 ) , h α + tr E 0 + tr G 0 = 0. (33) E_{F,t}=\begin{pmatrix}0&0\\C_0K_R&E_0\end{pmatrix},\qquad G_{F,t}=\begin{pmatrix}\alpha I_h&K_LB_0\\0&G_0\end{pmatrix},\qquad h\alpha+\operatorname{tr}E_0+\operatorname{tr}G_0=0. \tag*{(33)} E F , t = ( 0 C 0 K R 0 E 0 ) , G F , t = ( α I h 0 K L B 0 G 0 ) , h α + tr E 0 + tr G 0 = 0. (33)
For k = 0 k=0 k = 0 the same conclusion follows directly from (32) , with empty off-diagonal and lower blocks and with h α = 0 h\alpha=0 h α = 0 .
Conversely, every parameter choice in (33) defines a symmetry. Set
M 0 = ( 0 r × r 0 C 0 E 0 ) , N 0 = ( α I r B 0 0 G 0 ) . M_0=\begin{pmatrix}0_{r\times r}&0\\C_0&E_0\end{pmatrix},\qquad N_0=\begin{pmatrix}\alpha I_r&B_0\\0&G_0\end{pmatrix}. M 0 = ( 0 r × r C 0 0 E 0 ) , N 0 = ( α I r 0 B 0 G 0 ) .
They satisfy M 0 B R = B R E F , t M_0B_R=B_RE_{F,t} M 0 B R = B R E F , t and B L N 0 = G F , t B L B_LN_0=G_{F,t}B_L B L N 0 = G F , t B L . The constant linear operator W ↦ M 0 W + W N 0 W\mapsto M_0W+WN_0 W ↦ M 0 W + W N 0 consequently restricts on each dual fiber to (30) . Equivalently its dual is the quotient operator Y ↦ N 0 Y + Y M 0 Y\mapsto N_0Y+YM_0 Y ↦ N 0 Y + Y M 0 ; this is well defined because it sends L t o p L_{\mathrm{top}} L top into itself. In the chosen coordinates this is exactly the vector field in the statement. The trace condition in (33) makes it a stabilizer.
Finally, the fiber parameters are unique in the fixed gauge, the maps B 0 ↦ K L B 0 B_0\mapsto K_LB_0 B 0 ↦ K L B 0 and C 0 ↦ C 0 K R C_0\mapsto C_0K_R C 0 ↦ C 0 K R are injective, and the dual fibers span Y ∗ \mathcal{Y}^* Y ∗ . Thus the displayed parameters are independent as operators. Their dimension is
r k + k r + k 2 + k 2 + 1 − 1 = 2 r k + 2 k 2 , rk+kr+k^2+k^2+1-1=2rk+2k^2, r k + k r + k 2 + k 2 + 1 − 1 = 2 r k + 2 k 2 ,
as claimed.
Stabilizer dimension and distinctness
Write y = ( S , U , V , Z ) y=(S,U,V,Z) y = ( S , U , V , Z ) . Every stabilizer has δ x = T ( x ) \delta x=T(x) δ x = T ( x ) , as proved above. In the kernel of this projection, write δ y = B ( x ) + R ( y ) \delta y=B(x)+R(y) δ y = B ( x ) + R ( y ) . The two terms have different x x x -degrees after differentiation, namely 2 h + 1 2h+1 2 h + 1 and 2 h 2h 2 h . Thus they vanish separately. For a fixed generic x x x , surjectivity of compression and conciseness of the determinant give π x B ( x ) = 0 \pi_xB(x)=0 π x B ( x ) = 0 . These equations are precisely the matrix syzygy for B S B_S B S , the left and right vector syzygies for the k k k columns and rows of B U B_U B U , B V B_V B V , and B Z = 0 B_Z=0 B Z = 0 . By Lemma 9.3, their total dimension is 14 h 2 − 4 + 2 r k 14h^2-4+2rk 14 h 2 − 4 + 2 r k . Lemma 9.4 gives 2 k 2 + 2 r k 2k^2+2rk 2 k 2 + 2 r k for the pure y y y -action. Adding the image and these two kernels yields
dim Stab ( H h , k ) = ( 4 h 2 + 3 ) + ( 14 h 2 − 4 + 2 r k ) + ( 2 k 2 + 2 r k ) = 18 h 2 + 4 r k + 2 k 2 − 1 = 2 n 2 − 1. \begin{aligned}
\dim\operatorname{Stab}(H_{h,k})=(4h^2+3)+(14h^2-4+2rk)+(2k^2+2rk) \\
&=18h^2+4rk+2k^2-1=2n^2-1.
\end{aligned} dim Stab ( H h , k ) = ( 4 h 2 + 3 ) + ( 14 h 2 − 4 + 2 r k ) + ( 2 k 2 + 2 r k ) = 18 h 2 + 4 r k + 2 k 2 − 1 = 2 n 2 − 1.
Since H h , k H_{h,k} H h , k is a boundary point, the stabilizer dimension 2 n 2 − 1 2n^2-1 2 n 2 − 1 proves that its orbit closure is a boundary component.
The translation dimension in (25) recovers h h h , so it distinguishes the higher-corank components from one another. To distinguish them from all odd-skew components, use the codimension of the linear span of the largest singular component. It is 2 2 2 here. For r = 3 r=3 r = 3 in the odd-skew family it is 3 3 3 ; for every odd r ≥ 5 r \ge5 r ≥ 5 it is 0 0 0 , since the rank-≤ r − 3 \le r-3 ≤ r − 3 skew matrices contain all rank-two skew matrices and therefore span the full skew space. Thus the two families remain distinct even when their translation dimensions agree. Equal orbit closures would have equal dense orbits, so these orbit invariants distinguish the components. This finishes the proof of Theorem 9.1 .
Corollary 9.5. For every n ≥ 3 n \ge3 n ≥ 3 , the determinant boundary has at least
⌊ n − 1 2 ⌋ + ⌊ n 3 ⌋ − 1 \left\lfloor\frac{n-1}{2}\right\rfloor+\left\lfloor\frac{n}{3}\right\rfloor-1 ⌊ 2 n − 1 ⌋ + ⌊ 3 n ⌋ − 1
distinct concise components. Including the traceless component gives
# Irr ( ∂ O ‾ n ) ≥ ⌊ n − 1 2 ⌋ + ⌊ n 3 ⌋ . \#\operatorname{Irr}\left(\partial\overline{\mathcal O}_n\right)\ge\left\lfloor\frac{n-1}{2}\right\rfloor+\left\lfloor\frac{n}{3}\right\rfloor. # Irr ( ∂ O n ) ≥ ⌊ 2 n − 1 ⌋ + ⌊ 3 n ⌋ .
The h = 1 h=1 h = 1 block construction recovers the r = 3 r=3 r = 3 odd-skew family and is therefore counted only once. For n = 6 , 7 , 9 , 12 n=6,7,9,12 n = 6 , 7 , 9 , 12 , the displayed total lower bounds are respectively 4 , 5 , 7 , 9 4,5,7,9 4 , 5 , 7 , 9 . They remain lower bounds, not a classification of all components.
Boundary components from middle Koszul maps
The middle Koszul maps give a third family of boundary components. The first seed has size ten, so arbitrary padding supplies every size n ≥ 10 n \ge10 n ≥ 10 . These are the forms used for the exact-complexity bound in Theorem 1.1 .
We apply the leading-coefficient construction of [12 Section 6.7.1] to these classical spaces of exterior-multiplication matrices. We prove the component and padding statements uniformly in the exterior degree. Landsberg and Manivel [13 Section 3.1, Proposition 3] identify the common constant compression kernel in a more general rectangular setting. We include an elementary proof in the square setting used here.
Recall that Λ p V \Lambda^p V Λ p V has a basis indexed by p p p -element subsets of a chosen basis of V V V . The exterior product is alternating on vectors, so a product with a repeated vector is zero. Fix p ≥ 2 p \ge2 p ≥ 2 and put
d = 2 p + 1 , V = C d , E = Λ p V , r = ( 2 p + 1 p ) , R = ( 2 p p ) , c = r − R = ( 2 p p − 1 ) . d=2p+1,\qquad V=\mathbb{C}^{d},\qquad E=\Lambda^p V,\qquad r=\binom{2p+1}{p},\qquad R=\binom{2p}{p},\qquad c=r-R=\binom{2p}{p-1}. d = 2 p + 1 , V = C d , E = Λ p V , r = ( p 2 p + 1 ) , R = ( p 2 p ) , c = r − R = ( p − 1 2 p ) .
Here d d d is the dimension of the seed space, r r r is the matrix size, R R R is its rank at nonzero v v v , and c c c is its corank. For p = 2 p=2 p = 2 ,
d = 5 , r = 10 , R = 6 , c = 4. d=5,\qquad r=10,\qquad R=6,\qquad c=4. d = 5 , r = 10 , R = 6 , c = 4.
Choose a volume form on V V V . It identifies Λ p + 1 V \Lambda^{p+1}V Λ p + 1 V with E ∗ E^* E ∗ , so exterior multiplication by v ∈ V v\in V v ∈ V becomes a square matrix A ( v ) : E → E ∗ A(v):E\to E^* A ( v ) : E → E ∗ . Equivalently, up to a fixed sign,
A ( v ) ( α , β ) = vol ( v ∧ α ∧ β ) , α , β ∈ E . A(v)(\alpha,\beta)=\operatorname{vol}(v\wedge\alpha\wedge\beta),\qquad\alpha,\beta\in E. A ( v ) ( α , β ) = vol ( v ∧ α ∧ β ) , α , β ∈ E .
In the exterior-power bases, every entry is zero or a signed coordinate of v v v . The transpose is ( − 1 ) p A ( v ) (-1)^pA(v) ( − 1 ) p A ( v ) .
For v ≠ 0 v \ne0 v = 0 , choose a basis with first vector v v v . A basis wedge is killed by multiplication by v v v exactly when it already contains v v v . Counting wedges that do not contain it gives rank R R R ; counting those that do gives kernel dimension c c c . Thus the left and right kernels are both
U v = v ∧ Λ p − 1 V , dim U v = c . U_v = v \wedge\Lambda^{p-1} V,\qquad\dim U_v = c. U v = v ∧ Λ p − 1 V , dim U v = c .
Let L p = { A ( v ) : v ∈ V } \mathcal{L}_p = \{A(v): v \in V\} L p = { A ( v ) : v ∈ V } , a d d d -dimensional matrix space, and choose a complement S p \mathcal{S}_p S p in Mat r \operatorname{Mat}_r Mat r . For k ≥ 0 k \ge0 k ≥ 0 , set n = r + k n = r + k n = r + k and let S ∈ S p S \in\mathcal{S}_p S ∈ S p , U ∈ Mat r , k U \in\operatorname{Mat}_{r,k} U ∈ Mat r , k , V 0 ∈ Mat k , r V_0 \in\operatorname{Mat}_{k,r} V 0 ∈ Mat k , r , and Z ∈ Mat k Z \in\operatorname{Mat}_k Z ∈ Mat k be independent coefficient variables. Define
K p , k ( v , S , U , V 0 , Z ) = [ t c ] det ( A ( v ) + t S U t V 0 Z ) . (34) \mathcal{K}_{p,k}(v,S,U,V_0,Z) = [t^c]\det\begin{pmatrix} A(v)+tS & U \\ tV_0 & Z \end{pmatrix}. \tag*{(34)} K p , k ( v , S , U , V 0 , Z ) = [ t c ] det ( A ( v ) + tS t V 0 U Z ) . (34)
The notation [ t c ] [t^c] [ t c ] denotes the coefficient of t c t^c t c . The empty blocks are omitted for k = 0 k=0 k = 0 , and there are n 2 n^2 n 2 independent variables. Balancing the entries selected from block rows and block columns shows that this coefficient has degree R R R in v v v and degree c + k c+k c + k in the other variables. In particular its total degree is n n n .
Theorem 10.1. For every p ≥ 2 p \ge2 p ≥ 2 and k ≥ 0 k \ge0 k ≥ 0 , the polynomial K p , k \mathcal{K}_{p,k} K p , k is concise in n 2 n^2 n 2 variables and satisfies
dim Stab ( K p , k ) = 2 n 2 − 1. \dim\operatorname{Stab}(\mathcal{K}_{p,k}) = 2n^2 - 1. dim Stab ( K p , k ) = 2 n 2 − 1. Its orbit closure is a determinant-boundary component. For fixed n n n , the components obtained from all p ≥ 2 p \ge2 p ≥ 2 with ( 2 p + 1 p ) ≤ n \binom{2p+1}{p} \le n ( p 2 p + 1 ) ≤ n are pairwise distinct. They are different from every component in the odd-skew and higher-cork block families.
Constant and linear kernel relations
For each nonzero v v v , the coefficient matrix determines a bilinear form on E ⊕ C k E \oplus\mathbb{C}^k E ⊕ C k . Restricting both of its arguments to U v ⊕ C k U_v \oplus\mathbb{C}^k U v ⊕ C k gives a smaller matrix of size c + k c+k c + k . This is the compression used in the proof.
We determine the constant matrices, and the matrices linear in v v v , whose restrictions vanish for every v v v . The constant kernel is used to prove conciseness. The linear kernel determines the stabilizer maps from v v v to the coefficient variables.
Lemma 10.2. The common constant kernel of B ↦ B ∣ U v × U v B \mapsto B|_{U_v \times U_v} B ↦ B ∣ U v × U v , for B ∈ E ∗ ⊗ E ∗ B \in E^* \otimes E^* B ∈ E ∗ ⊗ E ∗ , is L p \mathcal{L}_p L p , of dimension d d d . The space of matrix-valued linear maps B ( v ) B(v) B ( v ) satisfying B ( v ) ∣ U v × U v = 0 B(v)|_{U_v \times U_v} = 0 B ( v ) ∣ U v × U v = 0 for every v v v has dimension 2 r 2 − 1 2r^2 - 1 2 r 2 − 1 . After quotienting by the L p \mathcal{L}_p L p -valued maps, its dimension is 2 r 2 − d 2 − 1 2r^2 - d^2 - 1 2 r 2 − d 2 − 1 . The space of linear maps b ( v ) ∈ E ∗ b(v) \in E^* b ( v ) ∈ E ∗ annihilating U v U_v U v has dimension r r r ; the analogous left and right assertions coincide.
Proof. We begin with the constant matrices. A bilinear form B B B on Λ p V \Lambda^p V Λ p V can be viewed as a function of 2 p 2p 2 p vector arguments, alternating within each of two groups of p p p arguments. Its vanishing on ( v ∧ α , v ∧ β ) (v \wedge\alpha,v \wedge\beta) ( v ∧ α , v ∧ β ) , polarized in v v v , makes it antisymmetric under exchanging the first arguments of the two blocks. Together with alternation within each block, this generates all transpositions. The tensor is therefore fully alternating. Conversely every fully alternating tensor vanishes when the two blocks share v v v . The common kernel is thus Λ 2 p V ∗ \Lambda^{2p}V^* Λ 2 p V ∗ , identified with L p \mathcal{L}_p L p by the volume form. The same polarization argument for a tensor in V ∗ ⊗ Λ p V ∗ V^* \otimes\Lambda^p V^* V ∗ ⊗ Λ p V ∗ proves the vector assertion: its kernel is Λ p + 1 V ∗ \Lambda^{p+1}V^* Λ p + 1 V ∗ , of dimension r r r .
For matrices B ( v ) B(v) B ( v ) linear in v v v , we compute the kernel by decomposing the relevant tensor spaces into irreducible GL ( V ) \operatorname{GL}(V) GL ( V ) -modules.
Put W = V ∗ W = V^* W = V ∗ . The map in question is the double exterior coproduct followed by symmetric multiplication,
W ⊗ Λ p W ⊗ Λ p W ⟶ Sym 3 W ⊗ Λ p − 1 W ⊗ Λ p − 1 W . (35) W \otimes\Lambda^p W \otimes\Lambda^p W \longrightarrow\operatorname{Sym}^3 W \otimes\Lambda^{p-1} W \otimes\Lambda^{p-1} W. \tag*{(35)} W ⊗ Λ p W ⊗ Λ p W ⟶ Sym 3 W ⊗ Λ p − 1 W ⊗ Λ p − 1 W . (35)
Entrywise this is exactly B ( v ) ↦ K ( v ) T B ( v ) K ( v ) B(v) \mapsto K(v)^T B(v)K(v) B ( v ) ↦ K ( v ) T B ( v ) K ( v ) , where K ( v ) : Λ p − 1 V → Λ p V K(v): \Lambda^{p-1}V \to\Lambda^p V K ( v ) : Λ p − 1 V → Λ p V is exterior multiplication by v v v . We write S λ W S_\lambda W S λ W for the irreducible GL ( W ) \operatorname{GL}(W) GL ( W ) representation with symmetry type indexed by the partition λ \lambda λ . The partition is pictured by rows of boxes of lengths λ i \lambda_i λ i ; its width is the largest row length. The Pieri rules describe what happens to these types when one tensors with an exterior or symmetric power [12 Section 8.1.3] . Applied here, they first give
Λ p W ⊗ Λ p W = ⨁ j = 0 p S ( 2 j , 1 2 p − 2 j ) W . \Lambda^p W \otimes\Lambda^p W = \bigoplus_{j=0}^{p} S_{(2^j,1^{2p-2j})}W. Λ p W ⊗ Λ p W = j = 0 ⨁ p S ( 2 j , 1 2 p − 2 j ) W .
Tensoring with W W W gives the multiplicity decomposition
⨁ j = 1 p A j ⊕ 2 ⨁ j = 1 p C j ⊕ C 0 , (36) \bigoplus_{j=1}^{p} A_j \oplus2\bigoplus_{j=1}^{p} C_j \oplus C_0, \tag*{(36)} j = 1 ⨁ p A j ⊕ 2 j = 1 ⨁ p C j ⊕ C 0 , (36)
where
A j = S ( 3 , 2 j − 1 , 1 2 p − 2 j ) W , C j = S ( 2 j , 1 2 p + 1 − 2 j ) W , C 0 = Λ d W . A_j = S_{(3,2^{j-1},1^{2p-2j})}W,\qquad C_j = S_{(2^j,1^{2p+1-2j})}W,\qquad C_0 = \Lambda^d W. A j = S ( 3 , 2 j − 1 , 1 2 p − 2 j ) W , C j = S ( 2 j , 1 2 p + 1 − 2 j ) W , C 0 = Λ d W .
The map is GL ( W ) \operatorname{GL}(W) GL ( W ) -equivariant, so an irreducible summand can have nonzero image only if the same representation occurs in the target. The target of (35) has no summand of width at most two: tensoring with Sym 3 W \operatorname{Sym}^3 W Sym 3 W adds a horizontal strip of three boxes, using three columns. Hence every C j C_j C j maps to zero.
Each A j A_j A j occurs once in the source. To show that its restriction is nonzero, choose a basis e 1 , … , e d e_1,\ldots,e_d e 1 , … , e d of W W W and put W ′ = ⟨ e 2 , … , e d ⟩ W'=\langle e_2,\ldots,e_d\rangle W ′ = ⟨ e 2 , … , e d ⟩ . Choose a nonzero highest-weight vector z j z_j z j of shape ( 2 j − 1 , 1 2 p − 2 j ) (2^{j-1},1^{2p-2j}) ( 2 j − 1 , 1 2 p − 2 j ) in Λ p − 1 W ′ ⊗ Λ p − 1 W ′ \Lambda^{p-1}W' \otimes\Lambda^{p-1}W' Λ p − 1 W ′ ⊗ Λ p − 1 W ′ , which occurs by Pieri. Wedging e 1 e_1 e 1 into each factor gives
w j = ( ( e 1 ∧ − ) ⊗ ( e 1 ∧ − ) ) z j . w_j = ((e_1 \wedge-) \otimes(e_1 \wedge-))z_j. w j = (( e 1 ∧ − ) ⊗ ( e 1 ∧ − )) z j .
This is a highest-weight vector of shape ( 2 j , 1 2 p − 2 j ) (2^j,1^{2p-2j}) ( 2 j , 1 2 p − 2 j ) : the raising operators within W ′ W' W ′ kill z j z_j z j , while a raising operator introducing e 1 e_1 e 1 creates a repeated factor. Consequently e 1 ⊗ w j e_1 \otimes w_j e 1 ⊗ w j is a highest-weight vector of type A j A_j A j . Its image under (35) has e 1 3 e_1^3 e 1 3 coefficient z j ≠ 0 z_j \ne0 z j = 0 . Every A j A_j A j occurs once in the source, so complete reducibility now identifies the kernel as C 0 ⊕ 2 ⨁ j = 1 p C j C_0 \oplus2\bigoplus_{j=1}^{p} C_j C 0 ⊕ 2 ⨁ j = 1 p C j . Finally, another application of Pieri gives
Λ p W ⊗ Λ p + 1 W = ⨁ j = 0 p C j . \Lambda^p W \otimes\Lambda^{p+1}W = \bigoplus_{j=0}^{p} C_j. Λ p W ⊗ Λ p + 1 W = j = 0 ⨁ p C j .
Both exterior powers have dimension r r r , and dim C 0 = 1 \dim C_0 = 1 dim C 0 = 1 . The kernel dimension is therefore 1 + 2 ( r 2 − 1 ) = 2 r 2 − 1 1 + 2(r^2 - 1) = 2r^2 - 1 1 + 2 ( r 2 − 1 ) = 2 r 2 − 1 . The space of linear maps V → L p V \to\mathcal{L}_p V → L p has dimension d 2 d^2 d 2 and is contained in this kernel, giving the quotient assertion.
In particular, the common constant kernel of restriction from Mat n \operatorname{Mat}_n Mat n to the spaces U v ⊕ C k U_v \oplus\mathbb{C}^k U v ⊕ C k is exactly L p \mathcal{L}_p L p in the upper-left block. Indeed the seed assertion was just proved, the spaces U v U_v U v span E E E , and the remaining Z Z Z block is unrestricted. Thus the coefficient space can be identified intrinsically for the calculation as
Y = Mat n / ( L p ) t o p . \mathcal{Y} = \operatorname{Mat}_n / (\mathcal{L}_p)_{\mathrm{top}}. Y = Mat n / ( L p ) top .
The choice of a complement in (34) does not affect the resulting equivalence class.
Local determinant form and multiplicity loci
The compressed determinant gives a local description of the polynomial and its multiplicity loci. The largest component of a suitable multiplicity locus recovers the seed directions, as the singular locus did for the preceding families.
All coefficients below t c t^c t c in (34) vanish. On det Z ≠ 0 \det Z \ne0 det Z = 0 , this follows by Schur elimination and the constant rank R R R of A ( v ) A(v) A ( v ) for v ≠ 0 v \ne0 v = 0 ; it then holds identically. The leading coefficient is nonzero. For t ≠ 0 t \ne0 t = 0 the displayed matrix parametrizes Mat n \operatorname{Mat}_n Mat n invertibly, so its normalized limit belongs to O ‾ n \overline{\mathcal{O}}_n O n .
Near a fixed nonzero v v v , choose bases on a Zariski open chart that reduce A ( v ) A(v) A ( v ) to an invertible R × R R \times R R × R block and a zero c × c c \times c c × c block. Eliminating the invertible block shows that
K p , k ( v , Y ) = u ( v ) det ( κ v ( Y ) ) , u ( v ) ≠ 0 , (37) \mathcal{K}_{p,k}(v,Y)=u(v)\det\bigl(\kappa_v(Y)\bigr),\qquad u(v)\ne0, \tag*{(37)} K p , k ( v , Y ) = u ( v ) det ( κ v ( Y ) ) , u ( v ) = 0 , (37)
where κ v \kappa_v κ v is restriction to the left and right kernels, including the padding coordinates. It is surjective: restriction of an arbitrary bilinear form to U v ⊕ C k U_v\oplus\mathbb{C}^k U v ⊕ C k is surjective, and ( L p ) t o p (\mathcal{L}_p)_{\mathrm{top}} ( L p ) top is in its kernel. The entries of κ v ( Y ) \kappa_v(Y) κ v ( Y ) are therefore part of a local coordinate system, in which the polynomial is a generic determinant times a unit. The multiplicity at a point is the least degree of a nonzero term after translating that point to the origin. For a determinant, this is the corank. Hence the multiplicity here is
( c + k ) − rank κ v ( Y ) . (c+k)-\operatorname{rank}\kappa_v(Y). ( c + k ) − rank κ v ( Y ) .
Since every monomial has degree R R R in v v v , the multiplicity is at least R R R along v = 0 v=0 v = 0 . Put N = n 2 N=n^2 N = n 2 . Thus the linear space
M p = { v = 0 } × Y , dim M p = N − d , M_p=\{v=0\}\times\mathcal{Y},\qquad\dim M_p=N-d, M p = { v = 0 } × Y , dim M p = N − d ,
is contained in its multiplicity-at-least-R R R locus. Outside M p M_p M p , the same multiplicity locus is empty if c + k < R c+k<R c + k < R , and otherwise has codimension R 2 R^2 R 2 by (37) and the usual rank codimension for a square matrix. Since R 2 > d R^2>d R 2 > d for p ≥ 2 p\ge2 p ≥ 2 , M p M_p M p is the unique largest irreducible component of this multiplicity locus. Every linear stabilizer preserves it. Thus every infinitesimal stabilizer satisfies δ v = T ( v ) \delta v=T(v) δ v = T ( v ) , with no coefficient-variable contribution to δ v \delta v δ v .
To prove conciseness, consider a constant directional derivative annihilating the polynomial. Its v v v -direction has v v v -degree R − 1 R-1 R − 1 and the coefficient directions have degree R R R , so they vanish separately. Changing the basis of V V V changes the exterior-power bases in a compatible way. In particular, it gives a linear GL ( V ) \operatorname{GL}(V) GL ( V ) action on the coefficient quotient, under which K p , k \mathcal{K}_{p,k} K p , k transforms by a character. Hence its space of annihilating constant v v v -directions is GL ( V ) \operatorname{GL}(V) GL ( V ) -stable. Since V V V is irreducible, that space is zero or all of V V V ; the latter would contradict its positive v v v -degree and nonvanishing. For the remaining direction, (37) and conciseness of the determinant imply κ v ( Y 0 ) = 0 \kappa_v(Y_0)=0 κ v ( Y 0 ) = 0 for every v ≠ 0 v\ne0 v = 0 . The common constant kernel in the quotient is zero, so Y 0 = 0 Y_0=0 Y 0 = 0 .
The same local form also gives
dim Sing ( K p , k ) = N − 4. (38) \dim\operatorname{Sing}(\mathcal{K}_{p,k})=N-4. \tag*{(38)} dim Sing ( K p , k ) = N − 4. (38)
Indeed on v ≠ 0 v\ne0 v = 0 the singular determinant locus has codimension four, since c + k ≥ 4 c+k\ge4 c + k ≥ 4 . The omitted locus M p M_p M p has codimension d ≥ 5 d\ge5 d ≥ 5 and cannot increase that dimension.
The coefficient stabilizer under padding
For each v v v , a coefficient stabilizer induces row and column operations on the compressed determinant. The induced operations must arise from a single linear map on the coefficient space. Intersections of the kernel spaces impose the compatibility conditions used below.
The next lemma also holds for p = 1 p=1 p = 1 . Fix p ≥ 1 p\ge1 p ≥ 1 and k ≥ 0 k\ge0 k ≥ 0 , and put
V = C 2 p + 1 , E = ⋀ p V , r = ( 2 p + 1 p ) , c = ( 2 p p − 1 ) , T = C k . V=\mathbb{C}^{2p+1},\qquad E=\bigwedge^{p}V,\qquad r=\binom{2p+1}{p},\qquad c=\binom{2p}{p-1},\qquad T=\mathbb{C}^{k}. V = C 2 p + 1 , E = ⋀ p V , r = ( p 2 p + 1 ) , c = ( p − 1 2 p ) , T = C k .
For 0 ≠ v ∈ V 0\ne v\in V 0 = v ∈ V , set
U v = v ∧ ⋀ p − 1 V ⊂ E , P v = U v ⊕ T . U_v=v\wedge\bigwedge^{p-1}V\subset E,\qquad P_v=U_v\oplus T. U v = v ∧ ⋀ p − 1 V ⊂ E , P v = U v ⊕ T .
Then dim U v = c \dim U_v=c dim U v = c . A volume form on V V V identifies the middle Koszul map A ( v ) : ⋀ p V → ⋀ p + 1 V A(v):\bigwedge^{p}V\to\bigwedge^{p+1}V A ( v ) : ⋀ p V → ⋀ p + 1 V with a bilinear form on E E E ; its left and right kernels are both U v U_v U v . Let L = A ( V ) \mathcal{L}=A(V) L = A ( V ) , embedded in the upper-left block of Mat r + k \operatorname{Mat}_{r+k} Mat r + k , and regard the coefficient space as
Y = Mat r + k / L . \mathcal{Y}=\operatorname{Mat}_{r+k}/\mathcal{L}. Y = Mat r + k / L .
By the common-kernel calculation above, restriction of bilinear forms to U v × U v U_v\times U_v U v × U v , for all v ≠ 0 v\ne0 v = 0 , has common kernel L \mathcal{L} L . Since the U v U_v U v span E E E , its padded version says that the dual compression spaces
W v = P v ⊗ P v ⊂ Y ∗ satisfy ∑ v ≠ 0 W v = Y ∗ . (39) W_v=P_v\otimes P_v\subset\mathcal{Y}^{*}\qquad\text{satisfy}\qquad\sum_{v\ne0}W_v=\mathcal{Y}^{*}. \tag*{(39)} W v = P v ⊗ P v ⊂ Y ∗ satisfy v = 0 ∑ W v = Y ∗ . (39)
Thus the images of the dual compression maps span the full dual coefficient space. Splitting off the padding directions gives four blocks. If W 0 = ∑ v ≠ 0 U v ⊗ U v W_0=\sum_{v\ne0}U_v\otimes U_v W 0 = ∑ v = 0 U v ⊗ U v , then
Y ∗ = W 0 ⊕ ( E ⊗ T ) ⊕ ( T ⊗ E ) ⊕ ( T ⊗ T ) . (40) \mathcal{Y}^{*}=W_0\oplus(E\otimes T)\oplus(T\otimes E)\oplus(T\otimes T). \tag*{(40)} Y ∗ = W 0 ⊕ ( E ⊗ T ) ⊕ ( T ⊗ E ) ⊕ ( T ⊗ T ) . (40)
For each v ≠ 0 v\ne0 v = 0 , compression gives a surjection π v : Y → P v ∗ ⊗ P v ∗ \pi_v:\mathcal{Y}\to P_v^{*}\otimes P_v^{*} π v : Y → P v ∗ ⊗ P v ∗ . The padded leading determinant coefficient, denoted by K p , k ( v , Y ) \mathcal{K}_{p,k}(v,Y) K p , k ( v , Y ) , is a nonzero scalar multiple of det c + k ( π v Y ) \det_{c+k}(\pi_vY) det c + k ( π v Y ) when v v v is fixed. The scalar can depend on the choice of bases and on v v v ; it has no effect on the following infinitesimal statement.
Lemma 10.3. The space of linear operators R : Y → Y R:\mathcal{Y}\to\mathcal{Y} R : Y → Y satisfying
D Y K p , k ( v , Y ) [ R ( Y ) ] = 0 for all v , Y D_Y\mathcal{K}_{p,k}(v,Y)[R(Y)]=0\qquad\text{for all }v,Y D Y K p , k ( v , Y ) [ R ( Y )] = 0 for all v , Y has dimension
dim p p , k = 2 k 2 + 2 r k . \dim\mathfrak{p}_{p,k}=2k^2+2rk. dim p p , k = 2 k 2 + 2 r k . More precisely, its elements are uniquely parametrized by
α ∈ C , C , D ∈ End ( T ) , Φ , Ψ ∈ Hom ( E , T ) , c α + tr C + tr D = 0. \alpha\in\mathbb{C},\qquad C,D\in\operatorname{End}(T),\qquad\Phi,\Psi\in\operatorname{Hom}(E,T),\qquad c\alpha+\operatorname{tr}C+\operatorname{tr}D=0. α ∈ C , C , D ∈ End ( T ) , Φ , Ψ ∈ Hom ( E , T ) , c α + tr C + tr D = 0. Their dual action on (43) is the restriction of
A ⊗ id + id ⊗ B , A = ( 0 0 Φ C ) , B = ( α I E 0 Ψ D ) . (41) A\otimes\operatorname{id}+\operatorname{id}\otimes B,\qquad A=\begin{pmatrix}0&0\\ \Phi&C\end{pmatrix},\qquad B=\begin{pmatrix}\alpha I_E&0\\ \Psi&D\end{pmatrix}. \tag*{(41)} A ⊗ id + id ⊗ B , A = ( 0 Φ 0 C ) , B = ( α I E Ψ 0 D ) . (41) where the matrices act on E ⊕ T E\oplus T E ⊕ T . For k = 0 k=0 k = 0 , the trace condition forces α = 0 \alpha=0 α = 0 , so the space is zero.
We use two elementary facts. First, the determinant Lie algebra consists of the left–right actions H ↦ E H + H F H\mapsto EH+HF H ↦ E H + H F with tr E + tr F = 0 \operatorname{tr}E+\operatorname{tr}F=0 tr E + tr F = 0 . The following argument gives the required description directly. If L L L is tangent to the determinant hypersurface, fix nonzero left and right null vectors a , b a,b a , b and consider
Z a , b = { H : a t H = 0 , H b = 0 } . Z_{a,b}=\{H:a^{\mathsf{t}}H=0,\ Hb=0\}. Z a , b = { H : a t H = 0 , H b = 0 } .
Rank-( m − 1 ) (m-1) ( m − 1 ) matrices are dense in this space. At these matrices the adjugate is a nonzero multiple of b a t ba^{\mathsf t} b a t , so tangency gives a t L ( H ) b = 0 a^{\mathsf t}L(H)b=0 a t L ( H ) b = 0 throughout Z a , b Z_{a,b} Z a , b . In particular,
L ( x y t ) ∈ x C m t + C m y t L(xy^{\mathsf t}) \in x\mathbb{C}^{m\mathsf t}+\mathbb{C}^{m}y^{\mathsf t} L ( x y t ) ∈ x C m t + C m y t
for all nonzero x , y x,y x , y . At a matrix unit E i j E_{ij} E ij , this means that L ( E i j ) L(E_{ij}) L ( E ij ) is supported on its row and column. Applying the same condition to ( e i + e s ) e j t (e_i+e_s)e_j^{\mathsf t} ( e i + e s ) e j t and e i ( e j + e t ) t e_i(e_j+e_t)^{\mathsf t} e i ( e j + e t ) t makes the corresponding off-diagonal coefficients independent of i i i and j j j , respectively. Subtracting the resulting left–right action leaves L ( E i j ) = q i j E i j L(E_{ij})=q_{ij}E_{ij} L ( E ij ) = q ij E ij . The condition at ( e i + e s ) ( e j + e t ) t (e_i+e_s)(e_j+e_t)^{\mathsf t} ( e i + e s ) ( e j + e t ) t gives q i j − q i t − q s j + q s t = 0 q_{ij}-q_{it}-q_{sj}+q_{st}=0 q ij − q i t − q s j + q s t = 0 , hence q i j = u i + w j q_{ij}=u_i+w_j q ij = u i + w j . These are the diagonal left–right actions. Finally, D det H [ E H + H F ] = ( tr E + tr F ) det H D\det H[EH+HF]=(\operatorname{tr}E+\operatorname{tr}F)\det H D det H [ E H + H F ] = ( tr E + tr F ) det H gives the trace condition. The size-one case is immediate. Taking duals gives the same description as E ⊗ id + id ⊗ F E\otimes\operatorname{id}+\operatorname{id}\otimes F E ⊗ id + id ⊗ F on the dual matrix space.
Second, a linear endomorphism of an exterior power that preserves every decomposable line is scalar. Indeed, the wedge products of basis vectors are eigenvectors. If two such products differ in exactly one factor, their sum, with an appropriate choice of sign, is also decomposable, so their eigenvalues agree. The graph of subsets of fixed cardinality joined by these replacements is connected. The assertion follows; exterior powers of dimension one are included.
Proof of Lemma 10.3. Let R R R satisfy the stated identity. For fixed v ≠ 0 v\ne0 v = 0 , it preserves ker π v \ker\pi_v ker π v . To see this, take Y 0 ∈ ker π v Y_0\in\ker\pi_v Y 0 ∈ ker π v , replace Y Y Y by Y + s Y 0 Y+sY_0 Y + s Y 0 in the identity, and extract the coefficient of s s s . It gives
D H det [ π v R ( Y 0 ) ] = 0 D_H\det[\pi_vR(Y_0)]=0 D H det [ π v R ( Y 0 )] = 0
for all H ∈ Mat c + k H\in\operatorname{Mat}_{c+k} H ∈ Mat c + k .
The first derivatives of the generic determinant are linearly independent, so π v R ( Y 0 ) = 0 \pi_vR(Y_0)=0 π v R ( Y 0 ) = 0 . Consequently R ∗ R^* R ∗ preserves W v W_v W v , and the determinant Lie-algebra calculation gives
R ∗ ∣ W v = a v ⊗ id + id ⊗ b v , tr a v + tr b v = 0 , (42) R^*|_{W_v}=a_v\otimes\operatorname{id}+\operatorname{id}\otimes b_v,\qquad\operatorname{tr}a_v+\operatorname{tr}b_v=0, \tag*{(42)} R ∗ ∣ W v = a v ⊗ id + id ⊗ b v , tr a v + tr b v = 0 , (42)
for a v , b v ∈ End ( P v ) a_v,b_v\in\operatorname{End}(P_v) a v , b v ∈ End ( P v ) .
We compare the actions a v , b v a_v,b_v a v , b v on intersections of the spaces W v W_v W v to obtain the constant block formula in the statement. For subspaces E i E_i E i of a common vector space,
⋂ i ( E i ⊗ E i ) = ( ⋂ i E i ) ⊗ ( ⋂ i E i ) . \bigcap_i(E_i\otimes E_i)=\left(\bigcap_iE_i\right)\otimes\left(\bigcap_iE_i\right). i ⋂ ( E i ⊗ E i ) = ( i ⋂ E i ) ⊗ ( i ⋂ E i ) .
This follows from E ⊗ E = ( E ⊗ H ) ∩ ( H ⊗ E ) E\otimes E=(E\otimes H)\cap(H\otimes E) E ⊗ E = ( E ⊗ H ) ∩ ( H ⊗ E ) , where H H H is the ambient space. Since ⋂ v ≠ 0 U v = 0 \bigcap_{v\ne0}U_v=0 ⋂ v = 0 U v = 0 , it follows that ⋂ v ≠ 0 W v = T ⊗ T \bigcap_{v\ne0}W_v=T\otimes T ⋂ v = 0 W v = T ⊗ T . Thus R ∗ R^* R ∗ preserves T ⊗ T T\otimes T T ⊗ T . In (42) , projection to ( P v / T ) ⊗ P v (P_v/T)\otimes P_v ( P v / T ) ⊗ P v and to P v ⊗ ( P v / T ) P_v\otimes(P_v/T) P v ⊗ ( P v / T ) separately shows that both a v a_v a v and b v b_v b v preserve T T T . When T = 0 T=0 T = 0 , this statement is automatic.
If v 1 , … , v p v_1,\ldots,v_p v 1 , … , v p are linearly independent and q = v 1 ∧ ⋯ ∧ v p q=v_1\wedge\cdots\wedge v_p q = v 1 ∧ ⋯ ∧ v p , then
⋂ i = 1 p U v i = C q , ⋂ i = 1 p W v i = ( C q ⊕ T ) ⊗ ( C q ⊕ T ) . \bigcap_{i=1}^{p}U_{v_i}=\mathbb{C}q,\qquad\bigcap_{i=1}^{p}W_{v_i}=(\mathbb{C}q\oplus T)\otimes(\mathbb{C}q\oplus T). i = 1 ⋂ p U v i = C q , i = 1 ⋂ p W v i = ( C q ⊕ T ) ⊗ ( C q ⊕ T ) .
The first equality says that a degree-p p p exterior form divisible by each of p p p independent vectors is divisible by their product. Hence, for each fixed v v v and each decomposable q ∈ U v q\in U_v q ∈ U v , both a v a_v a v and b v b_v b v preserve C q ⊕ T \mathbb{C}q\oplus T C q ⊕ T : apply the same two quotient projections to the invariant tensor square just displayed. On P v / T = U v ≅ ⋀ p − 1 ( V / C v ) P_v/T=U_v\cong\bigwedge^{p-1}(V/\mathbb{C}v) P v / T = U v ≅ ⋀ p − 1 ( V / C v ) , the induced operators therefore preserve every decomposable line. The second elementary fact makes them scalar. This also covers p = 1 p=1 p = 1 , when U v U_v U v is a line.
The induced actions on U v U_v U v are scalar. The decomposition in (42) is unique up to ( a v , b v ) ↦ ( a v + s I , b v − s I ) (a_v,b_v)\mapsto(a_v+sI,b_v-sI) ( a v , b v ) ↦ ( a v + s I , b v − s I ) . Use this freedom to make the U v U_v U v -block of a v a_v a v zero. We obtain unique local data
a v = ( 0 0 ϕ v C v ) , b v = ( α v I U v 0 ψ v D v ) . (43) a_v=\begin{pmatrix}0&0\\ \phi_v&C_v\end{pmatrix},\qquad b_v=\begin{pmatrix}\alpha_v I_{U_v}&0\\ \psi_v&D_v\end{pmatrix}. \tag*{(43)} a v = ( 0 ϕ v 0 C v ) , b v = ( α v I U v ψ v 0 D v ) . (43)
where ϕ v , ψ v : U v → T \phi_v,\psi_v:U_v\to T ϕ v , ψ v : U v → T .
The scalar α v \alpha_v α v is independent of v v v , also for p = 1 p=1 p = 1 . To prove this, project R ∗ ∣ W 0 R^*|_{W_0} R ∗ ∣ W 0 to W 0 W_0 W 0 in (40) . Every decomposable p p p -form q q q belongs to some U v U_v U v , and (43) gives
R 00 ∗ ( q ⊗ q ) = α v ( q ⊗ q ) . R^*_{00}(q\otimes q)=\alpha_v(q\otimes q). R 00 ∗ ( q ⊗ q ) = α v ( q ⊗ q ) .
The cone of these nonzero tensors q ⊗ q q\otimes q q ⊗ q is irreducible: it is the image of the irreducible cone of decomposable p p p -forms under q ↦ q ⊗ q q\mapsto q\otimes q q ↦ q ⊗ q . A linear operator has only finitely many eigenvalues, and its eigenspaces are closed. This irreducible cone must lie in a single eigenspace. Therefore every α v \alpha_v α v has the same value. No assertion about the span of decomposable squares in Sym 2 ( U v ) \operatorname{Sym}^2(U_v) Sym 2 ( U v ) is required here.
If k = 0 k=0 k = 0 , all remaining blocks in (43) are absent. The trace condition is c α = 0 c\alpha=0 c α = 0 , whence α = 0 \alpha=0 α = 0 and R ∗ = 0 R^*=0 R ∗ = 0 by (39) . We may henceforth assume k > 0 k>0 k > 0 .
We next show that C v , D v C_v,D_v C v , D v are constant. Project the restriction of R ∗ R^* R ∗ to E ⊗ T E\otimes T E ⊗ T back to E ⊗ T E\otimes T E ⊗ T . For q ∈ U v q\in U_v q ∈ U v and t ∈ T t\in T t ∈ T , the result is
q ⊗ D v t . q\otimes D_vt. q ⊗ D v t .
Choose a basis of T T T and express this projected operator as a matrix of linear operators E → E E\to E E → E . Each such operator preserves every decomposable line in E E E , since every decomposable q q q belongs to some U v U_v U v . The second elementary fact makes every matrix entry a scalar operator. Thus the projection is id E ⊗ D \operatorname{id}_E\otimes D id E ⊗ D for a single D ∈ End ( T ) D\in\operatorname{End}(T) D ∈ End ( T ) , and D v = D D_v=D D v = D . Similarly the projection of R ∗ ∣ T ⊗ E R^*|_{T\otimes E} R ∗ ∣ T ⊗ E onto T ⊗ E T\otimes E T ⊗ E acts locally by ( C v + α I T ) ⊗ id (C_v+\alpha I_T)\otimes\operatorname{id} ( C v + α I T ) ⊗ id . It follows that C v = C C_v=C C v = C for a single C C C .
Finally, the maps ϕ v \phi_v ϕ v and ψ v \psi_v ψ v extend to linear maps Φ , Ψ : E → T \Phi,\Psi:E\to T Φ , Ψ : E → T . The following argument applies uniformly for p ≥ 1 p\ge1 p ≥ 1 . Let F F F be the projection of R ∗ ∣ E ⊗ T R^*|_{E\otimes T} R ∗ ∣ E ⊗ T onto T ⊗ T T\otimes T T ⊗ T . Formula (43) gives
F ( q ⊗ t ) = ϕ v ( q ) ⊗ t ( q ∈ U v ) . F(q\otimes t)=\phi_v(q)\otimes t\qquad(q\in U_v). F ( q ⊗ t ) = ϕ v ( q ) ⊗ t ( q ∈ U v ) .
Fix 0 ≠ t 0 ∈ T 0\ne t_0\in T 0 = t 0 ∈ T and η ∈ T ∗ \eta\in T^* η ∈ T ∗ with η ( t 0 ) = 1 \eta(t_0)=1 η ( t 0 ) = 1 , and set
Φ ( u ) = ( id T ⊗ η ) F ( u ⊗ t 0 ) . \Phi(u)=(\operatorname{id}_T\otimes\eta)F(u\otimes t_0). Φ ( u ) = ( id T ⊗ η ) F ( u ⊗ t 0 ) .
This is linear in u u u and restricts to ϕ v \phi_v ϕ v on every U v U_v U v . Since the U v U_v U v span E E E , the identity F ( u ⊗ t ) = Φ ( u ) ⊗ t F(u\otimes t)=\Phi(u)\otimes t F ( u ⊗ t ) = Φ ( u ) ⊗ t holds on all of E ⊗ T E\otimes T E ⊗ T . The analogous projection from T ⊗ E T\otimes E T ⊗ E gives a global Ψ \Psi Ψ restricting to every ψ v \psi_v ψ v .
Thus R ∗ R^* R ∗ agrees with (41) on every W v W_v W v , and hence on all of Y ∗ Y^* Y ∗ by their spanning property. The local trace condition becomes the single equation
c α + tr C + tr D = 0. c\alpha+\operatorname{tr}C+\operatorname{tr}D=0. c α + tr C + tr D = 0.
Conversely, the operators in (41) preserve every P v ⊗ P v P_v\otimes P_v P v ⊗ P v , and this trace equation makes their induced actions annihilate every compressed determinant. They therefore give exactly the required pure coefficient stabilizers.
The parameters are unique. Indeed the local left–right ambiguity has already been removed by the zero U v U_v U v -block of a v a_v a v ; agreement on all the U v U_v U v determines Φ , Ψ \Phi,\Psi Φ , Ψ because these subspaces span E E E . There are 1 + 2 k 2 + 2 r k 1+2k^2+2rk 1 + 2 k 2 + 2 r k parameters and one nonzero linear trace condition, proving the dimension formula. ∙იდ
Stabilizer dimension and distinctness
The stabilizer dimension is the sum of three contributions: the action on v v v , the maps from v v v to the coefficient variables, and the coefficient action. The image of the stabilizer on the v v v -variables has dimension at most d 2 d^2 d 2 , and all of g l ( V ) \mathfrak{gl}(V) gl ( V ) occurs. Indeed, if G = ⋀ p g G=\bigwedge^p g G = ⋀ p g for g ∈ GL ( V ) g\in\operatorname{GL}(V) g ∈ GL ( V ) , exterior equivariance gives
A ( g v ) = L g A ( v ) R g , L g = ( det g ) G − T , R g = G − 1 . A(gv)=L_gA(v)R_g,\qquad L_g=(\det g)G^{-T},\qquad R_g=G^{-1}. A ( g v ) = L g A ( v ) R g , L g = ( det g ) G − T , R g = G − 1 .
The changes S ↦ L g S R g S\mapsto L_gSR_g S ↦ L g S R g , U ↦ L g U U\mapsto L_gU U ↦ L g U , V 0 ↦ V 0 R g V_0\mapsto V_0R_g V 0 ↦ V 0 R g , Z ↦ Z Z\mapsto Z Z ↦ Z descend to the coefficient quotient and multiply (34) by det L g det R g = ( det g ) r − 2 c \det L_g\det R_g=(\det g)^{r-2c} det L g det R g = ( det g ) r − 2 c . This scalar character is cancelled infinitesimally by a dilation of Y Y Y , because its coefficient degree is c + k > 0 c+k>0 c + k > 0 .
In the kernel of this projection, the pieces δ Y = B ( v ) \delta Y=B(v) δ Y = B ( v ) and δ Y = R 0 ( Y ) \delta Y=R_0(Y) δ Y = R 0 ( Y ) have different v v v -degrees in the stabilizer identity and are independent. By (37) , the first kind is a stabilizer exactly when κ v ( B ( v ) ) = 0 \kappa_v(B(v))=0 κ v ( B ( v )) = 0 for every v v v . Lemma 10.2 gives 2 r 2 − d 2 − 1 2r^2-d^2-1 2 r 2 − d 2 − 1 dimensions in the S S S block and r r r in each of the k k k added columns and rows. The Z Z Z block must vanish. This contributes 2 r 2 − d 2 − 1 + 2 r k 2r^2-d^2-1+2rk 2 r 2 − d 2 − 1 + 2 r k dimensions. Lemma 10.3 gives 2 k 2 + 2 r k 2k^2+2rk 2 k 2 + 2 r k for the pure coefficient-variable stabilizer. Therefore
dim Stab ( K p , k ) = d 2 + ( 2 r 2 − d 2 − 1 + 2 r k ) + ( 2 k 2 + 2 r k ) = 2 ( r + k ) 2 − 1. \dim\operatorname{Stab}(\mathcal{K}_{p,k})=d^2+(2r^2-d^2-1+2rk)+(2k^2+2rk)=2(r+k)^2-1. dim Stab ( K p , k ) = d 2 + ( 2 r 2 − d 2 − 1 + 2 r k ) + ( 2 k 2 + 2 r k ) = 2 ( r + k ) 2 − 1.
The orbit has codimension one in O ‾ n \overline{O}_n O n and is outside the determinant orbit, whose stabilizer has dimension 2 n 2 − 2 2n^2-2 2 n 2 − 2 . Its closure is therefore a boundary component.
The dimensions of the multiplicity loci distinguish the components for different parameter values. Fix n n n and let p < q p<q p < q be two allowed indices. Write R p = ( 2 p p ) R_p=\binom{2p}{p} R p = ( p 2 p ) and d p = 2 p + 1 d_p=2p+1 d p = 2 p + 1 . The multiplicity-at-least-R p R_p R p locus of the p p p -form has dimension N − d p N-d_p N − d p . For the q q q -form, its part with nonzero seed vector has codimension R p 2 R_p^2 R p 2 or is empty, by its local determinant form; the remaining part has dimension N − d q N-d_q N − d q . Thus its dimension is at most
max { N − d q , N − R p 2 } < N − d p . \max\{N-d_q,N-R_p^2\}<N-d_p. max { N − d q , N − R p 2 } < N − d p .
because d q > d p d_q>d_p d q > d p and R p 2 > d p R_p^2>d_p R p 2 > d p . This invariant separates the two linear-equivalence classes. If their orbit closures were the same component, their dense orbits would coincide, so the components are distinct. Finally (38) separates this family from the odd-skew and higher-corank block families, whose largest singular components have dimension N − 3 N-3 N − 3 .
Consequently this family adds
κ ( n ) = # { p ≥ 2 : ( 2 p + 1 p ) ≤ n } . \kappa(n)=\#\left\{p\ge2:\binom{2p+1}{p}\le n\right\}. κ ( n ) = # { p ≥ 2 : ( p 2 p + 1 ) ≤ n } .
distinct concise components. Its first seed sizes are 10, 35, 126, and 462.
The component count
The three families are pairwise distinct. The largest singular components of the odd-skew and block families have codimension three. The codimensions of their linear spans separate the two families, while their translation dimensions recover the parameter within each family. The Koszul family has singular locus of codimension four, and its multiplicity loci recover its parameter. These invariants are preserved by linear changes of variables. The omitted endpoint cases h = 1 h=1 h = 1 and p = 1 p=1 p = 1 both recover the 3 × 3 3\times3 3 × 3 skew construction.
Corollary 11.1. For every n ≥ 3 n \ge3 n ≥ 3 , put κ ( n ) = # { p ≥ 2 : ( 2 p + 1 p ) ≤ n } \kappa(n)=\#\{p \ge2 : \binom{2p+1}{p} \le n\} κ ( n ) = # { p ≥ 2 : ( p 2 p + 1 ) ≤ n } . The determinant boundary has at least
⌊ n − 1 2 ⌋ + ⌊ n 3 ⌋ + κ ( n ) − 1 \left\lfloor\frac{n-1}{2} \right\rfloor+\left\lfloor\frac{n}{3} \right\rfloor+\kappa(n)-1 ⌊ 2 n − 1 ⌋ + ⌊ 3 n ⌋ + κ ( n ) − 1 distinct concise components, and at least
# Irr ( ∂ O ‾ n ) ≥ ⌊ n − 1 2 ⌋ + ⌊ n 3 ⌋ + κ ( n ) \#\operatorname{Irr}(\partial\overline{\mathcal{O}}_n) \ge\left\lfloor\frac{n-1}{2} \right\rfloor+\left\lfloor\frac{n}{3} \right\rfloor+\kappa(n) # Irr ( ∂ O n ) ≥ ⌊ 2 n − 1 ⌋ + ⌊ 3 n ⌋ + κ ( n ) components in total after including the traceless component.
Table 11.1.
The count concerns explicitly constructed components; it does not assert that these exhaust the boundary. In particular, known extra components at small sizes need not be represented by this uniform lower bound.
A general component criterion
The preceding stabilizer calculations admit a general formulation. Two quantities determine the excess orbit codimension: first-order deformations of the singular matrix space beyond row and column operations, and additional block-diagonal symmetries of its leading determinant coefficient. Under the recognition and conciseness hypotheses below, the orbit codimension is one plus the sum of these two excesses.
Let L ⊂ Mat r L \subset\operatorname{Mat}_r L ⊂ Mat r be a d d d -dimensional linear space of singular matrices with generic rank R < r R<r R < r . Put c = r − R c=r-R c = r − R , and assume d < r R d<rR d < r R . Choose Mat r = L ⊕ S 0 \operatorname{Mat}_r=L\oplus S_0 Mat r = L ⊕ S 0 , write A A A and S S S for its two generic matrices, and set
P ( A , S ) = [ t c ] det ( A + t S ) . P(A,S)=[t^c]\det(A+tS). P ( A , S ) = [ t c ] det ( A + tS ) .
The form P P P has degree R R R in A A A and degree c c c in S S S . Assume that it is concise in its r 2 r^2 r 2 variables and that every linear vector field annihilating P P P preserves the S 0 S_0 S 0 -directions. Explicitly, its A A A -part has the form δ A = T ( A ) \delta A=T(A) δ A = T ( A ) , with no S S S -to-A A A term.
Let F d , R \mathcal{F}_{d,R} F d , R be the closed subscheme of the Grassmannian Gr ( d , Mat r ) \operatorname{Gr}(d,\operatorname{Mat}_r) Gr ( d , Mat r ) defined by the vanishing of all ( R + 1 ) (R+1) ( R + 1 ) -minors on the parametrized matrix space. Thus it parametrizes d d d -dimensional linear spaces of matrices of rank at most R R R . This is the Fano scheme used below.
A tangent vector at [ L ] [L] [ L ] is a first-order perturbation A ↦ A + ε F ( A ) A\mapsto A+\varepsilon F(A) A ↦ A + εF ( A ) , with ε 2 = 0 \varepsilon^2=0 ε 2 = 0 , satisfying the linearized rank equations. Define
f = dim T [ L ] F d , R , n ( L ) = { ( C , D ) ∈ Mat r ⊕ Mat r : C A + A D ∈ L for every A ∈ L } , ν = dim n ( L ) , b = dim { ( T , J ) ∈ End ( L ) ⊕ End ( S 0 ) : D A P [ T ( A ) ] + D S P [ J ( S ) ] = 0 } . \begin{aligned}
f&=\dim T_{[L]}\mathcal{F}_{d,R},\\
\mathfrak{n}(L)&=\{(C,D)\in\operatorname{Mat}_r\oplus\operatorname{Mat}_r:CA+AD\in L\text{ for every }A\in L\},\qquad\nu=\dim\mathfrak{n}(L),\\
b&=\dim\{(T,J)\in\operatorname{End}(L)\oplus\operatorname{End}(S_0):D_A P[T(A)]+D_S P[J(S)]=0\}.
\end{aligned} f n ( L ) b = dim T [ L ] F d , R , = {( C , D ) ∈ Mat r ⊕ Mat r : C A + A D ∈ L for every A ∈ L } , ν = dim n ( L ) , = dim {( T , J ) ∈ End ( L ) ⊕ End ( S 0 ) : D A P [ T ( A )] + D S P [ J ( S )] = 0 } .
The space n ( L ) \mathfrak{n}(L) n ( L ) is the normalizer Lie algebra for row and column operations. The number b b b is the dimension of the block-diagonal stabilizer, whose elements act separately on A A A and S S S .
There are 2 r 2 2r^{2} 2 r 2 parameters in an arbitrary pair of row and column changes. Thus the orbit of L L L under these changes has dimension f 0 = 2 r 2 − ν f_{0}=2r^{2}-\nu f 0 = 2 r 2 − ν . Since its tangent space is contained in the Fano tangent space, δ = f − f 0 ≥ 0 \delta=f-f_{0}\geq0 δ = f − f 0 ≥ 0 .
Theorem 12.1. Under the stated hypotheses, b ≥ ν − 1 b\geq\nu-1 b ≥ ν − 1 . Put ρ = b − ( ν − 1 ) ≥ 0 \rho=b-(\nu-1)\geq0 ρ = b − ( ν − 1 ) ≥ 0 . Then P P P belongs to the determinant boundary and
dim Stab ( P ) = f + b = 2 r 2 − 1 + δ + ρ , codim GL r ⋅ P ‾ = 1 + δ + ρ . \dim\operatorname{Stab}(P)=f+b=2r^{2}-1+\delta+\rho,\qquad\operatorname{codim}_{\overline{\operatorname{GL}_{r}\cdot P}}=1+\delta+\rho. dim Stab ( P ) = f + b = 2 r 2 − 1 + δ + ρ , codim GL r ⋅ P = 1 + δ + ρ . Its orbit closure is a boundary component if and only if δ = ρ = 0 \delta=\rho=0 δ = ρ = 0 . Equivalently, the Fano tangent space consists exactly of infinitesimal row and column changes, and the block-diagonal stabilizer has the minimum dimension ν − 1 \nu-1 ν − 1 .
Proof. The seed-to-coefficient stabilizers are the tangent vectors counted by f f f . At a generic A ∈ L A\in L A ∈ L of rank R R R , let κ A ( B ) \kappa_{A}(B) κ A ( B ) be the induced map ker A → coker A \ker A\to\operatorname{coker} A ker A → coker A , of matrix size c c c . A perturbation B B B preserves rank at most R R R to first order exactly when it maps ker A \ker A ker A into im A \operatorname{im} A im A . This is the condition κ A ( B ) = 0 \kappa_{A}(B)=0 κ A ( B ) = 0 , so the tangent space to the rank-at-most-R R R variety at A A A is ker κ A \ker\kappa_{A} ker κ A . Since L L L lies in that variety, κ A ( L ) = 0 \kappa_{A}(L)=0 κ A ( L ) = 0 , and hence κ A : S 0 → Mat c \kappa_{A}:S_{0}\to\operatorname{Mat}_{c} κ A : S 0 → Mat c is surjective. After choosing bases locally on the rank-R R R locus,
P ( A , S ) = u ( A ) det c κ A ( S ) , u ( A ) ≠ 0. P(A,S)=u(A)\det\nolimits_{c}\kappa_{A}(S),\qquad u(A)\neq0. P ( A , S ) = u ( A ) det c κ A ( S ) , u ( A ) = 0.
A tangent vector to Gr ( d , Mat r ) \operatorname{Gr}(d,\operatorname{Mat}_{r}) Gr ( d , Mat r ) is a map F : L → Mat r / L F:L\to\operatorname{Mat}_{r}/L F : L → Mat r / L , represented in S 0 S_{0} S 0 . Linearizing the defining minors shows that it lies in T [ L ] F d , R T_{[L]}F_{d,R} T [ L ] F d , R exactly when κ A ( F ( A ) ) = 0 \kappa_{A}(F(A))=0 κ A ( F ( A )) = 0 for every generic A ∈ L A\in L A ∈ L . It is enough to test the generic locus because these are polynomial identities; at smaller rank all first derivatives of the ( R + 1 ) (R+1) ( R + 1 ) -minors already vanish. By surjectivity of κ A \kappa_{A} κ A and conciseness of the determinant, the same condition is equivalent to D S P ( A , S ) [ F ( A ) ] = 0 D_{S}P(A,S)[F(A)]=0 D S P ( A , S ) [ F ( A )] = 0 for all A , S A,S A , S . Thus the A A A -to-S S S stabilizers have dimension exactly f f f .
By the recognition hypothesis, every stabilizer has δ A = T ( A ) \delta A=T(A) δ A = T ( A ) and δ S = F ( A ) + J ( S ) \delta S=F(A)+J(S) δ S = F ( A ) + J ( S ) . The F F F -part has bidegree ( R + 1 , c − 1 ) (R+1,c-1) ( R + 1 , c − 1 ) in the stabilizer identity, while the other two parts have bidegree ( R , c ) (R,c) ( R , c ) . They vanish separately, proving dim Stab ( P ) = f + b \dim\operatorname{Stab}(P)=f+b dim Stab ( P ) = f + b .
To prove b ≥ ν − 1 b\geq\nu-1 b ≥ ν − 1 , we combine the action of the trace-zero normalizer with the weighted Euler action. We first compute the kernel of the normalizer action. For ( C , D ) ∈ n ( L ) (C,D)\in\mathfrak{n}(L) ( C , D ) ∈ n ( L ) let N ( X ) = C X + X D N(X)=CX+XD N ( X ) = C X + X D and impose tr C + tr D = 0 \operatorname{tr}C+\operatorname{tr}D=0 tr C + tr D = 0 . This is one independent condition on n ( L ) \mathfrak{n}(L) n ( L ) , since ( I r , 0 ) (I_{r},0) ( I r , 0 ) belongs to it. Write
N ( A ) ∈ L , N ( S ) = H ( S ) + J ( S ) , H ( S ) ∈ L , J ( S ) ∈ S 0 . N(A)\in L,\qquad N(S)=H(S)+J(S),\qquad H(S)\in L,\quad J(S)\in S_{0}. N ( A ) ∈ L , N ( S ) = H ( S ) + J ( S ) , H ( S ) ∈ L , J ( S ) ∈ S 0 .
The determinant identity D det X [ N ( X ) ] = ( tr C + tr D ) det X = 0 D\det_{X}[N(X)]=(\operatorname{tr}C+\operatorname{tr}D)\det X=0 D det X [ N ( X )] = ( tr C + tr D ) det X = 0 , applied to X = A + t S X=A+tS X = A + tS , shows at its coefficient of t c t^{c} t c that δ A = N ( A ) \delta A=N(A) δ A = N ( A ) , δ S = J ( S ) \delta S=J(S) δ S = J ( S ) is a block-diagonal stabilizer of P P P . Suppose this induced field vanishes. Then N ( A ) = 0 N(A)=0 N ( A ) = 0 and N ( S ) = H ( S ) ∈ L N(S)=H(S)\in L N ( S ) = H ( S ) ∈ L . In that case
0 = D det A + t S [ t H ( S ) ] = t D A ( det ( A + t S ) ) [ H ( S ) ] . 0=D\det_{A+tS}[tH(S)]=tD_{A}(\det(A+tS))[H(S)]. 0 = D A + tS det [ t H ( S )] = t D A ( det ( A + tS )) [ H ( S )] .
The coefficient of t c + 1 t^{c+1} t c + 1 is D A P ( A , S ) [ H ( S ) ] = 0 D_{A}P(A,S)[H(S)]=0 D A P ( A , S ) [ H ( S )] = 0 . The recognition hypothesis forces the pure S S S -to-A A A field H H H to vanish. Hence N = 0 N=0 N = 0 on all of Mat r \operatorname{Mat}_{r} Mat r , whose only pair representatives are ( C , D ) = ( λ I r , − λ I r ) (C,D)=(\lambda I_{r},-\lambda I_{r}) ( C , D ) = ( λ I r , − λ I r ) . The trace-zero normalizer therefore has an image of dimension ν − 2 \nu-2 ν − 2 in the block-diagonal stabilizer.
Every such image has ordinary trace zero on L ⊕ S 0 = Mat r L \oplus S_0 = \operatorname{Mat}_r L ⊕ S 0 = Mat r : its trace equals that of the upper triangular ambient operator N N N , namely r ( tr C + tr D ) = 0 r(\operatorname{tr} C + \operatorname{tr} D) = 0 r ( tr C + tr D ) = 0 . The bidegree ( R , c ) (R,c) ( R , c ) of P P P also gives the weighted Euler stabilizer
δ A = c A , δ S = − R S . \delta A = cA,\qquad\delta S = -RS. δ A = c A , δ S = − R S .
Its ordinary trace is c d − R ( r 2 − d ) = r ( d − r R ) ≠ 0 cd - R(r^2-d) = r(d-rR) \ne0 c d − R ( r 2 − d ) = r ( d − r R ) = 0 . It is therefore independent of the normalizer image, giving b ≥ ( ν − 2 ) + 1 = ν − 1 b \ge(\nu-2)+1 = \nu-1 b ≥ ( ν − 2 ) + 1 = ν − 1 .
The normalized leading coefficient belongs to O ‾ r \overline{\mathcal{O}}_r O r , since A + t S A+tS A + tS is an invertible linear parametrization of Mat r \operatorname{Mat}_r Mat r for t ≠ 0 t \ne0 t = 0 . Using f = 2 r 2 − ν + δ f = 2r^2-\nu+\delta f = 2 r 2 − ν + δ and b = ν − 1 + ρ b = \nu-1+\rho b = ν − 1 + ρ gives dim Stab ( P ) = 2 r 2 − 1 + δ + ρ \dim\operatorname{Stab}(P) = 2r^2-1+\delta+\rho dim Stab ( P ) = 2 r 2 − 1 + δ + ρ . This is strictly greater than the determinant stabilizer dimension 2 r 2 − 2 2r^2-2 2 r 2 − 2 , so P P P is outside the orbit, and the displayed codimension formula follows. Finally, the determinant boundary is pure of codimension one [6 Corollary 5.4.1] , so the orbit closure is a component exactly when δ + ρ = 0 \delta+\rho=0 δ + ρ = 0 . Both terms are nonnegative, proving the criterion.
The equality δ = 0 \delta=0 δ = 0 expresses infinitesimal rigidity : every first-order deformation within the bounded-rank locus is induced by row and column operations. The equality ρ = 0 \rho=0 ρ = 0 gives the minimum block-diagonal stabilizer dimension. Both equalities are required, together with the stated conciseness and recognition hypotheses.
Infinitesimal rigidity of the block and Koszul spaces
We verify infinitesimal rigidity for the block and Koszul spaces. Every first-order deformation preserving the rank bound is induced by row and column operations. In the Koszul case, compare the rigidity of the associated homogeneous bundles discussed in [13 ] .
For this proof write F ~ d , R ( r ) \widetilde{\mathcal{F}}_{d,R}(r) F d , R ( r ) for the parameter scheme just described: it is the closed subscheme of Gr ( d , Mat r ) \operatorname{Gr}(d,\operatorname{Mat}_r) Gr ( d , Mat r ) parametrizing d d d -dimensional vector spaces of matrices of rank at most R R R . Thus its equations say that every ( R + 1 ) × ( R + 1 ) (R+1)\times(R+1) ( R + 1 ) × ( R + 1 ) minor vanishes identically on the parametrized space. The left–right group G = GL r × GL r G=\operatorname{GL}_r\times\operatorname{GL}_r G = GL r × GL r acts on this scheme by row and column operations.
Proposition 12.2. The following points of the indicated Fano schemes are smooth, and their left–right orbits are open in the unique irreducible components through them:
the block space L h \mathcal{L}_h L h of § 10 , with h ≥ 2 h \ge2 h ≥ 2 , r = 3 h r=3h r = 3 h , d = h 2 + 2 d=h^2+2 d = h 2 + 2 , and R = 2 h R=2h R = 2 h ;
the middle Koszul space L p \mathcal{L}_p L p of § 11 , with p ≥ 2 p \ge2 p ≥ 2 , r = ( 2 p + 1 p ) r=\binom{2p+1}{p} r = ( p 2 p + 1 ) , d = 2 p + 1 d=2p+1 d = 2 p + 1 , and R = ( 2 p p ) R=\binom{2p}{p} R = ( p 2 p ) .
Their tangent dimensions are, respectively,
dim T [ L h ] F ~ d , R ( r ) = 14 h 2 − 4 , dim T [ L p ] F ~ d , R ( r ) = 2 r 2 − d 2 − 1. \dim T_{[\mathcal{L}_h]}\widetilde{\mathcal{F}}_{d,R}(r)=14h^2-4,\qquad
\dim T_{[\mathcal{L}_p]}\widetilde{\mathcal{F}}_{d,R}(r)=2r^2-d^2-1. dim T [ L h ] F d , R ( r ) = 14 h 2 − 4 , dim T [ L p ] F d , R ( r ) = 2 r 2 − d 2 − 1. In particular, every first-order deformation of either seed as a bounded-rank linear space is induced by row and column operations.
Proof. For a matrix space L \mathcal{L} L of generic rank R R R , a Grassmannian tangent vector is a linear map B : L → Mat r / L B:\mathcal{L}\to\operatorname{Mat}_r/\mathcal{L} B : L → Mat r / L . Choose a lift and write the corresponding deformation as A ( x ) + ϵ B ( x ) A(x)+\epsilon B(x) A ( x ) + ϵ B ( x ) , where ϵ 2 = 0 \epsilon^2=0 ϵ 2 = 0 . At a rank-R R R matrix, the linearized determinantal equations are exactly
K L ( x ) B ( x ) K R ( x ) = 0 , (44) K_L(x)B(x)K_R(x)=0, \tag*{(44)} K L ( x ) B ( x ) K R ( x ) = 0 , (44)
where K R K_R K R and K L K_L K L span its right and left kernels. Equivalently, B ( x ) B(x) B ( x ) maps the right kernel into the image of A ( x ) A(x) A ( x ) . The rank-R R R locus is dense in L \mathcal{L} L , so these identities there are equivalent to the vanishing of the linearizations of all restricted ( R + 1 ) (R+1) ( R + 1 ) -minors as polynomials in x x x . Adding an L \mathcal{L} L -valued map to the lift of B B B only changes the parametrization of the seed. Consequently (44) computes the full scheme tangent space. Lemma 9.3 and Lemma 10.2 give the two dimensions in the statement.
It remains to show that row and column operations account for these entire tangent spaces. Write the infinitesimal left–right action as A ↦ C A + A D A \mapsto CA+AD A ↦ C A + A D , absorbing the usual sign of the second group factor into D D D . Its normalizer Lie algebra is
n ( L ) = { ( C , D ) ∈ M a t r 2 : C A + A D ∈ L for every A ∈ L } . \mathfrak{n}(\mathcal{L})=\{(C,D)\in\mathrm{Mat}_{r}^{2}:CA+AD\in\mathcal{L}\text{ for every }A\in\mathcal{L}\}. n ( L ) = {( C , D ) ∈ Mat r 2 : C A + A D ∈ L for every A ∈ L } .
The kernel of the induced action on L \mathcal{L} L consists of pairs with C A + A D = 0 CA+AD=0 C A + A D = 0 for every A ∈ L A\in\mathcal{L} A ∈ L . We show that this kernel is exactly
( C , D ) = ( − λ I r , λ I r ) , λ ∈ C . (45) (C,D)=(-\lambda I_r,\lambda I_r),\qquad\lambda\in\mathbb{C}. \tag*{(45)} ( C , D ) = ( − λ I r , λ I r ) , λ ∈ C . (45)
Such a pair must preserve every right kernel under D D D . For the block space, work on a = 1 a=1 a = 1 and put b = t b=t b = t . The right kernels are the graphs
im ( F − t I h I h ) . \operatorname{im}\begin{pmatrix}F\\-tI_h\\I_h\end{pmatrix}. im F − t I h I h .
Write D = ( D i j ) D=(D_{ij}) D = ( D ij ) in h × h h\times h h × h blocks. Preservation of these graphs gives
D 11 F − t D 12 + D 13 = F ( D 31 F − t D 32 + D 33 ) , D_{11}F-tD_{12}+D_{13}=F(D_{31}F-tD_{32}+D_{33}), D 11 F − t D 12 + D 13 = F ( D 31 F − t D 32 + D 33 ) ,
D 21 F − t D 22 + D 23 = − t ( D 31 F − t D 32 + D 33 ) . D_{21}F-tD_{22}+D_{23}=-t(D_{31}F-tD_{32}+D_{33}). D 21 F − t D 22 + D 23 = − t ( D 31 F − t D 32 + D 33 ) .
Comparison in F , t F,t F , t first gives D 21 = D 23 = D 31 = D 32 = 0 D_{21}=D_{23}=D_{31}=D_{32}=0 D 21 = D 23 = D 31 = D 32 = 0 and D 22 = D 33 D_{22}=D_{33} D 22 = D 33 . The first equation then gives D 12 = D 13 = 0 D_{12}=D_{13}=0 D 12 = D 13 = 0 and D 11 F = F D 33 D_{11}F=FD_{33} D 11 F = F D 33 for every F F F . Thus all three diagonal blocks are the same scalar matrix, and D = λ I r D=\lambda I_r D = λ I r .
For the Koszul space, D D D preserves every U v = v ∧ Λ p − 1 V U_v=v\wedge\Lambda^{p-1}V U v = v ∧ Λ p − 1 V . If v 1 , … , v p v_1,\ldots,v_p v 1 , … , v p are independent,
⋂ i = 1 p U v i = C ( v 1 ∧ ⋯ ∧ v p ) . \bigcap_{i=1}^{p}U_{v_i}=\mathbb{C}(v_1\wedge\cdots\wedge v_p). i = 1 ⋂ p U v i = C ( v 1 ∧ ⋯ ∧ v p ) .
Hence D D D preserves every decomposable line in Λ p V \Lambda^pV Λ p V and is scalar. One elementary verification is that the basis wedges are eigenvectors, and the sum of two basis wedges differing in one factor is decomposable; the connected graph of these replacements makes all eigenvalues equal.
In both cases the images of the matrices in L \mathcal{L} L span the entire target. For L h \mathcal{L}_h L h , the matrices obtained by varying a a a and b b b alone already have this property. For L p \mathcal{L}_p L p , the images v ∧ Λ p V v\wedge\Lambda^pV v ∧ Λ p V span Λ p + 1 V \Lambda^{p+1}V Λ p + 1 V before the volume-form identification. Thus ( C + λ I r ) A = 0 (C+\lambda I_r)A=0 ( C + λ I r ) A = 0 for every A A A implies C = − λ I r C=-\lambda I_r C = − λ I r , proving (45) .
The coordinate actions on the block seed admit explicit row–column lifts. Write E i j ( B ) E_{ij}(B) E ij ( B ) for a block matrix with B B B in position ( i , j ) (i,j) ( i , j ) . The change δ F = P F + F Q \delta F=PF+FQ δ F = P F + F Q is induced by C = diag ( − Q , P , P ) C=\operatorname{diag}(-Q,P,P) C = diag ( − Q , P , P ) and D = diag ( − P , Q , Q ) D=\operatorname{diag}(-P,Q,Q) D = diag ( − P , Q , Q ) in δ A = C A + A D \delta A=CA+AD δ A = C A + A D . The changes δ F = a E \delta F=aE δ F = a E and δ F = b G \delta F=bG δ F = b G are induced respectively by ( C , D ) = ( − E 31 ( E ) , − E 13 ( E ) ) (C,D)=(-E_{31}(E),-E_{13}(E)) ( C , D ) = ( − E 31 ( E ) , − E 13 ( E )) and ( C , D ) = ( E 21 ( G ) , E 12 ( G ) ) (C,D)=(E_{21}(G),E_{12}(G)) ( C , D ) = ( E 21 ( G ) , E 12 ( G )) . The off-diagonal generators δ a = b \delta a=b δ a = b and δ b = a \delta b=a δ b = a are induced by ( E 23 ( I h ) , E 32 ( I h ) ) (E_{23}(I_h),E_{32}(I_h)) ( E 23 ( I h ) , E 32 ( I h )) and ( E 32 ( I h ) , E 23 ( I h ) ) (E_{32}(I_h),E_{23}(I_h)) ( E 32 ( I h ) , E 23 ( I h )) . Finally, the diagonal generators δ a = a \delta a=a δ a = a and δ b = b \delta b=b δ b = b use ( diag ( 0 , 0 , − I h ) , diag ( I h , I h , 0 ) ) (\operatorname{diag}(0,0,-I_h),\operatorname{diag}(I_h,I_h,0)) ( diag ( 0 , 0 , − I h ) , diag ( I h , I h , 0 )) and ( diag ( 0 , − I h , 0 ) , diag ( I h , 0 , I h ) ) (\operatorname{diag}(0,-I_h,0),\operatorname{diag}(I_h,0,I_h)) ( diag ( 0 , − I h , 0 ) , diag ( I h , 0 , I h )) . Thus every coordinate action counted in (27) is induced by row and column operations.
For L h \mathcal{L}_h L h , the rank-drop-locus calculation and explicit row–column lifts leading to (27) show that the induced normalizer action on the seed coordinates has dimension 4 h 2 + 3 4h^2 + 3 4 h 2 + 3 . For L p \mathcal{L}_p L p , exterior equivariance realizes all of g l ( V ) \mathfrak{gl}(V) gl ( V ) , so the induced action has dimension d 2 d^2 d 2 . Adding the one-dimensional kernel in (45) gives
dim n ( L h ) = 4 h 2 + 4 , dim n ( L p ) = d 2 + 1. \dim\mathfrak{n}(\mathcal{L}_h) = 4h^2 + 4,\qquad\dim\mathfrak{n}(\mathcal{L}_p) = d^2 + 1. dim n ( L h ) = 4 h 2 + 4 , dim n ( L p ) = d 2 + 1.
The corresponding orbit dimensions are therefore
2 ( 3 h ) 2 − ( 4 h 2 + 4 ) = 14 h 2 − 4 , 2 r 2 − ( d 2 + 1 ) = 2 r 2 − d 2 − 1. 2(3h)^2 - (4h^2 + 4) = 14h^2 - 4,\qquad2r^2 - (d^2 + 1) = 2r^2 - d^2 - 1. 2 ( 3 h ) 2 − ( 4 h 2 + 4 ) = 14 h 2 − 4 , 2 r 2 − ( d 2 + 1 ) = 2 r 2 − d 2 − 1.
They equal the full Fano tangent dimensions. The differential of the orbit map is consequently surjective onto the tangent space.
Equality of these dimensions also proves smoothness. The smooth orbit through either point supplies a local dimension lower bound equal to the tangent dimension. The reverse inequality holds for every finite-type scheme, so the local dimension equals the embedding dimension, namely the tangent-space dimension. Its local ring is regular, and over C \mathbb{C} C the Fano scheme is smooth at the point. There is a unique irreducible component through it; the orbit has the same dimension as that component and is locally closed, hence open in it.
In the notation of Theorem 12.1, the proposition gives δ = 0 \delta= 0 δ = 0 for both seed families. Their block-diagonal stabilizer calculations give b = ν − 1 b = \nu- 1 b = ν − 1 , hence ρ = 0 \rho= 0 ρ = 0 as well. Thus both excess terms in the general component criterion vanish for these seeds.
A criterion for corank-one padding
For a seed of generic corank one, the orbit codimension after padding has an exact formula under the hypotheses below. The formula depends on the linear relations among the left and right kernel coordinates. Relations beyond those supplied by the seed matrix produce additional stabilizers for each added row and column. The resulting criterion applies both to the odd-skew family and to components whose positive paddings cease to be components.
Let L ⊂ Mat r L \subset\operatorname{Mat}_r L ⊂ Mat r , r ≥ 3 r \ge3 r ≥ 3 , consist of singular matrices and have generic rank r − 1 r - 1 r − 1 . Write A = A ( x ) A = A(x) A = A ( x ) for its generic matrix, choose a complement K K K with Mat r = L ⊕ K \operatorname{Mat}_r = L \oplus K Mat r = L ⊕ K , and let S S S be generic in K K K . Set
C ( x ) = adj A ( x ) = g ( x ) u ( x ) v ( x ) T , P ( x , S ) = tr ( C ( x ) S ) . C(x) = \operatorname{adj} A(x) = g(x)u(x)v(x)^{T},\qquad P(x,S) = \operatorname{tr}(C(x)S). C ( x ) = adj A ( x ) = g ( x ) u ( x ) v ( x ) T , P ( x , S ) = tr ( C ( x ) S ) .
Here g g g is a nonzero polynomial, and u u u , v v v are primitive right and left polynomial kernel vectors. Primitive means that the coordinates of each vector have no common nonconstant factor. We impose the following hypotheses:
(H1) P P P is concise in its r 2 r^2 r 2 coordinates: every one of these variables is essential.
(H2) The vectors u ( x ) u(x) u ( x ) , as x x x varies, span C r \mathbb{C}^r C r , and so do the vectors v ( x ) v(x) v ( x ) . Equivalently, both projective kernel images span P r − 1 \mathbb{P}^{r-1} P r − 1 .
(H3) Every linear vector field annihilating P P P has δ x = T ( x ) \delta x = T(x) δ x = T ( x ) , with no S S S -to-x x x part.
Condition (H3) requires the coefficient directions to be preserved by the seed stabilizer. In the preceding families, these directions were determined by the singular or multiplicity locus. No hypothesis on the padded stabilizer is imposed. Hypothesis (H1) and differentiation of det A = 0 \det A = 0 det A = 0 give
{ D ∈ Mat r : tr ( C ( x ) D ) = 0 for all x } = L . (46) \{D \in\operatorname{Mat}_r : \operatorname{tr}(C(x)D) = 0\text{ for all }x\} = L. \tag*{(46)} { D ∈ Mat r : tr ( C ( x ) D ) = 0 for all x } = L . (46)
Indeed this space contains L L L and intersects K K K trivially. Let π : Mat r → K \pi:\operatorname{Mat}_r\to K π : Mat r → K be projection along L L L ; thus tr ( C π ( D ) ) = tr ( C D ) \operatorname{tr}(C\pi(D))=\operatorname{tr}(CD) tr ( C π ( D )) = tr ( C D ) .
Define the left and right linear syzygy spaces by
Σ L = { b : L → C r linear : v ( x ) T b ( x ) = 0 } , Σ R = { c : L → ( C r ) ∗ linear : c ( x ) u ( x ) = 0 } . \Sigma_L=\{b:L\to\mathbb{C}^r\text{ linear}:v(x)^{\mathsf{T}}b(x)=0\},\qquad\Sigma_R=\{c:L\to(\mathbb{C}^r)^*\text{ linear}:c(x)u(x)=0\}. Σ L = { b : L → C r linear : v ( x ) T b ( x ) = 0 } , Σ R = { c : L → ( C r ) ∗ linear : c ( x ) u ( x ) = 0 } .
Put s L = dim Σ L s_L=\dim\Sigma_L s L = dim Σ L and s R = dim Σ R s_R=\dim\Sigma_R s R = dim Σ R , and define the excess syzygy dimension
ϵ = s L + s R − 2 r . \epsilon=s_L+s_R-2r. ϵ = s L + s R − 2 r .
The columns and rows of A A A provide r r r independent members of the respective spaces. For otherwise A A A would have a common kernel or cokernel vector, contrary to (H2). Consequently s L , s R ≥ r s_L,s_R\ge r s L , s R ≥ r and ϵ ≥ 0 \epsilon\ge0 ϵ ≥ 0 . The equality ϵ = 0 \epsilon=0 ϵ = 0 holds precisely when all linear syzygies are constant linear combinations of these columns or rows. Equivalently, adjoining a column or row linear in the same x x x variables without increasing generic rank gives only a redundant extension. This is the linear-syzygy condition related to unexpandability in [9 Section 2.3] ; no assertion about broader conventions for simultaneous extensions or additional variables is needed.
The other term in the codimension formula is the dimension of the seed stabilizers that extend to the padded polynomial. The extension condition is an identity involving only the seed. By (H3), every seed stabilizer has the form δ x = T ( x ) \delta x=T(x) δ x = T ( x ) , δ S = F ( x ) + R ( S ) \delta S=F(x)+R(S) δ S = F ( x ) + R ( S ) . Let h \mathfrak{h} h be the subspace for which there are constant E , F 0 ∈ Mat r E,F_0\in\operatorname{Mat}_r E , F 0 ∈ Mat r satisfying
d C x [ T ( x ) ] + F 0 C ( x ) + C ( x ) E = 0 , b l i f t = dim h . (47) dC_x[T(x)]+F_0C(x)+C(x)E=0,\qquad b_{\mathrm{lift}}=\dim\mathfrak{h}. \tag*{(47)} d C x [ T ( x )] + F 0 C ( x ) + C ( x ) E = 0 , b lift = dim h . (47)
For an element of h \mathfrak{h} h , hypothesis (H1) forces R ( S ) = π ( E S + S F 0 ) R(S)=\pi(ES+SF_0) R ( S ) = π ( E S + S F 0 ) ; the remaining F ( x ) F(x) F ( x ) satisfies tr ( C ( x ) F ( x ) ) = 0 \operatorname{tr}(C(x)F(x))=0 tr ( C ( x ) F ( x )) = 0 . All pure x x x -to-S S S seed stabilizers belong to h \mathfrak{h} h .
Theorem 13.1. Assume (H1)–(H3). For k ≥ 1 k\ge1 k ≥ 1 , set n = r + k n=r+k n = r + k and
Q k = det Z P ( x , S ) − tr ( C ( x ) U adj Z V ) , U ∈ Mat r , k , V ∈ Mat k , r , Z ∈ Mat k . (48) Q_k=\det ZP(x,S)-\operatorname{tr}(C(x)U\operatorname{adj}ZV),\qquad U\in\operatorname{Mat}_{r,k},\ V\in\operatorname{Mat}_{k,r},\ Z\in\operatorname{Mat}_k. \tag*{(48)} Q k = det Z P ( x , S ) − tr ( C ( x ) U adj Z V ) , U ∈ Mat r , k , V ∈ Mat k , r , Z ∈ Mat k . (48) Then Q k Q_k Q k is concise in n 2 n^2 n 2 variables, belongs to the determinant boundary, and
dim Stab ( Q k ) = b l i f t + 2 k 2 + 2 r k + k ( s L + s R ) , (49) \dim\operatorname{Stab}(Q_k)=b_{\mathrm{lift}}+2k^2+2rk+k(s_L+s_R), \tag*{(49)} dim Stab ( Q k ) = b lift + 2 k 2 + 2 r k + k ( s L + s R ) , (49) codim O n ‾ GL n 2 ⋅ Q k = b l i f t − 2 r 2 + 2 + k ϵ . (50) \operatorname{codim}_{\overline{\mathcal{O}_n}}\operatorname{GL}_{n^2}\cdot Q_k=b_{\mathrm{lift}}-2r^2+2+k\epsilon. \tag*{(50)} codim O n GL n 2 ⋅ Q k = b lift − 2 r 2 + 2 + k ϵ . (50) Here O n \mathcal{O}_n O n denotes the affine orbit of det n \det_n det n .
Proof. The coefficient of t t t in det ( A + t S U t V Z ) \det\begin{pmatrix}A+tS&U\\tV&Z\end{pmatrix} det ( A + tS t V U Z ) is Q k Q_k Q k . For t ≠ 0 t\ne0 t = 0 its entries parametrise Mat n \operatorname{Mat}_n Mat n invertibly, so the normalized limit lies in O n ‾ \overline{\mathcal{O}_n} O n . Directional derivatives distinguish Q k Q_k Q k from the determinant orbit. Every nonzero constant directional derivative of det n \det_n det n is irreducible. To see this, for D ≠ 0 D\ne0 D = 0 restrict f D ( X ) = tr ( adj ( X ) D ) f_D(X)=\operatorname{tr}(\operatorname{adj}(X)D) f D ( X ) = tr ( adj ( X ) D ) to GL n \operatorname{GL}_n GL n :
f D ( X ) = det ( X ) tr ( X − 1 D ) . f_D(X)=\det(X)\operatorname{tr}(X^{-1}D). f D ( X ) = det ( X ) tr ( X − 1 D ) .
Inversion identifies its zero locus there with the nonempty open part of a linear hyperplane, hence an irreducible reduced hypersurface. An additional component would have to be the determinant hypersurface, impossible since deg f D = n − 1 \deg f_D=n-1 deg f D = n − 1 ; reducedness on GL n \operatorname{GL}_n GL n also excludes a repeated factor. But a nonzero S S S -direction of Q k Q_k Q k has derivative det Z tr ( C D ) \det Z\operatorname{tr}(CD) det Z tr ( C D ) , a product of two positive-degree nonzero factors. Thus Q k ∉ O n Q_k\notin\mathcal{O}_n Q k ∈ / O n .
For conciseness, a constant x x x -directional derivative has x x x -degree r − 2 r-2 r − 2 , whereas all other directions have degree r − 1 r-1 r − 1 . Setting U = V = 0 U=V=0 U = V = 0 and using (H1) forces the x x x -direction to vanish. The coefficient involving S S S then forces the Z Z Z -direction to vanish. The coefficient linear in V V V , at Z = I k Z=I_k Z = I k , gives v ( x ) T U 0 V u ( x ) = 0 v(x)^{T}U_0V u(x)=0 v ( x ) T U 0 V u ( x ) = 0 for all x , V x,V x , V , forcing U 0 = 0 U_0=0 U 0 = 0 by (H2). Similarly V 0 = 0 V_0=0 V 0 = 0 , and then (H1) forces the S S S -direction to vanish.
The stabilizer calculation uses the same sequence of projections as in the odd-skew case. We determine the exact dimension of each image and kernel. Cofactor independence also applies for k = 1 k=1 k = 1 , when the unique cofactor is 1.
Step 1: preservation of the coefficient directions. Let δ Q k = 0 \delta Q_k=0 δ Q k = 0 . The part δ x = C S ( S ) + C U ( U ) + C V ( V ) + C Z ( Z ) \delta x=C_S(S)+C_U(U)+C_V(V)+C_Z(Z) δ x = C S ( S ) + C U ( U ) + C V ( V ) + C Z ( Z ) has x x x -degree r − 2 r-2 r − 2 and therefore satisfies a separate identity. At U = V = 0 U=V=0 U = V = 0 , its S 2 S^2 S 2 part says that the seed vector field δ x = C S ( S ) \delta x=C_S(S) δ x = C S ( S ) , δ S = 0 \delta S=0 δ S = 0 annihilates P P P . Hypothesis (H3) gives C S = 0 C_S=0 C S = 0 . Its S Z SZ S Z part and (H1) give C Z = 0 C_Z=0 C Z = 0 . Next set V = 0 V=0 V = 0 , and then U = 0 U=0 U = 0 , to obtain C U = C V = 0 C_U=C_V=0 C U = C V = 0 . Hence every padded stabilizer has δ x = T ( x ) \delta x=T(x) δ x = T ( x ) .
Step 2: the action on Z Z Z . Terms of x x x -degree r r r involving S S S force δ Z ( x ) = 0 \delta Z(x)=0 δ Z ( x ) = 0 ; terms quadratic in S S S force δ Z ( S ) = 0 \delta Z(S)=0 δ Z ( S ) = 0 . The S U SU S U terms are
P ( x , S ) tr ( adj Z δ Z ( U ) ) = g ( x ) v ( x ) T U adj Z δ V ( S ) u ( x ) . (51) P(x,S)\operatorname{tr}(\operatorname{adj}Z\,\delta Z(U))=g(x)v(x)^{T}U\operatorname{adj}Z\,\delta V(S)u(x). \tag*{(51)} P ( x , S ) tr ( adj Z δ Z ( U )) = g ( x ) v ( x ) T U adj Z δ V ( S ) u ( x ) . (51)
For generic x x x , restrict to v ( x ) T U = 0 v(x)^{T}U=0 v ( x ) T U = 0 and choose S S S with P ( x , S ) ≠ 0 P(x,S)\ne0 P ( x , S ) = 0 . The functional of U U U on the left annihilates this kernel for every such x x x . The annihilators corresponding to two nonproportional values of v v v are v 1 ⊗ ( C k ) ∗ v_1\otimes(\mathbb{C}^k)^* v 1 ⊗ ( C k ) ∗ and v 2 ⊗ ( C k ) ∗ v_2\otimes(\mathbb{C}^k)^* v 2 ⊗ ( C k ) ∗ , with zero intersection. Thus cofactor independence gives δ Z ( U ) = 0 \delta Z(U)=0 δ Z ( U ) = 0 . Equation (51) and (H2) then give δ V ( S ) = 0 \delta V(S)=0 δ V ( S ) = 0 . Similarly δ Z ( V ) = δ U ( S ) = 0 \delta Z(V)=\delta U(S)=0 δ Z ( V ) = δ U ( S ) = 0 . We have δ Z = J ( Z ) \delta Z=J(Z) δ Z = J ( Z ) , and the coefficient of S S S gives D ( det ) Z [ J ( Z ) ] = c det Z D(\det)Z[J(Z)]=c\det Z D ( det ) Z [ J ( Z )] = c det Z . The determinant similitude description gives J ( Z ) = L Z Z + Z R Z J(Z)=L_ZZ+ZR_Z J ( Z ) = L Z Z + Z R Z , an image of dimension 2 k 2 − 1 2k^2-1 2 k 2 − 1 . Every such image occurs: with c = tr L Z + tr R Z c=\operatorname{tr}L_Z+\operatorname{tr}R_Z c = tr L Z + tr R Z , take
δ Z = L Z Z + Z R Z , δ U = U R Z − c U , δ V = L Z V , δ S = − c S , δ x = 0. \delta Z=L_ZZ+ZR_Z,\qquad\delta U=UR_Z-cU,\qquad\delta V=L_ZV,\qquad\delta S=-cS,\qquad\delta x=0. δ Z = L Z Z + Z R Z , δ U = U R Z − c U , δ V = L Z V , δ S = − c S , δ x = 0.
Direct differentiation of det Z \det Z det Z and adj Z \operatorname{adj}Z adj Z verifies this lift.
Step 3: the kernel of the Z Z Z -projection. Set δ Z = 0 \delta Z=0 δ Z = 0 . The induced pair ( δ x , δ S ( x ) + δ S ( S ) ) (\delta x,\delta S(x)+\delta S(S)) ( δ x , δ S ( x ) + δ S ( S )) is a seed stabilizer. The x x x -linear pieces of δ U , δ V \delta U,\delta V δ U , δ V belong independently to Σ L \Sigma_L Σ L in each column and Σ R \Sigma_R Σ R in each row; they are genuine stabilizers and contribute k ( s L + s R ) k(s_L+s_R) k ( s L + s R ) dimensions. Subtract them. Terms involving only Z Z Z give δ S ( Z ) = 0 \delta S(Z)=0 δ S ( Z ) = 0 . The terms linear in V V V give, on invertible Z Z Z ,
δ S ( V ) = π ( δ U ( Z ) Z − 1 V ) . \delta S(V)=\pi\bigl(\delta U(Z)Z^{-1}V\bigr). δ S ( V ) = π ( δ U ( Z ) Z − 1 V ) .
The map B ↦ [ V ↦ π ( B V ) ] B\mapsto[V\mapsto\pi(BV)] B ↦ [ V ↦ π ( B V )] is injective: otherwise v ( x ) T B V u ( x ) = 0 v(x)^{T}BV u(x)=0 v ( x ) T B V u ( x ) = 0 for every x , V x,V x , V , and (H2) forces B = 0 B=0 B = 0 . Thus δ U ( Z ) Z − 1 = B 0 \delta U(Z)Z^{-1}=B_0 δ U ( Z ) Z − 1 = B 0 is constant, giving the stabilizers
δ U = B 0 Z , δ S = π ( B 0 V ) ; δ V = Z C 0 , δ S = π ( U C 0 ) , \delta U=B_0Z,\qquad\delta S=\pi(B_0V);\qquad\delta V=ZC_0,\qquad\delta S=\pi(UC_0), δ U = B 0 Z , δ S = π ( B 0 V ) ; δ V = Z C 0 , δ S = π ( U C 0 ) ,
with 2 r k 2rk 2 r k parameters in total. Subtract these as well. Terms quadratic in V V V give, by cofactor independence, ( v T δ U ( V ) ) a ( v u ) b = 0 (v^{T}\delta U(V))_a(v u)_b=0 ( v T δ U ( V ) ) a ( v u ) b = 0 for all a , b a,b a , b . The polynomial ring is a domain and (H2) holds, so δ U ( V ) = 0 \delta U(V)=0 δ U ( V ) = 0 . Similarly δ V ( U ) = 0 \delta V(U)=0 δ V ( U ) = 0 .
Only δ U = P U ( U ) \delta U=P_U(U) δ U = P U ( U ) and δ V = P V ( V ) \delta V=P_V(V) δ V = P V ( V ) remain in these blocks. The U V UV U V equation, after equating cofactors of Z Z Z , is the matrix identity
V d C x [ T ( x ) ] U + V C ( x ) P U ( U ) + P V ( V ) C ( x ) U = 0. (52) V\,dC_x[T(x)]U+VC(x)P_U(U)+P_V(V)C(x)U=0. \tag*{(52)} V d C x [ T ( x )] U + V C ( x ) P U ( U ) + P V ( V ) C ( x ) U = 0. (52)
Write the j j j th output column of P U P_U P U as ∑ ℓ E j ℓ U ℓ \sum_{\ell} E_{j\ell}U_\ell ∑ ℓ E j ℓ U ℓ . In entry ( i , j ) (i,j) ( i , j ) of (52) , a term involving U ℓ U_\ell U ℓ with ℓ ≠ j \ell\ne j ℓ = j occurs only as V i C E j ℓ U ℓ V_iCE_{j\ell}U_\ell V i C E j ℓ U ℓ . Thus C E j ℓ = 0 CE_{j\ell}=0 C E j ℓ = 0 for all x x x , and the spanning v v v -image forces E j ℓ = 0 E_{j\ell}=0 E j ℓ = 0 . Dually P V P_V P V acts separately on each row. Write P U ( U ) j = E j U j P_U(U)_j=E_jU_j P U ( U ) j = E j U j and P V ( V ) i = V i F i P_V(V)_i=V_iF_i P V ( V ) i = V i F i ; then
d C X [ T ( x ) ] + C E j + F i C = 0 for every i , j . dC_X[T(x)]+CE_j+F_iC=0 \qquad\text{for every } i,j. d C X [ T ( x )] + C E j + F i C = 0 for every i , j .
Taking differences and using the two spanning kernel images gives E j = E E_j=E E j = E and F i = F 0 F_i=F_0 F i = F 0 , independent of the indices. This is precisely (47) . Conversely every seed element satisfying that identity lifts by δ U = E U \delta U=EU δ U = E U , δ V = V F 0 \delta V=VF_0 δ V = V F 0 and δ Z = 0 \delta Z=0 δ Z = 0 . Hence the seed projection has image exactly h \mathfrak{h} h .
Step 4: the residual scalar action. After removing the k ( s L + s R ) k(s_L+s_R) k ( s L + s R ) syzygies and 2 r k 2rk 2 r k shears, the kernel of the seed projection satisfies F 0 C + C E = 0 F_0C+CE=0 F 0 C + C E = 0 . Since C = g u v T C=guv^T C = g u v T , every u ( x ) u(x) u ( x ) is an eigenvector of F 0 F_0 F 0 and every v ( x ) v(x) v ( x ) an eigenvector of E T E^T E T . Each projective kernel-image closure is irreducible. An irreducible spanning set of eigenvectors lies in one of the finitely many eigenspaces, which must be the entire space. Thus F 0 = − λ I r F_0=-\lambda I_r F 0 = − λ I r and E = λ I r E=\lambda I_r E = λ I r . The residual kernel is one-dimensional, with δ U = λ U \delta U=\lambda U δ U = λ U , δ V = − λ V \delta V=-\lambda V δ V = − λV . The successive images and kernels therefore have total dimension
( 2 k 2 − 1 ) + b l i f t + k ( s L + s R ) + 2 r k + 1 , (2k^2-1)+b_{\mathrm{lift}}+k(s_L+s_R)+2rk+1, ( 2 k 2 − 1 ) + b lift + k ( s L + s R ) + 2 r k + 1 ,
proving (49) . Subtracting the determinant stabilizer dimension 2 n 2 − 2 2n^2-2 2 n 2 − 2 gives (50) .
Suppose the seed orbit closure is a component, and let η \eta η be the codimension of the liftable seed stabilizers. The orbit codimension after padding is 1 + k ϵ − η 1+k\epsilon-\eta 1 + k ϵ − η , with 0 ≤ η ≤ ϵ 0\leq\eta\leq\epsilon 0 ≤ η ≤ ϵ . Since η \eta η is independent of k k k , componenthood for k = 2 k=2 k = 2 is equivalent to componenthood for every positive padding.
Corollary 13.2. Assume additionally that the seed orbit closure is a determinant-boundary component. Put η = 2 r 2 − 1 − b l i f t \eta=2r^2-1-b_{\mathrm{lift}} η = 2 r 2 − 1 − b lift . Then 0 ≤ η ≤ ϵ 0\leq\eta\leq\epsilon 0 ≤ η ≤ ϵ and
codim G L ‾ n 2 ⋅ Q k = 1 + k ϵ − η . \operatorname{codim}_{\overline{\mathrm{GL}}_{n^2}\cdot Q_k}=1+k\epsilon-\eta. codim GL n 2 ⋅ Q k = 1 + k ϵ − η . The following are equivalent:
Q 2 Q_2 Q 2 represents a boundary component;
s L = s R = r s_L=s_R=r s L = s R = r ;
every Q k Q_k Q k , k ≥ 1 k\geq1 k ≥ 1 , represents a boundary component.
If all seed stabilizers lift, the codimension is 1 + k ϵ 1+k\epsilon 1 + k ϵ .
Proof. The seed stabilizer has dimension 2 r 2 − 1 2r^2-1 2 r 2 − 1 , so η ≥ 0 \eta\geq0 η ≥ 0 . Theorem 13.1 gives the displayed formula. Since Q 1 Q_1 Q 1 is in the boundary, its codimension is at least one, giving η ≤ ϵ \eta\leq\epsilon η ≤ ϵ . The determinant boundary is pure of codimension one [6 Corollary 5.4.1] : its stabilizer is reductive, hence its orbit is affine, and the complement of a dense affine open has pure codimension one. Thus a contained orbit closure is a component exactly when its codimension is one. If ϵ = 0 \epsilon=0 ϵ = 0 , necessarily η = 0 \eta=0 η = 0 , and every padding has codimension one. Conversely componenthood of Q 2 Q_2 Q 2 gives 2 ϵ = η ≤ ϵ 2\epsilon=\eta\leq\epsilon 2 ϵ = η ≤ ϵ , forcing ϵ = 0 \epsilon=0 ϵ = 0 .
For the odd-skew seed, the Pfaffian kernel syzygies give s L = s R = r s_L=s_R=r s L = s R = r , and the intrinsic coefficient-space recognition already establishes verifies (H3). The corollary recovers all its paddings from the seed component theorem. The next example shows why the syzygy condition cannot be omitted.
A component whose positive paddings are not components
The following seed defines a boundary component and has excess syzygy dimension two. Its orbit codimension therefore increases by two for each added row and column.
Consider the known second concise 4 × 4 4 \times4 4 × 4 seed [10 Section 8.4 and Proposition 8.4.1] ,
A = ( a b 0 0 0 0 a b c 0 d 0 0 c 0 d ) , u = ( − b d a d b c − a c ) , v = ( c d − a − b ) . A = \begin{pmatrix} a & b & 0 & 0 \\ 0 & 0 & a & b \\ c & 0 & d & 0 \\ 0 & c & 0 & d \end{pmatrix}, \qquad u = \begin{pmatrix} -bd \\ ad \\ bc \\ -ac \end{pmatrix}, \qquad v = \begin{pmatrix} c \\ d \\ -a \\ -b \end{pmatrix}. A = a 0 c 0 b 0 0 c 0 a d 0 0 b 0 d , u = − b d a d b c − a c , v = c d − a − b .
Then adj A = u v T \operatorname{adj} A = uv^{T} adj A = u v T , and both kernel images span C 4 \mathbb{C}^{4} C 4 . Write E = ⟨ a , b ⟩ E = \langle a,b\rangle E = ⟨ a , b ⟩ and F = ⟨ c , d ⟩ F = \langle c,d\rangle F = ⟨ c , d ⟩ . The coefficients of P = v T S u P = v^{T}Su P = v T S u , viewed as a linear polynomial in S S S , are all cubic monomials containing two variables from one of E , F E,F E , F and one from the other. In tensor notation their span is
W = Sym 2 E ⊗ F ⊕ E ⊗ Sym 2 F , dim W = 12 , W = \operatorname{Sym}^{2} E \otimes F \oplus E \otimes\operatorname{Sym}^{2} F, \qquad\dim W = 12, W = Sym 2 E ⊗ F ⊕ E ⊗ Sym 2 F , dim W = 12 ,
so P P P is concise in sixteen variables. Its multiplicity-at-least-three locus is exactly x = 0 x = 0 x = 0 with arbitrary S S S : mixed S , x S,x S , x second derivatives include a 2 , b 2 , c 2 , d 2 a^{2}, b^{2}, c^{2}, d^{2} a 2 , b 2 , c 2 , d 2 , and the converse follows from x x x -degree three. This proves (H3).
The four entries of v v v are independent linear forms. Their usual pairwise alternating relations give six independent linear syzygies, so s L = 6 s_{L} = 6 s L = 6 . The entries of u u u are the four products of coordinates of E E E and F F F . Their syzygies split into ⋀ 2 E ⊗ F \bigwedge^{2} E \otimes F ⋀ 2 E ⊗ F and E ⊗ ⋀ 2 F E \otimes\bigwedge^{2} F E ⊗ ⋀ 2 F , each of dimension two; hence s R = 4 s_{R} = 4 s R = 4 and ϵ = 2 \epsilon= 2 ϵ = 2 . All seed stabilizers lift, as can be checked directly. The x x x -to-S S S stabilizers are the kernel of the surjective multiplication map
( E ⊕ F ) ⊗ W ⟶ Sym 3 E ⊗ F ⊕ Sym 2 E ⊗ Sym 2 F ⊕ E ⊗ Sym 3 F . (E \oplus F) \otimes W \longrightarrow\operatorname{Sym}^{3} E \otimes F \oplus\operatorname{Sym}^{2} E \otimes\operatorname{Sym}^{2} F \oplus E \otimes\operatorname{Sym}^{3} F. ( E ⊕ F ) ⊗ W ⟶ Sym 3 E ⊗ F ⊕ Sym 2 E ⊗ Sym 2 F ⊕ E ⊗ Sym 3 F .
Its dimension is 48 − ( 8 + 9 + 8 ) = 23 48 - (8 + 9 + 8) = 23 48 − ( 8 + 9 + 8 ) = 23 . The remaining x x x -actions preserve W W W : their off-diagonal blocks would produce forbidden E 3 E^{3} E 3 or F 3 F^{3} F 3 terms, so they are exactly g l ( E ) ⊕ g l ( F ) \mathfrak{gl}(E) \oplus\mathfrak{gl}(F) gl ( E ) ⊕ gl ( F ) , of dimension eight. Their action on both u u u and v v v is linear in the corresponding four coordinates, so they satisfy (47) . Thus b l i f t = 23 + 8 = 31 b_{\mathrm{lift}} = 23 + 8 = 31 b lift = 23 + 8 = 31 , and η = 0 \eta= 0 η = 0 . For every k ≥ 1 k \ge1 k ≥ 1 , with n = 4 + k n = 4 + k n = 4 + k , the exact conclusions are
dim Stab ( Q k ) = 2 n 2 − 1 + 2 k , codim G L n 2 ⋅ Q k ‾ = 2 k + 1. \dim\operatorname{Stab}(Q_{k}) = 2n^{2} - 1 + 2k, \qquad\operatorname{codim}_{\overline{\mathrm{GL}_{n^{2}}\cdot Q_{k}}} = 2k + 1. dim Stab ( Q k ) = 2 n 2 − 1 + 2 k , codim GL n 2 ⋅ Q k = 2 k + 1.
Consequently no positive padding of this concise seed component is a boundary component. This is a second application of the codimension formula.
Bounds from affine level sets
A nonzero level set of a polynomial can contain large affine spaces. Under an injective affine determinantal representation, each such space maps to an affine space of invertible matrices. Classical bounds on the dimension of matrix spaces therefore give lower bounds for exact determinantal complexity. This method does not require free cokernels and applies to all three component families.
We write an affine space as a + T a + T a + T , where T T T is its linear space of directions. Its dimension is dim T \dim T dim T .
Affine level sets and nilpotent matrix spaces
We use Gerstenhaber’s theorem [8 ] : a complex linear space of nilpotent m × m m \times m m × m matrices has dimension at most ( m 2 ) \binom{m}{2} ( 2 m ) . Recall that a matrix is nilpotent if some power of it is zero, equivalently if all of its eigenvalues are zero. The corresponding bounds for affine spaces of invertible matrices and their use with injective affine matrix maps appear in Meshulam [16 ] . The result below is this classical obstruction, stated in the form needed for our families.
Theorem 14.1. Let f f f be a nonzero polynomial concise in N N N variables. Suppose that an affine space a + T ⊆ C N a+T \subseteq\mathbb{C}^{N} a + T ⊆ C N satisfies f ∣ a + T = γ f|_{a+T}=\gamma f ∣ a + T = γ for a nonzero constant γ \gamma γ . Then every affine-linear determinantal representation of f f f of size m m m satisfies
dim T ≤ ( m 2 ) . \dim T \le\binom{m}{2}. dim T ≤ ( 2 m ) .
Proof. Write the representation as f ( x ) = det ( M 0 + L ( x ) ) f(x)=\det(M_0+L(x)) f ( x ) = det ( M 0 + L ( x )) , with L L L linear. First, L L L must be injective. If L ( h ) = 0 L(h)=0 L ( h ) = 0 , then the representing matrix, and hence f f f , is unchanged under every translation by t h th t h . Conciseness says that f f f has no such nonzero constant direction, so h = 0 h=0 h = 0 .
The matrix M a = M 0 + L ( a ) M_a=M_0+L(a) M a = M 0 + L ( a ) is invertible because f ( a ) = γ ≠ 0 f(a)=\gamma\ne0 f ( a ) = γ = 0 . Normalize the representation by multiplying by M a − 1 M_a^{-1} M a − 1 . For h ∈ T h\in T h ∈ T and every scalar t t t ,
det ( I m + t M a − 1 L ( h ) ) = f ( a + t h ) f ( a ) = 1. \det\left(I_m+tM_a^{-1}L(h)\right)=\frac{f(a+th)}{f(a)}=1. det ( I m + t M a − 1 L ( h ) ) = f ( a ) f ( a + t h ) = 1.
This identity says that the characteristic polynomial of M a − 1 L ( h ) M_a^{-1}L(h) M a − 1 L ( h ) is z m z^m z m . Thus every matrix in M a − 1 L ( T ) M_a^{-1}L(T) M a − 1 L ( T ) is nilpotent. This linear space has dimension dim T \dim T dim T , because L L L is injective. Gerstenhaber’s theorem supplies the stated bound. □ \square □
The use of conciseness is to make the linear map L L L injective. Without this hypothesis, the proof bounds the dimension only after quotienting by constant directions of f f f .
Applications to determinant fibers
For each component construction, fixing the seed variables at a point of maximal rank leaves a generic determinant after a surjective linear substitution in the coefficient variables. This gives the following consequence of the affine-level-set bound.
Corollary 14.2. Let f ( x , y ) f(x,y) f ( x , y ) be concise in d + D d+D d + D variables, with x ∈ C d x\in\mathbb{C}^{d} x ∈ C d and y ∈ C D y\in\mathbb{C}^{D} y ∈ C D . Suppose that for some x 0 x_0 x 0 there are a nonzero constant u u u and a surjective linear map κ : C D → Mat q \kappa:\mathbb{C}^{D}\to\operatorname{Mat}_{q} κ : C D → Mat q such that
f ( x 0 , y ) = u det κ ( y ) . f(x_0,y)=u\det\kappa(y). f ( x 0 , y ) = u det κ ( y ) . Then every affine-linear representation of f f f of size m m m satisfies
m ( m − 1 ) ≥ 2 D − q ( q + 1 ) . (53) m(m-1)\ge2D-q(q+1). \tag*{(53)} m ( m − 1 ) ≥ 2 D − q ( q + 1 ) . (53) In particular,
dc ( f ) ≥ ⌈ 1 + 1 + 8 D − 4 q ( q + 1 ) 2 ⌉ . (54) \operatorname{dc}(f)\ge\left\lceil\frac{1+\sqrt{1+8D-4q(q+1)}}{2}\right\rceil. \tag*{(54)} dc ( f ) ≥ ⌈ 2 1 + 1 + 8 D − 4 q ( q + 1 ) ⌉ . (54)
Proof. A matrix of the form I q + U I_q + U I q + U , with U U U strictly upper triangular, has determinant one. Let u q \mathfrak{u}_q u q be the linear space of these strictly upper triangular matrices; it has dimension ( q 2 ) \binom{q}{2} ( 2 q ) . Choose y 0 y_0 y 0 with κ ( y 0 ) = I q \kappa(y_0) = I_q κ ( y 0 ) = I q . Then f f f is the nonzero constant u u u on
{ x 0 } × ( y 0 + κ − 1 ( u q ) ) . \{x_0\} \times(y_0 + \kappa^{-1}(\mathfrak{u}_q)). { x 0 } × ( y 0 + κ − 1 ( u q )) .
There are D − q 2 D - q^2 D − q 2 directions in ker κ \ker\kappa ker κ and ( q 2 ) \binom{q}{2} ( 2 q ) additional directions mapping to u q \mathfrak{u}_q u q . The displayed affine space therefore has dimension
D − q 2 + ( q 2 ) = D − q ( q + 1 ) 2 . D - q^2 + \binom{q}{2} = D - \frac{q(q+1)}{2}. D − q 2 + ( 2 q ) = D − 2 q ( q + 1 ) .
Apply Theorem 14.1 and rearrange. Solving the quadratic inequality in m m m gives the integer bound. Surjectivity implies D ≥ q 2 D \ge q^2 D ≥ q 2 , so its square root is defined. □
The estimate compares the number D D D of coefficient variables with the order q q q of the determinant obtained after fixing the seed. The kernel of the compression map contributes D − q 2 D - q^2 D − q 2 independent directions to every exact representation of the original concise polynomial.
A bound from partial degree
The same inequality follows from a partial-degree hypothesis. Suppose f f f is concise, T T T is a D D D -dimensional space of input directions, f ( a ) ≠ 0 f(a) \ne0 f ( a ) = 0 , and
deg t f ( a + t h ) ≤ q for every h ∈ T . \deg_t f(a + th) \le q \qquad\text{for every } h \in T. deg t f ( a + t h ) ≤ q for every h ∈ T .
Then every size-m m m affine representation satisfies
D ≤ ( m 2 ) + ( q + 1 2 ) . D \le\binom{m}{2} + \binom{q+1}{2}. D ≤ ( 2 m ) + ( 2 q + 1 ) .
This is equivalent to (53) . Normalize the representing matrix at a a a to I m I_m I m , as in the proof of Theorem 14.1 . The images of the directions form a D D D -dimensional linear space of matrices B B B with deg t det ( I m + t B ) ≤ q \deg_t \det(I_m + tB) \le q deg t det ( I m + tB ) ≤ q . When 0 < q < m 0 < q < m 0 < q < m , every such matrix has zero as an eigenvalue of algebraic multiplicity at least m − q m - q m − q . Atkinson’s theorem [2 ] , also recorded in Loewy–Radwan [15 Theorem 2.7] , bounds the dimension of this matrix space by
m q + ( m − q ) ( m − q − 1 ) 2 = ( m 2 ) + ( q + 1 2 ) . mq + \frac{(m-q)(m-q-1)}{2} = \binom{m}{2} + \binom{q+1}{2}. m q + 2 ( m − q ) ( m − q − 1 ) = ( 2 m ) + ( 2 q + 1 ) .
For q = 0 q = 0 q = 0 , this is Gerstenhaber’s nilpotent-space bound. For q ≥ m q \ge m q ≥ m , injectivity gives D ≤ m 2 D \le m^2 D ≤ m 2 , which already implies the assertion.
Thus low degree in a large space of directions through a nonzero value suffices. In our constructions, the smaller determinant also makes the constant-value affine spaces explicit.
The trace-pairing bound
A short linear-algebra argument gives the slightly different bound
m 2 ≥ 2 D − q 2 . (55) m^2 \ge2D-q^2. \tag*{(55)} m 2 ≥ 2 D − q 2 . (55)
This bound uses the Hessian at a point where the polynomial is nonzero, rather than at a point of its zero set.
For f ( a ) ≠ 0 f(a) \ne0 f ( a ) = 0 , define
H log f ( a ) = Hess f ( a ) f ( a ) − ∇ f ( a ) ∇ f ( a ) T f ( a ) 2 . H_{\log f}(a)=\frac{\operatorname{Hess}f(a)}{f(a)}-\frac{\nabla f(a)\nabla f(a)^T}{f(a)^2}. H l o g f ( a ) = f ( a ) Hess f ( a ) − f ( a ) 2 ∇ f ( a ) ∇ f ( a ) T .
This rational expression is the logarithmic Hessian; it requires no choice of a logarithm. For a representation f = det ( M 0 + L ) f=\det(M_0+L) f = det ( M 0 + L ) , put M a = M 0 + L ( a ) M_a=M_0+L(a) M a = M 0 + L ( a ) . Differentiating the determinant twice gives
H log f ( a ) ( h , h ′ ) = − tr ( M a − 1 L ( h ) M a − 1 L ( h ′ ) ) . H_{\log f}(a)(h,h')=-\operatorname{tr}(M_a^{-1}L(h)M_a^{-1}L(h')). H l o g f ( a ) ( h , h ′ ) = − tr ( M a − 1 L ( h ) M a − 1 L ( h ′ )) .
The expression on the right is a restriction of the nondegenerate bilinear form ( B , C ) ↦ tr ( B C ) (B,C)\mapsto\operatorname{tr}(BC) ( B , C ) ↦ tr ( B C ) on Mat m \operatorname{Mat}_m Mat m .
If a nondegenerate bilinear form on an m 2 m^2 m 2 -dimensional space has rank s s s on a D D D -dimensional subspace W W W , then its kernel within W W W has dimension D − s D-s D − s . That kernel lies in the orthogonal complement W ⊥ W^\perp W ⊥ , which has dimension m 2 − D m^2-D m 2 − D . Hence
D − s ≤ m 2 − D , or m 2 ≥ 2 D − s . D-s\le m^2-D,\qquad\text{or}\qquad m^2\ge2D-s. D − s ≤ m 2 − D , or m 2 ≥ 2 D − s .
Use the coefficient directions at a = ( x 0 , y 0 ) a=(x_0,y_0) a = ( x 0 , y 0 ) from Corollary 14.2. Their image has dimension D D D , by conciseness. Since κ ( y 0 ) = I q \kappa(y_0)=I_q κ ( y 0 ) = I q , the logarithmic Hessian on these directions is the pullback under κ \kappa κ of the nondegenerate trace form on Mat q \operatorname{Mat}_q Mat q . Its rank is exactly q 2 q^2 q 2 . Substituting s = q 2 s=q^2 s = q 2 proves (55) .
The same reasoning without a restricted direction space gives
dc ( f ) 2 ≥ 2 N − rank H log f ( a ) . \operatorname{dc}(f)^2\ge2N-\operatorname{rank}H_{\log f}(a). dc ( f ) 2 ≥ 2 N − rank H l o g f ( a ) .
for every concise N N N -variable polynomial and every a a a with f ( a ) ≠ 0 f(a)\ne0 f ( a ) = 0 . A small rank on a large space forces a large ambient matrix space. Both this argument and the constant-value-space argument use an actual injective matrix representation of f f f . They do not assert these bounds for border determinantal complexity.
Bounds for the three component families
Each family is concise in n 2 n^2 n 2 variables. At a seed value of maximal rank, the coefficient variables give the smaller determinant required by Corollary 14.2. The relevant dimensions are
Table 1 (PDF p. 59).
For the first two rows, use the compressed determinants displayed in the constructions. For the third, use the local determinant formula (37) . The compression map is surjective in each case. Thus (54) gives an explicit integer lower bound for every parameter choice in all three families. Their border complexity is exactly n n n : the construction gives the upper bound, and their degree n n n gives the lower bound.
Odd-skew padding. For odd r ≥ 3 r \ge3 r ≥ 3 , k ≥ 0 k \ge0 k ≥ 0 , and n = r + k n=r+k n = r + k , substitution gives
dc ( Q r , k ) ( dc ( Q r , k ) − 1 ) ≥ 2 n 2 − r ( r − 1 ) − ( k + 1 ) ( k + 2 ) . (56) \operatorname{dc}(Q_{r,k})\bigl(\operatorname{dc}(Q_{r,k})-1\bigr) \ge2n^2-r(r-1)-(k+1)(k+2). \tag*{(56)} dc ( Q r , k ) ( dc ( Q r , k ) − 1 ) ≥ 2 n 2 − r ( r − 1 ) − ( k + 1 ) ( k + 2 ) . (56)
Choose an allowed odd r r r within a bounded distance of n / 2 n/2 n /2 , and put k = n − r k=n-r k = n − r . The leading term on the right is 3 2 n 2 \frac{3}{2}n^2 2 3 n 2 . Therefore
dc ( Q r , k ) ≥ 3 2 n − O ( 1 ) , \operatorname{dc}(Q_{r,k}) \ge\sqrt{\frac{3}{2}}\,n-O(1), dc ( Q r , k ) ≥ 2 3 n − O ( 1 ) ,
and
dc ( Q r , k ) − dc ‾ ( Q r , k ) ≥ ( 3 2 − 1 ) n − O ( 1 ) . \operatorname{dc}(Q_{r,k})-\underline{\operatorname{dc}}(Q_{r,k}) \ge\left(\sqrt{\frac{3}{2}}-1\right)n-O(1). dc ( Q r , k ) − dc ( Q r , k ) ≥ ( 2 3 − 1 ) n − O ( 1 ) .
The additive gap grows linearly even though these seeds have corank one. This supplies a bound for padded odd-skew forms to which the stronger free-cokernel theorem was not applied.
The block family. For h ≥ 2 h \ge2 h ≥ 2 , k ≥ 0 k \ge0 k ≥ 0 , and n = 3 h + k n=3h+k n = 3 h + k , the same substitution gives
dc ( H h , k ) ( dc ( H h , k ) − 1 ) ≥ 15 h 2 + 10 h k + k 2 − h − k − 4. (57) \operatorname{dc}(H_{h,k})\bigl(\operatorname{dc}(H_{h,k})-1\bigr) \ge15h^2+10hk+k^2-h-k-4. \tag*{(57)} dc ( H h , k ) ( dc ( H h , k ) − 1 ) ≥ 15 h 2 + 10 hk + k 2 − h − k − 4. (57)
Take h = ⌊ n / 3 ⌋ h=\lfloor n/3\rfloor h = ⌊ n /3 ⌋ and k = n − 3 h ∈ { 0 , 1 , 2 } k=n-3h\in\{0,1,2\} k = n − 3 h ∈ { 0 , 1 , 2 } . This choice is available for every n ≥ 6 n\ge6 n ≥ 6 . The right side of (57) is
5 3 n 2 − 1 3 n − 2 3 k ( k + 1 ) − 4. \frac{5}{3}n^2-\frac{1}{3}n-\frac{2}{3}k(k+1)-4. 3 5 n 2 − 3 1 n − 3 2 k ( k + 1 ) − 4.
Consequently
dc ( H h , k ) ≥ 5 3 n − O ( 1 ) , \operatorname{dc}(H_{h,k}) \ge\sqrt{\frac{5}{3}}\,n-O(1), dc ( H h , k ) ≥ 3 5 n − O ( 1 ) ,
and
dc ( H h , k ) − dc ‾ ( H h , k ) ≥ ( 5 3 − 1 ) n − O ( 1 ) . \operatorname{dc}(H_{h,k})-\underline{\operatorname{dc}}(H_{h,k}) \ge\left(\sqrt{\frac{5}{3}}-1\right)n-O(1). dc ( H h , k ) − dc ( H h , k ) ≥ ( 3 5 − 1 ) n − O ( 1 ) .
An explicit affine representation in Section 7 also gives dc ( H h , k ) ≤ n + 4 h \operatorname{dc}(H_{h,k})\le n+4h dc ( H h , k ) ≤ n + 4 h . The gap for this choice of parameters is therefore bounded both below and above by positive constant multiples of n n n for large n n n :
dc ( H h , k ) − dc ‾ ( H h , k ) = Θ ( n ) . \operatorname{dc}(H_{h,k})-\underline{\operatorname{dc}}(H_{h,k})=\Theta(n). dc ( H h , k ) − dc ( H h , k ) = Θ ( n ) .
Thus the additive gap for this family has linear order of growth.
The Koszul family. Put r = ( 2 p + 1 p ) r=\binom{2p+1}{p} r = ( p 2 p + 1 ) and c = ( 2 p p − 1 ) c=\binom{2p}{p-1} c = ( p − 1 2 p ) . For every k ≥ 0 k\ge0 k ≥ 0 , with n = r + k n=r+k n = r + k , we get
dc ( K p , k ) ( dc ( K p , k ) − 1 ) ≥ 2 n 2 − 2 ( 2 p + 1 ) − ( c + k ) ( c + k + 1 ) . (58) \operatorname{dc}(K_{p,k})\bigl(\operatorname{dc}(K_{p,k})-1\bigr) \ge2n^2-2(2p+1)-(c+k)(c+k+1). \tag*{(58)} dc ( K p , k ) ( dc ( K p , k ) − 1 ) ≥ 2 n 2 − 2 ( 2 p + 1 ) − ( c + k ) ( c + k + 1 ) . (58)
For the unpadded sequence n = r n=r n = r ,
c n = p 2 p + 1 ⟶ 1 2 , 2 p + 1 n 2 ⟶ 0. \frac{c}{n}=\frac{p}{2p+1}\longrightarrow\frac{1}{2},\qquad\frac{2p+1}{n^2}\longrightarrow0. n c = 2 p + 1 p ⟶ 2 1 , n 2 2 p + 1 ⟶ 0.
Hence
dc ( K p , 0 ) ≥ ( 7 2 − o ( 1 ) ) n , \operatorname{dc}(K_{p,0}) \ge\left(\frac{\sqrt{7}}{2}-o(1)\right)n, dc ( K p , 0 ) ≥ ( 2 7 − o ( 1 ) ) n ,
and
dc ( K p , 0 ) − dc ‾ ( K p , 0 ) ≥ ( 7 2 − 1 − o ( 1 ) ) n . \operatorname{dc}(K_{p,0})-\underline{\operatorname{dc}}(K_{p,0}) \ge\left(\frac{\sqrt{7}}{2}-1-o(1)\right)n. dc ( K p , 0 ) − dc ( K p , 0 ) ≥ ( 2 7 − 1 − o ( 1 ) ) n .
For this family the kernel-degree result of Corollary 6.3 is stronger. The present argument remains useful because the same simple calculation also covers every odd-skew padding and the block family.
Asymptotics for a fixed seed
Fix the seed and let the amount of padding grow. Let d d d be the number of seed variables and R R R the degree in those variables. In all three families, the smaller determinant has order q = n − R q=n-R q = n − R , while there are D = n 2 − d D=n^2-d D = n 2 − d coefficient variables. The common inequality becomes
m ( m − 1 ) ≥ n 2 + ( 2 R − 1 ) n − R 2 + R − 2 d . m(m-1) \ge n^2+(2R-1)n-R^2+R-2d. m ( m − 1 ) ≥ n 2 + ( 2 R − 1 ) n − R 2 + R − 2 d .
Compare its right side with the value for m = n + R − 1 m=n+R-1 m = n + R − 1 . Their difference is
2 ( n − ( R − 1 ) 2 − d ) . 2\left(n-(R-1)^2-d\right). 2 ( n − ( R − 1 ) 2 − d ) .
Thus, if
n > ( R − 1 ) 2 + d (59) n>(R-1)^2+d \tag*{(59)} n > ( R − 1 ) 2 + d (59)
exact size n + R − 1 n+R-1 n + R − 1 is impossible. Hence
dc ( f ) ≥ n + R when n > ( R − 1 ) 2 + d . (60) \operatorname{dc}(f) \ge n+R \quad\text{when } n>(R-1)^2+d. \tag*{(60)} dc ( f ) ≥ n + R when n > ( R − 1 ) 2 + d . (60)
The seed degrees are
R = r − 1 for Q r , k , R = 2 h for H h , k , R = ( 2 p p ) for K p , k . R=r-1\quad\text{for }Q_{r,k},\qquad R=2h\quad\text{for }H_{h,k},\qquad R=\binom{2p}{p}\quad\text{for }\mathcal{K}_{p,k}. R = r − 1 for Q r , k , R = 2 h for H h , k , R = ( p 2 p ) for K p , k .
This is a lower bound for arbitrary affine representations. It does not determine the exact complexity or claim the best possible threshold.
All bounds in this section are applications of classical nilpotent-space and affine-invertible-space theory [8 , 16 ] . The family-specific ingredient is the smaller determinant left after fixing the seed. Seed corank alone does not supply the lower bound; what matters here is the number of coefficient directions compared with the order of that remaining determinant.
Regular determinantal representations
Restriction to a nonzero seed value relates the exact complexity of a Koszul form to the regular determinantal complexity of a generic determinant. A representation g ( X ) = det M ( X ) g(X)=\det M(X) g ( X ) = det M ( X ) is regular at X = 0 X=0 X = 0 if M ( 0 ) M(0) M ( 0 ) has corank one, and rdc ( g ) \operatorname{rdc}(g) rdc ( g ) denotes the least size of such a representation. This is a condition at one point; the freeness hypothesis in Theorem 4.1 is a global condition.
Restriction to a generic determinant
For the Koszul family, write
d 0 = 2 p + 1 , r = ( 2 p + 1 p ) , c = ( 2 p p − 1 ) , n = r + k . d_0=2p+1,\qquad r=\binom{2p+1}{p},\qquad c=\binom{2p}{p-1},\qquad n=r+k. d 0 = 2 p + 1 , r = ( p 2 p + 1 ) , c = ( p − 1 2 p ) , n = r + k .
Fix a nonzero value of the seed vector v v v . As a polynomial in the remaining coefficient variables, K p , k \mathcal{K}_{p,k} K p , k becomes a nonzero scalar times the determinant of a ( c + k ) × ( c + k ) (c+k)\times(c+k) ( c + k ) × ( c + k ) matrix. The map from the coefficient variables to that matrix is onto. Choose a linear section of this map: this means choosing coefficient values that produce each desired small matrix, linearly in its entries. Restricting to this section leaves exactly a generic determinant, up to the nonzero scalar.
Let M M M be any exact representation of K p , k K_{p,k} K p , k . Theorem 6.2 gives corank one at every zero of the polynomial. In particular, after the restriction just described, it has corank one at the zero small matrix. Absorbing the nonzero scalar by a constant row rescaling gives a regular representation of the generic determinant. The matrix size has not increased. Therefore
dc ( K p , k ) ≥ rdc ( det c + k ) . (61) \operatorname{dc}(K_{p,k}) \ge\operatorname{rdc}(\det_{c+k}). \tag*{(61)} dc ( K p , k ) ≥ rdc ( c + k det ) . (61)
The inequality applies to arbitrary exact representations, with no assumption on their matrix form.
For a generic determinant, regularity at the zero matrix has a special consequence. The classification recalled in [4 Section 4.4] leaves three possible rank-one maximal Cohen–Macaulay modules: the free module and two nonfree classes. For a determinant of order q q q , each of the latter has a q q q -dimensional fiber at the zero matrix. A regular representation has a one-dimensional cokernel fiber there, so, for q ≥ 2 q \ge2 q ≥ 2 , only the free module is possible. The kernel-degree lemma consequently gives
rdc ( det q ) ≥ 2 q − 1. \operatorname{rdc}(\det_q) \ge2q-1. rdc ( q det ) ≥ 2 q − 1.
For q = 1 q=1 q = 1 the same bound is immediate. On the other hand, a known regular construction has size ( q 3 − q ) / 3 + 1 (q^3-q)/3+1 ( q 3 − q ) /3 + 1 [11 ] . Since c + k c+k c + k is a constant fraction of n = r + k n=r+k n = r + k , a superlinear lower bound for rdc ( det q ) \operatorname{rdc}(\det_q) rdc ( det q ) would give one for the Koszul family through (61) .
A spectral formulation
There is an equivalent formulation that uses only a linear map between matrix spaces. Suppose
det M ( X ) = det q X , M ( X ) = M 0 + L ( X ) , \det M(X)=\det_q X,\qquad M(X)=M_0+L(X), det M ( X ) = q det X , M ( X ) = M 0 + L ( X ) ,
where M M M has size m m m and L L L is linear. At X = I q X=I_q X = I q the determinant is one, so M ( I q ) M(I_q) M ( I q ) is invertible. Normalize by this matrix and define
B ( T ) = M ( I q ) − 1 L ( T ) . B(T)=M(I_q)^{-1}L(T). B ( T ) = M ( I q ) − 1 L ( T ) .
The determinant identity at I q + t T I_q+tT I q + tT then becomes
det ( I m + t B ( T ) ) = det ( I q + t T ) . \det(I_m+tB(T))=\det(I_q+tT). det ( I m + tB ( T )) = det ( I q + tT ) .
Equivalently, for every T T T ,
det ( λ I m − B ( T ) ) = λ m − q det ( λ I q − T ) . \det(\lambda I_m-B(T))=\lambda^{m-q}\det(\lambda I_q-T). det ( λ I m − B ( T )) = λ m − q det ( λ I q − T ) .
Thus B B B preserves the eigenvalues of T T T , with the additional m − q m-q m − q eigenvalues all zero, counting algebraic multiplicity.
The regularity condition also has a direct meaning. Since
M ( I q ) − 1 M ( 0 ) = I m − B ( I q ) , M(I_q)^{-1}M(0)=I_m-B(I_q), M ( I q ) − 1 M ( 0 ) = I m − B ( I q ) ,
regularity is equivalent to I m − B ( I q ) I_m-B(I_q) I m − B ( I q ) having corank one. The characteristic-polynomial identity says that the eigenvalue 1 1 1 of B ( I q ) B(I_q) B ( I q ) has algebraic multiplicity q q q . Corank one says that its eigenspace is one-dimensional. In Jordan normal form, it must therefore have one Jordan block, of length q q q .
Conversely, suppose a linear map B : Mat q → Mat m B:\operatorname{Mat}_q \to\operatorname{Mat}_m B : Mat q → Mat m satisfies the characteristic-polynomial identity above, and the eigenvalue 1 1 1 of B ( I q ) B(I_q) B ( I q ) has a single Jordan block. Then
M ( X ) = I m − B ( I q ) + B ( X ) = I m + B ( X − I q ) M(X)=I_m-B(I_q)+B(X)=I_m+B(X-I_q) M ( X ) = I m − B ( I q ) + B ( X ) = I m + B ( X − I q )
is regular at zero. Substituting T = X − I q T = X - I_q T = X − I q in the determinant identity shows that det M ( X ) = det q X \det M(X) = \det_q X det M ( X ) = det q X .
Thus regular determinantal complexity is equivalently the minimum m m m for which a linear map B B B satisfies this characteristic-polynomial identity and the stated Jordan-block condition. A superlinear lower bound for this minimum would give a superlinear bound for the Koszul family through (61) . Such a bound remains open.
A criterion for polynomial realizations of rational matrices
The Koszul construction admits a formula in terms of a rational projection matrix. We study polynomial realizations of this matrix, first proving a general criterion in terms of its residue module. The criterion concerns a prescribed rational matrix; its implications for scalar determinant representations are discussed below.
Let P = X / q P = X/q P = X / q , where X X X is a polynomial matrix and q q q is irreducible. A polynomial realization is an expression
P = A + C D − 1 B P = A + CD^{-1}B P = A + C D − 1 B
with polynomial matrices of compatible sizes and D D D square. The determinant of D D D controls the denominator of C D − 1 B CD^{-1}B C D − 1 B . We prescribe this determinant, but impose no bound on the size of D D D or on the degrees of the entries. These requirements are weaker than those for an affine determinantal representation of a scalar polynomial.
The residue module
Reduce X X X modulo q q q . Its columns now live on the hypersurface q = 0 q = 0 q = 0 , and their span records which output directions occur in the numerator at that hypersurface. Write this image as im X ‾ \operatorname{im}\overline{X} im X . Away from points where its rank drops, it is a vector bundle: locally it is a fixed number of independent columns varying algebraically with the point.
We will use its reflexive extension E E E . Concretely, take the double dual of the image module and regard it inside the ambient free module. Over a normal ring this is the extension determined by what happens away from subsets of codimension at least two. Thus isolated exceptional behavior in higher codimension does not change E E E . The relevant question is whether E E E has full depth at those exceptional points. The standard name for this condition is maximal Cohen–Macaulay (MCM), as defined earlier. It is exactly the condition needed for a short free resolution over the ambient polynomial ring.
Theorem 16.1. Let T = C [ v 1 , … , v d ] T = \mathbb{C}[v_1,\ldots,v_d] T = C [ v 1 , … , v d ] , let q ∈ T q \in T q ∈ T be irreducible, and suppose S = T / ( q ) S = T/(q) S = T / ( q ) is normal. Let X X X be an r × s r \times s r × s polynomial matrix and put P = X / q P = X/q P = X / q . Suppose X ‾ \overline{X} X has rank a > 0 a > 0 a > 0 and constant rank a a a away from a closed subset Z ⊂ Spec S Z \subset\operatorname{Spec} S Z ⊂ Spec S of codimension at least two. Let E ⊂ S r E \subset S^r E ⊂ S r be the reflexive extension of im X ‾ \operatorname{im}\overline{X} im X . The following are equivalent:
There are polynomial matrices A A A , B B B , C C C , D D D of compatible sizes, with D D D square, such that
P = A + C D − 1 B , det D = u q a , u ∈ C × . P = A + CD^{-1}B,\qquad\det D = uq^a,\qquad u \in\mathbb{C}^{\times}. P = A + C D − 1 B , det D = u q a , u ∈ C × . The module E E E is maximal Cohen–Macaulay at every local ring of S S S .
The matrices need not be homogeneous or affine linear. The equivalence is an existence criterion, with no upper bound on the size of D D D .
The exponent a a a is the minimum permitted by the generic residue rank. At a general point of q = 0 q=0 q = 0 , at least a a a directions contribute a pole. Equality in the exponent forces all of these poles to be simple. The proof uses this observation and the extension of the residue module across the exceptional set.
Proof. Step 1: Smith exponents. Suppose the realization in part (1) exists. Localize the polynomial ring at the prime ideal ( q ) (q) ( q ) . This is a discrete valuation ring: it measures the power of q q q dividing a polynomial. Smith normal form writes D D D , up to invertible row and column operations over this local ring, with diagonal entries q e i q^{e_i} q e i times units. The integers e i ≥ 0 e_i \ge0 e i ≥ 0 are its Smith exponents, and
∑ i e i = a \sum_i e_i=a i ∑ e i = a
because det D = u q a \det D=uq^a det D = u q a .
Let t t t be the number of positive exponents. Only these t t t directions can contribute singular terms to C D − 1 B CD^{-1}B C D − 1 B , so its singular part has rank at most t t t . Multiplying that part by q q q makes it regular: the whole expression q ( P − A ) = X − q A q(P-A)=X-qA q ( P − A ) = X − q A is regular, and the Smith blocks with exponent zero also contribute regular terms. Reducing modulo q q q therefore gives a matrix of rank at most t t t . Its rank is a a a , the rank of X ˉ \bar{X} X ˉ . We obtain a ≤ t ≤ ∑ i e i = a a \le t \le\sum_i e_i=a a ≤ t ≤ ∑ i e i = a . Consequently there are exactly a a a positive exponents, and all are one.
Step 2: the cokernel of D D D . Set N = q D − 1 N=qD^{-1} N = q D − 1 . This is a polynomial matrix. Indeed the only possible denominator factor of D − 1 D^{-1} D − 1 is q q q , while Step 1 shows that multiplication by q q q removes its poles at ( q ) (q) ( q ) . We have
D N = N D = q I . DN=ND=qI. D N = N D = q I .
It follows that F = coker T D F=\operatorname{coker}_{T}D F = coker T D is annihilated by q q q , so it is an S S S -module. Its rank over S S S is a a a . The exact sequence
0 ⟶ T m → D T m ⟶ F ⟶ 0 0\longrightarrow T^m \xrightarrow{D} T^m \longrightarrow F \longrightarrow0 0 ⟶ T m D T m ⟶ F ⟶ 0
shows, by the depth formula over the regular ring T T T , that F F F is MCM over S S S .
Step 3: identification with the residue module. The map
F ⟶ S r , [ z ] ⟼ C N z m o d q F\longrightarrow S^r,\qquad[z]\longmapsto CNz\mathbin{\bmod}q F ⟶ S r , [ z ] ⟼ C N z mod q
is well-defined, because C N D = q C CND=qC C N D = q C . Its image contains im X ˉ \operatorname{im}\bar{X} im X ˉ : multiplying the realization by q q q gives X = q A + C N B X=qA+CNB X = q A + C N B . Both F F F and im X ˉ \operatorname{im}\bar{X} im X ˉ have generic rank a a a , so this map is generically injective. An MCM module over our normal domain is torsion-free; hence the map is injective everywhere.
Away from Z Z Z , the constant-rank image of X ˉ \bar{X} X ˉ is a subbundle of S r S^r S r . In particular the quotient of S r S^r S r by that image is locally free, and therefore torsion-free. The quotient of the image of F F F by im X ˉ \operatorname{im}\bar{X} im X ˉ is a submodule of this torsion-free quotient. But it has rank zero, so it is torsion. It must vanish. Thus F F F and im X ˉ \operatorname{im}\bar{X} im X ˉ agree away from Z Z Z . Both F F F and E E E are reflexive, and the omitted set has codimension at least two. They therefore agree globally. This proves that E ≅ F E\cong F E ≅ F is MCM.
Step 4: construction from an MCM residue module. Conversely, assume part (2). The local Auslander–Buchsbaum formula over T T T gives a projective resolution of length one. Quillen–Suslin makes its projective terms free. Thus we can write
0 ⟶ T m → D T m ⟶ E ⟶ 0. 0\longrightarrow T^m \xrightarrow{D} T^m \longrightarrow E \longrightarrow0. 0 ⟶ T m D T m ⟶ E ⟶ 0.
The two ranks are equal because E E E is torsion as a T T T -module. Its support is q = 0 q=0 q = 0 , and its generic length there is a a a . Hence det D = u q a \det D=uq^a det D = u q a for a nonzero constant u u u . Since q E = 0 qE=0 q E = 0 , the matrix N = q D − 1 N=qD^{-1} N = q D − 1 is polynomial and satisfies D N = N D = q I DN=ND=qI D N = N D = q I . It remains to recover the desired numerator X X X . Lift the inclusion E ↪ S r E \hookrightarrow S^{r} E ↪ S r to a polynomial map H : T m → T r H:T^{m}\to T^{r} H : T m → T r . The composition H D HD H D is zero modulo q q q , so define
Then C N = H CN=H C N = H . Lift the columns of X ˉ \bar{X} X ˉ , viewed as elements of E E E , through the surjection T m → E T^{m}\to E T m → E ; collect these lifts in a polynomial matrix B B B . The matrix X − C N B X-CNB X − C N B is now divisible by q q q . Define
A = ( X − C N B ) / q . A=(X-CNB)/q. A = ( X − C N B ) / q .
Since N = q D − 1 N=qD^{-1} N = q D − 1 , these definitions give A + C D − 1 B = X / q A+CD^{-1}B=X/q A + C D − 1 B = X / q . All four matrices are polynomial, as required.
The residue obstruction for the Koszul projection
We apply the criterion to the unpadded Koszul form. Equip V = C 2 p + 1 V=\mathbb{C}^{2p+1} V = C 2 p + 1 with its standard nondegenerate symmetric bilinear form, and write
p ≥ 2 , q ( v ) = ∑ i v i 2 , r = ( 2 p + 1 p ) , a = ( 2 p − 1 p − 1 ) = 1 2 ( 2 p p ) . p\ge2,\qquad q(v)=\sum_i v_i^2,\qquad r=\binom{2p+1}{p},\qquad a=\binom{2p-1}{p-1}=\frac{1}{2}\binom{2p}{p}. p ≥ 2 , q ( v ) = i ∑ v i 2 , r = ( p 2 p + 1 ) , a = ( p − 1 2 p − 1 ) = 2 1 ( p 2 p ) .
There are two basic maps on exterior powers. Exterior multiplication by v v v , denoted K ( v ) K(v) K ( v ) , sends Λ p − 1 V \Lambda^{p-1}V Λ p − 1 V to Λ p V \Lambda^pV Λ p V . Contraction by v v v , denoted C 0 ( v ) C_0(v) C 0 ( v ) , goes in the opposite direction. Their product has the identities
X ( v ) = K ( v ) C 0 ( v ) , X ( v ) 2 = q ( v ) X ( v ) , P ( v ) = X ( v ) / q ( v ) . X(v)=K(v)C_0(v),\qquad X(v)^2=q(v)X(v),\qquad P(v)=X(v)/q(v). X ( v ) = K ( v ) C 0 ( v ) , X ( v ) 2 = q ( v ) X ( v ) , P ( v ) = X ( v ) / q ( v ) .
For q ( v ) ≠ 0 q(v)\ne0 q ( v ) = 0 , the last identity says that P 2 = P P^2=P P 2 = P : it is a projection. Its image is the kernel space U v = v ∧ Λ p − 1 V U_v=v\wedge\Lambda^{p-1}V U v = v ∧ Λ p − 1 V used in the Koszul construction.
The projection formula
Normalize the unpadded Koszul form by a fixed nonzero scalar. Then
f ( v , S ) = q ( v ) a det r ( I r + P ( v ) ( S − I r ) ) . (62) f(v,S)=q(v)^a\det\nolimits_r\left(I_r+P(v)(S-I_r)\right). \tag*{(62)} f ( v , S ) = q ( v ) a det r ( I r + P ( v ) ( S − I r ) ) . (62)
In particular, the value at S = I r S=I_r S = I r is q a q^a q a , and the dependence on the coefficient variables is expressed through P P P .
Here we may choose the coefficient complement to contain I r I_r I r . Alternatively let S S S range over all of Mat r \operatorname{Mat}_r Mat r : adding a member of the original seed space to S S S does not change the formula. To verify (62) , first work where q ( v ) ≠ 0 q(v)\ne0 q ( v ) = 0 . There is an orthogonal splitting
Λ p V = U v ⊕ U v ⊥ . \Lambda^pV=U_v\oplus U_v^\perp. Λ p V = U v ⊕ U v ⊥ .
On these summands P P P is respectively the identity and zero. The determinant on the right of (62) is therefore the determinant of the restriction of S S S to U v U_v U v . The middle Koszul map vanishes on U v U_v U v . On its complementary summand, which has dimension 2 a 2a 2 a , its determinant is a fixed nonzero scalar times q ( v ) a q(v)^a q ( v ) a . This last assertion follows by taking v = t e 2 p + 1 v=te_{2p+1} v = t e 2 p + 1 and then using equivariance under the orthogonal group. The leading coefficient of det ( A ( v ) + t S ) \det(A(v)+tS) det ( A ( v ) + tS ) is the product of these two determinants. Thus the equality holds on q ≠ 0 q\ne0 q = 0 , and hence as a polynomial identity everywhere.
Proposition 17.1. For p ≥ 2 p \ge2 p ≥ 2 , there is no polynomial realization
P = A + C D − 1 B , det D = u q a , P = A + CD^{-1}B,\qquad\det D = u q^{a}, P = A + C D − 1 B , det D = u q a , regardless of the size of the square matrix D D D .
Proof. We check the hypothesis and the obstruction in Theorem 16.1 . The only exceptional point will be the vertex v = 0 v=0 v = 0 of the quadric cone.
The residue has constant rank a a a away from the vertex. Let v ≠ 0 v \ne0 v = 0 satisfy q ( v ) = 0 q(v)=0 q ( v ) = 0 . Choose a hyperbolic partner w w w and a nondegenerate complement W W W of dimension 2 p − 1 2p-1 2 p − 1 . In the resulting exterior-power decomposition, K C 0 KC_0 K C 0 maps w ∧ Λ p − 1 W w \wedge\Lambda^{p-1}W w ∧ Λ p − 1 W isomorphically onto v ∧ Λ p − 1 W v \wedge\Lambda^{p-1}W v ∧ Λ p − 1 W and kills the other summands. Its rank is therefore
( 2 p − 1 p − 1 ) = a . \binom{2p-1}{p-1}=a. ( p − 1 2 p − 1 ) = a .
The associated bundle has an intermediate cohomology group. Pass from the punctured cone to the smooth projective quadric
Q = { q = 0 } ⊂ P 2 p . Q=\{q=0\}\subset\mathbb{P}^{2p}. Q = { q = 0 } ⊂ P 2 p .
Let L = O Q ( − 1 ) L=\mathcal{O}_Q(-1) L = O Q ( − 1 ) be the line spanned by v v v , and let H = L ⊥ / L \mathcal{H}=L^\perp/L H = L ⊥ / L . The image bundle just computed is
E = L ⊗ Λ p − 1 H = Λ p − 1 T Q ( − p ) = Ω Q p ( p − 1 ) . \mathcal{E}=L\otimes\Lambda^{p-1}\mathcal{H}=\Lambda^{p-1}T_Q(-p)=\Omega_Q^p(p-1). E = L ⊗ Λ p − 1 H = Λ p − 1 T Q ( − p ) = Ω Q p ( p − 1 ) .
The identities use T Q = L ∗ ⊗ H T_Q=L^*\otimes\mathcal{H} T Q = L ∗ ⊗ H and det T Q = O Q ( 2 p − 1 ) \det T_Q=\mathcal{O}_Q(2p-1) det T Q = O Q ( 2 p − 1 ) . Thus the bundle can be expressed in terms of differential forms on the quadric, whose cohomology supplies the required test.
In fact
H p ( Q , E ( 1 − p ) ) = H p ( Q , Ω Q p ) ≠ 0. H^p(Q,\mathcal{E}(1-p))=H^p(Q,\Omega_Q^p)\ne0. H p ( Q , E ( 1 − p )) = H p ( Q , Ω Q p ) = 0.
One way to see nonvanishing is to take the p p p -th power of the hyperplane class. If it were zero, its product with the remaining hyperplane powers would also be zero. That product has top degree two on a quadric. Since dim Q = 2 p − 1 \dim Q=2p-1 dim Q = 2 p − 1 and p ≥ 2 p\ge2 p ≥ 2 , the cohomology degree p p p is strictly between zero and the dimension.
This is a depth defect at the vertex. Let E = ( im X ˉ ) ∗ ∗ E=(\operatorname{im}\bar X)^{**} E = ( im X ˉ ) ∗∗ be the reflexive residue module on the affine cone, and let m \mathfrak{m} m be the vertex ideal. Its associated sheaf on Q Q Q is E \mathcal{E} E . The standard correspondence between sheaf cohomology and graded local cohomology gives
H m p + 1 ( E ) 1 − p ≃ H p ( Q , Ω Q p ) ≠ 0 , p + 1 < 2 p = dim S . H^{p+1}_{\mathfrak{m}}(E)_{1-p}\simeq H^p(Q,\Omega_Q^p)\ne0,\qquad p+1<2p=\dim S. H m p + 1 ( E ) 1 − p ≃ H p ( Q , Ω Q p ) = 0 , p + 1 < 2 p = dim S .
Full depth would force these local cohomology groups to vanish below dim S \dim S dim S . Hence E E E is not MCM. The realization criterion now proves the proposition. □
The restriction p ≥ 2 p\ge2 p ≥ 2 is necessary. When p = 1 p=1 p = 1 , the projection is P = v v T / q P=vv^{T}/q P = v v T / q , and the scalar matrix D = [ q ] D=[q] D = [ q ] already gives a realization.
Realizations with a larger denominator
The exponent in Proposition 17.1 is essential. With a larger exponent of q q q in det D \det D det D , the same projection has an explicit realization of size linear in r r r .
Put
b = ( 2 p + 1 p − 1 ) , h = 2 p , N = h ⌈ b / h ⌉ . b=\binom{2p+1}{p-1},\qquad h=2^p,\qquad N=h\lceil b/h\rceil. b = ( p − 1 2 p + 1 ) , h = 2 p , N = h ⌈ b / h ⌉ .
Choose complex Clifford matrices Γ 1 , … , Γ 2 p + 1 \Gamma_1,\ldots,\Gamma_{2p+1} Γ 1 , … , Γ 2 p + 1 of size h h h , with
Γ i Γ j + Γ j Γ i = 2 δ i j I h . \Gamma_i\Gamma_j+\Gamma_j\Gamma_i=2\delta_{ij}I_h. Γ i Γ j + Γ j Γ i = 2 δ ij I h .
For example, these follow from tensor products of the Pauli matrices. The anticommutation relation is the only property we use: it makes ( ∑ v i Γ i ) 2 = q ( v ) I h (\sum v_i\Gamma_i)^2=q(v)I_h ( ∑ v i Γ i ) 2 = q ( v ) I h . Repeating this pencil in diagonal blocks gives a size-N N N linear matrix Γ ( v ) \Gamma(v) Γ ( v ) with
Γ ( v ) 2 = q ( v ) I N . \Gamma(v)^2=q(v)I_N. Γ ( v ) 2 = q ( v ) I N .
Let J : C b ↪ C N J:\mathbb{C}^b\hookrightarrow\mathbb{C}^N J : C b ↪ C N be the coordinate inclusion. Consider the affine-linear matrix
M ( v , S ) = ( I r K ( v ) J T 0 0 0 Γ ( v ) − I N 0 0 0 Γ ( v ) J C 0 ( v ) S − I r 0 0 I r ) . \mathcal{M}(v,S)=
\begin{pmatrix}
I_r & K(v)J^{T} & 0 & 0 \\
0 & \Gamma(v) & -I_N & 0 \\
0 & 0 & \Gamma(v) & JC_0(v) \\
S-I_r & 0 & 0 & I_r
\end{pmatrix}. M ( v , S ) = I r 0 0 S − I r K ( v ) J T Γ ( v ) 0 0 0 − I N Γ ( v ) 0 0 0 J C 0 ( v ) I r .
The inverse of the middle triangular block has upper-right block Γ ( v ) − 2 = q ( v ) − 1 I N \Gamma(v)^{-2}=q(v)^{-1}I_N Γ ( v ) − 2 = q ( v ) − 1 I N . Eliminating that block leaves the Schur complement
( I r − P S − I r I r ) . \begin{pmatrix}
I_r & -P \\
S-I_r & I_r
\end{pmatrix}. ( I r S − I r − P I r ) .
The eliminated block has determinant q N q^N q N . We therefore obtain
det M ( v , S ) = q ( v ) N det ( I r + P ( v ) ( S − I r ) ) = q ( v ) N − a f ( v , S ) . (63) \det\mathcal{M}(v,S)=q(v)^N\det\left(I_r+P(v)(S-I_r)\right)=q(v)^{N-a}f(v,S). \tag*{(63)} det M ( v , S ) = q ( v ) N det ( I r + P ( v ) ( S − I r ) ) = q ( v ) N − a f ( v , S ) . (63)
The calculation first holds where q ≠ 0 q\ne0 q = 0 ; both sides are polynomials, so it holds everywhere. The size is 2 r + 2 N < 6 r 2r+2N<6r 2 r + 2 N < 6 r , because b < r b<r b < r and h ≤ r h\le r h ≤ r . Also N ≥ b > a N\ge b>a N ≥ b > a , so the added power of q q q is strictly positive.
This gives a linear-size representation of q N − a f q^{N-a}f q N − a f . It does not give a way to remove the extra factor. The same rational projection P P P has been realized with det D = q N \det D=q^N det D = q N , whereas Proposition 17.1 excludes det D = q a \det D=q^a det D = q a .
Relation to scalar determinantal representations
An exact determinant for the scalar polynomial f f f need not arise from this particular projection formula. The proposition rules out one prescribed rational-matrix realization; it does not rule out other exact representations of f f f .
Exact scalar representations of f f f exist at finite size. Therefore, no normal-form theorem can convert every such representation into a realization excluded by Proposition 17.1 . The next section constructs small scalar matrices with the same base value and first coefficient derivatives, without realizing the prescribed projection in that form.
First-order coefficient data
We compare the residue obstruction with the information obtained by differentiating an arbitrary scalar determinant representation at S = I S=I S = I . We then construct an affine matrix of linear size whose determinant has the same base value and all first coefficient derivatives as the Koszul form. For p ≥ 3 p\ge3 p ≥ 3 , the construction can also preserve the degree in each group of variables.
Throughout this section, use the unpadded Koszul parameters
d 0 = 2 p + 1 , r = ( 2 p + 1 p ) , a = ( 2 p − 1 p − 1 ) , c = ( 2 p p − 1 ) , r = 2 a + c . d_0=2p+1,\qquad r=\binom{2p+1}{p},\qquad a=\binom{2p-1}{p-1},\qquad c=\binom{2p}{p-1},\qquad r=2a+c. d 0 = 2 p + 1 , r = ( p 2 p + 1 ) , a = ( p − 1 2 p − 1 ) , c = ( p − 1 2 p ) , r = 2 a + c .
As before, q ( v ) = ∑ i v i 2 q(v)=\sum_i v_i^2 q ( v ) = ∑ i v i 2 , X = K C 0 X=KC_0 X = K C 0 , and P = X / q P=X/q P = X / q . The scalar identity (62) gives f ( v , I ) = q ( v ) a f(v,I)=q(v)^a f ( v , I ) = q ( v ) a . We first construct a regular representation of this base value and then incorporate the coefficient derivatives.
Regular representations of quadratic powers
Let h = 2 p h=2^p h = 2 p and N = h ⌈ a / h ⌉ N=h\lceil a/h\rceil N = h ⌈ a / h ⌉ . Use a block repetition of the Clifford pencil from the preceding section, so that
Γ ( v ) 2 = q ( v ) I N . \Gamma(v)^2=q(v)I_N. Γ ( v ) 2 = q ( v ) I N .
Let J : C a ↪ C N J:\mathbb{C}^a\hookrightarrow\mathbb{C}^N J : C a ↪ C N be the coordinate inclusion, and let T a T_a T a be the nilpotent Jordan block of order a a a . Define
M a ( v ) = ( − T a J T Γ ( v ) − Γ ( v ) J I N ) . (64) M_a(v)=\begin{pmatrix}
-T_a & J^{T}\Gamma(v)\\
-\Gamma(v)J & I_N
\end{pmatrix}. \tag*{(64)} M a ( v ) = ( − T a − Γ ( v ) J J T Γ ( v ) I N ) . (64)
Eliminate the identity block. Its Schur complement is q I a − T a qI_a-T_a q I a − T a , because J T J = I a J^{T}J=I_a J T J = I a and Γ 2 = q I N \Gamma^2=qI_N Γ 2 = q I N . Therefore
det M a = q a , size ( M a ) = a + N < 2 a + h . \det M_a=q^a,\qquad\operatorname{size}(M_a)=a+N<2a+h. det M a = q a , size ( M a ) = a + N < 2 a + h .
Moreover, at every zero of q q q , this Schur complement is − T a -T_a − T a , which has corank one. The full matrix has corank one there as well, including at the vertex v = 0 v=0 v = 0 .
If h h h divides a a a , the size is exactly 2 a 2a 2 a , equal to the degree of q a q^a q a . There is no conflict with the kernel-degree lemma: for a > 1 a>1 a > 1 the polynomial q a q^a q a is reducible, whereas that lemma requires irreducibility.
For the Koszul values above, the size is linear in r r r . The rounding overhead is only O ( 2 p ) = o ( r ) O(2^p)=o(r) O ( 2 p ) = o ( r ) . Thus the large multiplicity of the base value q a q^a q a does not, by itself, force an additional factor such as log r \log r log r or log r \sqrt{\log r} log r in the matrix size.
Derivatives of a scalar determinant representation
Suppose an arbitrary affine matrix M ( v , S ) M(v,S) M ( v , S ) represents the full scalar polynomial f ( v , S ) f(v,S) f ( v , S ) . Set D ( v ) = M ( v , I ) D(v)=M(v,I) D ( v ) = M ( v , I ) . Since the entries of M M M are affine linear, its change in a coefficient direction T T T is a constant linear map L ( T ) L(T) L ( T ) , independent of v v v . Differentiating the determinant at S = I S=I S = I gives
det D = q a , tr ( D − 1 L ( T ) ) = tr ( P ( v ) T ) . \det D=q^a,\qquad\operatorname{tr}\bigl(D^{-1}L(T)\bigr)=\operatorname{tr}\bigl(P(v)T\bigr). det D = q a , tr ( D − 1 L ( T ) ) = tr ( P ( v ) T ) .
These equalities hold as rational functions where q ≠ 0 q\ne0 q = 0 . By varying T T T , one recovers the entries of P P P as linear combinations of entries of D − 1 D^{-1} D − 1 . But arbitrary linear combinations of those entries need not have the factorized form C D − 1 B CD^{-1}B C D − 1 B . The latter form is the requirement in Theorem 16.1 .
One can see the difference at the generic point of q = 0 q=0 q = 0 . The global corank-one theorem for f f f makes D D D have corank one there. Its Smith normal form consequently has one positive exponent, equal to a a a . By contrast, a realization of the prescribed projection with det D = q a \det D=q^a det D = q a would require a a a positive exponents, all equal to one, as the proof of Theorem 16.1 showed.
Thus the trace measurements in a scalar representation can cancel higher-order poles. The residue criterion does not measure the cost of those cancellations. No lower bound for that cost is established here.
A simultaneous realization of the first derivatives
For a coefficient direction T T T , write the quadratic polynomial
G T ( v ) = tr ( X ( v ) T ) = v T H ( T ) v . G_T(v)=\operatorname{tr}(X(v)T)=v^{T}H(T)v. G T ( v ) = tr ( X ( v ) T ) = v T H ( T ) v .
where H ( T ) H(T) H ( T ) is symmetric and depends linearly on T T T . Equation (62) says that the required first-order expansion is
f ( v , I + T ) = q a + q a − 1 G T ( v ) + O ( T 2 ) . f(v,I+T)=q^{a}+q^{a-1}G_T(v)+O(T^{2}). f ( v , I + T ) = q a + q a − 1 G T ( v ) + O ( T 2 ) .
Here O ( T 2 ) O(T^{2}) O ( T 2 ) means terms of total coefficient degree at least two in T T T ; no analytic estimate is intended. Consider the small affine matrix and scalar factor
A ( v , T ) = ( 0 v T − v I d 0 − H ( T ) ) , l ( T ) = 1 + tr H ( T ) . A(v,T)=\begin{pmatrix}0&v^{T}\\-v&I_{d_0}-H(T)\end{pmatrix},\qquad l(T)=1+\operatorname{tr}H(T). A ( v , T ) = ( 0 − v v T I d 0 − H ( T ) ) , l ( T ) = 1 + tr H ( T ) .
Expanding its determinant to first order gives
det A = q + G T − q tr H + O ( T 2 ) , l ( T ) det A = q + G T + O ( T 2 ) . \det A=q+G_T-q\operatorname{tr}H+O(T^{2}),\qquad l(T)\det A=q+G_T+O(T^{2}). det A = q + G T − q tr H + O ( T 2 ) , l ( T ) det A = q + G T + O ( T 2 ) .
The scalar factor cancels the unwanted trace term. Multiplying this determinant by the regular q a − 1 q^{a-1} q a − 1 block from (64) now gives the right base value and the right first derivative.
We also preserve corank one along the entire base q = 0 q=0 q = 0 . The block diagonal product has corank two there, because both main blocks are singular. A constant upper block reduces this corank to one without changing the determinant. The following choice gives the required coupling. Take T a − 1 T_{a-1} T a − 1 with ones on the superdiagonal. On q = 0 q=0 q = 0 , a left kernel vector of its quadratic-power block has top coordinate a − 1 a-1 a − 1 equal to one. A right kernel vector of A ( v , 0 ) A(v,0) A ( v , 0 ) is ( 1 , v ) T (1,v)^{T} ( 1 , v ) T , with first coordinate one. Put a single constant 1 in the upper block joining these two coordinates. The pairing of the chosen left and right kernel vectors through this block is then one. This is exactly the condition that removes one of the two kernel dimensions.
Including the scalar block l l l , which equals one on the base, gives an affine matrix with
det M ( v , T ) = q a + q a − 1 G T ( v ) + O ( T 2 ) , size ( M ) < 2 a + d 0 + h = O ( r ) . \det M(v,T)=q^{a}+q^{a-1}G_T(v)+O(T^{2}),\qquad\operatorname{size}(M)<2a+d_0+h=O(r). det M ( v , T ) = q a + q a − 1 G T ( v ) + O ( T 2 ) , size ( M ) < 2 a + d 0 + h = O ( r ) .
It has corank one at every point with q = 0 q=0 q = 0 and T = 0 T=0 T = 0 . Because H H H is linear in the whole matrix T T T , this construction matches every coefficient direction simultaneously.
Preservation of bidegree
The construction can be modified to have the same bidegree as the Koszul form, while retaining the base value and first derivatives.
Assume p ≥ 3 p\geq3 p ≥ 3 , so that c ≥ d 0 c\geq d_0 c ≥ d 0 . Choose a linear functional ℓ \ell ℓ on coefficient matrices with ℓ ( I ) = 1 \ell(I)=1 ℓ ( I ) = 1 . Since G I = c q G_I=cq G I = c q , the symmetric matrix representing this quadratic form is H ( I ) = c I d 0 H(I)=cI_{d_0} H ( I ) = c I d 0 . Define
H ~ ( S ) = H ( S − ℓ ( S ) I ) , A ^ ( v , S ) = ( 0 v T − v ℓ ( S ) I d 0 − H ~ ( S ) ) \widetilde{H}(S)=H(S-\ell(S)I),\qquad\widehat{A}(v,S)=\begin{pmatrix}0&v^{T}\\-v&\ell(S)I_{d_0}-\widetilde{H}(S)\end{pmatrix} H ( S ) = H ( S − ℓ ( S ) I ) , A ( v , S ) = ( 0 − v v T ℓ ( S ) I d 0 − H ( S ) )
and
l ^ ( S ) = ℓ ( S ) + tr H ~ ( S ) . \widehat{l}(S)=\ell(S)+\operatorname{tr}\widetilde{H}(S). l ( S ) = ℓ ( S ) + tr H ( S ) .
Use the same constant coupling with a regular q a − 1 q^{a-1} q a − 1 block, and add c − d 0 c-d_0 c − d 0 scalar blocks ℓ ( S ) \ell(S) ℓ ( S ) . The resulting determinant is
g ( v , S ) = q a − 1 det A ^ ( v , S ) ℓ ^ ( S ) ℓ ( S ) c − d 0 . g(v,S)=q^{a-1}\det\widehat{A}(v,S)\widehat{\ell}(S)\ell(S)^{c-d_0}. g ( v , S ) = q a − 1 det A ( v , S ) ℓ ( S ) ℓ ( S ) c − d 0 .
Its degree in v v v is 2 a 2a 2 a , and its degree in S S S is c c c , exactly as for the Koszul form. These two degrees are its bidegree .
We check the first derivative as well. Write S = I + T S=I+T S = I + T and σ = ℓ ( T ) \sigma=\ell(T) σ = ℓ ( T ) . Homogenizing the preceding first-order construction changes its quadratic first derivative to G T + ( d 0 − c ) q σ G_T+(d_0-c)q\sigma G T + ( d 0 − c ) q σ . The first derivative of the additional factor ℓ c − d 0 \ell^{c-d_0} ℓ c − d 0 cancels the extra term. Consequently
g ( v , I + T ) = q a + q a − 1 G T ( v ) + O ( T 2 ) . (65) g(v,I+T)=q^a+q^{a-1}G_T(v)+O(T^2). \tag*{(65)} g ( v , I + T ) = q a + q a − 1 G T ( v ) + O ( T 2 ) . (65)
For the quadratic-power block put s = a − 1 s=a-1 s = a − 1 and N = h ⌈ s / h ⌉ N=h\lceil s/h\rceil N = h ⌈ s / h ⌉ . The full matrix size is
( d 0 + 1 ) + ( s + N ) + 1 + ( c − d 0 ) = c + a + 1 + N ≤ r + h − 1 = r + o ( r ) . (d_0+1)+(s+N)+1+(c-d_0)=c+a+1+N\le r+h-1=r+o(r). ( d 0 + 1 ) + ( s + N ) + 1 + ( c − d 0 ) = c + a + 1 + N ≤ r + h − 1 = r + o ( r ) .
The chosen coupling still gives corank one everywhere on q = 0 q=0 q = 0 at S = I S=I S = I .
The polynomial g g g need not equal the Koszul form: its higher coefficient terms may differ. No claim of irreducibility, conciseness, or corank one away from S = I S=I S = I is made. The construction shows that the scalar base value, all first coefficient derivatives, the target bidegree, and corank one along the base can hold simultaneously at matrix size r + o ( r ) r+o(r) r + o ( r ) . Hence these data alone do not imply a superlinear lower bound.
Limitations of two further reductions
Two further reductions would be relevant to extending the kernel-degree method: bounding the span of homogeneous pieces of kernel coordinates, and deducing simultaneous triangularizability from a homogeneous determinant identity. The following examples show that neither conclusion follows from the conditions considered here.
Homogeneous decompositions of kernel coordinates
Let
F = ∑ i = 1 d y i x i 2 , d ≥ 3. F=\sum_{i=1}^{d}y_i x_i^2,\qquad d\ge3. F = i = 1 ∑ d y i x i 2 , d ≥ 3.
Start with the size-2 d + 1 2d+1 2 d + 1 affine representation
( 0 x T 0 0 I − diag ( y ) − x 0 I ) . \begin{pmatrix}
0 & x^{T} & 0\\
0 & I & -\operatorname{diag}(y)\\
-x & 0 & I
\end{pmatrix}. 0 0 − x x T I 0 0 − diag ( y ) I .
Append a determinant-one triangular block that computes the prefix products ∏ i = 1 j ( 1 + x i ) \prod_{i=1}^{j}(1+x_i) ∏ i = 1 j ( 1 + x i ) . This gives a representation of size 3 d + 1 3d+1 3 d + 1 whose kernel along y = 0 y=0 y = 0 contains the coordinates 1 , x 1 , … , x d 1,x_1,\ldots,x_d 1 , x 1 , … , x d and those prefix products.
There are only linearly many displayed coordinates. Nevertheless, if one separates all of them into homogeneous pieces, their span has dimension
1 + d ( d + 1 ) 2 . 1+\frac{d(d+1)}{2}. 1 + 2 d ( d + 1 ) .
Indeed, in each degree j = 1 , … , d j = 1,\ldots,d j = 1 , … , d , the nonzero homogeneous parts of the prefix products contribute d − j + 1 d-j+1 d − j + 1 independent polynomials; summing these dimensions and adding the constant gives the formula. Thus taking homogeneous pieces can turn a linear coordinate count into a quadratic one.
The added states can be removed, so this is not a counterexample to an invariant minimized over all possible representations. It does show that coefficient purity and factoriality alone do not justify a bound of O ( m ) O(m) O ( m ) for the span obtained by grading an arbitrary size-m m m representation.
Homogeneity and simultaneous triangularization
Consider
N = ( 0 x 0 y 0 x 0 − y 0 ) , D = diag ( N , N ) , b T = ( z , 0 , 0 , 0 , 0 , s ) , c = ( 0 , 0 , w , t , 0 , 0 ) T . N = \begin{pmatrix} 0 & x & 0 \\ y & 0 & x \\ 0 & -y & 0 \end{pmatrix}, \qquad D = \operatorname{diag}(N,N), \qquad b^{T} = (z,0,0,0,0,s), \qquad c = (0,0,w,t,0,0)^{T}. N = 0 y 0 x 0 − y 0 x 0 , D = diag ( N , N ) , b T = ( z , 0 , 0 , 0 , 0 , s ) , c = ( 0 , 0 , w , t , 0 , 0 ) T .
The direct identities N 3 = 0 N^{3}=0 N 3 = 0 and b T c = b T D c = 0 b^{T}c=b^{T}Dc=0 b T c = b T D c = 0 give
det ( 0 b T c I 6 − D ) = s t y 2 − z w x 2 . \det\begin{pmatrix} 0 & b^{T} \\ c & I_{6}-D \end{pmatrix} = st y^{2} - zw x^{2}. det ( 0 c b T I 6 − D ) = s t y 2 − z w x 2 .
Thus the determinant is a concise irreducible homogeneous quartic, represented by an affine matrix of size seven. For example, irreducibility follows by treating it as a primitive polynomial linear in s s s ; the coefficients t y 2 ty^{2} t y 2 and − z w x 2 -zwx^{2} − z w x 2 are relatively prime.
The two coefficient matrices of N N N have no common invariant line. They are nilpotent, and their kernels are distinct coordinate lines, so no line can be invariant under both. The same absence of a common invariant line holds for their doubled matrices in D D D . In particular these coefficient matrices cannot be simultaneously triangularized. A pure-degree determinant identity therefore does not, by itself, supply a common invariant flag.
Computational verification
Finite-field calculations over F 1009 \mathbb{F}_{1009} F 1009 check the quadratic-power determinants and corank-one assertions for p = 1 , 2 , 3 , 4 p=1,2,3,4 p = 1 , 2 , 3 , 4 . The corresponding pairs ( a , matrix size ) (a,\text{matrix size}) ( a , matrix size ) are
( 1 , 3 ) , ( 3 , 7 ) , ( 10 , 26 ) , ( 35 , 83 ) . (1,3),\quad(3,7),\quad(10,26),\quad(35,83). ( 1 , 3 ) , ( 3 , 7 ) , ( 10 , 26 ) , ( 35 , 83 ) .
For p = 3 p=3 p = 3 , the construction matching the first derivative and bidegree has size 42. Checks in three independent coefficient directions verify the base value and the complete derivative in each of those directions, as well as the bidegree scaling. A nonzero isotropic point checks corank one on the base. The source bundle contains the code and its results.
These computations check the explicit matrix identities in small cases. They do not replace the proofs above or the commutative-algebra arguments for the main lower bound.
Open questions
The exact complexities of the forms in Theorem 1.1 remain undetermined. Even the padded odd-skew family with starting size three leaves the interval
n + 2 ≤ dc ( Q 3 , n − 3 ) ≤ n + 4. n+2 \le\operatorname{dc}(Q_{3,n-3}) \le n+4. n + 2 ≤ dc ( Q 3 , n − 3 ) ≤ n + 4.
Determining these values would give a concrete comparison between the geometry of a boundary component and the size of its smallest affine representation.
The principal asymptotic question is whether there are explicit forms with
dc ( f n ) = n , dc ( f n ) = ω ( n ) . \operatorname{dc}(f_n)=n,\qquad\operatorname{dc}(f_n)=\omega(n). dc ( f n ) = n , dc ( f n ) = ω ( n ) .
The kernel-degree bound is sharp for individual representations with free cokernel, so its hypotheses alone do not imply superlinear growth. A stronger argument must use additional properties of the target forms, such as the number of coordinates required by compatible kernel vectors or relations among them.
The rational-matrix criterion in Theorem 16.1 gives an obstruction for a prescribed realization of the Koszul projector. An arbitrary scalar determinantal representation need not induce that realization. Moreover, linear-size constructions can agree with the Koszul form in its base value, all first coefficient derivatives, and, in the stated range, its bidegree. These data therefore do not suffice for a superlinear scalar lower bound. An extension would need to use further simultaneous coefficient identities or a quantitative constraint on the corresponding kernel modules.
A sufficient intermediate result would be a superlinear lower bound on rdc ( det q ) \operatorname{rdc}(\det_q) rdc ( det q ) , the minimum size of a representation of det q \det_q det q that is regular at the zero matrix. § 18 transfers such a bound to padded Koszul forms and gives an equivalent spectral formulation. This problem remains open.
The component count also leaves a separate classification problem. Theorems 12.1 and 13.1 give criteria for further singular matrix spaces to produce boundary components. Finding additional spaces satisfying their recognition and symmetry hypotheses may improve the count. The count does not provide a classification of the entire determinant boundary.
Acknowledgment of assistance
ChatGPT (OpenAI) assisted with mathematical exploration, proof drafting, symbolic and finite-field checks, and preparation of the manuscript.
Computational verification
The source archive contains exact computations verifying the formulas and symmetry dimensions in several small cases. This appendix describes the certificates and their verifiers. These computations are independent checks of the constructions; the theorems for all parameters are proved by the algebraic arguments in the text.
Odd-skew examples
To measure the infinitesimal orbit of a polynomial Q Q Q , form the polynomials X j ∂ X i Q X_j\partial_{X_i}Q X j ∂ X i Q and list their coefficients in a matrix. Its rank is the dimension of the orbit. For two padded odd-skew forms, the resulting matrices and certified ranks are
Table A.1.
Here a nonzero minor modulo 101 proves that the same integer minor is nonzero over C. It therefore gives a rigorous lower bound on the complex rank. The fact that the form lies on the determinant boundary gives the matching upper bound, so the rank is determined exactly.
The archive includes the selected row monomials and column pairs, the integer polynomial generator, and a verifier using only the Python standard library. The verifier uses row elimination; the original certificate search used column elimination. For n = 5 n = 5 n = 5 , a separate calculation expands the full determinant and performs a dense modular verification as well.
Block-space examples
Two further certificates concern the h = 2 h = 2 h = 2 block seed, written in coordinates obtained from Huang–Landsberg’s case (III) by row operations. They check both the orbit dimension and dependence on all variables:
Table 1 (PDF p. 73).
“Partial rank” is the rank of the first partial derivatives. Its being equal to the number of variables certifies conciseness. The corresponding minors are 16 and 85 modulo 101.
The independent verifier expands the entire determinant using monomial exponent vectors. It checks that all coefficients before the claimed leading coefficient vanish, compares that leading coefficient exactly, and verifies the chosen orbit minors by row elimination. These checks are separate from the uniform proof over C.
Koszul examples
For the exterior-algebra construction, exact rational calculations and finite-field calculations check the constant and degree-one compression maps. The numbers in the first two cases are
Table 2.
The numerator records the rank of a linear map; the denominator records the dimension of its domain. The last column counts the remaining linear relations after passing to the quotient used in the component proof. These are the same dimensions obtained in that proof for all p p p by representation theory. Splitting the calculation into weight spaces makes the second example manageable: its largest block has only 140 columns.
These certificates do not replace the proofs for general parameters, and they do not check the commutative-algebra arguments used for the exact-complexity lower bound. Their role is to make the displayed constructions and several decisive dimension counts independently reproducible.