Dedicated to Professor Francisco Marcellán on the occasion of his 75th birthday
Introduction
The objective is to calculate orthogonal polynomials after independent rational matrix transformations on the two arguments of a pairing. The calculation must preserve the prescribed degrees of each polynomial component and the equations defining its algebraic support. The results give the transformed polynomials, their reproducing kernel, the complete freedom of added masses, and the sizes of the determinants required to compute them.
In mixed-type multiple orthogonality, a polynomial row is required to be orthogonal to prescribed polynomial columns. A typical pairing is
β μ ( f , g ) : = ∫ Z f ( x ) d μ ( x ) g ( x ) , \beta_{\mu}(f,g) := \int_{Z} f(x)\,\mathrm{d}\mu(x)\,g(x), β μ ( f , g ) := ∫ Z f ( x ) d μ ( x ) g ( x ) ,
where Z Z Z is the specified integration support, μ \mu μ is a q q q -by-p p p matrix of measures, f f f is a polynomial row with q q q components, and g g g is a polynomial column with p p p components. The product under the integral is a scalar. Each component of f f f and g g g is assigned its own degree bound. To construct a row polynomial, the column space supplies the orthogonality tests; to construct a column polynomial, the row space supplies them. In either case, the unknown polynomial is required to pair to zero with every element of the opposite test space, and its remaining coefficients are fixed by a prescribed normalization.
More generally, moments can be prescribed by linear functionals instead of measures. The algebraic formulation by a matrix of linear forms belongs to the vector orthogonality framework of Sorokin and Van Iseghem [33 ] ; the mixed-type orthogonality conditions extend types I and II as in Daems and Kuijlaars [34 ] . This notation is bilinear, hence linear in both arguments. When a sesquilinear pairing is used, the second argument is conjugated explicitly.
Polynomial coefficients and matrix entries may be complex even when the integration variables are real. In the real-support examples, the integration domain is specified separately from the complex coordinate algebra used for the polynomial calculations. Added evaluation or derivative functionals may instead be supported at complex points; their support is stated when they are introduced.
The transformations of Christoffel and Geronimus multiply or divide the measure by polynomial factors, and Uvarov modifications add masses [29 , 30 , 31 ] . The transformations considered here use independent rational matrix factors on both sides of the pairing. The order of the matrix factors matters: the coordinate algebra is commutative, but these matrices generally do not commute. The row and column may have different component counts and component degree bounds. The variables satisfy the equations of an algebraic support of any dimension. Reducible supports, algebraic multiplicities and pairings depending on two independent support points are included.
We construct the perturbed mixed-type multiple orthogonal polynomials and their reproducing kernel , a finite sum of polynomial column–row products in two sets of variables. Pairing in one set of variables reproduces polynomials from the chosen space in the other. Division leaves homogeneous masses , terms vanishing after both denominator actions. Our Christoffel–Geronimus–Uvarov problem is to determine all these masses and compute the transformed polynomials and kernels while retaining the prescribed component degree bounds and the defining relations of the support.
More precisely, the data are two q q q -by-q q q polynomial matrices L 0 L_0 L 0 , L 1 L_1 L 1 and two p p p -by-p p p polynomial matrices R 0 R_0 R 0 , R 1 R_1 R 1 . The new pairing β ^ \widehat{\beta} β is required to satisfy
β ^ ( f L 1 , R 1 g ) = β μ ( f L 0 , R 0 g ) \widehat{\beta}(fL_1,R_1g)=\beta_{\mu}(fL_0,R_0g) β ( f L 1 , R 1 g ) = β μ ( f L 0 , R 0 g )
for every polynomial row f f f and column g g g .
The identity is formulated using multiplication. Under the injectivity hypotheses stated below, the rational factors are S L : = L 1 − 1 L 0 S_L:=L_1^{-1}L_0 S L := L 1 − 1 L 0 and S R : = R 0 R 1 − 1 S_R:=R_0R_1^{-1} S R := R 0 R 1 − 1 . Evaluating the original pairing on f S L , S R g fS_L,S_Rg f S L , S R g gives a particular solution; its rational moments are additional input whenever they are not fixed by the original polynomial moments. Every other solution adds a mass that vanishes on f L 1 , R 1 g fL_1,R_1g f L 1 , R 1 g . Section 2.1 proves these assertions and also treats the consistency question when the matrices are not injective.
The calculation begins with finite polynomial spaces and an original biorthogonal block on which the pairing matrix is invertible. Multiply the target rows and columns by the rational factors and enlarge the original spaces until all these products can be expressed in them. Their coordinate matrices determine linear equations whose solutions are exactly the admissible images. These equations impose both polynomial divisibility and the required component degrees. Pair the added functions with the original polynomials, insert the chosen mass, and impose the normalization. Eliminating the coefficients already determined by the original orthogonality gives the reduced matrix of the construction. The scalar calculation in Subsection 3.1 carries out every step before the general matrix formula is introduced.
The final step is block elimination. Its inputs and admissible subspaces are supplied by the construction: the ordered rational matrix actions, the complete degree and divisibility equations on the algebraic support, and the mass coordinates. In particular, the constraints can be computed by polynomial reduction followed by a nullspace calculation. Values and derivatives, and residues for the finite quotients treated below, give alternative coordinates for the divisibility conditions. The degree or section-space conditions are imposed alongside them. Pairings with the added rational functions play the role of functions of the second kind; for measures, Subsubsection 3.2.3 gives their Cauchy and Fantappiè integral representations.
The algebraic support affects this calculation through its polynomial relations. Two expressions that differ by a defining relation represent the same function, so they must not be counted as independent directions. At singular points or on supports with multiplicities, these relations also restrict which derivative data can occur. We retain them throughout the construction. Subsubsection 3.2.6 calculates all the blocks on a curve, with noncommuting matrix factors on both sides and a mass that cannot be split into unilateral masses, and verifies the resulting polynomial and kernel directly.
The main construction can be summarized as follows. A square pairing is called normal when its matrix in bases of the chosen spaces is invertible. A normalized orthogonal block consists of polynomial vectors orthogonal to the opposite test space and satisfying specified linear normalization conditions.
Main construction. Let A = C [ x 1 , … , x M ] / I A=\mathbb{C}[x_1,\ldots,x_M]/I A = C [ x 1 , … , x M ] / I be the coordinate algebra of the support. Assume that the four polynomial matrices L 0 L_0 L 0 , L 1 L_1 L 1 , R 0 R_0 R 0 , R 1 R_1 R 1 have regular determinants, meaning that multiplication by each determinant is injective on A A A . Fix a rational extension of the original pairing, a homogeneous mass, and finite polynomial row and column spaces. For a normalized problem, take k > 0 k>0 k > 0 , a candidate space of dimension r + k r+k r + k , an opposite test space of dimension r r r , and a prescribed C k C^k C k -valued linear normalization. Choose an original normal block and finite rational enlargements as in Section 3.
The original biorthogonal data determine reduced square matrices for the right and left normalized problems. Each problem has a unique solution if and only if its reduced determinant is nonzero. Every entry of the solution is a bordered determinant divided by that reduced determinant. For equal-dimensional target spaces, the same reduced matrix that tests normality also gives their reproducing kernel.
The orders of these matrices are computed from dimensions of polynomial images, complete constraint spaces and the restricted mass. For fixed factors and original degree-ordered blocks normal at the levels considered, take target degree-ordered flags or component bounds of bounded dispersion on each side. If the normalization dimension is of order at most n s − 1 n^{s-1} n s − 1 , the degree-based construction has order O ( n s − 1 ) O(n^{s-1}) O ( n s − 1 ) as n → ∞ n \to\infty n → ∞ , where s = dim A ≥ 1 s = \dim A \ge1 s = dim A ≥ 1 and n n n is the largest target component degree.
The determinant identities and solution formulas are proved in Theorem 3.3 , Corollary 3.4 and the kernel construction (24) ; Theorem 3.6 and Proposition 3.9 give the size statements. A tau determinant, denoted by τ \tau τ , is the determinant of the reduced system. Bordering means adding one row and one column to that matrix. The prescribed component degree bounds and normalization are imposed directly, so the construction applies without normality assumptions on intermediate transformations. On each finite space, Proposition 3.2 also supplies a separating list of value and derivative conditions for division; their concrete coordinates are computed in the examples and specializations.
The order of the determinant is part of the output. Subsection 3.3 calculates the orders for specified constructions, including all four matrix factors. Theorem 3.6 gives an explicit degree count and a reduced count obtained from two polynomial coefficient matrices. Corollary 3.8 gives the exact stable count on curves, including the row and column component counts, masses and the position within the last degree block. The bounds of Proposition 3.9 explain the change from bounded orders on curves to at most linear growth on surfaces. The comparison with formulas using a fixed range of consecutive polynomial degrees states which determinant representation is being counted. It also explains how matrix division and component bounds contribute information beyond the degree of a matrix determinant.
The mass equation supplies the remaining input to these matrices. Theorem 2.2 describes every solution when numerator and denominator actions are injective: no nonzero polynomial row or column on the support is sent to zero. Proposition 2.1 also decides existence when this fails. Cancelling common rational factors can discard masses. For pairings allowing polynomial factors to move between arguments, Theorem 2.4 decides whether a bilateral mass splits into left and right masses. Shared pieces of the denominator zero sets can prevent splitting. At their intersections, a mass killed by each denominator separately can be added to one summand and subtracted from the other without changing the sum.
The geometric part of the article calculates particular inputs to the reduced matrices. A divisibility condition is a condition on a remainder class. A mass also evaluates such classes, because it vanishes on the image of the denominator action. Finite quotient coordinates therefore provide a common method for both parts, with the quotient chosen for the matrix division or denominator mass being calculated. These quotients can differ: a Christoffel numerator imposes divisibility conditions even when the denominators are identity matrices and there is no homogeneous mass. Section 4 constructs dual coordinates for the relevant quotients using residues. On an integral curve, a geometric Cauchy kernel gives a matrix divided difference that reproduces the remainder. At a singular point the residue is interpreted using the module of dualizing differentials, which retains the relations between the branches. On surfaces the construction concerns finite quotients obtained by imposing a matrix equation and an additional scalar equation. These coordinates supply the linear tests and mass entries of the Christoffel-type determinant formulas. The rational moments and the freely chosen mass parameters remain part of the input; polynomial moments alone need not determine them.
These residue coordinates are then used in an explicit family. The curve family in Section 5 gives the factors, residue functionals and finite Grams explicitly. Its normality is verified on the stated component bounds by exploiting the triangular matrix structure. The residue quotients remain perfectly paired when denominator zeros reach a non-Gorenstein singular point. A constant nonzero minor also proves that the matrix mass keeps its full rank, so no mass direction is lost at the collision. Once this property and invertibility of the finite Grams are established, the inverse Gram formulas give regular limits of the normalized polynomials and kernels.
The same construction also yields the standard specializations. On the line, Theorem 6.1 and Proposition 6.2 express the division conditions through principal parts, complete Jordan chains and forbidden coefficients at infinity. The Cauchy data include shared poles and their confluent limits. On Laurent supports, conjugation changes the second argument and the degree bounds depend on the chosen embedding; the finite construction retains both features. For symmetric congruences, Proposition 7.1 describes the complete mass quotient and classifies positive balanced masses on real curves. The examples distinguish these evaluation masses from positive bivariate pairings that couple distinct points. Theorem 7.2 eliminates the common space by orthogonal projection and gives the normalized polynomials and kernel from two positive matrices of the same explicit order. Only the mass on the overlap increases that order. Under the coprimality and degree conditions of Corollary 7.3, every positive homogeneous mass leaves the order unchanged. Proposition 7.4 constructs a noncommutative matrix family on a nodal curve, with a positive Sobolev mass involving values and derivatives at the node and exact counts for unequal component bounds. This derivative mass belongs to the positive, generally nonbalanced theory. The coordinate identities below apply when both pairings are balanced, as happens for the evaluation-mass specialization.
Once the polynomial bases have been constructed, they determine coordinate multiplication matrices. For balanced pairings on nested normal spaces that together contain every polynomial vector, multiplication by a coordinate has a matrix in the orthogonal polynomial bases. The four factors give connection matrices between the original and transformed bases. Theorem 8.1 proves four-factor intertwining identities simultaneously for all coordinates. The matrix identities preserve the equations defining the support. For the stated positive congruences, orthonormal bases make the multiplication matrices symmetric.
Relation to previous work and organization
The scalar rational transformation problem has an established spectral formulation in [32 ] . The algebraic theory of vector orthogonality with a matrix of linear forms is developed by Sorokin and Van Iseghem [33 ] ; the mixed-type formulation and its Christoffel–Darboux formula are treated by Daems and Kuijlaars [34 ] . Matrix rational transformations and bivariate pairings on the line are developed in [1 ] , including noncommutative factors and bilateral specializations; the matrix Christoffel case appears in [2 ] . Christoffel, Geronimus and Uvarov transformations for mixed-type orthogonality on the step-line are treated in [3 , 4 , 5 ] . Kozhan and Vaktnäs [6, Theorems 3.1 and 3.3] give rational perturbation formulas for types I and II on the line with general multi-indices and componentwise scalar factors; their companion paper [44 ] treats Christoffel transformations and their relation to nearest-neighbour recurrences. Bueno and Marcellán [42 ] formulate scalar polynomial perturbations through separate factors in the two arguments of a bilinear functional. Their detailed factorization results concern one-sided modifications and the same modification in both arguments. The construction here retains independent matrix actions on both polynomial spaces over the coordinate algebra of the support, with the prescribed component bounds.
For equal component bounds and a square positive pairing on the line, Corollary 7.5 gives the monic matrix orthogonal polynomials, norm matrices and Christoffel–Darboux kernel after a rational congruence of the measure. This connects the construction with the matrix weights of Durán and Grünbaum [26 , 27 , 28 ] . Subsection 7.2 recovers their two Hermite matrix weights by polynomial triangular factors, computes the reduced orders and displays a monic formula in scalar Hermite polynomials. Their differential equations impose additional conditions on the weights; the rational construction supplies the polynomial and kernel transformation.
In several variables, Fernández, Pérez, Piñar and Xu [43 ] give polynomial and reproducing-kernel formulas for the addition of a Dirac mass to a positive definite functional. The formulas of [35 ] treat finitely many point-mass additions. Christoffel modifications of total degree two and Uvarov modifications are treated in [36 ] ; scalar Geronimus transformations in two variables, their connection formulas and block Jacobi factorizations appear in [37 ] . Multivariate rational formulas in [7 ] use scalar polynomial factors. Related multivariate Christoffel constructions appear in [8 , 9 ] . Here the support relations, the full bilateral mass equation and the finite component degree bounds enter one determinant construction. Polynomial classes on singular or nonreduced supports are included, subject to the stated regularity and finite normality assumptions. Thus the extension is the joint construction with independent matrix actions on the two polynomial spaces over a quotient algebra, together with complete mass data and determinant counts for those spaces.
Fasondini, Olver and Xu [10 , 11 ] construct orthogonal bases and computational recurrences on classes of planar algebraic curves, including singular cubics. The present problem starts from an original biorthogonal system and computes the systems produced by independent rational matrix factors on its two arguments. The scalar calculation in Subsection 3.1 and the full mixed-type example in Subsubsection 3.2.6 illustrate the construction of the finite matrices from the original data.
The geometric ingredients have their own established background. Bertola [12 ] relates scalar section biorthogonality on smooth curves to matrix biorthogonality through a finite projection. His Proposition 3.8 also gives rational left and right changes induced by changes of geometric data. The curve residues below use the associated Cauchy-kernel framework. On smooth complex curves, the geometric Szegő kernel has the classical theta-function description of Fay [46 ] and the bundle formulation of Ben-Zvi and Biswas [18 ] . For smooth projective curves over finite fields, Hu and Yu [13 Theorems 6.2 and 6.3] identify the finite residual Casimir tensor with the intrinsic principal part of a normalized Szegő kernel and prove its reproduction property. This is the direct scalar, smooth antecedent of the residue-quotient construction below. Here that construction is carried out on integral curves with singular points, including non-Gorenstein points, and for matrix quotients with the explicit order of multiplication required by the bilateral formulas. Rosenlicht’s residue description of regular differentials, presented in [45 Chapter IV] , supplies the dualizing module at those points; see also [39 ] . The finite quotient construction on surfaces is stated separately with its own hypotheses. The recurrence interpretation is connected with the functional and Darboux transformations of [38 ] and with the Hessenberg-matrix factorizations for polynomial perturbations of bilinear functionals in [42 ] . Here the same four connection matrices intertwine every coordinate action and preserve the polynomial relations of the support.
Section 2 develops the functional equation and its mass space. Section 3 proves the determinant formulas and explains their Gram realization. Section 4 provides the curve and surface residue coordinates. Section 5 develops an explicit bilateral family on singular curves and proves regularity of its normalized polynomials. Sections 6 and 7 treat standard, Laurent and positive specializations. Section 8 proves the simultaneous coordinate Darboux identities.
Functional equation and homogeneous masses
The functional equation prescribes the new pairing only on arguments multiplied by the two denominator matrices. The first task is to decide whether these prescribed values are consistent. The second is to describe every value that remains free. These free values are the homogeneous masses. On a curve their description reduces to finitely many polynomial remainders, although a bivariate mass can still involve arbitrary functionals in the opposite argument. The last part of the section decides when a mass can be written as the sum of a left mass and a right mass, each annihilated by just one denominator.
The functional equation and all its masses
The functional equation and all its masses. We encode the support equations in the coordinate algebra A : = C [ x 1 , … , x M ] / I A := \mathbb{C}[x_1,\ldots,x_M]/I A := C [ x 1 , … , x M ] / I , where I I I is a proper ideal (1 ∉ I 1 \notin I 1 ∈ / I ). Thus two polynomials represent the same element of A A A when their difference belongs to I I I ; addition and multiplication are performed subject to these relations. Write Z o Z^o Z o for this algebraic support, retaining multiplicities.1 Here dim A \dim A dim A is the dimension of the support, whereas vector-space dimensions are over C \mathbb{C} C . For instance, C [ x ] \mathbb{C}[x] C [ x ] has support dimension one but infinitely many linearly independent powers of x x x .
The number M M M counts coordinate variables. The row and column sizes q q q , p p p introduced below count polynomial components, not variables.
A scalar means an element a ∈ A a \in A a ∈ A , possibly a nonconstant polynomial, as opposed to a matrix over A A A . On a vector it multiplies every component by a a a . It is regular if a b = 0 ab = 0 ab = 0 forces b = 0 b = 0 b = 0 . Thus multiplication by a regular scalar is injective, and cancellation by that scalar is valid. Let K K K consist of fractions with regular denominators. For integral A A A this is the usual fraction field; on A = C [ x , y ] / ( x y ) A = \mathbb{C}[x,y]/(xy) A = C [ x , y ] / ( x y ) , x x x is not regular because x y = 0 xy = 0 x y = 0 with y ≠ 0 y \ne0 y = 0 , whereas 1 + x + y 1 + x + y 1 + x + y is regular.
A linear functional is a linear rule assigning a complex number to each vector; evaluations and moments are examples. We write W ∗ : = Hom C ( W , C ) W^* := \operatorname{Hom}_{\mathbb{C}}(W,\mathbb{C}) W ∗ := Hom C ( W , C ) for the space of all such rules on W W W , called its algebraic dual. Fix positive integers q q q , p p p and write Mat a × b ( B ) \operatorname{Mat}_{a\times b}(B) Mat a × b ( B ) for a a a -by-b b b matrices with entries in B B B , with Mat a ( B ) : = Mat a × a ( B ) \operatorname{Mat}_a(B) := \operatorname{Mat}_{a\times a}(B) Mat a ( B ) := Mat a × a ( B ) . Matrix entries have comma-separated indices, as in H = [ H i , j ] H = [H_{i,j}] H = [ H i , j ] . The superscript T \mathrm{T} T denotes transpose; semicolons separate block rows in inline matrices.
We keep track of which argument a polynomial belongs to. For an arbitrary bivariate pairing use A : = A ⊗ C A \mathcal{A} := A \otimes_{\mathbb{C}} A A := A ⊗ C A , the algebra of two independent copies of the support. The maps ι L \iota_L ι L and ι R \iota_R ι R place a polynomial in the left or right copy: ι L ( a ) : = a ⊗ 1 \iota_L(a) := a \otimes1 ι L ( a ) := a ⊗ 1 and ι R ( a ) : = 1 ⊗ a \iota_R(a) := 1 \otimes a ι R ( a ) := 1 ⊗ a . For A = C [ t ] A = \mathbb{C}[t] A = C [ t ] , this means A = C [ x , y ] \mathcal{A} = \mathbb{C}[x,y] A = C [ x , y ] , ι L ( a ) = a ( x ) \iota_L(a) = a(x) ι L ( a ) = a ( x ) and ι R ( a ) = a ( y ) \iota_R(a) = a(y) ι R ( a ) = a ( y ) , so ι L ( a ) ι R ( b ) = a ( x ) b ( y ) \iota_L(a)\iota_R(b) = a(x)b(y) ι L ( a ) ι R ( b ) = a ( x ) b ( y ) . A balanced pairing allows a polynomial factor to be moved from one argument to the other. For this case one copy suffices: A : = A \mathcal{A} := A A := A and ι L ( a ) : = ι R ( a ) : = a \iota_L(a) := \iota_R(a) := a ι L ( a ) := ι R ( a ) := a . Both maps act entry by entry on matrices.
For a row f ∈ A 1 × q f \in A^{1 \times q} f ∈ A 1 × q and a column g ∈ A p × 1 g \in A^{p \times1} g ∈ A p × 1 , the basic scalar tests are the products ι L ( f i ) ι R ( g j ) \iota_L(f_i)\iota_R(g_j) ι L ( f i ) ι R ( g j ) . In the bivariate convention these are f i ( P ) g j ( Q ) f_i(P)g_j(Q) f i ( P ) g j ( Q ) , with independent support points P , Q P,Q P , Q ; in the balanced convention they are products in the same algebra A A A . Choose a scalar functional μ i , j \mu_{i,j} μ i , j for each component pair and define
β μ ( f , g ) : = ∑ i = 1 q ∑ j = 1 p μ i , j [ ι L ( f i ) ι R ( g j ) ] . \beta_\mu(f,g) := \sum_{i=1}^{q}\sum_{j=1}^{p}\mu_{i,j}[\iota_L(f_i)\iota_R(g_j)]. β μ ( f , g ) := i = 1 ∑ q j = 1 ∑ p μ i , j [ ι L ( f i ) ι R ( g j )] .
For a balanced matrix measure this is the familiar integral ∫ f d μ g \int f\,d\mu\,g ∫ f d μ g . The balanced identity β μ ( a f , g ) = β μ ( f , a g ) \beta_\mu(af,g) = \beta_\mu(f,ag) β μ ( a f , g ) = β μ ( f , a g ) follows because multiplication of scalar entries in A A A is commutative. For A = C [ x ] A = \mathbb{C}[x] A = C [ x ] , the matrix of linear forms in this balanced pairing is the framework of Sorokin and Van Iseghem [33 ] . The coordinate algebra and the two-copy construction above allow the same functional notation to express support relations and pairings with independent arguments.
It is useful to collect the q p qp q p tests in a matrix. Put V : = Mat q × p ( A ) V := \operatorname{Mat}_{q \times p}(\mathcal{A}) V := Mat q × p ( A ) and write
H f , g : = [ ι L ( f i ) ι R ( g j ) ] i ∈ { 1 , … , q } , j ∈ { 1 , … , p } = ι L ( f ) T ι R ( g ) T , μ [ H ] : = ∑ i , j μ i , j ( H i , j ) . H_{f,g} := [\iota_L(f_i)\iota_R(g_j)]_{i \in\{1,\ldots,q\},\,j \in\{1,\ldots,p\}} = \iota_L(f)^{\mathsf{T}}\iota_R(g)^{\mathsf{T}}, \qquad\mu[H] := \sum_{i,j}\mu_{i,j}(H_{i,j}). H f , g := [ ι L ( f i ) ι R ( g j ) ] i ∈ { 1 , … , q } , j ∈ { 1 , … , p } = ι L ( f ) T ι R ( g ) T , μ [ H ] := i , j ∑ μ i , j ( H i , j ) .
Then β μ ( f , g ) = μ [ H f , g ] \beta_\mu(f,g) = \mu[H_{f,g}] β μ ( f , g ) = μ [ H f , g ] . A general test matrix H ∈ V H \in V H ∈ V need not come from a single pair f , g f,g f , g . It is a finite sum of such matrices: use matrix units for the component indices and, in the bivariate case, express each tensor as a finite sum of elementary tensors. Consequently, specifying the pairing specifies μ \mu μ on all of V V V . In the bivariate convention, any complex bilinear pairing is represented in this way: take μ i , j ( a ⊗ b ) \mu_{i,j}(a \otimes b) μ i , j ( a ⊗ b ) to be its value on the row with component a a a at i i i and the column with component b b b at j j j , and extend linearly. In the balanced convention the same rule depends only on the product a b ab ab , because balance gives β ( a e i T , b e j ) = β ( e i T , a b e j ) \beta(ae_i^{\mathsf{T}},be_j) = \beta(e_i^{\mathsf{T}},abe_j) β ( a e i T , b e j ) = β ( e i T , ab e j ) , where e i , e j e_i,e_j e i , e j are coordinate columns of the appropriate sizes. This explains precisely which pairings are represented by each algebra.
Fix L 0 , L 1 ∈ Mat q ( A ) L_0,L_1 \in\operatorname{Mat}_q(A) L 0 , L 1 ∈ Mat q ( A ) and R 0 , R 1 ∈ Mat p ( A ) R_0,R_1 \in\operatorname{Mat}_p(A) R 0 , R 1 ∈ Mat p ( A ) . We seek a new pairing whose value on f L 1 , R 1 g fL_1,R_1g f L 1 , R 1 g equals the old pairing on f L 0 , R 0 g fL_0,R_0g f L 0 , R 0 g , for every polynomial row f f f and column g g g . This prescription makes sense using multiplication alone, even when some factors cannot be inverted. Subscripts 0 0 0 and 1 1 1 distinguish the factors on the original and perturbed sides; under regularity they become the numerator and denominator factors of rational ratios.
To express the prescription on arbitrary tests, first observe that
H f L i , R i g = ι L ( L i ) T H f , g ι R ( R i ) T . H_{fL_i,R_i g} = \iota_L(L_i)^{\mathsf{T}} H_{f,g}\iota_R(R_i)^{\mathsf{T}}. H f L i , R i g = ι L ( L i ) T H f , g ι R ( R i ) T .
To verify the matrix order, fix a transformation index ϵ ∈ { 0 , 1 } \epsilon\in\{0,1\} ϵ ∈ { 0 , 1 } . The ( a , b ) (a,b) ( a , b ) entry of the transformed test is
∑ u , v ι L ( ( L ϵ ) u , a ) ( H f , g ) u , v ι R ( ( R ϵ ) b , v ) . \sum_{u,v}\iota_L((L_\epsilon)_{u,a})(H_{f,g})_{u,v}\iota_R((R_\epsilon)_{b,v}). u , v ∑ ι L (( L ϵ ) u , a ) ( H f , g ) u , v ι R (( R ϵ ) b , v ) .
The left matrix is therefore ι L ( L ϵ ) T \iota_L(L_\epsilon)^{\mathsf{T}} ι L ( L ϵ ) T and the right matrix is ι R ( R ϵ ) T \iota_R(R_\epsilon)^{\mathsf{T}} ι R ( R ϵ ) T . This expansion checks both transposes directly from ordinary row and column multiplication. Extend this transformation linearly to every H ∈ V H \in V H ∈ V by setting
D i ( H ) : = ι L ( L i ) T H ι R ( R i ) T , ( L i μ R i ) [ H ] : = μ [ D i ( H ) ] . D_i(H) := \iota_L(L_i)^{\mathsf{T}}H\iota_R(R_i)^{\mathsf{T}}, \qquad(L_i\mu R_i)[H] := \mu[D_i(H)]. D i ( H ) := ι L ( L i ) T H ι R ( R i ) T , ( L i μ R i ) [ H ] := μ [ D i ( H )] .
Thus multiplication of a functional means applying the corresponding multiplication to its test before evaluation. Given μ \mu μ , the functional equation for μ ^ ∈ V ∗ \widehat{\mu} \in V^* μ ∈ V ∗ is
L 1 μ ^ R 1 = L 0 μ R 0 . (1) L_1\widehat{\mu}R_1 = L_0\mu R_0. \tag*{(1)} L 1 μ R 1 = L 0 μ R 0 . (1)
Equivalently, the two pairings satisfy
β μ ^ ( f L 1 , R 1 g ) = β μ ( f L 0 , R 0 g ) for every f , g . (2) \beta_{\widehat{\mu}}(fL_1,R_1g) = \beta_\mu(fL_0,R_0g) \qquad\text{for every } f,g. \tag*{(2)} β μ ( f L 1 , R 1 g ) = β μ ( f L 0 , R 0 g ) for every f , g . (2)
The next criterion allows factors that kill nonzero classes. For a pairing β \beta β , the left radical consists of rows f f f with β ( f , g ) = 0 \beta(f,g)=0 β ( f , g ) = 0 for every column g g g ; the right radical consists of columns g g g for which this holds for every row f f f . Nondegeneracy means that these sets contain only the zero row and zero column, respectively.
There is one possible obstruction to this prescription. Distinct tests H 1 , H 2 H_1,H_2 H 1 , H 2 may give the same transformed test D 1 ( H 1 ) = D 1 ( H 2 ) D_1(H_1)=D_1(H_2) D 1 ( H 1 ) = D 1 ( H 2 ) . They must then prescribe the same value of μ ^ \hat{\mu} μ ^ . The next proposition states that this consistency requirement is sufficient, and identifies all the values that remain free.
Proposition 2.1 (Existence for arbitrary factors). Equation (2.1) has a solution if and only if
μ [ D 0 ( H ) ] = 0 for every H ∈ ker D 1 . (3) \mu[D_0(H)]=0 \qquad\text{for every } H\in\ker D_1. \tag*{(3)} μ [ D 0 ( H )] = 0 for every H ∈ ker D 1 . (3) All solutions are obtained from one by adding a functional on V / D 1 V V/D_1V V / D 1 V . For a , b ∈ A a,b\in A a , b ∈ A and a scalar balanced functional u ∈ A ∗ u\in A^* u ∈ A ∗ with nondegenerate pairing, the equation b v = a u bv=au b v = a u has a solution if and only if Ann ( b ) ⊆ Ann ( a ) \operatorname{Ann}(b)\subseteq\operatorname{Ann}(a) Ann ( b ) ⊆ Ann ( a ) , where Ann ( b ) : = { f ∈ A : b f = 0 } \operatorname{Ann}(b):=\{f\in A:bf=0\} Ann ( b ) := { f ∈ A : b f = 0 } . The solution freedom is ( A / ( b ) ) ∗ (A/(b))^* ( A / ( b ) ) ∗ . If v v v is also nondegenerate, the two annihilators are equal.
Proof. Equation (2.1) asks for a linear functional μ ^ \hat{\mu} μ ^ satisfying μ ^ [ D 1 ( H ) ] = μ [ D 0 ( H ) ] \hat{\mu}[D_1(H)]=\mu[D_0(H)] μ ^ [ D 1 ( H )] = μ [ D 0 ( H )] for all H ∈ V H\in V H ∈ V . If D 1 ( H ) = 0 D_1(H)=0 D 1 ( H ) = 0 , the left-hand side is zero, proving necessity.
Conversely, assume (2.3). If D 1 ( H 1 ) = D 1 ( H 2 ) D_1(H_1)=D_1(H_2) D 1 ( H 1 ) = D 1 ( H 2 ) , then H 1 − H 2 ∈ ker D 1 H_1-H_2\in\ker D_1 H 1 − H 2 ∈ ker D 1 , so μ [ D 0 ( H 1 ) ] = μ [ D 0 ( H 2 ) ] \mu[D_0(H_1)]=\mu[D_0(H_2)] μ [ D 0 ( H 1 )] = μ [ D 0 ( H 2 )] . The prescribed value therefore depends only on D 1 ( H ) D_1(H) D 1 ( H ) . More explicitly, the rule
λ ( D 1 ( H ) ) : = μ [ D 0 ( H ) ] \lambda(D_1(H)):=\mu[D_0(H)] λ ( D 1 ( H )) := μ [ D 0 ( H )]
is well defined and linear on D 1 V D_1V D 1 V . Linearity follows by replacing H H H by a H + b K aH+bK a H + b K for complex scalars a , b a,b a , b . Complete a basis of D 1 V D_1V D 1 V to one of V V V and choose arbitrary values on the added vectors. This extends the functional to a solution μ ^ \hat{\mu} μ ^ on V V V .
The difference of two solutions vanishes on D 1 V D_1V D 1 V , so it gives the same value to tests differing by an element of D 1 V D_1V D 1 V . It therefore defines a functional on V / D 1 V V/D_1V V / D 1 V . Conversely, adding any such functional preserves the equation.
In the scalar case, b v = a u bv=au b v = a u means v ( b h ) = u ( a h ) v(bh)=u(ah) v ( bh ) = u ( ah ) for every h ∈ A h\in A h ∈ A . If b f = 0 bf=0 b f = 0 , then u ( ( a f ) g ) = v ( b f g ) = 0 u((af)g)=v(bfg)=0 u (( a f ) g ) = v ( b f g ) = 0 for every g ∈ A g\in A g ∈ A . Thus a f af a f lies in the radical of the pairing of u u u ; nondegeneracy forces a f = 0 af=0 a f = 0 . This proves Ann ( b ) ⊆ Ann ( a ) \operatorname{Ann}(b)\subseteq\operatorname{Ann}(a) Ann ( b ) ⊆ Ann ( a ) . Conversely, this inclusion gives b f = 0 bf=0 b f = 0 implies a f = 0 af=0 a f = 0 , hence u ( a f ) = 0 u(af)=0 u ( a f ) = 0 : precisely (2.3) for D 1 ( h ) = b h D_1(h)=bh D 1 ( h ) = bh and D 0 ( h ) = a h D_0(h)=ah D 0 ( h ) = ah . The image of multiplication by b b b is the ideal ( b ) (b) ( b ) , so the freedom is ( A / ( b ) ) ∗ (A/(b))^* ( A / ( b ) ) ∗ . Finally, if v v v is also nondegenerate, a f = 0 af=0 a f = 0 implies v ( ( b f ) g ) = u ( a f g ) = 0 v((bf)g)=u(afg)=0 v (( b f ) g ) = u ( a f g ) = 0 for every g g g , hence b f = 0 bf=0 b f = 0 . This gives the reverse inclusion and proves equality. □
A filtration is a nested sequence of finite-dimensional spaces; it is exhaustive when their union is A A A . Invertibility of its Gram matrices at every level (quasi-definiteness) implies this nondegeneracy. Indeed, a nonzero element belongs to one finite level. The invertible Gram matrix at that level supplies an opposite element with nonzero pairing. It therefore cannot belong to either radical of the full pairing. For pure division, a = 1 a=1 a = 1 gives Ann ( a ) = 0 \operatorname{Ann}(a)=0 Ann ( a ) = 0 , so a solution exists precisely when b b b kills no nonzero class. Nonregular factors are allowed if Ann ( b ) ⊆ Ann ( a ) \operatorname{Ann}(b)\subseteq\operatorname{Ann}(a) Ann ( b ) ⊆ Ann ( a ) ; equality is necessary if both pairings are nondegenerate. Subsubsection 6.2.2 uses fixed moments without inverting the factors.
Assume each of the four determinants is regular: multiplication by it kills no nonzero element of A A A . Adjugate formulas give inverses over K K K . Put S L : = L 1 − 1 L 0 S_L:=L_1^{-1}L_0 S L := L 1 − 1 L 0 and S R : = R 0 R 1 − 1 S_R:=R_0R_1^{-1} S R := R 0 R 1 − 1 ; these send polynomial rows f f f and columns g g g to rational rows f S L fS_L f S L and columns S R g S_Rg S R g .
Set d : = ι L ( det L 1 ) ι R ( det R 1 ) d:=\iota_L(\det L_1)\iota_R(\det R_1) d := ι L ( det L 1 ) ι R ( det R 1 ) . The notation A [ 1 / d ] \mathcal{A}[1/d] A [ 1/ d ] means the algebra of fractions h / d m h/d^m h / d m , with h ∈ A h\in\mathcal{A} h ∈ A and m ∈ { 0 , 1 , … } m\in\{0,1,\ldots\} m ∈ { 0 , 1 , … } ; m = 0 m=0 m = 0 includes the original elements.2 For A = C [ x ] \mathcal{A}=\mathbb{C}[x] A = C [ x ] and d = x − 1 d=x-1 d = x − 1 , this includes ( x + 2 ) / ( x − 1 ) 2 (x+2)/(x-1)^2 ( x + 2 ) / ( x − 1 ) 2 . Thus Mat q × p ( A [ 1 / d ] ) \operatorname{Mat}_{q\times p}(\mathcal{A}[1/d]) Mat q × p ( A [ 1/ d ]) consists of matrices with these fractional entries.
The functional μ \mu μ acts only on polynomial tests. To evaluate rational tests, extend a complex basis of V V V to one of the larger matrix space above, keeping the values prescribed by μ \mu μ and choosing values on the added basis elements. Linearity then determines an extension μ ~ \tilde{\mu} μ ~ on the whole space. Denote its pairing by β ~ μ \widetilde{\beta}_{\mu} β μ , and write β ^ : = β μ ~ \widehat{\beta}:=\beta_{\widetilde{\mu}} β := β μ for the perturbed pairing. If μ \mu μ is represented by a measure, integration provides such an extension whenever every rational test is integrable. The polynomial moments alone do not prescribe the new values: for example, a value on H / d H/d H / d must be supplied in addition unless that fraction is already a polynomial test. A finite construction below only uses finitely many such additional values.
The choice of extension fixes a particular solution, rather than the set of solutions. If μ ~ 1 , μ ~ 2 \widetilde{\mu}_1,\widetilde{\mu}_2 μ 1 , μ 2 are two extensions, their induced difference is
ν 12 ( f , g ) : = ( β ~ μ , 1 − β ~ μ , 2 ) ( f S L , S R g ) . \nu_{12}(f,g):=(\widetilde{\beta}_{\mu,1}-\widetilde{\beta}_{\mu,2})(fS_L,S_Rg). ν 12 ( f , g ) := ( β μ , 1 − β μ , 2 ) ( f S L , S R g ) .
On f L 1 , R 1 g fL_1,R_1g f L 1 , R 1 g it becomes the difference of the two extensions on the polynomial arguments f L 0 , R 0 g fL_0,R_0g f L 0 , R 0 g , hence is zero. Thus changing the extension only changes which homogeneous mass is called zero.
A homogeneous mass assigns arbitrary values to the classes of Q : = V / D 1 V Q:=V/D_1V Q := V / D 1 V . As a functional ν \nu ν on V V V , it satisfies ν [ D 1 ( H ) ] = 0 \nu[D_1(H)]=0 ν [ D 1 ( H )] = 0 ; we also write ν ( f , g ) \nu(f,g) ν ( f , g ) as a pairing.3 For a balanced scalar pairing on C [ t ] \mathbb{C}[t] C [ t ] , take q = p = 1 q=p=1 q = p = 1 and L 1 = R 1 = t L_1=R_1=t L 1 = R 1 = t . Then D 1 ( h ) = t 2 h D_1(h)=t^2h D 1 ( h ) = t 2 h , so the remainders are a + b t a+bt a + b t . A mass is therefore a linear combination of evaluation and first derivative at zero.
When dim A = 1 \dim A=1 dim A = 1 , put d L : = dim C ( A 1 × q / ( A 1 × q L 1 ) ) d_L:=\dim_{\mathbb{C}}(A^{1\times q}/(A^{1\times q}L_1)) d L := dim C ( A 1 × q / ( A 1 × q L 1 )) and d R : = dim C ( A p × 1 / ( R 1 A p × 1 ) ) d_R:=\dim_{\mathbb{C}}(A^{p\times1}/(R_1A^{p\times1})) d R := dim C ( A p × 1 / ( R 1 A p × 1 )) . Curves may be reducible or nonreduced.
The next theorem separates the construction into a particular solution and the complete freedom left by division. Its first part gives the particular solution by rational substitution. The second identifies the masses without choosing coordinates. The last two parts describe their size on curves: balanced masses have finitely many scalar parameters, whereas a bivariate mass still has finite rank but can depend on arbitrary functionals in the opposite argument. Here finite rank means a representation by finitely many products of one row functional and one column functional, as in (2.5).
Theorem 2.2 (All rational bilateral perturbations). (i) A functional μ ^ \widehat{\mu} μ solves (2.1) if and only if, for some ν \nu ν ,
β ^ ( f , g ) = β ~ μ ( f S L , S R g ) + ν ( f , g ) , L 1 ν R 1 = 0. (4) \widehat{\beta}(f,g)=\widetilde{\beta}_{\mu}(fS_L,S_Rg)+\nu(f,g),\qquad L_1\nu R_1=0. \tag*{(4)} β ( f , g ) = β μ ( f S L , S R g ) + ν ( f , g ) , L 1 ν R 1 = 0. (4) (ii) The full mass space is Q ∗ Q^* Q ∗ ; Q Q Q is finitely generated over A \mathcal{A} A and annihilated by d d d .
(iii) For balanced pairings with dim A = 1 \dim A=1 dim A = 1 , Q Q Q is finite dimensional over C \mathbb{C} C and satisfies dim Q = p d L + q d R \dim Q=p\,d_L+q\,d_R dim Q = p d L + q d R .
(iv) For balanced or bivariate pairings with dim A = 1 \dim A=1 dim A = 1 , each mass has a representation with scalar linear functionals ℓ a : A 1 × q → C \ell_a:A^{1\times q}\to\mathbb{C} ℓ a : A 1 × q → C and r a : A p × 1 → C r_a:A^{p\times1}\to\mathbb{C} r a : A p × 1 → C :
ν ( f , g ) = ∑ a = 1 h ℓ a ( f ) r a ( g ) , h ≤ d L + d R . (5) \nu(f,g)=\sum_{a=1}^{h}\ell_a(f)r_a(g),\qquad h\le d_L+d_R. \tag*{(5)} ν ( f , g ) = a = 1 ∑ h ℓ a ( f ) r a ( g ) , h ≤ d L + d R . (5)
The bivariate mass space can be infinite dimensional even on a curve.
Proof. (i) Apply the rational term in (2.4) to f L 1 , R 1 g fL_1,R_1g f L 1 , R 1 g . Since L 1 S L = L 0 L_1S_L=L_0 L 1 S L = L 0 and S R R 1 = R 0 S_RR_1=R_0 S R R 1 = R 0 , its value is β ~ μ ( f L 0 , R 0 g ) = β μ ( f L 0 , R 0 g ) \widetilde{\beta}_{\mu}(fL_0,R_0g)=\beta_{\mu}(fL_0,R_0g) β μ ( f L 0 , R 0 g ) = β μ ( f L 0 , R 0 g ) : both arguments are now polynomial, where the extension agrees with the original pairing. Thus it solves (2.2). Subtracting it from any solution leaves a homogeneous mass, and adding any homogeneous mass gives another solution.
(ii) Since ν [ D 1 ( H ) ] = 0 \nu[D_1(H)]=0 ν [ D 1 ( H )] = 0 , tests differing by an element of D 1 V D_1V D 1 V have the same value. Thus ν \nu ν defines a functional on Q Q Q . Conversely, a functional on Q Q Q gives a mass by evaluating quotient classes. Hence the mass space is Q ∗ Q^* Q ∗ . Since V V V has q p qp q p matrix-unit generators over A \mathcal{A} A , its quotient Q Q Q is finitely generated. Applying D 1 D_1 D 1 to ι L ( adj L 1 ) T ι R ( adj R 1 ) T \iota_L(\operatorname{adj}L_1)^{\mathsf{T}}\iota_R(\operatorname{adj}R_1)^{\mathsf{T}} ι L ( adj L 1 ) T ι R ( adj R 1 ) T gives d H dH d H by the adjugate identities. Hence d V ⊂ D 1 V dV\subset D_1V d V ⊂ D 1 V , so multiplication by d d d kills every quotient class.
This also locates the freedom geometrically: away from the zeros of the denominator determinants, d d d is invertible and the quotient is zero. All homogeneous mass data therefore belong to the part where at least one denominator ceases to be invertible.
(iii) A regular element belongs to no minimal prime of A A A : the minimal primes of a Noetherian ring are associated primes, and an element in an associated prime is a zero divisor. Thus a regular determinant does not vanish identically on any irreducible curve component. Its zero set is zero dimensional. The quotient of the finitely generated complex algebra A A A by that determinant is consequently Artinian and finite dimensional over C \mathbb{C} C . Here Artinian means that descending chains of ideals terminate; for an affine complex algebra of dimension zero it is equivalent to having only finitely many independent complex coefficients. Isolated zero-dimensional components cause no exception: on a finite-dimensional component, injective multiplication is surjective, so a regular scalar is a unit there and its quotient is zero. The argument also retains multiplicities and embedded point contributions. It is the precise reason that only finitely many remainder coefficients survive on a curve. Adjugates show that the row and column remainder spaces are annihilated by their respective determinants det L 1 \det L_1 det L 1 and det R 1 \det R_1 det R 1 . These finitely generated remainder spaces therefore have finite dimensions d L , d R d_L,d_R d L , d R . For balanced pairings, A = A \mathcal{A}=A A = A . Use the chain D 1 V = L 1 T V R 1 T ⊂ L 1 T V ⊂ V D_1V=L_1^TVR_1^T\subset L_1^TV\subset V D 1 V = L 1 T V R 1 T ⊂ L 1 T V ⊂ V . The quotient V / ( L 1 T V ) V/(L_1^TV) V / ( L 1 T V ) has p p p columns, each contributing d L d_L d L remainders: transposition gives the row quotient defining d L d_L d L . Multiplying L 1 T H = 0 L_1^TH=0 L 1 T H = 0 by adj ( L 1 ) T \operatorname{adj}(L_1)^T adj ( L 1 ) T gives ( det L 1 ) H = 0 (\det L_1)H=0 ( det L 1 ) H = 0 ; regularity forces H = 0 H=0 H = 0 . This injective multiplication identifies ( L 1 T V ) / D 1 V (L_1^TV)/D_1V ( L 1 T V ) / D 1 V with V / ( V R 1 T ) V/(VR_1^T) V / ( V R 1 T ) , which has q q q row remainders of dimension d R d_R d R . The two successive quotients fit into
0 ⟶ V / ( V R 1 T ) → [ H ] ↦ [ L 1 T H ] Q ⟶ V / ( L 1 T V ) ⟶ 0. 0\longrightarrow V/(VR_1^T)\xrightarrow{[H]\mapsto[L_1^TH]} \mathcal{Q}\longrightarrow V/(L_1^TV)\longrightarrow0. 0 ⟶ V / ( V R 1 T ) [ H ] ↦ [ L 1 T H ] Q ⟶ V / ( L 1 T V ) ⟶ 0.
The first map is injective because L 1 T L_1^T L 1 T is injective. Its image is exactly the kernel of the last quotient map. Choosing bases of that kernel and of the final quotient gives a basis of Q \mathcal{Q} Q , and therefore dim C Q = q d R + p d L \dim_{\mathbb{C}}\mathcal{Q}=qd_R+pd_L dim C Q = q d R + p d L .
(iv) For either convention on a curve, put F : = A 1 × q \mathcal{F}:=A^{1\times q} F := A 1 × q , G : = A p × 1 \mathcal{G}:=A^{p\times1} G := A p × 1 , F 0 : = F L 1 \mathcal{F}_0:=\mathcal{F}L_1 F 0 := F L 1 and G 0 : = R 1 G \mathcal{G}_0:=R_1\mathcal{G} G 0 := R 1 G . Choose complements of dimensions d L , d R d_L,d_R d L , d R to F 0 , G 0 \mathcal{F}_0,\mathcal{G}_0 F 0 , G 0 , with bases ϕ a , ψ b \phi_a,\psi_b ϕ a , ψ b and remainder coordinates ℓ a , η b \ell_a,\eta_b ℓ a , η b . Write f = f 0 + r L f=f_0+r_L f = f 0 + r L , g = g 0 + r R g=g_0+r_R g = g 0 + r R , where f 0 ∈ F 0 f_0\in\mathcal{F}_0 f 0 ∈ F 0 , g 0 ∈ G 0 g_0\in\mathcal{G}_0 g 0 ∈ G 0 , r L : = ∑ a ℓ a ( f ) ϕ a r_L:=\sum_a\ell_a(f)\phi_a r L := ∑ a ℓ a ( f ) ϕ a and r R : = ∑ b η b ( g ) ψ b r_R:=\sum_b\eta_b(g)\psi_b r R := ∑ b η b ( g ) ψ b ; put π F f : = r L \pi_{\mathcal{F}}f:=r_L π F f := r L . Homogeneity gives ν ( f 0 , g 0 ) = 0 \nu(f_0,g_0)=0 ν ( f 0 , g 0 ) = 0 , hence
ν ( f , g ) = ν ( r L , g ) + ν ( f 0 , r R ) . \nu(f,g)=\nu(r_L,g)+\nu(f_0,r_R). ν ( f , g ) = ν ( r L , g ) + ν ( f 0 , r R ) .
The first term already includes the pairing of both remainders. Define T a ( g ) : = ν ( ϕ a , g ) T_a(g):=\nu(\phi_a,g) T a ( g ) := ν ( ϕ a , g ) and S b ( f 0 ) : = ν ( f 0 , ψ b ) S_b(f_0):=\nu(f_0,\psi_b) S b ( f 0 ) := ν ( f 0 , ψ b ) . Expanding the two remainders gives
ν ( f , g ) = ∑ a = 1 d L ℓ a ( f ) T a ( g ) + ∑ b = 1 d R S b ( f − π F f ) η b ( g ) . (6) \nu(f,g)=\sum_{a=1}^{d_L}\ell_a(f)T_a(g)+\sum_{b=1}^{d_R}S_b(f-\pi_{\mathcal{F}}f)\eta_b(g). \tag*{(6)} ν ( f , g ) = a = 1 ∑ d L ℓ a ( f ) T a ( g ) + b = 1 ∑ d R S b ( f − π F f ) η b ( g ) . (6)
The data are recovered uniquely by evaluating the mass on ( ϕ a , g ) (\phi_a,g) ( ϕ a , g ) and ( f 0 , ψ b ) (f_0,\psi_b) ( f 0 , ψ b ) , respectively. Arbitrary T a ∈ G ∗ T_a\in\mathcal{G}^* T a ∈ G ∗ and S b ∈ F 0 ∗ S_b\in\mathcal{F}_0^* S b ∈ F 0 ∗ give all bivariate masses: when f ∈ F 0 f\in\mathcal{F}_0 f ∈ F 0 and g ∈ G 0 g\in\mathcal{G}_0 g ∈ G 0 , every term vanishes. For balanced masses the data must also satisfy ν ( a f , g ) = ν ( f , a g ) \nu(af,g)=\nu(f,ag) ν ( a f , g ) = ν ( f , a g ) . The same d L + d R d_L+d_R d L + d R products prove (2.5) for either convention. For A = C [ z ] A=\mathbb{C}[z] A = C [ z ] and d L + d R > 0 d_L+d_R>0 d L + d R > 0 , pairing one nonzero remainder coordinate with arbitrary opposite functionals gives an infinite-dimensional bivariate mass space. □
For balanced masses on a curve, at a smooth point with parameter t t t , use local Smith normal forms.4 Write L 1 = U L diag ( t α j ) V L L_1=U_L\operatorname{diag}(t^{\alpha_j})V_L L 1 = U L diag ( t α j ) V L and R 1 = U R diag ( t β j ) V R R_1=U_R\operatorname{diag}(t^{\beta_j})V_R R 1 = U R diag ( t β j ) V R , where U L , V L , U R , V R U_L,V_L,U_R,V_R U L , V L , U R , V R and their inverses have no poles at this point. These changes of basis reduce the diagonal action on η j : = V L U R \eta_j:=V_LU_R η j := V L U R to multiplication of η i , j \eta_{i,j} η i , j by t α i + β j t^{\alpha_i+\beta_j} t α i + β j . Thus that entry depends only on Taylor coefficients c k ( a ) : = [ t k ] a c_k(a):=[t^k]a c k ( a ) := [ t k ] a with k ∈ { 0 , … , α i + β j − 1 } k\in\{0,\ldots,\alpha_i+\beta_j-1\} k ∈ { 0 , … , α i + β j − 1 } , or equivalently derivatives of these orders. At a singular point the remainder module Q \mathcal{Q} Q supplies the tests; the following cusp shows why a smooth parameter can miss one.
Let e i e_i e i be the i i i th coordinate column and E i , j E_{i,j} E i , j the matrix unit with entry 1 1 1 at ( i , j ) (i,j) ( i , j ) , of the required size. For A = C [ t 2 , t 3 ] A=\mathbb{C}[t^2,t^3] A = C [ t 2 , t 3 ] , write x = t 2 x=t^2 x = t 2 and take L 1 = diag ( x , 1 ) L_1=\operatorname{diag}(x,1) L 1 = diag ( x , 1 ) , R 1 = diag ( x , 1 , 1 ) R_1=\operatorname{diag}(x,1,1) R 1 = diag ( x , 1 , 1 ) and ν = E 1 , 1 c 5 \nu=E_{1,1}c_5 ν = E 1 , 1 c 5 . The algebra contains 1 , t 2 , t 3 , t 4 , t 5 , … 1,t^2,t^3,t^4,t^5,\ldots 1 , t 2 , t 3 , t 4 , t 5 , … , but not t t t . Consequently c 5 ( x 2 h ) = c 5 ( h ) c_5(x^2h)=c_5(h) c 5 ( x 2 h ) = c 5 ( h ) : A A A a term of degree five would require a degree-one term in h h h . On the other hand, c 5 ( x h ) = c 3 ( h ) c_5(xh)=c_3(h) c 5 ( x h ) = c 3 ( h ) can be nonzero. Thus x 2 c 5 = 0 x^2c_5=0 x 2 c 5 = 0 but x c 5 = c 3 ≠ 0 xc_5=c_3\ne0 x c 5 = c 3 = 0 . The quotient A / ( x 2 ) A/(x^2) A / ( x 2 ) has basis 1 , t 2 , t 3 , t 5 1,t^2,t^3,t^5 1 , t 2 , t 3 , t 5 . Pairing two basis elements by extracting the coefficient of t 5 t^{5} t 5 gives the matrix
( 0 0 0 1 0 0 1 0 0 1 0 0 1 0 0 0 ) \begin{pmatrix} 0 & 0 & 0 & 1 \\ 0 & 0 & 1 & 0 \\ 0 & 1 & 0 & 0 \\ 1 & 0 & 0 & 0 \end{pmatrix} 0 0 0 1 0 0 1 0 0 1 0 0 1 0 0 0
It is invertible, so this mass attains rank d L + d R = 4 d_{L}+d_{R}=4 d L + d R = 4 . Normalization adjoins the missing parameter t t t , giving C [ t ] \mathbb{C}[t] C [ t ] . There, masses annihilated by t 4 t^{4} t 4 use only derivatives of orders zero through three. Since t 5 = t 4 t t^{5}=t^{4}t t 5 = t 4 t , they vanish at t 5 t^{5} t 5 , whereas c 5 ( t 5 ) = 1 c_{5}(t^{5})=1 c 5 ( t 5 ) = 1 ; none restricts to this cusp mass.
For the curve defined by the two circle equations, A = C [ x , a , b ] / ( x 2 + a 2 − 1 , x 2 + b 2 − 1 ) A=\mathbb{C}[x,a,b]/(x^{2}+a^{2}-1,x^{2}+b^{2}-1) A = C [ x , a , b ] / ( x 2 + a 2 − 1 , x 2 + b 2 − 1 ) is free over C [ x ] \mathbb{C}[x] C [ x ] with basis 1 , a , b , a b 1,a,b,ab 1 , a , b , ab . Hence 1 − x 1-x 1 − x is regular and A / ( 1 − x ) = C [ a , b ] / ( a 2 , b 2 ) A/(1-x)=\mathbb{C}[a,b]/(a^{2},b^{2}) A / ( 1 − x ) = C [ a , b ] / ( a 2 , b 2 ) retains four classes, including the one detected by ∂ a ∂ b \partial_{a}\partial_{b} ∂ a ∂ b .
Indeed, after setting x = 1 x=1 x = 1 , every class has a unique representative c 0 , 0 + c 1 , 0 a + c 0 , 1 b + c 1 , 1 a b c_{0,0}+c_{1,0}a+c_{0,1}b+c_{1,1}ab c 0 , 0 + c 1 , 0 a + c 0 , 1 b + c 1 , 1 ab . Evaluation and the three derivatives ∂ a \partial_{a} ∂ a , ∂ b \partial_{b} ∂ b , ∂ a ∂ b \partial_{a}\partial_{b} ∂ a ∂ b at a = b = 0 a=b=0 a = b = 0 recover these four coefficients. Point evaluation alone would retain only c 0 , 0 c_{0,0} c 0 , 0 and would miss three allowed mass data.
Remark 2.3 (Cancellations preserve a ratio, not an equation ). The rational term depends on S L , S R S_{L},S_{R} S L , S R and the extension; Q ∗ Q^{*} Q ∗ depends on the denominator factors. Cancelling common factors in (1) loses their mass freedom unless it is retained separately. For example, take q = p = 1 q=p=1 q = p = 1 , A = C [ t ] A=\mathbb{C}[t] A = C [ t ] and L 0 = L 1 = R 0 = R 1 = t L_{0}=L_{1}=R_{0}=R_{1}=t L 0 = L 1 = R 0 = R 1 = t . Both rational ratios equal one, but the equation is t 2 μ ^ = t 2 μ t^{2}\widehat{\mu}=t^{2}\mu t 2 μ = t 2 μ . Its solutions are μ ^ = μ + c 0 c 0 + c 1 c 1 \widehat{\mu}=\mu+c_{0}c_{0}+c_{1}c_{1} μ = μ + c 0 c 0 + c 1 c 1 . Cancelling the factors and replacing the equation by μ ^ = μ \widehat{\mu}=\mu μ = μ would remove both free constants.
Finite normality and reproducing kernels
We record the finite Gram construction used throughout the paper. For the comparison in this subsubsection, both target spaces will have the same dimension as the original spaces. The original row space E E E and column space H H H have the same dimension N N N . Choose bases e 1 , … , e N e_{1},\ldots,e_{N} e 1 , … , e N and h 1 , … , h N h_{1},\ldots,h_{N} h 1 , … , h N , and set
G 0 : = [ β μ ( e i , h j ) ] i , j ∈ { 1 , … , N } . G_{0}:=[\beta_{\mu}(e_{i},h_{j})]_{i,j\in\{1,\ldots,N\}}. G 0 := [ β μ ( e i , h j ) ] i , j ∈ { 1 , … , N } .
The pairing is normal on these spaces when G 0 G_{0} G 0 is invertible. This is a finite matrix condition; no inverse of a semi-infinite moment matrix is assumed. Set e ( Q ) : = [ e i ( Q ) ] i ∈ { 1 , … , N } e(Q):=[e_{i}(Q)]_{i\in\{1,\ldots,N\}} e ( Q ) := [ e i ( Q ) ] i ∈ { 1 , … , N } (rows stacked) and h ( P ) : = [ h 1 ( P ) ⋯ h N ( P ) ] h(P):=[h_{1}(P)\ \cdots\ h_{N}(P)] h ( P ) := [ h 1 ( P ) ⋯ h N ( P )] . The reproducing kernel is
K E ( P , Q ) : = h ( P ) G 0 − 1 e ( Q ) . (7) K_{E}(P,Q):=h(P)G_{0}^{-1}e(Q). \tag*{(7)} K E ( P , Q ) := h ( P ) G 0 − 1 e ( Q ) . (7)
To check reproduction, write g = ∑ j h j c j g=\sum_{j}h_{j}c_{j} g = ∑ j h j c j . The column of values [ β μ ( e i , g ) ] i [\beta_{\mu}(e_{i},g)]_{i} [ β μ ( e i , g ) ] i is G 0 c G_{0}c G 0 c ; multiplying by h ( P ) G 0 − 1 h(P)G_{0}^{-1} h ( P ) G 0 − 1 returns h ( P ) c = g ( P ) h(P)c=g(P) h ( P ) c = g ( P ) . For a row f = a e f=ae f = a e , the row of its pairings with the column basis is a G 0 aG_{0} a G 0 . Multiplication by G 0 − 1 e ( Q ) G_{0}^{-1}e(Q) G 0 − 1 e ( Q ) gives a e ( Q ) = f ( Q ) ae(Q)=f(Q) a e ( Q ) = f ( Q ) , which is reproduction on E E E . For independence of bases, replace e e e by C e Ce C e and h h h by h D hD h D , with C , D C,D C , D invertible. The Gram becomes C G 0 D CG_{0}D C G 0 D , and the new kernel is
h ( P ) D ( C G 0 D ) − 1 C e ( Q ) = h ( P ) G 0 − 1 e ( Q ) . h(P)D(CG_{0}D)^{-1}Ce(Q)=h(P)G_{0}^{-1}e(Q). h ( P ) D ( C G 0 D ) − 1 C e ( Q ) = h ( P ) G 0 − 1 e ( Q ) .
Thus the kernel depends on the spaces and their pairing, not on the chosen bases.
Choose target spaces X ⊂ A 1 × q X\subset A^{1\times q} X ⊂ A 1 × q and Y ⊂ A p × 1 Y\subset A^{p\times1} Y ⊂ A p × 1 of dimension N N N , normal for β ^ \widehat{\beta} β , and let K ^ \widehat{K} K be their kernel. Multiplication by the rational factors gives the following kernel and its difference from K E K_{E} K E :
F ( P , Q ) : = S R ( P ) K ^ ( P , Q ) S L ( Q ) , Δ : = F − K E . \mathcal{F}(P,Q):=S_{R}(P)\widehat{K}(P,Q)S_{L}(Q),\qquad\Delta:=\mathcal{F}-K_{E}. F ( P , Q ) := S R ( P ) K ( P , Q ) S L ( Q ) , Δ := F − K E .
They belong to the tensor product of W R : = H + S R Y W_{R}:=H+S_{R}Y W R := H + S R Y and W L : = E + X S L W_{L}:=E+XS_{L} W L := E + X S L , over C \mathbb{C} C . Here P , Q P,Q P , Q label independent arguments, even for balanced pairings. The separation rank of Δ \Delta Δ is the smallest r r r for which Δ = ∑ j = 1 r u j ( P ) v j ( Q ) \Delta=\sum_{j=1}^{r}u_{j}(P)v_{j}(Q) Δ = ∑ j = 1 r u j ( P ) v j ( Q ) with u j ∈ W R , v j ∈ W L u_{j}\in W_{R},v_{j}\in W_{L} u j ∈ W R , v j ∈ W L . In bases of these two spaces it is the ordinary rank of the coefficient matrix: each separated product gives an outer product of coordinate vectors, and a rank factorization of that matrix gives the converse decomposition. We also call it tensor rank. On a nonreduced support this is an identity of polynomial classes; point values alone need not distinguish the classes.
We shall also use finite spaces of sections of vector bundles. Their coordinates are polynomial or rational vectors in each local basis, and the same Gram-matrix construction defines their kernels. A balanced mass acts on the tensor product over A A A of the two section modules. All finite-dimensional arguments above use complex bases of section spaces and are therefore unchanged by this extension.
When a bilateral mass splits into unilateral terms
Can a bilateral mass split into two masses, each killed by one denominator? For arbitrary bivariate masses, formula (6) already supplies such a decomposition. Its first sum is killed by L 1 L_1 L 1 , because all its left remainder coordinates vanish on F 0 \mathcal{F}_0 F 0 . Its second sum is killed by R 1 R_1 R 1 , because the right remainder coordinates vanish on G 0 \mathcal{G}_0 G 0 . Requiring each summand to be balanced introduces an additional condition: these two sums need not separately allow scalar factors to move between arguments. Put P L : = A 1 × q \mathcal{P}_L := A^{1\times q} P L := A 1 × q , P R : = A p × 1 \mathcal{P}_R := A^{p\times1} P R := A p × 1 , D L : = P L L 1 \mathcal{D}_L := \mathcal{P}_L L_1 D L := P L L 1 , D R : = R 1 P R \mathcal{D}_R := R_1\mathcal{P}_R D R := R 1 P R , and C L : = P L / D L \mathcal{C}_L := \mathcal{P}_L/\mathcal{D}_L C L := P L / D L , C R : = P R / D R \mathcal{C}_R := \mathcal{P}_R/\mathcal{D}_R C R := P R / D R . Tensor products in this subsection are over A A A , and stars denote full complex algebraic duals.5 Identify V V V with P L ⊗ P R \mathcal{P}_L \otimes\mathcal{P}_R P L ⊗ P R by f ⊗ g ↦ f t g t f \otimes g \mapsto f^{\mathsf{t}}g^{\mathsf{t}} f ⊗ g ↦ f t g t . Let M \mathcal{M} M be the space of balanced homogeneous bilateral masses and let M L , M R \mathcal{M}_L,\mathcal{M}_R M L , M R be its subspaces satisfying L 1 ν = 0 L_1\nu= 0 L 1 ν = 0 and ν R 1 = 0 \nu R_1 = 0 ν R 1 = 0 , respectively. The concrete obstruction is the quotient
( D L ⊗ P R ) ∩ ( P L ⊗ D R ) D L ⊗ D R \frac{(\mathcal{D}_L \otimes\mathcal{P}_R) \cap(\mathcal{P}_L \otimes\mathcal{D}_R)}{\mathcal{D}_L \otimes\mathcal{D}_R} D L ⊗ D R ( D L ⊗ P R ) ∩ ( P L ⊗ D R )
The intersection consists of tests divisible on each side separately; the denominator discards those divisible on both sides at once. The remaining classes form the obstruction, denoted here by Tor 1 A ( C L , C R ) \operatorname{Tor}^{A}_{1}(\mathcal{C}_L,\mathcal{C}_R) Tor 1 A ( C L , C R ) . 6
For scalar denominators a , b ≠ 0 a,b \ne0 a , b = 0 , the obstruction is ( ( a ) ∩ ( b ) ) / ( a b ) ((a) \cap(b))/(ab) (( a ) ∩ ( b )) / ( ab ) : separate divisibility need not imply divisibility by the product. For a = b = t a=b=t a = b = t in C [ t ] \mathbb{C}[t] C [ t ] , the extra mass direction is represented by ν [ f ] = f ′ ( 0 ) \nu[f]=f'(0) ν [ f ] = f ′ ( 0 ) . It annihilates t 2 A t^2A t 2 A but not t A tA t A , whereas unilateral masses are multiples of evaluation at zero. The determinant divisors are the codimension-one zero sets of det L 1 \det L_1 det L 1 and det R 1 \det R_1 det R 1 , with their vanishing orders. The local formula uses normality of the support.7
The exact sequence below records two separate issues. Its leftmost term measures bilateral masses that cannot be written as sums of unilateral masses. Its rightmost term measures the ambiguity of a decomposition when one exists. Thus failure of existence and nonuniqueness of the two summands have different algebraic sources.
Theorem 2.4 (The obstruction to unilateral decomposition). For balanced masses, with regular denominator determinants, there is a canonical exact sequence
0 → Tor 1 A ( C L , C R ) → Q → ( C L ⊗ P R ) ⊕ ( P L ⊗ C R ) → C L ⊗ C R → 0 (8) 0 \to\operatorname{Tor}^{A}_{1}(\mathcal{C}_L,\mathcal{C}_R) \to\mathcal{Q} \to(\mathcal{C}_L \otimes\mathcal{P}_R) \oplus(\mathcal{P}_L \otimes\mathcal{C}_R) \to\mathcal{C}_L \otimes\mathcal{C}_R \to0 \tag*{(8)} 0 → Tor 1 A ( C L , C R ) → Q → ( C L ⊗ P R ) ⊕ ( P L ⊗ C R ) → C L ⊗ C R → 0 (8) *The middle maps take [ H ] [H] [ H ] to ( [ H ] , [ H ] ) ([H],[H]) ([ H ] , [ H ]) and ( [ H 1 ] , [ H 2 ] ) ([H_1],[H_2]) ([ H 1 ] , [ H 2 ]) to [ H 1 − H 2 ] [H_1-H_2] [ H 1 − H 2 ] in the indicated quotients. Consequently M / ( M L + M R ) = Tor 1 A ( C L , C R ) \mathcal{M}/(\mathcal{M}_L+\mathcal{M}_R)=\operatorname{Tor}^{A}_{1}(\mathcal{C}_L,\mathcal{C}_R) M / ( M L + M R ) = Tor 1 A ( C L , C R ) , and M L ∩ M R = ( C L ⊗ C R ) \mathcal{M}_L\cap\mathcal{M}_R=(\mathcal{C}_L\otimes\mathcal{C}_R) M L ∩ M R = ( C L ⊗ C R ) . If Z ∘ Z^\circ Z ∘ is smooth and the two determinant divisors have no common irreducible component, every balanced bilateral mass is a sum of balanced unilateral masses. For the following local formula, assume that Z ∘ Z^\circ Z ∘ is normal at the generic point η \eta η of the common component. Its local ring B : = O Z ∘ , η B:=\mathcal{O}_{Z^\circ,\eta} B := O Z ∘ , η inverts functions not identically zero on that component. It is a discrete valuation ring: every nonzero element is a unit times a power of a local parameter t t t . With denominator Smith orders α i , β j \alpha_i,\beta_j α i , β j , Tor 1 A ( C L , C R ) ⊗ A B ≅ ⨁ i , j B / ( t min ( α i , β j ) ) \operatorname{Tor}^{A}_{1}(\mathcal{C}_L,\mathcal{C}_R)\otimes_A B \cong\bigoplus_{i,j} B/(t^{\min(\alpha_i,\beta_j)}) Tor 1 A ( C L , C R ) ⊗ A B ≅ ⨁ i , j B / ( t m i n ( α i , β j ) ) .
Proof. Inside P L ⊗ P R \mathcal{P}_L\otimes\mathcal{P}_R P L ⊗ P R write I : = D L ⊗ P R I:=\mathcal{D}_L\otimes\mathcal{P}_R I := D L ⊗ P R , J : = P L ⊗ D R J:=\mathcal{P}_L\otimes\mathcal{D}_R J := P L ⊗ D R , K 0 : = D L ⊗ D R K_0:=\mathcal{D}_L\otimes\mathcal{D}_R K 0 := D L ⊗ D R . All these embeddings are injective, since P L , D L , P R , D R \mathcal{P}_L,\mathcal{D}_L,\mathcal{P}_R,\mathcal{D}_R P L , D L , P R , D R are free. A class in Q = V / K 0 \mathcal{Q}=V/K_0 Q = V / K 0 maps to its two classes in V / I V/I V / I and V / J V/J V / J . Both images vanish exactly when its representative lies in I ∩ J I\cap J I ∩ J , giving kernel ( I ∩ J ) / K 0 (I\cap J)/K_0 ( I ∩ J ) / K 0 . Next, the difference of the two images is taken modulo I + J I+J I + J . To compute its kernel, suppose H 1 − H 2 = i + j H_1-H_2=i+j H 1 − H 2 = i + j , with i ∈ I i\in I i ∈ I and j ∈ J j\in J j ∈ J . The element H : = H 1 − i = H 2 + j H:=H_1-i=H_2+j H := H 1 − i = H 2 + j has class [ H 1 ] [H_1] [ H 1 ] modulo $I and class [ H 2 ] [H_2] [ H 2 ] modulo $J. Thus every pair in the kernel is represented by one common element of V V V , which is exactly the image of Q \mathcal{Q} Q . The difference map is surjective, with quotient V / ( I + J ) = C L ⊗ C R V/(I+J)=\mathcal{C}_L\otimes\mathcal{C}_R V / ( I + J ) = C L ⊗ C R . This verifies exactness directly at the three quotient terms. For the leftmost term, tensor the free resolution 0 → D L → P L → C L → 0 0\to\mathcal{D}_L\to\mathcal{P}_L\to\mathcal{C}_L\to0 0 → D L → P L → C L → 0 with C R \mathcal{C}_{R} C R . Both modules in the resolution are free: multiplication by the regular matrix L 1 L_{1} L 1 identifies P L \mathcal{P}_{L} P L with its image D L \mathcal{D}_{L} D L . By the definition of the first Tor group, the new kernel is
ker ( D L ⊗ C R ⟶ P L ⊗ C R ) . \ker(\mathcal{D}_{L} \otimes\mathcal{C}_{R} \longrightarrow\mathcal{P}_{L} \otimes\mathcal{C}_{R}). ker ( D L ⊗ C R ⟶ P L ⊗ C R ) .
The source is I / K 0 I/K_{0} I / K 0 and the target is V / J V/J V / J , because tensoring the presentation of C R \mathcal{C}_{R} C R with either free module preserves exactness. This map sends a representative i ∈ I i \in I i ∈ I to its class modulo J J J . Its kernel consists exactly of i ∈ I ∩ J i \in I \cap J i ∈ I ∩ J , modulo K 0 K_{0} K 0 . This identifies the kernel with ( I ∩ J ) / K 0 (I \cap J)/K_{0} ( I ∩ J ) / K 0 and proves the left end of (2.8) [14 ] .
The obstruction can be evaluated without further homological notation. Given a mass ν \nu ν , restrict it to I ∩ J I \cap J I ∩ J :
obs ( ν ) ( [ H ] ) : = ν [ H ] , [ H ] ∈ ( I ∩ J ) / K 0 . \operatorname{obs}(\nu)([H]) := \nu[H], \qquad[H] \in(I \cap J)/K_{0}. obs ( ν ) ([ H ]) := ν [ H ] , [ H ] ∈ ( I ∩ J ) / K 0 .
It is well defined because every mass vanishes on K 0 K_{0} K 0 . A sum of unilateral masses vanishes on I ∩ J I \cap J I ∩ J , so its obstruction is zero. Conversely, if this restriction is zero, ν \nu ν factors through the image of Q Q Q in ( V / I ) ⊕ ( V / J ) (V/I) \oplus(V/J) ( V / I ) ⊕ ( V / J ) . Extend that functional to the whole direct sum by completing a complex basis. Its two restrictions are the required unilateral masses. This proves the existence criterion directly. Any functional on ( I ∩ J ) / K 0 (I \cap J)/K_{0} ( I ∩ J ) / K 0 extends to Q Q Q by the same basis argument, so every obstruction value occurs.
Equivalently, duality over C \mathbb{C} C is exact and reverses arrows. Since masses are functionals on quotient tests, dualizing gives the addition map M L ⊕ M R → M \mathcal{M}_{L} \oplus\mathcal{M}_{R} \to\mathcal{M} M L ⊕ M R → M , whose kernel consists of pairs ( ω , − ω ) (\omega,-\omega) ( ω , − ω ) from ( C L ⊗ C R ) ∗ (\mathcal{C}_{L} \otimes\mathcal{C}_{R})^{*} ( C L ⊗ C R ) ∗ . Its cokernel is M / ( M L + M R ) \mathcal{M}/(\mathcal{M}_{L}+\mathcal{M}_{R}) M / ( M L + M R ) , proving both assertions.
It remains to relate this obstruction to the denominator zero sets. We first prove vanishing when the support is smooth and the two determinant divisors have no common component.
There is a direct divisibility proof. Smoothness makes the local ring regular, and a regular local ring is a unique factorization domain by [14 Theorem 19.19] . Thus a nonzero function factors into prime elements, with multiplicities, up to a unit. These prime factors describe the local codimension-one components of its zero set. Put s : = det L 1 s := \det L_{1} s := det L 1 and r : = det R 1 r := \det R_{1} r := det R 1 in that ring. The hypothesis on components says that s s s , r r r have no common prime factor. Suppose that a column class [ v ] ∈ C R [v] \in\mathcal{C}_{R} [ v ] ∈ C R is killed by s s s . This means s v = R 1 w sv=R_{1}w s v = R 1 w for some column w w w . Multiplying by the adjugate gives
s adj ( R 1 ) v = r w . s\,\operatorname{adj}(R_{1})v = rw. s adj ( R 1 ) v = r w .
If a prime element occurs to exponent m m m in r r r , it does not divide s s s . The displayed equality therefore forces that prime to occur to at least exponent m m m in every entry of adj ( R 1 ) v \operatorname{adj}(R_{1})v adj ( R 1 ) v . Doing this for all prime factors of r r r proves that r r r divides each entry. Thus z : = r − 1 adj ( R 1 ) v z := r^{-1}\operatorname{adj}(R_{1})v z := r − 1 adj ( R 1 ) v is a column over the local ring, and R 1 z = v R_{1}z=v R 1 z = v . Hence [ v ] = 0 [v]=0 [ v ] = 0 . We have proved that multiplication by s s s is injective on C R \mathcal{C}_{R} C R . The adjugate of L 1 L_{1} L 1 now shows that the map defining the Tor kernel is injective. The kernel vanishes at every point and is therefore zero.
At the generic point of a common component, the discrete valuation ring B B B admits Smith bases. They split the two cokernels into cyclic modules B / ( t α i ) B/(t^{\alpha_i}) B / ( t α i ) and B / ( t β j ) B/(t^{\beta_j}) B / ( t β j ) . For a pair of exponents a , b a,b a , b , use the free resolution
0 ⟶ B → t a B ⟶ B / ( t a ) ⟶ 0. 0 \longrightarrow B \xrightarrow{t^{a}} B \longrightarrow B/(t^{a}) \longrightarrow0. 0 ⟶ B t a B ⟶ B / ( t a ) ⟶ 0.
After tensoring with B / ( t b ) B/(t^{b}) B / ( t b ) , the Tor module is the kernel of multiplication by t a t^{a} t a on B / ( t b ) B/(t^{b}) B / ( t b ) . If a ≥ b a \ge b a ≥ b , that multiplication is zero, so its kernel is B / ( t b ) B/(t^{b}) B / ( t b ) . If a < b a < b a < b , a class [ w ] [w] [ w ] belongs to the kernel exactly when t a w ∈ ( t b ) t^{a}w \in(t^{b}) t a w ∈ ( t b ) . A nonzero element of the discrete valuation ring is a unit times a power of t t t ; hence this condition means w ∈ ( t b − a ) w \in(t^{b-a}) w ∈ ( t b − a ) . The kernel is therefore t b − a B / ( t b ) t^{b-a}B/(t^{b}) t b − a B / ( t b ) , and the map
B / ( t a ) ⟶ t b − a B / ( t b ) , [ w ] ⟼ [ t b − a w ] B/(t^{a}) \longrightarrow t^{b-a}B/(t^{b}), \qquad[w] \longmapsto[t^{b-a}w] B / ( t a ) ⟶ t b − a B / ( t b ) , [ w ] ⟼ [ t b − a w ]
is onto and has zero kernel. Both cases give B / ( t min ( a , b ) ) B/(t^{\min(a,b)}) B / ( t m i n ( a , b ) ) ; if either exponent is zero, the module is zero. Localization preserves the free presentations used here, and tensor products and their kernels distribute over these finite direct sums. Applying the calculation to every pair of Smith components yields
Tor 1 A ( C L , C R ) ⊗ A B ≃ ⨁ i ∈ { 1 , … , q } ⨁ j ∈ { 1 , … , p } B / ( t min ( α i , β j ) ) . \operatorname{Tor}_{1}^{A}(\mathcal{C}_{L},\mathcal{C}_{R}) \otimes_{A} B \simeq\bigoplus_{i \in\{1,\ldots,q\}} \bigoplus_{j \in\{1,\ldots,p\}} B/(t^{\min(\alpha_i,\beta_j)}). Tor 1 A ( C L , C R ) ⊗ A B ≃ i ∈ { 1 , … , q } ⨁ j ∈ { 1 , … , p } ⨁ B / ( t m i n ( α i , β j ) ) .
This proves the local formula.
The no-common-component assertion also has a homological interpretation. The vanishing argument above uses divisibility; the following interpretation records how smoothness controls the associated primes. The support of a finitely generated module is the set of prime ideals where its localization is nonzero. For the cokernel of a square matrix with regular determinant, it lies on the determinant zero set: where the determinant is invertible the adjugate gives an inverse. An associated prime is the annihilator of a nonzero module element. The zero divisors on a finitely generated module over a Noetherian ring form the union of its associated primes [14 Chapter 3] . Thus, to prove that a scalar acts injectively, it suffices to show that it belongs to none of those primes. We apply this criterion to det L 1 \det L_1 det L 1 on C R \mathcal{C}_R C R .
Depth counts the successive parameters that act injectively after quotienting by the preceding ones. Over a local ring, projective dimension is the length of a shortest finite free resolution, when such a resolution exists. The relation needed here is the Auslander–Buchsbaum formula for a finite module of finite projective dimension over a regular local ring. Here the injective square presentation gives projective dimension at most one. At a point of the support the cokernel is a nonzero torsion module, so it is not free and its projective dimension is exactly one. The formula therefore subtracts precisely one from the depth of the local ring. These are the only homological inputs in the no-common-component assertion.
Smoothness makes each local ring A p A_{\mathfrak p} A p regular: its maximal ideal is generated by dim A p \dim A_{\mathfrak p} dim A p parameters. This ring property differs from a regular scalar (a nonzerodivisor). For nonzero C R \mathcal{C}_R C R , its injective square presentation and the depth formula give depth ( ( C R ) p ) = dim A p − 1 \operatorname{depth}((\mathcal{C}_R)_{\mathfrak p}) = \dim A_{\mathfrak p} - 1 depth (( C R ) p ) = dim A p − 1 at primes in its support. At an associated prime, depth is zero. The displayed formula forces the local ring to have dimension one, so the prime has height one: it is the generic point of a codimension-one component. These primes belong to the determinant divisor of R 1 R_1 R 1 ; the disjoint-component hypothesis therefore makes det L 1 \det L_1 det L 1 a nonzerodivisor on C R \mathcal{C}_R C R . By the adjugate identity, D L ⊗ C R → P L ⊗ C R D_L \otimes\mathcal{C}_R \to P_L \otimes\mathcal{C}_R D L ⊗ C R → P L ⊗ C R is injective, so its Tor kernel vanishes.
For A = C [ x , y ] A = \mathbb{C}[x,y] A = C [ x , y ] , scalar denominators x , y x,y x , y give Q = A / ( x y ) Q = A/(xy) Q = A / ( x y ) and zero Tor. A polynomial divisible by both x x x and y y y is divisible by x y xy x y , so ( x ) ∩ ( y ) = ( x y ) (x) \cap(y) = (xy) ( x ) ∩ ( y ) = ( x y ) . The final quotient in (2.8) is A / ( x , y ) = C A/(x,y) = \mathbb{C} A / ( x , y ) = C . A mass splits into functionals on the two axes; transferring a multiple of evaluation at their intersection between the two summands is the only ambiguity. For denominators x , x x,x x , x , Q = A / ( x 2 ) Q = A/(x^2) Q = A / ( x 2 ) and the Tor submodule is x A / ( x 2 ) xA/(x^2) x A / ( x 2 ) . The missing unilaterial classes have the form a ↦ λ ( ∂ x a ( 0 , y ) ) a \mapsto\lambda(\partial_x a(0,y)) a ↦ λ ( ∂ x a ( 0 , y )) , with arbitrary λ ∈ C [ y ] ∗ \lambda\in\mathbb{C}[y]^* λ ∈ C [ y ] ∗ . For n ≥ 0 n \ge0 n ≥ 0 , a basis of the classes in Q Q Q represented by polynomials of total degree at most n n n is
1 , y , … , y n , x , x y , … , x y n − 1 1,y,\ldots,y^n,\quad x,xy,\ldots,xy^{n-1} 1 , y , … , y n , x , x y , … , x y n − 1
where the second list is empty for n = 0 n = 0 n = 0 . The submodule x A / ( x 2 ) xA/(x^2) x A / ( x 2 ) has the second list as its basis. Its dimension is n n n , while the whole quotient at this degree has dimension 2 n + 1 2n + 1 2 n + 1 . Thus a proper crossing and a common component have different effects. The full-dual statement concerns algebraic functionals; imposing positivity or a topology requires additional hypotheses.
Finite formulas for polynomials and kernels
We now calculate the transformed polynomials and the reproducing kernel. The inputs are finite lists of original dual polynomials, moments of rational functions, the four transformation matrices, and the prescribed mass values. The rational moments are values of the extension fixed in Theorem 2.2; a finite list of them is supplied together with the polynomial moments. They are evaluated by integration when a measure and integrable rational tests are given, or specified as additional functional data. The unknown polynomial must satisfy three requirements: its components have the allowed bounds, it is orthogonal to the specified lower space, and its leading coefficients have the chosen normalization.
The method writes the rational images of these polynomials in finite bases. Linear equations select exactly those combinations that have allowed polynomial preimages. Original duality then gives an identity block in the pairing matrix; eliminating that block leaves the reduced system used in all subsequent determinant formulas, with the original coefficients eliminated. Its order is calculated from the dimensions of the added spaces and the constraints; comparison with the direct normalized Gram order is made below. We first carry out the elimination for a scalar example, then prove it for arbitrary component sizes and polynomial bounds. The final two subsections count the determinant orders and give direct finite-rank update formulas.
A scalar calculation with a three-by-three determinant
We begin with a scalar problem in one variable. The calculation shows how original polynomials and Cauchy moments become the entries of a small system, and how its solution gives the transformed polynomial. Every step will then be used in the general mixed-type orthogonality problem. Take q = p = 1 q=p=1 q = p = 1 and choose
L 0 = z − a , L 1 = z − λ , R 0 = z − b , R 1 = z − ξ , a ≠ λ , b ≠ ξ . L_0=z-a,\qquad L_1=z-\lambda,\qquad R_0=z-b,\qquad R_1=z-\xi,\qquad a\ne\lambda,\quad b\ne\xi. L 0 = z − a , L 1 = z − λ , R 0 = z − b , R 1 = z − ξ , a = λ , b = ξ .
Let μ \mu μ be a balanced scalar functional, and choose its extension μ ~ \widetilde{\mu} μ to the rational tests with these denominator poles. With zero homogeneous mass, the transformed functional is
μ ^ [ h ] = μ ~ [ ( z − a ) ( z − b ) ( z − λ ) ( z − ξ ) h ( z ) ] . \widehat{\mu}[h]=\widetilde{\mu}\left[\frac{(z-a)(z-b)}{(z-\lambda)(z-\xi)}h(z)\right]. μ [ h ] = μ [ ( z − λ ) ( z − ξ ) ( z − a ) ( z − b ) h ( z ) ] .
For n ≥ 1 n\ge1 n ≥ 1 , the problem is to find the monic polynomial B ^ n \widehat{B}_n B n of degree n n n , with leading coefficient one, such that
μ ^ [ f B ^ n ] = 0 \widehat{\mu}[f\widehat{B}_n]=0 μ [ f B n ] = 0
for every f ∈ C [ z ] < n f\in\mathbb{C}[z]_{<n} f ∈ C [ z ] < n . Thus X = C [ z ] < n X=\mathbb{C}[z]_{<n} X = C [ z ] < n and Y = C [ z ] ≤ n Y=\mathbb{C}[z]_{\le n} Y = C [ z ] ≤ n , with one more unknown coefficient than orthogonality conditions.
Original data and the rational basis. Assume that μ \mu μ has monic orthogonal polynomials B 0 , … , B n B_0,\ldots,B_n B 0 , … , B n with h i : = μ [ B i 2 ] ≠ 0 h_i:=\mu[B_i^2]\ne0 h i := μ [ B i 2 ] = 0 . Define A i : = B i / h i A_i:=B_i/h_i A i := B i / h i ; then μ [ B i A j ] = δ i , j \mu[B_iA_j]=\delta_{i,j} μ [ B i A j ] = δ i , j . Choose E = H = C [ z ] ≤ n E=H=\mathbb{C}[z]_{\le n} E = H = C [ z ] ≤ n as the original spaces, of dimension N = n + 1 N=n+1 N = n + 1 , and arrange their dual bases as
B ( z ) : = [ B 0 ( z ) ⋯ B n ( z ) ] T , A ( z ) : = [ A 0 ( z ) ⋯ A n ( z ) ] . \mathcal{B}(z):=[B_0(z)\ \cdots\ B_n(z)]^{\mathsf T},\qquad\mathcal{A}(z):=[A_0(z)\ \cdots\ A_n(z)]. B ( z ) := [ B 0 ( z ) ⋯ B n ( z ) ] T , A ( z ) := [ A 0 ( z ) ⋯ A n ( z )] .
Append 1 / ( z − λ ) 1/(z-\lambda) 1/ ( z − λ ) on the left and 1 / ( z − ξ ) 1/(z-\xi) 1/ ( z − ξ ) on the right. Pairing the two enlarged lists gives
[ I N C D E 0 ] , C i : = μ ~ [ B i / ( z − ξ ) ] , D i : = μ ~ [ A i / ( z − λ ) ] , E 0 : = μ ~ [ ( ( z − λ ) ( z − ξ ) ) − 1 ] . \begin{bmatrix}
I_N & C\\
D & E_0
\end{bmatrix},
\qquad
C_i:=\widetilde{\mu}[B_i/(z-\xi)],
\qquad
D_i:=\widetilde{\mu}[A_i/(z-\lambda)],
\qquad
E_0:=\widetilde{\mu}[((z-\lambda)(z-\xi))^{-1}]. [ I N D C E 0 ] , C i := μ [ B i / ( z − ξ )] , D i := μ [ A i / ( z − λ )] , E 0 := μ [(( z − λ ) ( z − ξ ) ) − 1 ] .
Here C C C is the column with entries C i C_i C i and D D D the row with entries D i D_i D i , for i ∈ { 0 , … , n } i\in\{0,\ldots,n\} i ∈ { 0 , … , n } . They are Cauchy values of the original polynomials. The scalar E 0 E_0 E 0 pairs the two added pole functions. If the poles coincide, it is the moment μ ~ [ ( z − λ ) − 2 ] \widetilde{\mu}[(z-\lambda)^{-2}] μ [( z − λ ) − 2 ] ; the construction is unchanged.
Image conditions and normalization. Let e N e_N e N be the last coordinate column in C N \mathbb{C}^N C N and put
v : = B ( a ) , w : = A ( b ) T , η : = h n − 1 e N T . v:=\mathcal{B}(a),\qquad w:=\mathcal{A}(b)^{\mathsf T},\qquad\eta:=h_n^{-1}e_N^{\mathsf T}. v := B ( a ) , w := A ( b ) T , η := h n − 1 e N T .
A general left ambient function is ϕ ( z ) = u T B ( z ) + u p / ( z − λ ) \phi(z)=u^{\mathsf T}\mathcal{B}(z)+u_p/(z-\lambda) ϕ ( z ) = u T B ( z ) + u p / ( z − λ ) . Its only possible polynomial preimage is
ϕ ( z ) S L − 1 ( z ) = z − λ z − a ϕ ( z ) . \phi(z)S_L^{-1}(z)=\frac{z-\lambda}{z-a}\phi(z). ϕ ( z ) S L − 1 ( z ) = z − a z − λ ϕ ( z ) .
Multiplication by z − λ z-\lambda z − λ removes the original pole. The result is divisible by z − a z-a z − a precisely when ϕ ( a ) = 0 \phi(a)=0 ϕ ( a ) = 0 , since a ≠ λ a\ne\lambda a = λ . After this division, the degree is at most n n n . The coefficient of degree n n n is e N T u e_N^{\mathsf T}u e N T u , because B n B_n B n is monic and S L − 1 S_L^{-1} S L − 1 tends to one at infinity. Requiring this coefficient to vanish gives exactly the bound defining X X X . Hence the two left constraints are
Γ L [ u u p ] = 0 , Γ L = [ v T ( a − λ ) − 1 e N T 0 ] . \Gamma_L\begin{bmatrix}u\\u_p\end{bmatrix}=0,\qquad
\Gamma_L=\begin{bmatrix}v^{\mathsf T}&(a-\lambda)^{-1}\\e_N^{\mathsf T}&0\end{bmatrix}. Γ L [ u u p ] = 0 , Γ L = [ v T e N T ( a − λ ) − 1 0 ] .
On the right, a general ambient function has the form ψ ( z ) = A ( z ) z 0 + z p / ( z − ξ ) \psi(z)=\mathcal{A}(z)z_0+z_p/(z-\xi) ψ ( z ) = A ( z ) z 0 + z p / ( z − ξ ) . Its preimage is ( z − ξ ) ψ ( z ) / ( z − b ) (z-\xi)\psi(z)/(z-b) ( z − ξ ) ψ ( z ) / ( z − b ) . Vanishing at b b b is the only image condition, since Y Y Y permits degree n n n . Its leading coefficient is η z 0 \eta z_0 η z 0 . Thus the right image condition and monic normalization are, respectively,
w T z 0 + ( b − ξ ) − 1 z p = 0 , η z 0 = 1. w^{\mathsf T}z_0+(b-\xi)^{-1}z_p=0,\qquad\eta z_0=1. w T z 0 + ( b − ξ ) − 1 z p = 0 , η z 0 = 1.
Eliminating the original coefficients. The vector of pairings of the enlarged left basis with ψ \psi ψ is
[ z 0 + C z p D z 0 + E 0 z p ] . \begin{bmatrix}
z_0+Cz_p\\
Dz_0+E_0z_p
\end{bmatrix}. [ z 0 + C z p D z 0 + E 0 z p ] .
Orthogonality requires its scalar product with every vector in ker Γ L \ker\Gamma_L ker Γ L to be zero. Because Γ L \Gamma_L Γ L has full row rank, its transpose has image equal to this space of vectors orthogonal to ker Γ L \ker\Gamma_L ker Γ L . Therefore there are unique scalars α 1 , α 2 \alpha_1,\alpha_2 α 1 , α 2 such that
z 0 + C z p = v α 1 + e N α 2 , D z 0 + E 0 z p = ( a − λ ) − 1 α 1 . z_0+Cz_p=v\alpha_1+e_N\alpha_2,\qquad Dz_0+E_0z_p=(a-\lambda)^{-1}\alpha_1. z 0 + C z p = v α 1 + e N α 2 , D z 0 + E 0 z p = ( a − λ ) − 1 α 1 .
The first equation gives z 0 = − C z p + v α 1 + e N α 2 z_0=-Cz_p+v\alpha_1+e_N\alpha_2 z 0 = − C z p + v α 1 + e N α 2 . Substituting this into the second equation, the right image condition and the normalization leaves three equations for z p , α 1 , α 2 z_p,\alpha_1,\alpha_2 z p , α 1 , α 2 :
T [ z p α 1 α 2 ] = e 3 , T : = [ E 0 − D C D v − ( a − λ ) − 1 D e N ( b − ξ ) − 1 − w T C w T v A n ( b ) − η C η v η e N ] . T\begin{bmatrix}z_p\\ \alpha_1\\ \alpha_2\end{bmatrix}=e_3,\qquad
T:=\begin{bmatrix}
E_0-DC & Dv-(a-\lambda)^{-1}De_N\\
(b-\xi)^{-1}-w^{T}C & w^{T}v & A_n(b)\\
-\eta C & \eta v & \eta e_N
\end{bmatrix}. T z p α 1 α 2 = e 3 , T := E 0 − D C ( b − ξ ) − 1 − w T C − η C D v − ( a − λ ) − 1 D e N w T v η v A n ( b ) η e N .
Here e 3 = [ 0 0 1 ] T ∈ C 3 e_3=[0\ 0\ 1]^{T}\in\mathbb{C}^{3} e 3 = [ 0 0 1 ] T ∈ C 3 . The first row is the remaining orthogonality equation, the second enforces polynomiality on the right, and the third fixes the leading coefficient. Its order is three for every n n n : the original polynomial coefficients have been eliminated. The general matrix (15) will implement the same elimination with vector-valued constraints and masses.
Recovering the polynomial. For det T ≠ 0 \det T\ne0 det T = 0 , solve the preceding system and multiply ψ \psi ψ by S R − 1 = ( z − ξ ) / ( z − b ) S_R^{-1}=(z-\xi)/(z-b) S R − 1 = ( z − ξ ) / ( z − b ) . With
r ( z ) : = [ 1 − ( z − ξ ) A ( z ) C ( z − ξ ) A ( z ) v ( z − ξ ) A n ( z ) ] , r(z):=[1-(z-\xi)\mathcal{A}(z)C\quad(z-\xi)\mathcal{A}(z)v\quad(z-\xi)A_n(z)], r ( z ) := [ 1 − ( z − ξ ) A ( z ) C ( z − ξ ) A ( z ) v ( z − ξ ) A n ( z )] ,
the result is B ^ n ( z ) = r ( z ) T − 1 e 3 / ( z − b ) \widehat{B}_n(z)=r(z)T^{-1}e_3/(z-b) B n ( z ) = r ( z ) T − 1 e 3 / ( z − b ) . Write T [ r ] T[r] T [ r ] for T T T with its last row replaced by r r r . Expansion of its determinant along that row gives det T [ r ] / det T = r T − 1 e 3 \det T[r]/\det T=rT^{-1}e_3 det T [ r ] / det T = r T − 1 e 3 , and hence
B ^ n ( z ) = det T [ r ( z ) ] ( z − b ) det T . \widehat{B}_n(z)=\frac{\det T[r(z)]}{(z-b)\det T}. B n ( z ) = ( z − b ) det T det T [ r ( z )] .
All objects in this formula are original polynomials, their values at a a a , b b b , and the rational moments C , D , E 0 C,D,E_0 C , D , E 0 .
The apparent pole at b b b cancels directly: r ( b ) r(b) r ( b ) is ( b − ξ ) (b-\xi) ( b − ξ ) times the second row of T T T , so the numerator determinant vanishes there. More explicitly, its three entries at b b b are
1 − ( b − ξ ) A ( b ) C , ( b − ξ ) A ( b ) v , ( b − ξ ) A n ( b ) , 1-(b-\xi)\mathcal{A}(b)C,\qquad(b-\xi)\mathcal{A}(b)v,\qquad(b-\xi)A_n(b), 1 − ( b − ξ ) A ( b ) C , ( b − ξ ) A ( b ) v , ( b − ξ ) A n ( b ) ,
which are precisely ( b − ξ ) (b-\xi) ( b − ξ ) times the second row of T T T . The numerator has degree at most n + 1 n+1 n + 1 , and division by z − b z-b z − b gives degree at most n n n . Since the degree-n n n coefficient of A \mathcal{A} A is η \eta η , the degree-n + 1 n+1 n + 1 coefficient of r r r is [ − η C η v η e N ] [-\eta C\ \eta v\ \eta e_N] [ − η C η v η e N ] , the last row of T T T . The leading numerator coefficient is therefore det T \det T det T , proving that the quotient has leading coefficient one. The pairing equations above give orthogonality of S R B ^ n S_R\widehat{B}_n S R B n to every product f S L fS_L f S L , with f ∈ X f\in X f ∈ X , for the chosen rational pairing. Therefore μ ^ [ f B ^ n ] = 0 \widehat{\mu}[f\widehat{B}_n]=0 μ [ f B n ] = 0 for every f ∈ X f\in X f ∈ X . They also show uniqueness, since all elimination steps are reversible and T T T is invertible.
The finite construction for prescribed component degree bounds
We now extend the scalar calculation to polynomial rows and columns with arbitrary prescribed finite bounds. The objective is to determine a polynomial column from prescribed orthogonality and normalization conditions, using data of the original pairing. We first choose the polynomial spaces, then express the rational products in enlarged bases, and finally eliminate the coefficients belonging to the original dual polynomials. Throughout this subsection, the four determinants are regular and the rational extension in Theorem 2.2 is fixed. Let X ⊂ A 1 × q X\subset A^{1\times q} X ⊂ A 1 × q and Y ⊂ A p × 1 Y\subset A^{p\times1} Y ⊂ A p × 1 have dimensions r r r and r + k r+k r + k , respectively, where k ≥ 0 k\ge0 k ≥ 0 . A square pairing is called normal when its Gram matrix is invertible.8 For k > 0 k>0 k > 0 , prescribe a linear map λ : Y → C k \lambda:Y\to\mathbb{C}^{k} λ : Y → C k and seek a p × k p\times k p × k matrix a ^ \widehat{a} a with columns in Y Y Y satisfying β ^ ( X , a ^ ) = 0 \widehat{\beta}(X,\widehat{a})=0 β ( X , a ) = 0 and λ ( a ^ ) = I k \lambda(\widehat{a})=I_k λ ( a ) = I k , where λ \lambda λ acts columnwise. Each column has r + k r+k r + k unknown coefficients, r r r orthogonality equations and k k k normalization equations. In one variable, X , Y X,Y X , Y impose the component degree bounds and λ \lambda λ selects leading coefficients. On a variety, we count independent polynomial classes with the prescribed bounds. The restriction ν ∣ X × Y \nu|_{X\times Y} ν ∣ X × Y is represented by a finite matrix, so it has a rank factorization with h h h terms of the form (5) . This statement uses only the finite target spaces. The bound h ≤ d L + d R h\le d_L+d_R h ≤ d L + d R is specific to the curve case of Theorem 2.2 ; no such bound is being assumed in higher dimension.
Polynomial spaces and additional rational vectors
Choose original polynomial spaces E ⊂ A 1 × q E \subset A^{1 \times q} E ⊂ A 1 × q and H ⊂ A p × 1 H \subset A^{p \times1} H ⊂ A p × 1 of dimension N N N , normal for β μ \beta_\mu β μ . Their dual bases form matrices B : = [ b i ] i ∈ { 1 , … , N } \mathcal{B} := [b_i]_{i \in\{1,\ldots,N\}} B := [ b i ] i ∈ { 1 , … , N } of size N × q N \times q N × q and A : = [ a 1 ⋯ a N ] \mathcal{A} := [a_1 \ \cdots\ a_N] A := [ a 1 ⋯ a N ] of size p × N p \times N p × N , with β μ ( b i , a j ) = δ i , j \beta_\mu(b_i,a_j)=\delta_{i,j} β μ ( b i , a j ) = δ i , j . We use P , Q P,Q P , Q for independent copies of the support variables. The spaces E , H E,H E , H provide the original data, while X , Y X,Y X , Y specify the perturbed problem. They need not coincide. For a chosen basis of X X X , multiply every row by S L S_L S L ; for a chosen basis of Y Y Y , multiply every column by S R S_R S R . Since these maps are linear, the resulting lists span all the rational products that can occur in the problem.
On the left, start with the original basis of E E E and add independent vectors from the first list until it spans E + X S L E+XS_L E + X S L . On the right, do the same with H H H and S R Y S_RY S R Y . Independence is tested over C \mathbb{C} C and subject to the equations of the support. For example, in A = C [ x , y ] / ( y 2 − x 3 ) A=\mathbb{C}[x,y]/(y^2-x^3) A = C [ x , y ] / ( y 2 − x 3 ) , the fractions y 2 / d y^2/d y 2 / d and x 3 / d x^3/d x 3 / d are the same element whenever the regular denominator d d d is inverted. They cannot contribute two independent directions.
Denote the added rows by ϕ α \phi_\alpha ϕ α and the added columns by ψ β \psi_\beta ψ β , and collect them in Φ L , Φ R \Phi_L,\Phi_R Φ L , Φ R . The minimal choice of ambient spaces and the corresponding numbers of added vectors are
W L min : = E + X S L , W R min : = H + S R Y , W_L^{\min}:=E+XS_L,\qquad W_R^{\min}:=H+S_RY, W L m i n := E + X S L , W R m i n := H + S R Y ,
s L min : = dim ( W L min / E ) , s R min : = dim ( W R min / H ) . s_L^{\min}:=\dim(W_L^{\min}/E),\qquad s_R^{\min}:=\dim(W_R^{\min}/H). s L m i n := dim ( W L m i n / E ) , s R m i n := dim ( W R m i n / H ) .
These spaces and dimensions are determined by the factors and the chosen polynomial spaces. The individual added vectors are choices of basis. Some may be polynomial, for example when multiplication raises a degree beyond the original space.
For a scalar example, take S L ( z ) = ( z − a ) / ( z − ξ ) S_L(z)=(z-a)/(z-\xi) S L ( z ) = ( z − a ) / ( z − ξ ) with a ≠ ξ a\ne\xi a = ξ . Polynomial division gives
f ( z ) S L ( z ) = f ( z ) + ( ξ − a ) f ( z ) − f ( ξ ) z − ξ + ( ξ − a ) f ( ξ ) z − ξ . f(z)S_L(z)=f(z)+(\xi-a)\frac{f(z)-f(\xi)}{z-\xi}+\frac{(\xi-a)f(\xi)}{z-\xi}. f ( z ) S L ( z ) = f ( z ) + ( ξ − a ) z − ξ f ( z ) − f ( ξ ) + z − ξ ( ξ − a ) f ( ξ ) .
The first two terms are polynomial. If their span is already in E E E , the single added function 1 / ( z − ξ ) 1/(z-\xi) 1/ ( z − ξ ) suffices for the rational part. With matrix factors, the component directions also matter; on an algebraic support, the defining relations determine which of the resulting fractions are independent.
Here an ambient space means a finite space containing both the original basis and the rational image to be calculated. Its extra directions need not themselves be rational images of allowed polynomials. The theorem permits larger ambient spaces when they are more convenient for computation. Choose W L ⊃ E + X S L W_L\supset E+XS_L W L ⊃ E + X S L and W R ⊃ H + S R Y W_R\supset H+S_RY W R ⊃ H + S R Y , with dimensions N + s L , N + s R N+s_L,N+s_R N + s L , N + s R , and bases
B + : = [ B Φ L ] , A + : = [ A Φ R ] . \mathcal{B}^{+}:=\begin{bmatrix}\mathcal{B}\\\Phi_L\end{bmatrix},\qquad\mathcal{A}^{+}:=[\mathcal{A}\ \Phi_R]. B + := [ B Φ L ] , A + := [ A Φ R ] .
All products used in the pairings must lie in the chosen rational extension domain. The minimal ambient spaces lie componentwise in A [ 1 / ( det L 1 ) ] A[1/(\det L_1)] A [ 1/ ( det L 1 )] on the left and A [ 1 / ( det R 1 ) ] A[1/(\det R_1)] A [ 1/ ( det R 1 )] on the right, as follows from the adjugate expressions for the inverses.
The image equations can now be constructed from polynomial coefficients. Fix ordered bases f 1 , … , f r f_1,\ldots,f_r f 1 , … , f r of X X X and g 1 , … , g r + k g_1,\ldots,g_{r+k} g 1 , … , g r + k of Y Y Y . Choose common regular denominators for the finite lists consisting of the ambient basis vectors and the rational images f i S L , S R g j f_iS_L,S_Rg_j f i S L , S R g j . Multiplication by these denominators turns equality of rational vectors into equality of polynomial classes in A A A . Reduce all entries modulo I I I , using a Gröbner basis, and collect the finitely many standard monomials that occur. Coefficients in the component copies of these monomials give ordinary complex column vectors.
On the left, let A L A_L A L have as columns the coefficient vectors of the cleared ambient basis, and let B L B_L B L have as columns those of the cleared images f i S L f_iS_L f i S L . Use the same scalar denominator and the same polynomial coordinates in both matrices. The ambient basis is independent and denominator multiplication is injective, so A L A_L A L has full column rank N + s L N+s_L N + s L . Each column of B L B_L B L belongs to its image. Solving the coefficient equations therefore determines a unique matrix Z L Z_L Z L by
A L Z L = B L , Z L ∈ Mat ( N + s L ) × r ( C ) . A_LZ_L=B_L,\qquad Z_L\in\operatorname{Mat}_{(N+s_L)\times r}(\mathbb{C}). A L Z L = B L , Z L ∈ Mat ( N + s L ) × r ( C ) .
Its columns are the coordinates of f i S L f_iS_L f i S L in B + \mathcal{B}^{+} B + . Multiplication by S L S_L S L is injective, so these columns have rank r r r . The identical calculation on columns gives
A R Z R = B R , Z R ∈ Mat ( N + s R ) × ( r + k ) ( C ) , rank Z R = r + k , A_RZ_R=B_R,\qquad Z_R\in\operatorname{Mat}_{(N+s_R)\times(r+k)}(\mathbb{C}),\qquad\operatorname{rank}Z_R=r+k, A R Z R = B R , Z R ∈ Mat ( N + s R ) × ( r + k ) ( C ) , rank Z R = r + k ,
where the ambient columns and the images S R g j S_R g_j S R g j are cleared by the same regular denominator on the right.
Row reduction of Z L T Z_L^{\mathsf{T}} Z L T gives a basis of its kernel. Transposing those basis vectors produces independent equations on the ambient coordinates; their common kernel is im Z L = X S L \operatorname{im} Z_L = X S_L im Z L = X S L . Indeed they annihilate every column of Z L Z_L Z L , and their number is N + s L − r N+s_L-r N + s L − r , leaving a kernel of dimension r r r . The right calculation leaves a kernel of dimension r + k r+k r + k . Splitting the columns according to the original and added basis vectors gives the constraint matrices
Γ L : = [ X 0 X p ] , Γ R : = [ Y 0 Y p ] , ker Γ L = X S L , ker Γ R = S R Y , (9) \Gamma_L := [X_0\ X_p], \qquad\Gamma_R := [Y_0\ Y_p], \qquad\ker\Gamma_L = X S_L, \qquad\ker\Gamma_R = S_R Y, \tag*{(9)} Γ L := [ X 0 X p ] , Γ R := [ Y 0 Y p ] , ker Γ L = X S L , ker Γ R = S R Y , (9)
where kernels are understood in the chosen coordinates. The rows are independent, so their numbers are c L : = N + s L − r c_L := N+s_L-r c L := N + s L − r and c R : = N + s R − r − k c_R := N+s_R-r-k c R := N + s R − r − k . In particular, X 0 X_0 X 0 , X p X_p X p have sizes c L × N c_L \times N c L × N , c L × s L c_L \times s_L c L × s L , and Y 0 Y_0 Y 0 , Y p Y_p Y p have sizes c R × N c_R \times N c R × N , c R × s R c_R \times s_R c R × s R . Thus the factors, the chosen finite spaces and polynomial coefficient reduction determine every entry of the constraints. The division construction below separates their polynomial meaning into remainder conditions and bounds on the resulting quotients. On their kernels, π L ( b ) : = b S L − 1 \pi_L(b) := bS_L^{-1} π L ( b ) := b S L − 1 and π R ( c ) : = S R − 1 c \pi_R(c) := S_R^{-1}c π R ( c ) := S R − 1 c recover the polynomial rows in X X X and columns in Y Y Y , respectively. Extend these maps linearly to W L W_L W L , W R W_R W R , with values in X X X , Y Y Y ; only their values on the kernels enter the answers.
Such extensions exist because the spaces are finite dimensional: complete a basis of each kernel and prescribe any values on the added basis vectors, for example zero. A single enlarged-basis vector need not itself correspond to an allowed polynomial. The maps π L , π R \pi_L,\pi_R π L , π R provide polynomial values for completing the tables below; the constraint equations ensure that the final combinations lie in the kernels, where the maps are the actual inverse multiplications. This is why the final answers do not depend on the chosen extensions.
Constructing the divisibility conditions
For a rational row b b b , solving b = f S L b=fS_L b = f S L requires two checks: f = b S L − 1 f=bS_L^{-1} f = b S L − 1 must be polynomial on the support, and it must belong to the prescribed space X X X . A polynomial quotient can still violate a component degree bound. For columns the order is reversed. Choose regular scalars a L , a R ∈ A a_L,a_R \in A a L , a R ∈ A clearing both the ambient spaces and the rational matrices, and put
P L : = a L S L , P R : = a R S R , a L W L ⊂ A 1 × q , a R W R ⊂ A p × 1 . (10) P_L := a_L S_L,\qquad P_R := a_R S_R,\qquad a_L W_L \subset A^{1\times q},\qquad a_R W_R \subset A^{p\times1}. \tag*{(10)} P L := a L S L , P R := a R S R , a L W L ⊂ A 1 × q , a R W R ⊂ A p × 1 . (10)
For the minimal ambient spaces, the explicit choices a L = det L 1 a_L=\det L_1 a L = det L 1 , P L = adj ( L 1 ) L 0 P_L=\operatorname{adj}(L_1)L_0 P L = adj ( L 1 ) L 0 , a R = det R 1 a_R=\det R_1 a R = det R 1 and P R = R 0 adj ( R 1 ) P_R=R_0\operatorname{adj}(R_1) P R = R 0 adj ( R 1 ) suffice; an enlarged ambient space may require additional scalar factors. These lifts represent the rational images; the functional equation and its mass module retain the original four factors.
There are two distinct sources of constraints. A remainder detects failure of polynomial divisibility. Once that remainder is zero, the quotient may still lie outside the prescribed degree or section space. The proposition counts these two groups separately. This distinction prevents a divisible rational expression from being accepted as a candidate with the wrong polynomial bounds.
Proposition 3.1 (Polynomial division and the target quotient). For b ∈ W L b \in W_L b ∈ W L and c ∈ W R c \in W_R c ∈ W R ,
b ∈ X S L if and only if a L b ∈ X P L , b \in X S_L \quad\text{if and only if} \quad a_L b \in X P_L, b ∈ X S L if and only if a L b ∈ X P L , c ∈ S R Y if and only if a R c ∈ P R Y . c \in S_R Y \quad\text{if and only if} \quad a_R c \in P_R Y. c ∈ S R Y if and only if a R c ∈ P R Y . Each condition consists of polynomial matrix divisibility, followed by membership of the unique quotient in X X X or Y Y Y . Define C ( P L ) : = A 1 × q / A 1 × q P L C(P_L) := A^{1\times q}/A^{1\times q}P_L C ( P L ) := A 1 × q / A 1 × q P L and ρ L : = dim im [ a L W L ⟶ C ( P L ) ] \rho_L := \dim\operatorname{im}[a_L W_L \longrightarrow C(P_L)] ρ L := dim im [ a L W L ⟶ C ( P L )] , and define C ( P R ) : = A p × 1 / P R A p × 1 C(P_R) := A^{p\times1}/P_R A^{p\times1} C ( P R ) := A p × 1 / P R A p × 1 and ρ R \rho_R ρ R on the right. Exactly ρ L \rho_L ρ L independent scalar divisibility tests and c L − ρ L c_L-\rho_L c L − ρ L quotient-space tests give Γ L \Gamma_L Γ L ; the corresponding numbers for Γ R \Gamma_R Γ R are ρ R , c R − ρ R \rho_R,c_R-\rho_R ρ R , c R − ρ R .
Proof. Since a L a_L a L is regular, multiplying an equality by a L a_L a L neither loses nor introduces solutions. Thus b = f S L b=fS_L b = f S L with f ∈ X f\in X f ∈ X is equivalent to a L b = f P L a_Lb=fP_L a L b = f P L with the same f f f . The right identity follows by the same argument, keeping the order a R c = P R g a_Rc=P_Rg a R c = P R g .
Consider first the map from a L W L a_LW_L a L W L to the quotient C ( P L ) C(P_L) C ( P L ) . Its image has dimension ρ L \rho_L ρ L . Choosing a basis of the dual of this image gives exactly ρ L \rho_L ρ L independent scalar conditions for its remainder to vanish. Their common kernel therefore has dimension N + s L − ρ L N+s_L-\rho_L N + s L − ρ L .
On that kernel, division by P L P_L P L is well defined and unique, because P L P_L P L is invertible over K K K . It identifies the kernel with a polynomial row space of dimension N + s L − ρ L N+s_L-\rho_L N + s L − ρ L containing X X X . Since dim X = r \dim X=r dim X = r , a further N + s L − ρ L − r = c L − ρ L N+s_L-\rho_L-r=c_L-\rho_L N + s L − ρ L − r = c L − ρ L independent conditions select precisely X X X . Choose these conditions on the divisible subspace and extend them linearly to the ambient space. Together with the remainder conditions, they form Γ L \Gamma_L Γ L . To verify independence, restrict a putative relation among all these tests to the divisible subspace. Independence of the quotient-space tests makes their coefficients zero. What remains is a relation among the independent remainder tests, whose coefficients are also zero. The common kernel is X S L X S_L X S L , and the number of independent tests is therefore exactly c L c_L c L .
For columns, the remainder map has image dimension ρ R \rho_R ρ R and kernel dimension N + s R − ρ R N+s_R-\rho_R N + s R − ρ R . Division identifies this kernel with a space containing Y Y Y , of dimension r + k r+k r + k . Thus the number of further conditions is N + s R − ρ R − ( r + k ) = c R − ρ R N+s_R-\rho_R-(r+k)=c_R-\rho_R N + s R − ρ R − ( r + k ) = c R − ρ R .
Remainders can be tested by values and derivatives. For example, modulo t 2 t^2 t 2 the remainder is a + b t a+bt a + b t : its value at zero detects a a a , and its first derivative detects b b b . On an algebraic support the derivatives must also respect its equations. The following construction imposes this compatibility, including at singular points. For a closed point ξ \xi ξ let m ξ \mathfrak{m}_\xi m ξ be its maximal ideal.9 A functional on C ( P L ) / m ξ h ξ C ( P L ) C(P_L)/\mathfrak{m}_\xi^{h_\xi}C(P_L) C ( P L ) / m ξ h ξ C ( P L ) acts on polynomial row representatives as a finite jet of order below h ξ ≥ 1 h_\xi\geq1 h ξ ≥ 1 . For α ∈ N M \alpha\in\mathbb{N}^M α ∈ N M put ∣ α ∣ : = ∑ j α j |\alpha|:=\sum_j\alpha_j ∣ α ∣ := ∑ j α j , α ! : = ∏ j α j ! \alpha!:=\prod_j\alpha_j! α ! := ∏ j α j ! and ∂ α : = ∏ j ∂ x j α j \partial^\alpha:=\prod_j\partial_{x_j}^{\alpha_j} ∂ α := ∏ j ∂ x j α j ; then
γ ( p ) : = ∑ i = 1 q ∑ ∣ α ∣ < h ξ c i , α ∂ α p i ( ξ ) α ! . (11) \gamma(p):=\sum_{i=1}^{q}\sum_{|\alpha|<h_\xi}c_{i,\alpha}\frac{\partial^\alpha p_i(\xi)}{\alpha!}. \tag*{(11)} γ ( p ) := i = 1 ∑ q ∣ α ∣ < h ξ ∑ c i , α α ! ∂ α p i ( ξ ) . (11)
If A = C [ x 1 , … , x M ] / I A=\mathbb{C}[x_1,\ldots,x_M]/I A = C [ x 1 , … , x M ] / I , the coefficients must annihilate I A a m b 1 × q I A_{\mathrm{amb}}^{1\times q} I A amb 1 × q and every polynomial row multiple of a representative of P L P_L P L , where A a m b : = C [ x 1 , … , x M ] A_{\mathrm{amb}}:=\mathbb{C}[x_1,\ldots,x_M] A amb := C [ x 1 , … , x M ] . These relations make the jet independent of representatives on the support. To see why these are precisely the necessary compatibility equations, expand each polynomial about ξ \xi ξ and discard all monomials of total degree at least h ξ h_\xi h ξ . Its remaining coefficients are ∂ α p i ( ξ ) / α ! \partial^\alpha p_i(\xi)/\alpha! ∂ α p i ( ξ ) / α ! . Every linear functional on this truncated space has the displayed form. It descends first to the support and then to the remainder module exactly when it vanishes on the truncated support relations and on the truncated polynomial matrix multiples. Only finitely many such truncated relations are independent. The column definition uses column multiples and the same Taylor coefficients.
The value-only criterion uses radical ideals.10
The quotient description is intrinsic, but to fill a constraint matrix we need scalar tests. The next result shows that finitely many values and derivatives suffice on each prescribed ambient space. It does not require the entire remainder module to be finite dimensional: only its image from a L W L a_LW_L a L W L or a R W R a_RW_R a R W R is finite dimensional. The compatibility conditions in (11) ensure that the tests respect both the support equations and multiplication by the polynomial matrix.
The next result proves that a finite separating list of jets exists for the chosen spaces. The actual points, orders and independent rows are part of the constraint data; no uniform jet-order bound is asserted. The coefficient construction above already computes the constraints from any supplied polynomial presentation of the finite spaces. The jet description identifies those equations with local values and derivatives. Subsubsection 3.2.6 computes the finite image conditions directly from its rational basis vectors.
Proposition 3.2 (Complete finite jets on a prescribed finite-dimensional space). The divisibility conditions in Proposition 3.1 can be chosen as ρ L \rho_L ρ L and ρ R \rho_R ρ R independent compatible jets. Directional values p ( ξ ) v p(\xi)v p ( ξ ) v , v ∈ ker P L ( ξ ) v\in\ker P_L(\xi) v ∈ ker P L ( ξ ) , suffice on the left in either of the following cases: ( det P L ) (\det P_L) ( det P L ) is radical in A A A ; or the support is a smooth affine curve and every local Smith exponent of P L P_L P L at a determinant zero is zero or one. The right version uses left null vectors of P R ( ξ ) P_R(\xi) P R ( ξ ) under the corresponding hypothesis. The second criterion allows repeated zeros of the determinant.
Proof. Let u u u be a nonzero remainder and choose a maximal ideal m \mathfrak{m} m containing its proper annihilator. Its localization u m u_{\mathfrak{m}} u m is nonzero: otherwise some a ∉ m a\notin\mathfrak{m} a ∈ / m would satisfy a u = 0 au=0 a u = 0 , contradicting containment of the annihilator in m \mathfrak{m} m . The localized remainder module is finite over the Noetherian local ring A m A_{\mathfrak{m}} A m . Krull’s intersection theorem gives
⋂ h ≥ 1 m h C ( P L ) m = 0. \bigcap_{h \ge1} \mathfrak{m}^{h}C(P_L)_{\mathfrak{m}} = 0. h ≥ 1 ⋂ m h C ( P L ) m = 0.
Hence u m u_{\mathfrak{m}} u m has a nonzero image modulo some power m h \mathfrak{m}^{h} m h . The quotient by this power is supported at the single closed point, so localizing it does not change it. It is a finite-dimensional vector space, and a linear functional on it can be chosen nonzero on the image of u u u . Thus a finite Taylor truncation detects u u u [14 ] . Its linear functionals are compatible jets (11) . Restrict these jets to the image of a L W L a_LW_L a L W L in C ( P L ) C(P_L) C ( P L ) . Their common kernel is zero: every nonzero class is detected by at least one of them. In a vector space of dimension ρ L \rho_L ρ L , a family that detects every nonzero vector spans the dual: if its span had smaller dimension, the common kernel would have positive dimension by rank-nullity. We can therefore select ρ L \rho_L ρ L independent jets, whose simultaneous vanishing is equivalent to zero remainder. This proves finiteness for the prescribed space, although the required points and orders may change when the space is enlarged. The right construction is obtained by transposition.
Where det P L = 0 \det P_L = 0 det P L = 0 , the columns of adj ( P L ) \operatorname{adj}(P_L) adj ( P L ) lie in ker P L \ker P_L ker P L , so the directional conditions make p adj ( P L ) p\operatorname{adj}(P_L) p adj ( P L ) vanish there. Since ( det P L ) (\det P_L) ( det P L ) is radical, the Nullstellensatz makes every entry divisible by det P L \det P_L det P L . Thus p P L − 1 = p adj ( P L ) / det P L pP_L^{-1} = p\operatorname{adj}(P_L)/\det P_L p P L − 1 = p adj ( P L ) / det P L is polynomial. Conversely, if p = h P L p = hP_L p = h P L is divisible, then p ( ξ ) v = h ( ξ ) P L ( ξ ) v = 0 p(\xi)v = h(\xi)P_L(\xi)v = 0 p ( ξ ) v = h ( ξ ) P L ( ξ ) v = 0 for every v ∈ ker P L ( ξ ) v \in\ker P_L(\xi) v ∈ ker P L ( ξ ) . Thus the directional tests are both necessary and sufficient. Transposition proves the column assertion.
On a smooth curve write locally P L = U diag ( t θ i ) V P_L = U\operatorname{diag}(t^{\theta_i})V P L = U diag ( t θ i ) V , with U , V U,V U , V invertible. Here regular at a point means without a pole; regular scalars in Subsection 2.1 are instead elements that are not zero divisors. If θ i ∈ { 0 , 1 } \theta_i \in\{0,1\} θ i ∈ { 0 , 1 } , the quotient p P L − 1 = p V − 1 diag ( t − θ i ) U − 1 pP_L^{-1} = pV^{-1}\operatorname{diag}(t^{-\theta_i})U^{-1} p P L − 1 = p V − 1 diag ( t − θ i ) U − 1 has no pole precisely when p ( ξ ) V ( ξ ) − 1 e i = 0 p(\xi)V(\xi)^{-1}e_i = 0 p ( ξ ) V ( ξ ) − 1 e i = 0 for each θ i = 1 \theta_i = 1 θ i = 1 . These vectors form a basis of ker P L ( ξ ) \ker P_L(\xi) ker P L ( ξ ) . A rational function without poles on the affine curve belongs to A A A , so the quotient is polynomial. This proves the second criterion and its transposed version. Differential descriptions of module membership in greater generality are developed in [16 ] . □
For a rational matrix with local Smith exponents ( − 1 , 0 , … , 0 ) (-1,0,\ldots,0) ( − 1 , 0 , … , 0 ) , a minimal local scalar lift has Smith form diag ( 1 , t , … , t ) \operatorname{diag}(1,t,\ldots,t) diag ( 1 , t , … , t ) . Its determinant vanishes to order q − 1 q-1 q − 1 , but the new criterion still uses values only. This illustrates why determinant multiplicity alone does not specify the derivative order needed by the constraint tests. A larger scalar lift may introduce redundant derivative tests.
For example, over C [ x , y ] \mathbb{C}[x,y] C [ x , y ] , take P = [ x 2 0 − x y 1 ] P=\begin{bmatrix}x^2&0\\-xy&1\end{bmatrix} P = [ x 2 − x y 0 1 ] . Its inverse over C ( x , y ) \mathbb{C}(x,y) C ( x , y ) is
P − 1 = [ x − 2 0 y / x 1 ] , p P − 1 = [ ( p 1 + x y p 2 ) / x 2 p 2 ] . P^{-1}=\begin{bmatrix}x^{-2}&0\\y/x&1\end{bmatrix},\qquad pP^{-1}=[(p_1+xyp_2)/x^2\quad p_2]. P − 1 = [ x − 2 y / x 0 1 ] , p P − 1 = [( p 1 + x y p 2 ) / x 2 p 2 ] .
Write p 1 = A 0 ( y ) + x A 1 ( y ) + x 2 R 1 ( x , y ) p_1=A_0(y)+xA_1(y)+x^2R_1(x,y) p 1 = A 0 ( y ) + x A 1 ( y ) + x 2 R 1 ( x , y ) and p 2 = B 0 ( y ) + x R 2 ( x , y ) p_2=B_0(y)+xR_2(x,y) p 2 = B 0 ( y ) + x R 2 ( x , y ) . The first numerator becomes A 0 + x ( A 1 + y B 0 ) + x 2 ( R 1 + y R 2 ) A_0+x(A_1+yB_0)+x^2(R_1+yR_2) A 0 + x ( A 1 + y B 0 ) + x 2 ( R 1 + y R 2 ) . It is divisible by x 2 x^2 x 2 exactly when A 0 = 0 A_0=0 A 0 = 0 and A 1 + y B 0 = 0 A_1+yB_0=0 A 1 + y B 0 = 0 , or equivalently
p 1 ( 0 , y ) = 0 , ∂ x p 1 ( 0 , y ) + y p 2 ( 0 , y ) = 0. p_1(0,y)=0,\qquad\partial_xp_1(0,y)+yp_2(0,y)=0. p 1 ( 0 , y ) = 0 , ∂ x p 1 ( 0 , y ) + y p 2 ( 0 , y ) = 0.
For total degree at most n n n , these two polynomials in y y y have degrees at most n n n and n + 1 n+1 n + 1 , respectively. Testing them at n + 2 n+2 n + 2 distinct values is complete, since a nonzero polynomial of degree at most n + 1 n+1 n + 1 has at most n + 1 n+1 n + 1 roots; the tests may be redundant. Here the divisor is the line x = 0 x=0 x = 0 : ∂ x \partial_x ∂ x differentiates away from it, while the coefficient y y y varies along it.
The same constraints can be calculated by polynomial division, without first choosing jet points. Fix a monomial order and a module Gröbner basis for the submodule generated by the rows of P L P_L P L and the support relations I e j Ie_j I e j . Reduce each polynomial row a L b i + a_Lb_i^+ a L b i + by this basis. Collect the coefficients of its remainder as column i i i of a matrix R L R_L R L . Linearity of the normal form gives
R L u = 0 if and only if a L ∑ i u i b i + is divisible by P L in A . R_Lu=0\quad\text{if and only if}\quad a_L\sum_i u_ib_i^+\text{ is divisible by }P_L\text{ in }A. R L u = 0 if and only if a L i ∑ u i b i + is divisible by P L in A .
Thus the kernel gives all divisible combinations of the ambient rows.
Take a basis of this kernel and divide each associated polynomial row by P L P_L P L . The quotients are unique in A A A under the regularity hypothesis. Express them in polynomial coordinates modulo the support ideal I I I . One more linear calculation now selects those combinations whose quotients belong to the prescribed space X X X . For component degree bounds this means setting every forbidden quotient coefficient to zero; in a projective description it includes the conditions at infinity. These are the quotient conditions in Proposition 3.1 , in addition to divisibility. Finally take a basis of the linear equations annihilating the selected coefficient subspace. It gives exactly the c L c_L c L independent rows of Γ L \Gamma_L Γ L . Transposing the entire calculation gives the column algorithm for Γ R \Gamma_R Γ R . Both polynomial reduction and compatible jets therefore compute the same image constraints.
If γ a \gamma_a γ a denotes the combined tests on a L W L a_LW_L a L W L , the actual entries are ( Γ L ) a , i = γ a ( a L b i + ) (\Gamma_L)_{a,i}=\gamma_a(a_Lb_i^+) ( Γ L ) a , i = γ a ( a L b i + ) ; similarly ( Γ R ) b , j = δ b ( a R a j + ) (\Gamma_R)_{b,j}=\delta_b(a_Ra_j^+) ( Γ R ) b , j = δ b ( a R a j + ) .
In particular, the first N N N columns of Γ L \Gamma_L Γ L contain these tests on the original polynomials b i b_i b i , and the remaining s L s_L s L columns contain the same tests on the added rational rows after clearing denominators. This is the meaning of the split Γ L = [ X 0 X p ] \Gamma_L=[X_0\ X_p] Γ L = [ X 0 X p ] ; Γ R = [ Y 0 Y p ] \Gamma_R=[Y_0\ Y_p] Γ R = [ Y 0 Y p ] is constructed in exactly the same order on the right. All these entries depend on the factors and the polynomial spaces, before any perturbed orthogonal polynomial has been found.
§ 4 realizes such remainder functionals by residues on the curve and surface quotients specified there. Evaluating those functionals on the cleared ambient basis provides the divisibility rows of the constraint matrix. The remaining rows impose the prescribed quotient space, as in Proposition 3.1 . Residue coordinates also evaluate a homogeneous mass once its values on a quotient basis have been chosen. In this way the geometric formulas supply concrete entries for the finite construction. The rational moments in the next subsection are the values of the chosen extension, and the mass coefficients specify the chosen solution of the functional equation.
Second-kind data on the algebraic support
The extra vectors Φ L , Φ R \Phi_L,\Phi_R Φ L , Φ R require moments with denominators. These are the second-kind data. For measure pairings they are expressed through Cauchy transforms in one variable and Fantappiè transforms in several variables, as made explicit below. Once the finite ambient bases are fixed, only the three finite tables below are needed from the rational extension. Thus their entries are known inputs to the elimination, whether evaluated from a measure or supplied as additional moments consistently with that extension. Write their rows and columns as ϕ α : = φ α / a L \phi_\alpha:=\varphi_\alpha/a_L ϕ α := φ α / a L and ψ β : = χ β / a R \psi_\beta:=\chi_\beta/a_R ψ β := χ β / a R , with polynomial numerators. The moment blocks are
( C 0 ) i , β : = ∑ u , v μ ~ u , v [ ι L ( b i , u ) ι R ( χ β , v / a R ) ] , ( D 0 ) α , j : = ∑ u , v μ ~ u , v [ ι L ( φ α , u / a L ) ι R ( a j , v ) ] , ( E 0 ) α , β : = ∑ u , v μ ~ u , v [ ι L ( φ α , u / a L ) ι R ( χ β , v / a R ) ] . (12) \begin{aligned}
(C_0)_{i,\beta}&:=\sum_{u,v}\widetilde{\mu}_{u,v}\left[\iota_L(b_{i,u})\iota_R(\chi_{\beta,v}/a_R)\right],\\
(D_0)_{\alpha,j}&:=\sum_{u,v}\widetilde{\mu}_{u,v}\left[\iota_L(\varphi_{\alpha,u}/a_L)\iota_R(a_{j,v})\right],\\
(E_0)_{\alpha,\beta}&:=\sum_{u,v}\widetilde{\mu}_{u,v}\left[\iota_L(\varphi_{\alpha,u}/a_L)\iota_R(\chi_{\beta,v}/a_R)\right]. \tag*{(12)}
\end{aligned} ( C 0 ) i , β ( D 0 ) α , j ( E 0 ) α , β := u , v ∑ μ u , v [ ι L ( b i , u ) ι R ( χ β , v / a R ) ] , := u , v ∑ μ u , v [ ι L ( φ α , u / a L ) ι R ( a j , v ) ] , := u , v ∑ μ u , v [ ι L ( φ α , u / a L ) ι R ( χ β , v / a R ) ] . (12)
Thus C 0 C_0 C 0 pairs original rows with extra columns, D 0 D_0 D 0 pairs extra rows with original columns, and E 0 E_0 E 0 pairs both extra lists. Their sizes are N × s R N\times s_R N × s R , s L × N s_L\times N s L × N and s L × s R s_L\times s_R s L × s R . Balanced pairings use one point in these moments; bivariate pairings keep two arguments.
Equivalently, the three tables are
( C 0 ) i , β = μ ~ ( b i , ψ β ) , ( D 0 ) α , j = μ ~ ( ϕ α , a j ) , ( E 0 ) α , β = μ ~ ( ϕ α , ψ β ) . (C_0)_{i,\beta}=\widetilde{\mu}(b_i,\psi_\beta),\qquad
(D_0)_{\alpha,j}=\widetilde{\mu}(\phi_\alpha,a_j),\qquad
(E_0)_{\alpha,\beta}=\widetilde{\mu}(\phi_\alpha,\psi_\beta). ( C 0 ) i , β = μ ( b i , ψ β ) , ( D 0 ) α , j = μ ( ϕ α , a j ) , ( E 0 ) α , β = μ ( ϕ α , ψ β ) .
For a scalar balanced measure and added functions ϕ ( z ) = 1 / ( z − ξ L ) \phi(z)=1/(z-\xi_L) ϕ ( z ) = 1/ ( z − ξ L ) , ψ ( z ) = 1 / ( z − ξ R ) \psi(z)=1/(z-\xi_R) ψ ( z ) = 1/ ( z − ξ R ) , these become
C i = ∫ b i ( z ) z − ξ R d μ ( z ) , D j = ∫ a j ( z ) z − ξ L d μ ( z ) , E = ∫ d μ ( z ) ( z − ξ L ) ( z − ξ R ) . C_i=\int\frac{b_i(z)}{z-\xi_R}\,d\mu(z),\qquad
D_j=\int\frac{a_j(z)}{z-\xi_L}\,d\mu(z),\qquad
E=\int\frac{d\mu(z)}{(z-\xi_L)(z-\xi_R)}. C i = ∫ z − ξ R b i ( z ) d μ ( z ) , D j = ∫ z − ξ L a j ( z ) d μ ( z ) , E = ∫ ( z − ξ L ) ( z − ξ R ) d μ ( z ) .
Thus C 0 , D 0 C_0,D_0 C 0 , D 0 are the familiar Cauchy data of the original polynomials; E 0 E_0 E 0 pairs the additional functions with one another. Repeated poles give derivatives with respect to pole parameters. For example,
1 ( m − 1 ) ! ∂ ξ m − 1 1 z − ξ = 1 ( z − ξ ) m , m ≥ 1 , \frac{1}{(m-1)!}\partial_{\xi}^{m-1}\frac{1}{z-\xi}=\frac{1}{(z-\xi)^m},\qquad m\ge1, ( m − 1 )! 1 ∂ ξ m − 1 z − ξ 1 = ( z − ξ ) m 1 , m ≥ 1 ,
so the factorial and sign are fixed by differentiating the denominator with respect to ξ \xi ξ , rather than with respect to z z z . For an abstract functional, these rational moments are part of the chosen extension μ ~ \widetilde{\mu} μ . Polynomial moments alone need not fix them. When a measure supplies convergent integrals, those integrals specify the extension values used here.
For a finite matrix measure on a compact support, the rational moments depend holomorphically on denominator coefficients as long as the denominators have no zeros on that support. Indeed, near any such choice their absolute values have a common positive lower bound, which justifies differentiation under the integral against the total variation of the measure. For a fixed polynomial h h h , q t : = ∑ a t a g a q_t := \sum_a t_a g_a q t := ∑ a t a g a and a multi-index α \alpha α , this gives
∂ t α ∫ h q t d μ u , v = ( − 1 ) ∣ α ∣ ∣ α ∣ ! ∫ h ∏ a g a α a q t ∣ α ∣ + 1 d μ u , v . \partial_t^\alpha\int\frac{h}{q_t}\,d\mu_{u,v} = (-1)^{|\alpha|}|\alpha|! \int\frac{h\prod_a g_a^{\alpha_a}}{q_t^{|\alpha|+1}}\,d\mu_{u,v}. ∂ t α ∫ q t h d μ u , v = ( − 1 ) ∣ α ∣ ∣ α ∣ ! ∫ q t ∣ α ∣ + 1 h ∏ a g a α a d μ u , v .
The same argument applies to both denominators in the last block of (12) , also when they belong to independent bivariate arguments. Thus repeated poles are obtained by parameter derivatives. To exhibit the multivariate transform explicitly, fix a denominator q 0 q_0 q 0 with no zeros on the support and write a nearby one as q 0 − ∑ a t a g a q_0-\sum_a t_a g_a q 0 − ∑ a t a g a . Then
∫ h q 0 − ∑ a t a g a d μ u , v = ∫ h / q 0 1 − ∑ a t a ( g a / q 0 ) d μ u , v . \int\frac{h}{q_0-\sum_a t_a g_a}\,d\mu_{u,v} = \int\frac{h/q_0}{1-\sum_a t_a(g_a/q_0)}\,d\mu_{u,v}. ∫ q 0 − ∑ a t a g a h d μ u , v = ∫ 1 − ∑ a t a ( g a / q 0 ) h / q 0 d μ u , v .
The right-hand side is a Fantappiè transform: the denominator is one minus a linear combination of the functions g a / q 0 g_a/q_0 g a / q 0 , integrated against the weighted measure ( h / q 0 ) μ u , v (h/q_0)\mu_{u,v} ( h / q 0 ) μ u , v [17 ] . Thus the multivariate blocks use Fantappiè transforms. The two-denominator block uses the corresponding iterated transform, with one parameter family for each denominator. In the univariate specialization, partial fractions recover the Cauchy values and their derivatives displayed above.
Eliminating the original polynomial coefficients
We now eliminate the coefficients of the original dual polynomials. The coefficients of the extra rational vectors remain, together with the mass and division constraints. The required moment blocks are
G r a t : = [ I N C 0 D 0 E 0 ] , F : = [ F 0 F p ] , H ν : = [ H 0 H p ] , G : = G r a t + F H ν . (13) G_{\mathrm{rat}}:=\begin{bmatrix}I_N&C_0\\D_0&E_0\end{bmatrix},\qquad F:=\begin{bmatrix}F_0\\F_p\end{bmatrix},\qquad H_\nu:=[H_0\ H_p],\qquad G:=G_{\mathrm{rat}}+FH_\nu. \tag*{(13)} G rat := [ I N D 0 C 0 E 0 ] , F := [ F 0 F p ] , H ν := [ H 0 H p ] , G := G rat + F H ν . (13)
Here F i , a : = ℓ a ( π L b i + ) F_{i,a}:=\ell_a(\pi_L b_i^+) F i , a := ℓ a ( π L b i + ) and ( H ν ) a , j : = r a ( π R a j + ) (H_\nu)_{a,j}:=r_a(\pi_R a_j^+) ( H ν ) a , j := r a ( π R a j + ) , using (5) . Thus F F F has size ( N + s L ) × h (N+s_L)\times h ( N + s L ) × h , H ν H_\nu H ν has size h × ( N + s R ) h\times(N+s_R) h × ( N + s R ) , and G G G agrees with the target pairing on the two rational images. Indeed, for admissible coordinates v , z v,z v , z , the added entry is
v T F H ν z = ∑ a = 1 h ℓ a ( π L ( v T B + ) ) r a ( π R ( A + z ) ) = v ( π L ( v T B + ) , π R ( A + z ) ) . v^{\mathsf T}FH_\nu z=\sum_{a=1}^{h}\ell_a\bigl(\pi_L(v^{\mathsf T}\mathcal{B}^+)\bigr)r_a\bigl(\pi_R(\mathcal{A}^+z)\bigr)=v\bigl(\pi_L(v^{\mathsf T}\mathcal{B}^+),\pi_R(\mathcal{A}^+z)\bigr). v T F H ν z = a = 1 ∑ h ℓ a ( π L ( v T B + ) ) r a ( π R ( A + z ) ) = v ( π L ( v T B + ) , π R ( A + z ) ) .
This identity explains the factorization of the mass table: its entries are sums of products of a left scalar value and a right scalar value. The rational part supplies the other term of the transformed pairing. Write η : = [ η 0 η p ] = λ π R \eta:=[\eta_0\ \eta_p]=\lambda\pi_R η := [ η 0 η p ] = λ π R , a k × ( N + s R ) k\times(N+s_R) k × ( N + s R ) matrix in column coordinates.
The split of the mass tables can be written entry by entry as
( F 0 ) i , a = ℓ a ( π L b i ) , ( F p ) α , a = ℓ a ( π L ϕ α ) , (F_0)_{i,a}=\ell_a(\pi_L b_i),\qquad(F_p)_{\alpha,a}=\ell_a(\pi_L\phi_\alpha), ( F 0 ) i , a = ℓ a ( π L b i ) , ( F p ) α , a = ℓ a ( π L ϕ α ) ,
( H 0 ) a , j = r a ( π R a j ) , ( H p ) a , β = r a ( π R ψ β ) . (H_0)_{a,j}=r_a(\pi_R a_j),\qquad(H_p)_{a,\beta}=r_a(\pi_R\psi_\beta). ( H 0 ) a , j = r a ( π R a j ) , ( H p ) a , β = r a ( π R ψ β ) .
For point and derivative masses, these entries are the corresponding values and derivatives of the recovered polynomial rows and columns. If there is no mass, h = 0 h=0 h = 0 and these blocks are absent. The entries of η 0 , η p \eta_0,\eta_p η 0 , η p are obtained by applying the prescribed normalization to π R a j , π R ψ β \pi_R a_j,\pi_R\psi_\beta π R a j , π R ψ β . For a monic scalar problem, this means extracting the coefficient of the prescribed highest power.
The identity block in G r a t G_{\mathrm{rat}} G rat comes from the original duality β μ ( b i , a j ) = δ i , j \beta_\mu(b_i,a_j)=\delta_{i,j} β μ ( b i , a j ) = δ i , j . Every other block is therefore already known from the original polynomials, their second-kind data, the factors, the masses and normalization.
For the normalized problem k > 0 k>0 k > 0 , write the rational product of a candidate as S R a ^ = A + z S_R\widehat{a}=\mathcal{A}^+z S R a = A + z , where z : = [ z 0 T z p T ] T z:=[z_0^{\mathsf T}\ z_p^{\mathsf T}]^{\mathsf T} z := [ z 0 T z p T ] T has k k k columns. Its conditions are G z = Γ L T α Gz=\Gamma_L^{\mathsf T}\alpha G z = Γ L T α , Γ R z = 0 \Gamma_Rz=0 Γ R z = 0 and η z = I k \eta z=I_k η z = I k . Here α \alpha α is a c L × k c_L\times k c L × k matrix of multipliers. The first equation expresses orthogonality: a linear functional vanishing on ker Γ L \ker\Gamma_L ker Γ L is a linear combination of the rows of Γ L \Gamma_L Γ L . The second enforces division and the target bounds; the third fixes the normalization. Introduce u : = H ν z u := H_{\nu}z u := H ν z , the h h h mass coordinates. In blocks, the full system is
z 0 + C 0 z p + F 0 u = X 0 T α , D 0 z 0 + E 0 z p + F p u = X p T α , H 0 z 0 + H p z p − u = 0 , Y 0 z 0 + Y p z p = 0 , η 0 z 0 + η p z p = I k . \begin{aligned}
z_0 + C_0z_p + F_0u &= X_0^{\mathsf T}\alpha,\\
D_0z_0 + E_0z_p + F_pu &= X_p^{\mathsf T}\alpha,\\
H_0z_0 + H_pz_p - u &= 0,\\
Y_0z_0 + Y_pz_p &= 0,\\
\eta_0z_0 + \eta_pz_p &= I_k.
\end{aligned} z 0 + C 0 z p + F 0 u D 0 z 0 + E 0 z p + F p u H 0 z 0 + H p z p − u Y 0 z 0 + Y p z p η 0 z 0 + η p z p = X 0 T α , = X p T α , = 0 , = 0 , = I k .
The first line can be solved directly because its coefficient of z 0 z_0 z 0 is the identity. It gives
z 0 = − C 0 z p − F 0 u + X 0 T α . z_0=-C_0z_p-F_0u+X_0^{\mathsf T}\alpha. z 0 = − C 0 z p − F 0 u + X 0 T α .
For example, substitution into the second line gives
( E 0 − D 0 C 0 ) z p + ( F p − D 0 F 0 ) u + ( D 0 X 0 T − X p T ) α = 0. (E_0-D_0C_0)z_p+(F_p-D_0F_0)u+(D_0X_0^{\mathsf T}-X_p^{\mathsf T})\alpha=0. ( E 0 − D 0 C 0 ) z p + ( F p − D 0 F 0 ) u + ( D 0 X 0 T − X p T ) α = 0.
For completeness, the other three substituted equations are
( H p − H 0 C 0 ) z p + ( − I h − H 0 F 0 ) u + H 0 X 0 T α = 0 , (H_p-H_0C_0)z_p+(-I_h-H_0F_0)u+H_0X_0^{\mathsf T}\alpha=0, ( H p − H 0 C 0 ) z p + ( − I h − H 0 F 0 ) u + H 0 X 0 T α = 0 ,
( Y p − Y 0 C 0 ) z p − Y 0 F 0 u + Y 0 X 0 T α = 0 , (Y_p-Y_0C_0)z_p-Y_0F_0u+Y_0X_0^{\mathsf T}\alpha=0, ( Y p − Y 0 C 0 ) z p − Y 0 F 0 u + Y 0 X 0 T α = 0 ,
( η p − η 0 C 0 ) z p − η 0 F 0 u + η 0 X 0 T α = I k . (\eta_p-\eta_0C_0)z_p-\eta_0F_0u+\eta_0X_0^{\mathsf T}\alpha=I_k. ( η p − η 0 C 0 ) z p − η 0 F 0 u + η 0 X 0 T α = I k .
These are, respectively, the mass-coordinate identity, the right image constraints and the normalization after elimination. Thus its four row groups come, in order, from the remaining pairing equations, the definition of the mass coordinates, the right divisibility and degree conditions, and normalization. No perturbed polynomial is needed to form them. For instance,
( E 0 − D 0 C 0 ) α , β = β ~ μ ( ϕ α , ψ β ) − ∑ j = 1 N β ~ μ ( ϕ α , a j ) β ~ μ ( b j , ψ β ) , (E_0-D_0C_0)_{\alpha,\beta}=\widetilde{\beta}_{\mu}(\phi_{\alpha},\psi_{\beta})-\sum_{j=1}^{N}\widetilde{\beta}_{\mu}(\phi_{\alpha},a_j)\widetilde{\beta}_{\mu}(b_j,\psi_{\beta}), ( E 0 − D 0 C 0 ) α , β = β μ ( ϕ α , ψ β ) − j = 1 ∑ N β μ ( ϕ α , a j ) β μ ( b j , ψ β ) ,
which subtracts the contribution of the original polynomial block from the pairing of two added rational vectors.
Retain the ordered bases f 1 , … , f r f_1,\ldots,f_r f 1 , … , f r of X X X and g 1 , … , g r + k g_1,\ldots,g_{r+k} g 1 , … , g r + k of Y Y Y used to construct the image coordinates. The target coefficient matrix is
M ^ R , k : = [ [ β ^ ( f i , g j ) ] i ∈ { 1 , … , r } , j ∈ { 1 , … , r + k } [ λ ( g 1 ) ⋯ λ ( g r + k ) ] ] (14) \widehat{M}_{R,k}:=
\begin{bmatrix}
[\widehat{\beta}(f_i,g_j)]_{i\in\{1,\ldots,r\},\,j\in\{1,\ldots,r+k\}}\\
[\lambda(g_1)\ \cdots\ \lambda(g_{r+k})]
\end{bmatrix} \tag*{(14)} M R , k := [ [ β ( f i , g j ) ] i ∈ { 1 , … , r } , j ∈ { 1 , … , r + k } [ λ ( g 1 ) ⋯ λ ( g r + k )] ] (14)
The remaining coefficient matrix is the square matrix
T k : = [ E 0 − D 0 C 0 F p − D 0 F 0 D 0 X 0 T − X p T H p − H 0 C 0 − I h − H 0 F 0 H 0 X 0 T Y p − Y 0 C 0 − Y 0 F 0 Y 0 X 0 T η p − η 0 C 0 − η 0 F 0 η 0 X 0 T ] . (15) \mathbb{T}_k:=
\begin{bmatrix}
E_0-D_0C_0 & F_p-D_0F_0 & D_0X_0^{\mathsf T}-X_p^{\mathsf T}\\
H_p-H_0C_0 & -I_h-H_0F_0 & H_0X_0^{\mathsf T}\\
Y_p-Y_0C_0 & -Y_0F_0 & Y_0X_0^{\mathsf T}\\
\eta_p-\eta_0C_0 & -\eta_0F_0 & \eta_0X_0^{\mathsf T}
\end{bmatrix}. \tag*{(15)} T k := E 0 − D 0 C 0 H p − H 0 C 0 Y p − Y 0 C 0 η p − η 0 C 0 F p − D 0 F 0 − I h − H 0 F 0 − Y 0 F 0 − η 0 F 0 D 0 X 0 T − X p T H 0 X 0 T Y 0 X 0 T η 0 X 0 T . (15)
Its column sizes are s R s_R s R , h h h , c L c_L c L , and its row sizes are s L s_L s L , h h h , c R c_R c R , k k k . These give the same total:
s R + h + c L = s L + h + c R + k = N − r + s L + s R + h . s_R+h+c_L=s_L+h+c_R+k=N-r+s_L+s_R+h. s R + h + c L = s L + h + c R + k = N − r + s L + s R + h .
Define this order as t : = N − r + s L + s R + h t:=N-r+s_L+s_R+h t := N − r + s L + s R + h and set τ R , k : = det T k \tau_{R,k}:=\det\mathbb{T}_k τ R , k := det T k . The term tau determinant means the ordinary determinant of the equations left after elimination. The theorem below relates its nonvanishing to the prescribed orthogonality and normalization. The normalized Gram matrix (14) has order r + k r+k r + k , whereas T k \mathbb{T}_k T k has order t t t . Which one is smaller depends on the original spaces and the transformation.
The dimensions in this count are computed using polynomial classes in A = C [ x 1 , … , x M ] / I A=\mathbb{C}[x_1,\ldots,x_M]/I A = C [ x 1 , … , x M ] / I . For component degree bounds, one counts the independent allowed classes in each of the q q q row components and p p p column components; the relations in I I I identify dependent monomials. The rational factors then determine the extra dimensions s L , s R s_L,s_R s L , s R relative to the chosen original polynomial spaces, while h h h is the number of terms in the chosen mass representation on these finite spaces. Thus the number M M M of ambient variables and the support dimension dim A \dim A dim A enter through the polynomial-space dimensions. The rectangular size and these two dimensions alone do not fix the order: the component bounds, the support relations, the rational factors and the mass data also enter the count. Subsection 3.3 makes this count numerical: it specifies the ambient spaces by degrees, gives their exact determinant order, and calculates the reduction obtained from polynomial coefficient ranks.
The passage from the full equations to T k \mathbb{T}_k T k is block elimination with the known invertible block I N I_N I N , or equivalently a Schur-complement calculation. Its application here uses the image equations already constructed, the supplied rational moment tables, and the mass and normalization tables. The following theorem records the resulting determinant relation and reconstructs the normalized polynomial from the remaining coefficients. Since the eliminated block belongs to the original pairing, the calculation applies directly to the final bilateral problem without solving an intermediate transformed orthogonality problem.
Theorem 3.3 (Determinant criterion and normalized orthogonality). There is a nonzero factor c c c , determined by the ordered bases, constraint complements and block orders, such that
det M ^ R , k = c τ R , k . (16) \det\widehat{M}_{R,k}=c\,\tau_{R,k}. \tag*{(16)} det M R , k = c τ R , k . (16) For k > 0 k>0 k > 0 , the following conditions are equivalent:
a unique normalized orthogonal block exists
if and only if det M ^ R , k ≠ 0 \det\widehat{M}_{R,k}\ne0 det M R , k = 0 if and only if τ R , k ≠ 0 \tau_{R,k}\ne0 τ R , k = 0 .
The left-normalized problem has the same criterion for the determinant τ L , k \tau_{L,k} τ L , k of its transposed reduced construction. Consequently, for any prescribed collection Ω \Omega Ω of spaces and normalizations to which this construction applies,
the entire family is normal if and only if
τ ω ≠ 0 for every ω ∈ Ω . (17) \tau_{\omega}\ne0 \quad\text{for every }\omega\in\Omega. \tag*{(17)} τ ω = 0 for every ω ∈ Ω. (17) When τ R , k ≠ 0 \tau_{R,k}\ne0 τ R , k = 0 , put
V ∗ : = [ − C 0 − F 0 X 0 T I S R 0 0 ] , J k : = [ 0 ( t − k ) × k I k ] . V_\ast:=\begin{bmatrix}
-C_0 & -F_0 & X_0^{\mathsf T}\\
I_{SR} & 0 & 0
\end{bmatrix},
\qquad
J_k:=\begin{bmatrix}
0_{(t-k)\times k}\\
I_k
\end{bmatrix}. V ∗ := [ − C 0 I S R − F 0 0 X 0 T 0 ] , J k := [ 0 ( t − k ) × k I k ] . The solution is
a ^ = π R ( A + V ∗ T k − 1 J k ) . (18) \widehat{a}=\pi_R\left(\mathcal{A}+V_\ast\mathbb{T}_k^{-1}J_k\right). \tag*{(18)} a = π R ( A + V ∗ T k − 1 J k ) . (18) For k = 0 k=0 k = 0 , omit the last block row of (15) . The target square pairing is normal exactly when det T 0 ≠ 0 \det\mathbb{T}_0\ne0 det T 0 = 0 .
Proof. We first compare solutions, then compare determinants. The preceding substitution gives the reduced system
T k [ z p u α ] = J k . \mathbb{T}_k
\begin{bmatrix}
z_p\\
u\\
\alpha
\end{bmatrix}
=J_k. T k z p u α = J k .
Given a solution of the original normalized problem, its coordinates z z z determine u = H ν z u=H_\nu z u = H ν z . Orthogonality determines α \alpha α uniquely because Γ L \Gamma_L Γ L has independent rows. These coordinates solve the reduced system. Conversely, a solution of the reduced system determines z 0 z_0 z 0 by the displayed elimination formula; the four remaining equations then recover all the original conditions. In particular Γ R z = 0 \Gamma_Rz=0 Γ R z = 0 , so inverse multiplication gives a polynomial in Y Y Y with the required normalization.
The two systems therefore have the same solutions and the same homogeneous freedom. For k > 0 k>0 k > 0 , a normalized block is equivalent to a solution of
M ^ R , k C = [ 0 r × k I k ] . \widehat{M}_{R,k}C=
\begin{bmatrix}
0_{r\times k}\\
I_k
\end{bmatrix}. M R , k C = [ 0 r × k I k ] .
If the square coefficient matrix is invertible, this equation has exactly one solution. Conversely, suppose a solution is unique. A nonzero vector in its homogeneous kernel could be added to any one of the k k k columns of that solution, giving another solution. Thus the homogeneous kernel is zero, and the matrix is invertible. The correspondence just proved transfers the same conclusion to T k \mathbb{T}_k T k . This proves the uniqueness criterion. For k = 0 k=0 k = 0 , the same correspondence identifies the right radicals of the two square pairings and proves the normality criterion. Reconstructing z z z from the reduced unknowns is exactly multiplication by V ∗ V_\ast V ∗ ; applying π R \pi_R π R gives (3.10).
To compare the determinants, use the coordinate matrices Z L , Z R Z_L,Z_R Z L , Z R constructed from f i S L , S R g j f_iS_L,S_Rg_j f i S L , S R g j . Their columns are bases of the respective constraint kernels. Choose matrices Q L , Q R Q_L,Q_R Q L , Q R with Γ L Q L = I c L \Gamma_LQ_L=I_{c_L} Γ L Q L = I c L and Γ R Q R = I c R \Gamma_RQ_R=I_{c_R} Γ R Q R = I c R . Then
U L : = [ Z L Q L ] , U R : = [ Z R Q R ] U_L := [Z_L\ Q_L], \qquad U_R := [Z_R\ Q_R] U L := [ Z L Q L ] , U R := [ Z R Q R ]
are invertible changes of coordinates: they separate the admissible directions from the directions detected by the constraints. The coefficient matrix before eliminating z 0 z_0 z 0 is
Q : = [ G r a t F − Γ L T H ν − I h 0 Γ R 0 0 η 0 0 ] Q := \begin{bmatrix}
G_{\mathrm{rat}} & F & -\Gamma_L^{\mathsf T}\\
H_\nu& -I_h & 0\\
\Gamma_R & 0 & 0\\
\eta& 0 & 0
\end{bmatrix} Q := G rat H ν Γ R η F − I h 0 0 − Γ L T 0 0 0
Elimination with the leading I N I_N I N block uses determinant-preserving row operations and leaves T k T_k T k . Thus det Q = τ R , k \det Q=\tau_{R,k} det Q = τ R , k . Alternatively, eliminate the mass variables using − I h -I_h − I h . The pairing block becomes G r a t + F H ν = G G_{\mathrm{rat}}+FH_\nu=G G rat + F H ν = G . Change its row and column coordinates by U L T U_L^{\mathsf T} U L T and U R U_R U R . In these coordinates, Γ L U L = [ 0 I c L ] \Gamma_LU_L=[0\ I_{c_L}] Γ L U L = [ 0 I c L ] and Γ R U R = [ 0 I c R ] \Gamma_RU_R=[0\ I_{c_R}] Γ R U R = [ 0 I c R ] . After the mass elimination the remaining matrix is
[ Z L T G Z R Z L T G Q R 0 Q L T G Z R Q L T G Q R − I c L 0 I c R 0 η Z R η Q R 0 ] \begin{bmatrix}
Z_L^{\mathsf T}GZ_R & Z_L^{\mathsf T}GQ_R & 0\\
Q_L^{\mathsf T}GZ_R & Q_L^{\mathsf T}GQ_R & -I_{c_L}\\
0 & I_{c_R} & 0\\
\eta Z_R & \eta Q_R & 0
\end{bmatrix} Z L T G Z R Q L T G Z R 0 η Z R Z L T G Q R Q L T G Q R I c R η Q R 0 − I c L 0 0
Expand first along the multiplier columns, whose only nonzero block is − I c L -I_{c_L} − I c L , removing the complementary left rows. Then expand along the right-constraint rows, whose only nonzero block is I c R I_{c_R} I c R , removing the complementary right columns. The remaining matrix is
[ Z L T G Z R η Z R ] = M ^ R , k . \begin{bmatrix}
Z_L^{\mathsf T}GZ_R\\
\eta Z_R
\end{bmatrix}
=\widehat{M}_{R,k}. [ Z L T G Z R η Z R ] = M R , k .
We can record the sign as well. Moving the h h h mass rows and columns to the end of Q Q Q gives the sign ( − 1 ) h ( c R + k + c L ) (-1)^{h(c_R+k+c_L)} ( − 1 ) h ( c R + k + c L ) ; eliminating their block − I h -I_h − I h gives another factor ( − 1 ) h (-1)^h ( − 1 ) h . After the changes of coordinates, reorder the remaining row groups from ( r , c L , c R , k ) (r,c_L,c_R,k) ( r , c L , c R , k ) to ( r , k , c R , c L ) (r,k,c_R,c_L) ( r , k , c R , c L ) . This contributes ( − 1 ) k ( c L + c R ) + c L c R (-1)^{k(c_L+c_R)+c_Lc_R} ( − 1 ) k ( c L + c R ) + c L c R . The final two diagonal blocks are I c R I_{c_R} I c R and − I c L -I_{c_L} − I c L , with determinant ( − 1 ) c L (-1)^{c_L} ( − 1 ) c L . Consequently, for precisely the block orders displayed above,
c = ( − 1 ) ϵ det U L det U R , c=(-1)^\epsilon\det U_L\det U_R, c = ( − 1 ) ϵ det U L det U R ,
ϵ : = h ( 1 + c L + c R + k ) + k ( c L + c R ) + c L c R + c L . \epsilon:=h(1+c_L+c_R+k)+k(c_L+c_R)+c_Lc_R+c_L. ϵ := h ( 1 + c L + c R + k ) + k ( c L + c R ) + c L c R + c L .
Both coordinate matrices are invertible, so c ≠ 0 c\ne0 c = 0 . This proves (16) , including its determinant normalization. Transposing exchanges the two arguments and proves the left criterion. Applying the criterion at each prescribed index proves (17) . □
By (16) , nonvanishing is independent of bases and complete constraints.
The inverse in (18) can be replaced by scalar determinants, one for each polynomial entry. The following corollary records the required rows and columns, including the left-normalized problem. Its last assertion also verifies polynomial divisibility of the determinant numerators, so that the formulas produce polynomials with the prescribed bounds.
Corollary 3.4. (Bordered determinants for both normalized blocks.) Put W R : = π R ( Q + V ∗ ) \mathcal{W}_R:=\pi_R(\mathcal{Q}^+V_*) W R := π R ( Q + V ∗ ) . For the left problem dim X = r + k \dim X=r+k dim X = r + k , dim Y = r \dim Y=r dim Y = r , prescribe λ L : X → C k \lambda_L:X\to\mathbb{C}^k λ L : X → C k , set ζ = λ π L \zeta=\lambda\pi_L ζ = λ π L and take T k L \mathsf{T}_k^L T k L from (3.7) by
( C 0 , D 0 , E 0 ) ↦ ( D 0 T , C 0 T , E 0 T ) , ( F 0 , F p , H 0 , H p ) ↦ ( H 0 T , H p T , F 0 T , F p T ) , (C_0,D_0,E_0)\mapsto(D_0^{\mathsf T},C_0^{\mathsf T},E_0^{\mathsf T}),\qquad
(F_0,F_p,H_0,H_p)\mapsto(H_0^{\mathsf T},H_p^{\mathsf T},F_0^{\mathsf T},F_p^{\mathsf T}), ( C 0 , D 0 , E 0 ) ↦ ( D 0 T , C 0 T , E 0 T ) , ( F 0 , F p , H 0 , H p ) ↦ ( H 0 T , H p T , F 0 T , F p T ) , ( X 0 , X p , Y 0 , Y p , η ) ↦ ( Y 0 , Y p , X 0 , X p , ζ ) . (19) (X_0,X_p,Y_0,Y_p,\eta)\mapsto(Y_0,Y_p,X_0,X_p,\zeta). \tag*{(19)} ( X 0 , X p , Y 0 , Y p , η ) ↦ ( Y 0 , Y p , X 0 , X p , ζ ) . (19) Set V ∗ L : = [ − D 0 T − H 0 T Y 0 T ; I S L 0 ] V_*^L:=[-D_0^{\mathsf T}-H_0^{\mathsf T}Y_0^{\mathsf T};I_{S_L}\ 0] V ∗ L := [ − D 0 T − H 0 T Y 0 T ; I S L 0 ] and W L : = π L ( ( V ∗ L ) T B + ) \mathcal{W}_L:=\pi_L((V_*^L)^{\mathsf T}\mathcal{B}^+) W L := π L (( V ∗ L ) T B + ) . Writing T k R : = T k \mathsf{T}_k^R:=\mathsf{T}_k T k R := T k , τ σ , k : = det T k σ \tau_{\sigma,k}:=\det\mathsf{T}_k^\sigma τ σ , k := det T k σ , for a row u u u of length t t t and j ∈ { 1 , … , k } j\in\{1,\ldots,k\} j ∈ { 1 , … , k } , define the scalar border
B σ ( u , j ) : = − τ σ , k − 1 det [ T k σ e t − k + j u 0 ] . \mathcal{B}_\sigma(u,j):=-\tau_{\sigma,k}^{-1}\det\begin{bmatrix}
\mathsf{T}_k^\sigma& e_{t-k+j}\\
u & 0
\end{bmatrix}. B σ ( u , j ) := − τ σ , k − 1 det [ T k σ u e t − k + j 0 ] .
Then the right and left normalized blocks are
a ^ i , j ( P ) = B R ( ( W R ( P ) ) i , : , j ) , b ^ i , j ( Q ) = B L ( ( W L ( Q ) ) : , j , i T ) . (20) \widehat{a}_{i,j}(P)=\mathcal{B}_{R}\left((\mathcal{W}_{R}(P))_{i,:,j}\right),\qquad\widehat{b}_{i,j}(Q)=\mathcal{B}_{L}\left((\mathcal{W}_{L}(Q))^{\mathsf{T}}_{:,j,i}\right). \tag*{(20)} a i , j ( P ) = B R ( ( W R ( P ) ) i , : , j ) , b i , j ( Q ) = B L ( ( W L ( Q ) ) : , j , i T ) . (20)
Here i , : i,: i , : selects row i i i and : , j :,j : , j selects column j j j ; a ^ \widehat{a} a and b ^ \widehat{b} b have sizes p × k p\times k p × k and k × q k\times q k × q . Their denominators are nonzero exactly when the respective normalized problems have unique solutions.
Divisibility is also explicit inside the borders. Replace W R \mathcal{W}_{R} W R by a R A † + V ∗ a_{R}\mathcal{A}^{\dagger}+V_{*} a R A † + V ∗ and W L \mathcal{W}_{L} W L by a L ( V ∗ L ) T B + a_{L}(V_{*}^{L})^{\mathsf{T}}\mathcal{B}^{+} a L ( V ∗ L ) T B + in (3.12), and multiply the results by τ R , k \tau_{R,k} τ R , k and τ L , k \tau_{L,k} τ L , k , respectively. The resulting polynomial matrices satisfy
N R = τ R , k P R a ^ , N L = τ L , k b ^ P L . (21) N_{R}=\tau_{R,k}P_{R}\widehat{a},\qquad N_{L}=\tau_{L,k}\widehat{b}P_{L}. \tag*{(21)} N R = τ R , k P R a , N L = τ L , k b P L . (21)
Thus their columns and rows lie in P R Y P_{R}Y P R Y and X P L XP_{L} X P L .
Proof. For an invertible square matrix T T T , a column v v v and a row u u u , subtracting u T − 1 uT^{-1} u T − 1 times the upper block row from the last row gives
det [ T v u 0 ] = − det ( T ) u T − 1 v . \det\begin{bmatrix}T&v\\u&0\end{bmatrix}=-\det(T)\,uT^{-1}v. det [ T u v 0 ] = − det ( T ) u T − 1 v .
The j j j th column of J k J_{k} J k is e t − k + j e_{t-k+j} e t − k + j . Apply this identity with T = T k T=T_{k} T = T k , this column as v v v , and the i i i th row of W R ( P ) \mathcal{W}_{R}(P) W R ( P ) as u u u . The resulting scalar is exactly the ( i , j ) (i,j) ( i , j ) entry of (3.10), which proves the right formula. Exchanging rows and columns transposes the pairing and all its data. The substitutions in (3.11) list these exchanges explicitly, so the same calculation proves the left formula.
For the divisibility assertion, apply the border identity before recovering the polynomial candidate, and multiply by the common scalar denominator. The numerator is then N R = τ R , k a R S R a ^ N_{R}=\tau_{R,k}a_{R}S_{R}\widehat{a} N R = τ R , k a R S R a on the right and N L = τ L , k a L b ^ S L N_{L}=\tau_{L,k}a_{L}\widehat{b}S_{L} N L = τ L , k a L b S L on the left. Since a R S R = P R a_{R}S_{R}=P_{R} a R S R = P R and a L S L = P L a_{L}S_{L}=P_{L} a L S L = P L , these are (3.13). Membership of a ^ \widehat{a} a in Y Y Y and of b ^ \widehat{b} b in X X X also proves the stated quotient bounds. □
The kernel from the same reduced matrix
When the target pairing is square, write T : = T 0 T:=T_{0} T := T 0 and put
U ∗ : = [ − D 0 I S L − H 0 0 − Y 0 0 ] , J : = [ I N 0 0 0 ] . U_{*}:=\begin{bmatrix}-D_{0}&I_{S_{L}}\\-H_{0}&0\\-Y_{0}&0\end{bmatrix},\qquad
J:=\begin{bmatrix}I_{N}&0\\0&0\end{bmatrix}. U ∗ := − D 0 − H 0 − Y 0 I S L 0 0 , J := [ I N 0 0 0 ] .
Here J J J has size ( N + s R ) × ( N + s L ) (N+s_{R})\times(N+s_{L}) ( N + s R ) × ( N + s L ) . Let K E ( P , Q ) : = A ( P ) B ( Q ) K_{E}(P,Q):=\mathcal{A}(P)\mathcal{B}(Q) K E ( P , Q ) := A ( P ) B ( Q ) , and let K ^ \widehat{K} K be the kernel of the target pairing, with its column factor at P P P and its row factor at Q Q Q . This is the dual-basis form of (2.7). The reproduction identities and their meaning on a nonreduced support are given in Subsubsection 2.1.1.
The separation rank from Subsubsection 2.1.1 equals the rank of the coefficient matrix in independent ambient bases. The present theorem constructs that matrix.
For the kernel we solve the same orthogonality equations with an arbitrary right-hand side, instead of a fixed normalization. The resulting inverse pairing reproduces all members of the target space simultaneously. This is why the same reduced matrix suffices for both the polynomial and kernel formulas.
Theorem 3.5 (Simultaneous kernel formula). If the target pairing is normal, then
F ( P , Q ) : = S R ( P ) K ^ ( P , Q ) S L ( Q ) = A + ( P ) [ J + V ∗ T − 1 U ∗ ] B + ( Q ) , (22) \mathcal{F}(P,Q):=S_{R}(P)\widehat{K}(P,Q)S_{L}(Q)=\mathcal{A}^{+}(P)\left[J+V_{*}T^{-1}U_{*}\right]\mathcal{B}^{+}(Q), \tag*{(22)} F ( P , Q ) := S R ( P ) K ( P , Q ) S L ( Q ) = A + ( P ) [ J + V ∗ T − 1 U ∗ ] B + ( Q ) , (22) F − K E = A + V ∗ T − 1 U ∗ B + . (23) \mathcal{F}-K_{E}=\mathcal{A}^{+}V_{*}T^{-1}U_{*}\mathcal{B}^{+}. \tag*{(23)} F − K E = A + V ∗ T − 1 U ∗ B + . (23) The correction has tensor rank rank ( V ∗ T − 1 U ∗ ) ≤ t \operatorname{rank}(V_{*}T^{-1}U_{*})\leq t rank ( V ∗ T − 1 U ∗ ) ≤ t . The full tensor F \mathcal{F} F has rank r r r and factor spaces S R Y S_{R}Y S R Y and X S L XS_{L} X S L .
Proof. Choose coordinate inclusion matrices Z L , Z R Z_{L},Z_{R} Z L , Z R for bases of X S L , S R Y XS_{L},S_{R}Y X S L , S R Y in the enlarged bases. Their pairing matrix is Z L T G Z R Z_{L}^{\mathsf{T}}GZ_{R} Z L T G Z R , which is invertible by target normality. The coefficient matrix of S R ( P ) K ^ ( P , Q ) S L ( Q ) S_{R}(P)\widehat{K}(P,Q)S_{L}(Q) S R ( P ) K ( P , Q ) S L ( Q ) in the enlarged bases is therefore
F : = Z R ( Z L T G Z R ) − 1 Z L T . \mathcal{F}:=Z_{R}(Z_{L}^{\mathsf{T}}GZ_{R})^{-1}Z_{L}^{\mathsf{T}}. F := Z R ( Z L T G Z R ) − 1 Z L T . Indeed, write an admissible coefficient column as z = Z R w z=Z_{R}w z = Z R w . Its pairing values against the target row basis form Z L T G Z R w Z_{L}^{\mathsf{T}}GZ_{R}w Z L T G Z R w . To reproduce prescribed values, multiply by the inverse of this square Gram
matrix and then by Z R Z_R Z R . The factor Z L T Z_L^{\mathsf T} Z L T converts an ambient right-hand side to the prescribed row values. This gives the displayed coefficient matrix with the basis inclusions in the stated order.
We calculate G v \mathcal{G}v G v for an arbitrary column v = [ v 0 T v p T ] T v = [v_0^{\mathsf T}\ v_p^{\mathsf T}]^{\mathsf T} v = [ v 0 T v p T ] T . The vector z = G v z = \mathcal{G}v z = G v is characterized by
G z − Γ L T α = v , Γ R z = 0. Gz-\Gamma_L^{\mathsf T}\alpha=v,\qquad\Gamma_Rz=0. G z − Γ L T α = v , Γ R z = 0.
The second equation gives z = Z R w z=Z_Rw z = Z R w . Multiplying the first by Z L T Z_L^{\mathsf T} Z L T yields ( Z L T G Z R ) w = Z L T v (Z_L^{\mathsf T}GZ_R)w=Z_L^{\mathsf T}v ( Z L T G Z R ) w = Z L T v , whose unique solution gives z = G v z=\mathcal{G}v z = G v . Conversely, for this solution, Z L T ( G z − v ) = 0 Z_L^{\mathsf T}(Gz-v)=0 Z L T ( G z − v ) = 0 . The vectors annihilated by Z L T Z_L^{\mathsf T} Z L T are exactly the columns of Γ L T \Gamma_L^{\mathsf T} Γ L T ; hence there is a unique α \alpha α with G z − v = Γ L T α Gz-v=\Gamma_L^{\mathsf T}\alpha G z − v = Γ L T α . This proves both implications of the characterization.
Introduce u = H ν z u=H_\nu z u = H ν z and eliminate the original polynomial coefficients as before. The first block now gives z 0 = v 0 − C 0 z p − F 0 u + X 0 T α z_0=v_0-C_0z_p-F_0u+X_0^{\mathsf T}\alpha z 0 = v 0 − C 0 z p − F 0 u + X 0 T α . The remaining pairing, mass and image equations have right-hand sides v p − D 0 v 0 v_p-D_0v_0 v p − D 0 v 0 , − H 0 v 0 -H_0v_0 − H 0 v 0 and − Y 0 v 0 -Y_0v_0 − Y 0 v 0 , respectively. Stacking these gives
[ z p u α ] = [ v p − D 0 v 0 − H 0 v 0 − Y 0 v 0 ] = U ∗ v . \begin{bmatrix}z_p\\u\\\alpha\end{bmatrix}
=
\begin{bmatrix}v_p-D_0v_0\\-H_0v_0\\-Y_0v_0\end{bmatrix}
=U_*v. z p u α = v p − D 0 v 0 − H 0 v 0 − Y 0 v 0 = U ∗ v .
Here U ∗ U_* U ∗ has size t × ( N + s L ) t\times(N+s_L) t × ( N + s L ) and V ∗ V_* V ∗ has size ( N + s R ) × t (N+s_R)\times t ( N + s R ) × t , so their product through T − 1 T^{-1} T − 1 has the same size as G \mathcal{G} G . The original contribution is J v = [ v 0 T 0 ] T Jv=[v_0^{\mathsf T}\ 0]^{\mathsf T} J v = [ v 0 T 0 ] T , while the remaining contribution is V ∗ T − 1 U ∗ v V_*T^{-1}U_*v V ∗ T − 1 U ∗ v . Since this holds for every v v v , it proves G = J + V ∗ T − 1 U ∗ \mathcal{G}=J+V_*T^{-1}U_* G = J + V ∗ T − 1 U ∗ . Multiplication by the two ambient basis matrices proves (3.14) and (3.15).
The ambient basis functions are linearly independent over C \mathbb{C} C . Consequently the smallest number of separated products equals the rank of their coefficient matrix. This proves the correction-rank assertion. In the expression defining G \mathcal{G} G , the middle matrix is invertible and both inclusion matrices have rank r r r . Thus G \mathcal{G} G has rank r r r , with column and row spaces exactly those of Z R Z_R Z R and Z L T Z_L^{\mathsf T} Z L T , respectively. □
Set τ 0 : = det T \tau_0:=\det T τ 0 := det T , U ( P ) : = A + ( P ) V ∗ \mathcal{U}(P):=\mathcal{A}^{+}(P)V_* U ( P ) := A + ( P ) V ∗ and V ( Q ) : = U ∗ B + ( Q ) \mathcal{V}(Q):=U_*\mathcal{B}^{+}(Q) V ( Q ) := U ∗ B + ( Q ) . For every matrix entry,
F i , j ( P , Q ) = det [ T V : , j ( Q ) − U i , : ( P ) ( K E ) i , j ( P , Q ) ] τ 0 . (24) \mathcal{F}_{i,j}(P,Q)=
\frac{\det\begin{bmatrix}
T & \mathcal{V}_{:,j}(Q)\\
-\mathcal{U}_{i,:}(P) & (K_E)_{i,j}(P,Q)
\end{bmatrix}}{\tau_0}. \tag*{(24)} F i , j ( P , Q ) = τ 0 det [ T − U i , : ( P ) V : , j ( Q ) ( K E ) i , j ( P , Q ) ] . (24)
The numerator determinant has order t + 1 t+1 t + 1 . Add U i , : T − 1 \mathcal{U}_{i,:}T^{-1} U i , : T − 1 times its upper block row to its last row. The lower-left block becomes zero, and the determinant is
det T ( ( K E ) i , j + U i , : T − 1 V : , j ) . \det T\left((K_E)_{i,j}+\mathcal{U}_{i,:}T^{-1}\mathcal{V}_{:,j}\right). det T ( ( K E ) i , j + U i , : T − 1 V : , j ) .
The minus sign in the lower-left block therefore produces the plus sign in (3.14). Replacing the lower-right entry by zero gives ( F − K E ) i , j (\mathcal{F}-K_E)_{i,j} ( F − K E ) i , j . An entrywise determinant for K ^ \widehat K K itself is obtained in (3.16) by replacing U ↦ S R − 1 U \mathcal{U}\mapsto S_R^{-1}\mathcal{U} U ↦ S R − 1 U , V ↦ V S L − 1 \mathcal{V}\mapsto\mathcal{V}S_L^{-1} V ↦ V S L − 1 and K E ↦ S R − 1 K E S L − 1 K_E\mapsto S_R^{-1}K_ES_L^{-1} K E ↦ S R − 1 K E S L − 1 . Theorem 3.5 places the factor spaces in S R Y S_RY S R Y and X S L XS_L X S L ; multiplying by the inverse factors therefore cancels all poles and gives a tensor in Y ⊗ C X Y\otimes_{\mathbb C}X Y ⊗ C X .
Empty matrices are permitted; the determinant of the zero-order square matrix is one.
A complete bilateral calculation for mixed-type orthogonality on a curve
We now construct every input block and recover the transformed polynomial on the curve Λ = C [ x , y ] / ( x 2 + y 2 − 1 ) \Lambda=\mathbb{C}[x,y]/(x^2+y^2-1) Λ = C [ x , y ] / ( x 2 + y 2 − 1 ) . Rows have two components and columns have three. Take
L 1 = [ x 0 0 1 ] , L 0 = [ 1 y 1 1 ] , R 1 = diag ( x , 1 , 1 ) , R 0 = [ 1 1 0 0 1 y 1 0 1 ] . L_1=\begin{bmatrix}x&0\\0&1\end{bmatrix},\qquad
L_0=\begin{bmatrix}1&y\\1&1\end{bmatrix},\qquad
R_1=\operatorname{diag}(x,1,1),\qquad
R_0=\begin{bmatrix}1&1&0\\0&1&y\\1&0&1\end{bmatrix}. L 1 = [ x 0 0 1 ] , L 0 = [ 1 1 y 1 ] , R 1 = diag ( x , 1 , 1 ) , R 0 = 1 0 1 1 1 0 0 y 1 .
Their determinants are x x x , 1 − y 1-y 1 − y , x x x , 1 + y 1+y 1 + y , respectively, all nonzero in this integral algebra. The products do not commute: ( L 1 L 0 − L 0 L 1 ) 1 , 2 = ( x − 1 ) y (L_1L_0-L_0L_1)_{1,2}=(x-1)y ( L 1 L 0 − L 0 L 1 ) 1 , 2 = ( x − 1 ) y and ( R 1 R 0 − R 0 R 1 ) 1 , 2 = x − 1 (R_1R_0-R_0R_1)_{1,2}=x-1 ( R 1 R 0 − R 0 R 1 ) 1 , 2 = x − 1 . Keeping the specified order gives
S L = [ x − 1 y / x 1 1 ] , S R = [ x − 1 1 0 0 1 y x − 1 0 1 ] . (25) S_L=\begin{bmatrix}x^{-1}&y/x\\1&1\end{bmatrix},\qquad
S_R=\begin{bmatrix}x^{-1}&1&0\\0&1&y\\x^{-1}&0&1\end{bmatrix}. \tag*{(25)} S L = [ x − 1 1 y / x 1 ] , S R = x − 1 0 x − 1 1 1 0 0 y 1 . (25)
For the original pairing use the two points P ± = ( ± 1 , 0 ) P_{\pm}=(\pm1,0) P ± = ( ± 1 , 0 ) and the rectangular matrix W = [ 1 0 0 0 1 1 ] W=\begin{bmatrix}1&0&0\\0&1&1\end{bmatrix} W = [ 1 0 0 1 0 1 ] :
β μ ( f , g ) : = 1 2 ∑ P ∈ { P + , P − } f ( P ) W g ( P ) . \beta_{\mu}(f,g):=\frac{1}{2}\sum_{P\in\{P_{+},P_{-}\}}f(P)Wg(P). β μ ( f , g ) := 2 1 P ∈ { P + , P − } ∑ f ( P ) W g ( P ) .
Evaluation at the same points defines its rational extension, since x x x does not vanish there. Choose the polynomial spaces with ordered bases
X = span { [ 1 0 ] , [ x 0 ] , [ 0 1 ] } , Y = span { [ 1 0 0 ] T , [ x 0 0 ] T , [ 0 1 0 ] T , [ 0 0 1 ] T } . X=\operatorname{span}\{[1\ 0],[x\ 0],[0\ 1]\},\qquad Y=\operatorname{span}\{[1\ 0\ 0]^{\mathsf T},[x\ 0\ 0]^{\mathsf T},[0\ 1\ 0]^{\mathsf T},[0\ 0\ 1]^{\mathsf T}\}. X = span {[ 1 0 ] , [ x 0 ] , [ 0 1 ]} , Y = span {[ 1 0 0 ] T , [ x 0 0 ] T , [ 0 1 0 ] T , [ 0 0 1 ] T } .
The normalization λ ( g ) = g 3 \lambda(g)=g_{3} λ ( g ) = g 3 prescribes the constant third component to be one. For the original dual pair take E = X E=X E = X and H H H spanned by the first three displayed columns of Y Y Y . The averages of x x x and x 2 x^{2} x 2 at P ± P_{\pm} P ± are zero and one; direct multiplication by W W W therefore gives G 0 = I 3 G_{0}=I_{3} G 0 = I 3 . The original pairing has global bilinear rank four: evaluation at P + P_{+} P + and P − P_{-} P − can be prescribed independently, and each copy of W W W has rank two. Its normality in this calculation is the nondegeneracy of the specified three-dimensional pair E × H E\times H E × H . The target is likewise the single mixed-type orthogonality problem on X × Y X\times Y X × Y with the displayed normalization.
At Q = ( 0 , 1 ) Q=(0,1) Q = ( 0 , 1 ) the tangent derivative is ∂ x \partial_{x} ∂ x : it kills x 2 + y 2 − 1 x^{2}+y^{2}-1 x 2 + y 2 − 1 at Q Q Q and is well defined on classes. For a parameter γ ∈ C \gamma\in\mathbb{C} γ ∈ C , choose the homogeneous mass
ν γ ( f , g ) = γ ∂ x ( f 1 g 1 ) ( Q ) , β ^ ( f , g ) = β ~ μ ( f S L , S R g ) + ν γ ( f , g ) . \nu_{\gamma}(f,g)=\gamma\partial_{x}(f_{1}g_{1})(Q),\qquad\widehat{\beta}(f,g)=\widetilde{\beta}_{\mu}(fS_{L},S_{R}g)+\nu_{\gamma}(f,g). ν γ ( f , g ) = γ ∂ x ( f 1 g 1 ) ( Q ) , β ( f , g ) = β μ ( f S L , S R g ) + ν γ ( f , g ) .
The first components of f L 1 fL_{1} f L 1 and R 1 g R_{1}g R 1 g are x f 1 xf_{1} x f 1 and x g 1 xg_{1} x g 1 . Their product has a double zero at Q Q Q , so L 1 ν γ R 1 = 0 L_{1}\nu_{\gamma}R_{1}=0 L 1 ν γ R 1 = 0 . For γ ≠ 0 \gamma\ne0 γ = 0 , either unilateral multiplication instead leaves γ f 1 ( Q ) g 1 ( Q ) \gamma f_{1}(Q)g_{1}(Q) γ f 1 ( Q ) g 1 ( Q ) , which is nonzero for constant first components. Thus this mass is allowed by the bilateral equation and is lost if unilateral vanishing is imposed separately. It cannot even be a sum of balanced unilateral masses: the ( 1 , 1 ) (1,1) ( 1 , 1 ) entry of either unilateral term is annihilated by x x x , whereas x ( ν γ ) 1 , 1 = γ δ Q ≠ 0 x(\nu_{\gamma})_{1,1}=\gamma\delta_{Q}\ne0 x ( ν γ ) 1 , 1 = γ δ Q = 0 . This is the obstruction of Theorem 2.4 in the present curve calculation.
The finite rank of the whole transformed pairing can also be read from these data. Both rational factors are invertible at P ± P_{\pm} P ± , so the rational part still has rank four. In the coordinates given by the value and the tangent derivative of the first component at Q Q Q , the mass has matrix
[ 0 γ γ 0 ] . \begin{bmatrix}0&\gamma\\\gamma&0\end{bmatrix}. [ 0 γ γ 0 ] .
The polynomial classes x 2 ( 1 + x ) / 2 x^{2}(1+x)/2 x 2 ( 1 + x ) /2 and x 2 ( 1 − x ) / 2 x^{2}(1-x)/2 x 2 ( 1 − x ) /2 independently prescribe the two values at P ± P_{\pm} P ± and have zero value and tangent derivative at Q Q Q . Conversely, y y y and x y xy x y vanish at P ± P_{\pm} P ± ; their respective value–derivative pairs at Q Q Q are [ 1 0 ] [1\ 0] [ 1 0 ] and [ 0 1 ] [0\ 1] [ 0 1 ] . Hence the two-point evaluation data and the two jet coordinates are independent. For γ ≠ 0 \gamma\ne0 γ = 0 the global rank of β ^ \widehat{\beta} β is therefore 4 + 2 = 6 4+2=6 4 + 2 = 6 . Normality of the finite spaces verified below is compatible with this rank bound; an exhaustive sequence of normal square pairings would require unbounded rank.
Rational vectors and image conditions. Append to the original bases the rows and columns
ϕ 1 = [ x − 1 y / x ] , ϕ 2 = [ 0 y ] , ψ 1 = [ x − 1 0 x − 1 ] T , ψ 2 = [ 0 0 1 ] T , ψ 3 = [ 0 y 0 ] T . \phi_{1}=[x^{-1}\ y/x],\qquad\phi_{2}=[0\ y],\qquad\psi_{1}=[x^{-1}\ 0\ x^{-1}]^{\mathsf T},\qquad\psi_{2}=[0\ 0\ 1]^{\mathsf T},\qquad\psi_{3}=[0\ y\ 0]^{\mathsf T}. ϕ 1 = [ x − 1 y / x ] , ϕ 2 = [ 0 y ] , ψ 1 = [ x − 1 0 x − 1 ] T , ψ 2 = [ 0 0 1 ] T , ψ 3 = [ 0 y 0 ] T .
These are independent additions in the coordinate algebra, giving N = r = 3 N=r=3 N = r = 3 , s L = 2 s_{L}=2 s L = 2 , s R = 3 s_{R}=3 s R = 3 and k = 1 k=1 k = 1 . Multiplication of the three basis rows of X X X by S L S_{L} S L gives ϕ 1 , b 1 + ϕ 2 , b 1 + b 3 \phi_{1},b_{1}+\phi_{2},b_{1}+b_{3} ϕ 1 , b 1 + ϕ 2 , b 1 + b 3 . Multiplication of the four basis columns of Y Y Y by S R S_{R} S R gives ψ 1 , a 1 + ψ 2 , a 1 + a 3 , ψ 2 + ψ 3 \psi_{1},a_{1}+\psi_{2},a_{1}+a_{3},\psi_{2}+\psi_{3} ψ 1 , a 1 + ψ 2 , a 1 + a 3 , ψ 2 + ψ 3 . Consequently the complete image constraints are
Γ L = [ 0 1 0 0 0 1 0 − 1 0 − 1 ] , Γ R = [ 0 1 0 0 0 0 1 0 − 1 0 − 1 1 ] . \Gamma_{L}=\begin{bmatrix}0&1&0&0&0\\1&0&-1&0&-1\end{bmatrix},\qquad\Gamma_{R}=\begin{bmatrix}0&1&0&0&0&0\\1&0&-1&0&-1&1\end{bmatrix}. Γ L = [ 0 1 1 0 0 − 1 0 0 0 − 1 ] , Γ R = [ 0 1 1 0 0 − 1 0 0 0 − 1 0 1 ] .
For example, left coordinates t t t must satisfy t 2 = 0 t_{2}=0 t 2 = 0 and t 1 = t 3 + t 5 t_{1}=t_{3}+t_{5} t 1 = t 3 + t 5 ; the resulting vector is t 4 ϕ 1 + t 5 ( b 1 + ϕ 2 ) + t 3 ( b 1 + b 3 ) t_{4}\phi_{1}+t_{5}(b_{1}+\phi_{2})+t_{3}(b_{1}+b_{3}) t 4 ϕ 1 + t 5 ( b 1 + ϕ 2 ) + t 3 ( b 1 + b 3 ) . Its polynomial preimage is [ t 4 + x t 5 t 3 ] [t_{4}+xt_{5}\ t_{3}] [ t 4 + x t 5 t 3 ] . On the right, coordinates v v v satisfy v 2 = 0 v_{2}=0 v 2 = 0 and v 1 = v 3 + v 5 − v 6 v_{1}=v_{3}+v_{5}-v_{6} v 1 = v 3 + v 5 − v 6 . This gives the recovery maps
π L ( t ) = [ t 4 + x t 5 t 3 ] , π R ( v ) = [ v 4 + x ( v 5 − v 6 ) v 3 v 6 ] T . \pi_{L}(t)=[t_{4}+xt_{5}\ t_{3}],\qquad\pi_{R}(v)=[v_{4}+x(v_{5}-v_{6})\ v_{3}\ v_{6}]^{\mathsf T}. π L ( t ) = [ t 4 + x t 5 t 3 ] , π R ( v ) = [ v 4 + x ( v 5 − v 6 ) v 3 v 6 ] T .
They enforce the stated component bounds as well as polynomiality. Using the same expressions off the two kernels specifies their linear extensions and gives η = [ 0 0 0 0 0 1 ] \eta=[0\ 0\ 0\ 0\ 0\ 1] η = [ 0 0 0 0 0 1 ] .
Moments, mass and the reduced matrix. All rational moments are averages at P ± P_{\pm} P ± . Since y = 0 y=0 y = 0 and x − 1 = x x^{-1}=x x − 1 = x there, their blocks are
C 0 = [ 0 0 0 1 0 0 0 1 0 ] , D 0 = [ 0 1 0 0 0 0 0 0 0 ] , E 0 = [ 1 0 0 0 0 0 0 0 0 ] . C_0=\begin{bmatrix}0&0&0\\1&0&0\\0&1&0\end{bmatrix},\qquad D_0=\begin{bmatrix}0&1&0\\0&0&0\\0&0&0\end{bmatrix},\qquad E_0=\begin{bmatrix}1&0&0\\0&0&0\\0&0&0\end{bmatrix}. C 0 = 0 1 0 0 0 1 0 0 0 , D 0 = 0 0 0 1 0 0 0 0 0 , E 0 = 1 0 0 0 0 0 0 0 0 .
For instance, the first column of C 0 C_0 C 0 is the pairing of b 1 , b 2 , b 3 b_1,b_2,b_3 b 1 , b 2 , b 3 with ψ 1 \psi_1 ψ 1 , whose values are the averages of x − 1 , 1 , x − 1 x^{-1},1,x^{-1} x − 1 , 1 , x − 1 . The other entries follow from the same rule. For the recovered polynomials the mass is
γ ( π L ι L , π R ι V ) = γ { t 5 v 4 + t 4 ( v 5 − v 6 ) } . \gamma(\pi_L\iota_L,\pi_R\iota_V)=\gamma\{t_5v_4+t_4(v_5-v_6)\}. γ ( π L ι L , π R ι V ) = γ { t 5 v 4 + t 4 ( v 5 − v 6 )} .
For γ ≠ 0 \gamma\ne0 γ = 0 the mass has rank two. Taking the left mass coordinates to be [ t 4 t 5 ] [t_4\ t_5] [ t 4 t 5 ] gives h = 2 h=2 h = 2 , F 0 = 0 F_0=0 F 0 = 0 , F p = I 2 F_p=I_2 F p = I 2 , H 0 = 0 H_0=0 H 0 = 0 and H p = γ [ 0 1 1 0 ] H_p=\gamma\begin{bmatrix}0&1\\1&0\end{bmatrix} H p = γ [ 0 1 1 0 ] . Here h = 2 h=2 h = 2 is the chosen factorization size; at γ = 0 \gamma=0 γ = 0 the mass vanishes and its two coordinates can be omitted. Substitution in (15) , with column order ( z 1 , z 2 , z 3 , u 1 , u 2 , α 1 , α 2 ) (z_1,z_2,z_3,u_1,u_2,\alpha_1,\alpha_2) ( z 1 , z 2 , z 3 , u 1 , u 2 , α 1 , α 2 ) , gives
T 1 = [ 0 0 0 1 0 1 0 0 0 0 0 1 0 1 0 γ − γ − 1 0 0 0 γ 0 0 0 − 1 0 0 − 1 0 0 0 0 1 0 0 0 1 0 0 0 2 0 0 1 0 0 0 0 ] , det T 1 = − 2 γ 2 . (26) \mathbb{T}_1=\begin{bmatrix}
0&0&0&1&0&1&0\\
0&0&0&0&1&0&1\\
0&\gamma&-\gamma&-1&0&0&0\\
\gamma&0&0&0&-1&0&0\\
-1&0&0&0&0&1&0\\
0&0&1&0&0&0&2\\
0&0&1&0&0&0&0
\end{bmatrix},\qquad\det\mathbb{T}_1=-2\gamma^2. \tag*{(26)} T 1 = 0 0 0 γ − 1 0 0 0 0 γ 0 0 0 0 0 0 − γ 0 0 1 1 1 0 − 1 0 0 0 0 0 1 0 − 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 2 0 , det T 1 = − 2 γ 2 . (26)
Its first two rows are the remaining pairing equations; the next two define the mass coordinates. The last three impose the two right image conditions and the normalization. For an explicit determinant calculation, expand along its last row, then along the resulting row with the sole entry 2 2 2 in the α 2 \alpha_2 α 2 column. Expanding next along the u 2 u_2 u 2 entry in the second row and the remaining γ \gamma γ entry in the z 1 z_1 z 1 column gives
det T 1 = 2 γ det [ 0 1 1 γ − 1 0 0 0 1 ] = − 2 γ 2 . \det\mathbb{T}_1=2\gamma\det\begin{bmatrix}0&1&1\\\gamma&-1&0\\0&0&1\end{bmatrix}=-2\gamma^2. det T 1 = 2 γ det 0 γ 0 1 − 1 0 1 0 1 = − 2 γ 2 .
Reconstruction and direct verification. For γ ≠ 0 \gamma\ne0 γ = 0 , solve T 1 w = e 7 \mathbb{T}_1w=e_7 T 1 w = e 7 . The last two rows give z 3 = 1 z_3=1 z 3 = 1 and α 2 = − 1 / 2 \alpha_2=-1/2 α 2 = − 1/2 . The second and fourth then give u 2 = 1 / 2 u_2=1/2 u 2 = 1/2 and z 1 = 1 / ( 2 γ ) z_1=1/(2\gamma) z 1 = 1/ ( 2 γ ) . The fifth and first give α 1 = z 1 \alpha_1=z_1 α 1 = z 1 and u 1 = − z 1 u_1=-z_1 u 1 = − z 1 ; the third gives z 2 = 1 − 1 / ( 2 γ 2 ) z_2=1-1/(2\gamma^2) z 2 = 1 − 1/ ( 2 γ 2 ) . The eliminated original coordinates are z 0 = − C 0 z p + X 0 T α z_0=-C_0z_p+X_0^{\mathsf T}\alpha z 0 = − C 0 z p + X 0 T α . Applying π R \pi_R π R yields the polynomial column
a ^ = [ ( γ − x ) / ( 2 γ 2 ) ( 1 − γ 2 ) / ( 2 γ 2 ) 1 ] . (27) \hat{a}=\begin{bmatrix}(\gamma-x)/(2\gamma^2)\\(1-\gamma^2)/(2\gamma^2)\\1\end{bmatrix}. \tag*{(27)} a ^ = ( γ − x ) / ( 2 γ 2 ) ( 1 − γ 2 ) / ( 2 γ 2 ) 1 . (27)
It has exactly the required component bounds and normalization. As an independent check, evaluation of the transformed pairing on the prescribed bases gives
[ β ^ ( f i , g j ) ] i , j = [ 1 γ 0 0 γ 1 1 0 0 2 2 1 ] . [\widehat{\beta}(f_i,g_j)]_{i,j}=\begin{bmatrix}1&\gamma&0&0\\\gamma&1&1&0\\0&2&2&1\end{bmatrix}. [ β ( f i , g j ) ] i , j = 1 γ 0 γ 1 2 0 1 2 0 0 1 .
Its product with the coefficient column of (27) is zero. To verify the entire table in one calculation, write
f = [ a + b x c ] , g = [ u + v x w t ] T . f=[a+bx\ c],\qquad g=[u+v x\ w\ t]^{\mathsf T}. f = [ a + b x c ] , g = [ u + v x w t ] T .
At P ± P_{\pm} P ± , the products in (25) give
f S L = [ a / x + b + c c ] , S R g = [ u / x + v + w w u / x + v + t ] T . fS_L=[a/x+b+c\ c],\qquad S_Rg=[u/x+v+w\ w\ u/x+v+t]^{\mathsf T}. f S L = [ a / x + b + c c ] , S R g = [ u / x + v + w w u / x + v + t ] T .
The averages of 1 / x 1/x 1/ x and 1 / x 2 1/x^2 1/ x 2 are zero and one. Multiplying by W W W and averaging therefore gives a u + b v + b w + 2 c v + c x + c t au+bv+bw+2cv+cx+ct a u + b v + b w + 2 c v + c x + c t . The mass adds γ ( b u + a v ) \gamma(bu+av) γ ( b u + a v ) , which proves every entry of the displayed Gram matrix. For the column in (27) , its coefficients are
u = 1 2 γ , v = − 1 2 γ 2 , w = 1 − γ 2 2 γ 2 , t = 1. u=\frac{1}{2\gamma},\qquad v=-\frac{1}{2\gamma^2},\qquad w=\frac{1-\gamma^2}{2\gamma^2},\qquad t=1. u = 2 γ 1 , v = − 2 γ 2 1 , w = 2 γ 2 1 − γ 2 , t = 1.
The three orthogonality expressions are respectively u + γ v = 0 u+\gamma v=0 u + γ v = 0 , γ u + v + w = 0 \gamma u+v+w=0 γ u + v + w = 0 and 2 v + 2 w + t = 0 2v+2w+t=0 2 v + 2 w + t = 0 . Appending [ 0 0 0 1 ] [0\ 0\ 0\ 1] [ 0 0 0 1 ] to the Gram matrix gives determinant 1 ( 2 − 2 ) − γ ( 2 γ ) = − 2 γ 2 1(2-2)-\gamma(2\gamma)=-2\gamma^2 1 ( 2 − 2 ) − γ ( 2 γ ) = − 2 γ 2 , exactly the reduced determinant. At γ = 0 \gamma=0 γ = 0 the first two equations force the first coefficient to vanish and the second to be the negative of the third; the last pairing equation then contradicts the prescribed fourth coefficient one. Thus the mass restores solvability in this finite mixed-type orthogonality problem.
The square problem using the first three columns of Y Y Y has Gram matrix G ^ \widehat{G} G given by the first three columns of the last display. Its determinant is also − 2 γ 2 -2\gamma^2 − 2 γ 2 . The additional right image condition is v 6 = 0 v_6=0 v 6 = 0 : replacing the normalization by this image condition leaves the same reduced matrix (26) . With the original lists A , B \mathcal{A},\mathcal{B} A , B , its kernel is K ^ ( P , Q ) = A ( P ) G ^ − 1 B ( Q ) \widehat{K}(P,Q)=\mathcal{A}(P)\widehat{G}^{-1}\mathcal{B}(Q) K ( P , Q ) = A ( P ) G − 1 B ( Q ) , where
G ^ − 1 = [ 0 γ − 1 − ( 2 γ ) − 1 γ − 1 − γ − 2 ( 2 γ 2 ) − 1 − γ − 1 γ − 2 ( γ 2 − 1 ) / ( 2 γ 2 ) ] . \widehat{G}^{-1}=
\begin{bmatrix}
0 & \gamma^{-1} & -(2\gamma)^{-1}\\
\gamma^{-1} & -\gamma^{-2} & (2\gamma^2)^{-1}\\
-\gamma^{-1} & \gamma^{-2} & (\gamma^2-1)/(2\gamma^2)
\end{bmatrix}. G − 1 = 0 γ − 1 − γ − 1 γ − 1 − γ − 2 γ − 2 − ( 2 γ ) − 1 ( 2 γ 2 ) − 1 ( γ 2 − 1 ) / ( 2 γ 2 ) .
This also gives an explicit check of the inverse pairing constructed in Theorem 3.5 . The calculation uses the curve relation, independent matrix actions and a strictly bilateral mass throughout.
Determinant orders: degrees, components and support
The order of a Christoffel determinant is part of the constructive result. It specifies how many original polynomial and rational data must be assembled. This subsection gives two explicit counts for the four-factor transformation: a formula obtained directly from degrees, and a smaller formula obtained by polynomial row reduction. The order is the number of rows of the square matrix whose determinant is taken. It is distinct from the number of components of a polynomial and from the degree of that polynomial. Both counts refer to specified constructions. A determinant can be enlarged by an identity block, or shortened after an invertible block has been eliminated; there is therefore no representation-independent determinant order.
For example, the rational scalar perturbation in § 3.1 has numerator and denominator of degree two. The usual formula retains as unknowns the coefficients of a finite list of consecutive original polynomials. For this scalar example that system has order four, whereas the determinant displayed there has order three. The latter uses sums of products of the original polynomials and rational moments. Its entries already contain the coefficients eliminated from the larger system.
Counting polynomial functions on the support
Let F j A F_j\mathcal{A} F j A be the classes represented by polynomials of total degree at most j j j , and set F j A = 0 F_j\mathcal{A}=0 F j A = 0 for j < 0 j<0 j < 0 . Define
H A ( j ) : = dim C F j A , ρ A ( j ) : = H A ( j ) − H A ( j − 1 ) . H_{\mathcal{A}}(j):=\dim_{\mathbb{C}}F_j\mathcal{A},\qquad\rho_{\mathcal{A}}(j):=H_{\mathcal{A}}(j)-H_{\mathcal{A}}(j-1). H A ( j ) := dim C F j A , ρ A ( j ) := H A ( j ) − H A ( j − 1 ) .
Thus H A ( j ) H_{\mathcal{A}}(j) H A ( j ) counts all available polynomial functions through degree j j j , while ρ A ( j ) \rho_{\mathcal{A}}(j) ρ A ( j ) counts the new functions at degree j j j . The relations of the support must be imposed before counting. For example, on y = x 2 y=x^2 y = x 2 , the two expressions y y y and x 2 x^2 x 2 represent the same function, but the inherited degree of that function is one.
The following formulas make the dependence on the support explicit. Binomial terms with negative degree arguments are taken as zero.
Table 1.
For a hypersurface, multiplication by its defining polynomial identifies the relations of degree at most j j j with the polynomials of degree at most j − d j-d j − d . Subtracting their dimensions proves the displayed formula; smoothness is not required. For the torus, count the Laurent monomials z α z^\alpha z α with ∑ i ∣ α i ∣ ≤ j \sum_i|\alpha_i|\le j ∑ i ∣ α i ∣ ≤ j .
In D D D Laurent variables the same count is
H A ( j ) = ∑ i = 0 D 2 i ( D i ) ( j i ) . H_A(j)=\sum_{i=0}^{D}2^i\binom{D}{i}\binom{j}{i}. H A ( j ) = i = 0 ∑ D 2 i ( i D ) ( i j ) .
Indeed choose the i i i nonzero coordinates, their signs, and their positive absolute values whose sum is at most j j j . For fixed coordinates and signs, subtract one from each positive absolute value. The resulting i i i nonnegative integers have sum at most j − i j-i j − i , giving ( j i ) \binom{j}{i} ( i j ) possibilities. The case i = 0 i=0 i = 0 gives the constant monomial.
For any affine support, a Gröbner basis with an order compatible with total degree provides a basis of standard monomials (those not divisible by a leading monomial of a relation) and hence H A H_A H A . The eventual polynomial count can also be obtained by homogenizing the support equations. Write A = C [ x 1 , … , x M ] / I A=\mathbb{C}[x_1,\ldots,x_M]/I A = C [ x 1 , … , x M ] / I and let I h I^{\mathrm{h}} I h be the ideal generated by the homogenizations of all elements of I I I , using one new variable x 0 x_0 x 0 . Put
A h : = C [ x 0 , x 1 , … , x M ] / I h . A^{\mathrm{h}}:=\mathbb{C}[x_0,x_1,\ldots,x_M]/I^{\mathrm{h}}. A h := C [ x 0 , x 1 , … , x M ] / I h .
Setting x 0 = 1 x_0=1 x 0 = 1 identifies the homogeneous degree-j j j part of A h A^{\mathrm{h}} A h with F j A F_jA F j A . Surjectivity follows by homogenizing each polynomial of degree at most j j j to degree j j j . For injectivity, a homogeneous polynomial whose dehomogenization belongs to I I I is a power of x 0 x_0 x 0 times the homogenization of that element of I I I , and therefore belongs to I h I^{\mathrm{h}} I h . Hence the two spaces have the same dimension.
The Hilbert-polynomial theorem applied to A h A^{\mathrm{h}} A h gives, for sufficiently large j j j , a polynomial P A ( j ) P_A(j) P A ( j ) of degree s = dim A s=\dim A s = dim A . Its leading coefficient is d / s ! d/s! d / s ! , where d d d is the degree of the projective closure with its multiplicities; see [14 Section 1.9 and Chapter 12] . Write P A ( j ) = d j s / s ! + ∑ ℓ = 0 s − 1 c ℓ j ℓ P_A(j)=dj^s/s!+\sum_{\ell=0}^{s-1}c_\ell j^\ell P A ( j ) = d j s / s ! + ∑ ℓ = 0 s − 1 c ℓ j ℓ . For a fixed nonnegative integer a a a , subtraction gives
P A ( j + a ) − P A ( j ) = d s ! ( ( j + a ) s − j s ) + ∑ ℓ = 0 s − 1 c ℓ ( ( j + a ) ℓ − j ℓ ) . P_A(j+a)-P_A(j)=\frac{d}{s!}\left((j+a)^s-j^s\right)+\sum_{\ell=0}^{s-1}c_\ell\left((j+a)^\ell-j^\ell\right). P A ( j + a ) − P A ( j ) = s ! d ( ( j + a ) s − j s ) + ℓ = 0 ∑ s − 1 c ℓ ( ( j + a ) ℓ − j ℓ ) .
The binomial expansion of the first difference starts with s a j s − 1 sa\,j^{s-1} s a j s − 1 ; each remaining difference has degree at most s − 2 s-2 s − 2 . Thus, in the polynomial range,
H A ( j + a ) − H A ( j ) = d a ( s − 1 ) ! j s − 1 + O ( j s − 2 ) ( j ⟶ ∞ ) , s ≥ 2. H_A(j+a)-H_A(j)=\frac{da}{(s-1)!}j^{s-1}+O\left(j^{s-2}\right)\quad(j\longrightarrow\infty),\qquad s\geq2. H A ( j + a ) − H A ( j ) = ( s − 1 )! d a j s − 1 + O ( j s − 2 ) ( j ⟶ ∞ ) , s ≥ 2.
For a curve, P A ( j ) = d j + c 0 P_A(j)=dj+c_0 P A ( j ) = d j + c 0 , so this difference is exactly d a da d a for sufficiently large j j j . The degree and the filtration are separate data: the section spaces H 0 ( C , F ( n P ) ) H^0(C,F(nP)) H 0 ( C , F ( n P )) used in Section 5 add one dimension per step at a smooth point P P P in their stable range. They must not be counted as ambient total-degree blocks of size d d d .
An explicit order for all four matrix factors
Choose component bounds u i , v j ∈ { 0 , 1 , … } u_i,v_j\in\{0,1,\ldots\} u i , v j ∈ { 0 , 1 , … } and take
X = ⨁ i = 1 q F u i − 1 A e i T , Y = ⨁ j = 1 p F v j − 1 A e j . X=\bigoplus_{i=1}^{q}F_{u_i-1}Ae_i^{\mathsf{T}},\qquad Y=\bigoplus_{j=1}^{p}F_{v_j-1}Ae_j. X = i = 1 ⨁ q F u i − 1 A e i T , Y = j = 1 ⨁ p F v j − 1 A e j .
The numbers of coefficients are
r : = ∑ i = 1 q H A ( u i − 1 ) , N : = ∑ j = 1 p H A ( v j − 1 ) , k = N − r ≥ 0. r:=\sum_{i=1}^{q}H_A(u_i-1),\qquad N:=\sum_{j=1}^{p}H_A(v_j-1),\qquad k=N-r\geq0. r := i = 1 ∑ q H A ( u i − 1 ) , N := j = 1 ∑ p H A ( v j − 1 ) , k = N − r ≥ 0.
A zero bound gives a zero component. Choose once and for all a degree-ordered basis of A A A , and order its component copies by degree. Let E N , H N E_N,H_N E N , H N be the spans of the first N N N row and column vectors in these orders. Their nested lists as N N N increases will be called the row and column flags; thus a flag here simply adds one specified independent vector at each step. The only original normality assumption needed below is that β μ \beta_\mu β μ is nondegenerate on E N × H N E_N\times H_N E N × H N . For c c c components, the largest degree occurring among the first N N N vectors is
ℓ c ( N ) : = min { j : c H A ( j ) ≥ N } . \ell_c(N):=\min\{j:cH_A(j)\geq N\}. ℓ c ( N ) := min { j : c H A ( j ) ≥ N } .
We take N > 0 N>0 N > 0 ; the zero-dimensional system is interpreted with the usual empty determinant convention.
The four matrices enter the count through actual polynomial degrees:
δ L : = det L 1 , B L : = adj ( L 1 ) L 0 , e L : = deg δ L , γ L : = deg B L , (28) \delta_L := \det L_1,\qquad B_L := \operatorname{adj}(L_1)L_0,\qquad e_L := \deg\delta_L,\qquad\gamma_L := \deg B_L, \tag*{(28)} δ L := det L 1 , B L := adj ( L 1 ) L 0 , e L := deg δ L , γ L := deg B L , (28)
δ R : = det R 1 , B R : = R 0 adj ( R 1 ) , e R : = deg δ R , γ R : = deg B R . \delta_R := \det R_1,\qquad B_R := R_0\operatorname{adj}(R_1),\qquad e_R := \deg\delta_R,\qquad\gamma_R := \deg B_R. δ R := det R 1 , B R := R 0 adj ( R 1 ) , e R := deg δ R , γ R := deg B R .
The degree of a matrix is the largest degree of its entries in A A A . Thus cancellations caused by the support relations are included. Write U : = max i : u i > 0 ( u i − 1 ) U := \max_{i:u_i>0}(u_i-1) U := max i : u i > 0 ( u i − 1 ) and V : = max j : v j > 0 ( v j − 1 ) V := \max_{j:v_j>0}(v_j-1) V := max j : v j > 0 ( v j − 1 ) . If X = 0 X=0 X = 0 , omit the term containing U U U in the next maximum. Set
K L : = max { ℓ q ( N ) + e L , U + γ L } , K R : = max { ℓ p ( N ) + e R , V + γ R } . (29) K_L := \max\{\ell_q(N)+e_L,\ U+\gamma_L\},\qquad K_R := \max\{\ell_p(N)+e_R,\ V+\gamma_R\}. \tag*{(29)} K L := max { ℓ q ( N ) + e L , U + γ L } , K R := max { ℓ p ( N ) + e R , V + γ R } . (29)
Finally, let h h h be the rank of the matrix of the specified mass ν \nu ν on X × Y X\times Y X × Y . This rank is calculated from the mass data; it is zero when the mass is absent.
Theorem 3.6 (Degree count for bilateral determinants in mixed-type orthogonality). Assume the four determinants are regular and the original pair E N × H N E_N\times H_N E N × H N is normal. There is a reduced determinant in Theorem 3.3 whose exact order is
t deg = q H A ( K L ) + p H A ( K R ) − 2 r − k + h . (30) t_{\deg}=qH_A(K_L)+pH_A(K_R)-2r-k+h. \tag*{(30)} t d e g = q H A ( K L ) + p H A ( K R ) − 2 r − k + h . (30) Its nonvanishing is equivalent to existence and uniqueness of the prescribed normalized block. Each scalar polynomial entry has the bordered determinant formula of order t deg + 1 t_{\deg}+1 t d e g + 1 .
The same data give a smaller construction by polynomial rank computations. In bases of ( F K L A ) 1 × q (F_{K_L}A)^{1\times q} ( F K L A ) 1 × q and ( F K R A ) p × 1 (F_{K_R}A)^{p\times1} ( F K R A ) p × 1 , form the coefficient matrices
C L : = [ δ L E N X B L ] , C R : = [ δ R H N B R Y ] . (31) C_L := [\delta_L E_N\ \ XB_L],\qquad C_R := [\delta_R H_N\ \ B_RY]. \tag*{(31)} C L := [ δ L E N X B L ] , C R := [ δ R H N B R Y ] . (31)
Here a displayed space means the list obtained from any of its bases; these are ordinary complex coefficient matrices, with one column for each vector on the list. The exact order using the spans of these lists is
t span = rank C L + rank C R − 2 r − k + h ≤ t deg . (32) t_{\operatorname{span}}=\operatorname{rank}C_L+\operatorname{rank}C_R-2r-k+h\le t_{\deg}. \tag*{(32)} t span = rank C L + rank C R − 2 r − k + h ≤ t d e g . (32)
This is the smallest ambient-space count for the fixed original pair E N , H N E_N,H_N E N , H N and a mass factorization of rank h h h .
Proof. First choose the explicit rational spaces
W L = δ L − 1 ( F K L A ) 1 × q , W R = δ R − 1 ( F K R A ) p × 1 . W_L=\delta_L^{-1}(F_{K_L}A)^{1\times q},\qquad W_R=\delta_R^{-1}(F_{K_R}A)^{p\times1}. W L = δ L − 1 ( F K L A ) 1 × q , W R = δ R − 1 ( F K R A ) p × 1 .
Regularity of the denominators makes multiplication by their inverses injective, so these spaces have dimensions q H A ( K L ) qH_A(K_L) q H A ( K L ) and p H A ( K R ) pH_A(K_R) p H A ( K R ) . They contain the original spaces: for example, an element of E N E_N E N has degree at most ℓ q ( N ) \ell_q(N) ℓ q ( N ) , and multiplying it by δ L \delta_L δ L raises degree by at most e L e_L e L . They also contain the transformed spaces, because X S L = δ L − 1 X B L XS_L=\delta_L^{-1}XB_L X S L = δ L − 1 X B L and S R Y = δ R − 1 B R Y S_RY=\delta_R^{-1}B_RY S R Y = δ R − 1 B R Y . The two maxima in (29) are exactly the bounds needed for these inclusions.
Thus s L = q H A ( K L ) − N s_L=qH_A(K_L)-N s L = q H A ( K L ) − N and s R = p H A ( K R ) − N s_R=pH_A(K_R)-N s R = p H A ( K R ) − N . Substitution in the order N − r + s L + s R + h N-r+s_L+s_R+h N − r + s L + s R + h of (15) gives
N − r + ( q H A ( K L ) − N ) + ( p H A ( K R ) − N ) + h = q H A ( K L ) + p H A ( K R ) − N − r + h . N-r+(qH_A(K_L)-N)+(pH_A(K_R)-N)+h=qH_A(K_L)+pH_A(K_R)-N-r+h. N − r + ( q H A ( K L ) − N ) + ( p H A ( K R ) − N ) + h = q H A ( K L ) + p H A ( K R ) − N − r + h .
Using N = r + k N=r+k N = r + k proves (30) . A rank factorization of the mass matrix has exactly h h h terms, as required in that construction. The normality criterion and the bordered formulas follow from Theorem 3.3.
For the smaller spaces, clearing the regular denominator identifies E N + X S L E_N+XS_L E N + X S L with the span of δ L E N \delta_LE_N δ L E N and X B L XB_L X B L . Its dimension is therefore rank C L \operatorname{rank}C_L rank C L . The same argument gives rank C R \operatorname{rank}C_R rank C R on the right. Thus the numbers of added vectors are rank C L − N \operatorname{rank}C_L-N rank C L − N and rank C R − N \operatorname{rank}C_R-N rank C R − N , and their substitution gives
N − r + ( rank C L − N ) + ( rank C R − N ) + h = rank C L + rank C R − 2 r − k + h , N-r+(\operatorname{rank}C_L-N)+(\operatorname{rank}C_R-N)+h=\operatorname{rank}C_L+\operatorname{rank}C_R-2r-k+h, N − r + ( rank C L − N ) + ( rank C R − N ) + h = rank C L + rank C R − 2 r − k + h ,
which is (32) . Any ambient space containing the required two lists contains their span, which proves the stated minimality. This does not assert minimality among different original pairs or among determinants obtained by further elimination.
All quantities in (3.22) are now specified by polynomial degrees, the component bounds, the Hilbert function of the support and the given mass matrix. Formula (3.24) replaces the potentially redundant degree bounds by two ordinary rank calculations. Numerator factors equal to the identity give pure division; denominator factors equal to the identity give bilateral Christoffel multiplication. A factor equal to the identity on each side gives the corresponding unilateral specialization. No separate order theorem is needed for these cases.
If a different original normal pair has dimension N 0 N_0 N 0 , use its actual largest row and column degrees in place of ℓ q ( N ) \ell_q(N) ℓ q ( N ) , ℓ p ( N ) \ell_p(N) ℓ p ( N ) in (3.21). The same proof gives
t d e g = q H A ( K L ) + p H A ( K R ) − N 0 − r + h , t s p a n = rank C L + rank C R − N 0 − r + h . t_{\mathrm{deg}}=qH_A(K_L)+pH_A(K_R)-N_0-r+h,\qquad t_{\mathrm{span}}=\operatorname{rank}C_L+\operatorname{rank}C_R-N_0-r+h. t deg = q H A ( K L ) + p H A ( K R ) − N 0 − r + h , t span = rank C L + rank C R − N 0 − r + h .
Thus an original flag that fails to be normal at r + k r+k r + k does not invalidate a construction based on another normal pair. The worked mixed-type example on a curve uses precisely this option, with N 0 = r = 3 N_0=r=3 N 0 = r = 3 .
For row and column sections of modules, replace H A H_A H A on the left and right by the respective dimensions of the section spaces, and compute the coefficient matrices in bases of those spaces. The proof is unchanged when multiplication by a degree-a a a function maps a section of level j j j to level at most j + a j+a j + a and the denominator actions are injective. Use actual filtration shifts, such as allowed pole orders, in place of ambient polynomial degrees. In particular, for line bundles filtered by allowed poles along a divisor D ∞ D_\infty D ∞ , each stable increment has size deg D ∞ \deg D_\infty deg D ∞ , not necessarily the degree of an ambient equation of the curve. In the family of Section 5, the stated bases have dimension N N N on both sides; their increments are one.
For a zero-dimensional coordinate algebra, H A ( j ) H_A(j) H A ( j ) eventually equals dim C A \dim_{\mathbb{C}} A dim C A , and the same finite formulas apply; there is no infinite sequence of independent higher-degree blocks. If a factor is nonregular, the inverse-based formulas are not invoked: Proposition 2.1 first decides existence of the transformed functional, and its direct normalized Gram system, when specified, has order N N N .
Mass ranks and their contribution to the order
The number of scalar parameters describing all possible masses is different from the rank of one chosen mass. Only the latter is added to (3.22). It can be computed without a choice of global mass coordinates. Define
d X : = dim X − dim ( X ∩ A 1 × q L 1 ) , d_X:=\dim X-\dim\left(X\cap A^{1\times q}L_1\right), d X := dim X − dim ( X ∩ A 1 × q L 1 ) ,
d Y : = dim Y − dim ( Y ∩ R 1 A p × 1 ) . d_Y:=\dim Y-\dim\left(Y\cap R_1A^{p\times1}\right). d Y := dim Y − dim ( Y ∩ R 1 A p × 1 ) .
These are polynomial module membership computations, over the same coordinate algebra used to calculate H A H_A H A .
Proposition 3.7 (Finite mass rank bound). For every admissible balanced or bivariate mass,
h ≤ min { r , N , d X + d Y } . (33) h\le\min\{r,N,d_X+d_Y\}. \tag*{(33)} h ≤ min { r , N , d X + d Y } . (33) If a 1 : = deg L 1 a_1:=\deg L_1 a 1 := deg L 1 and b 1 : = deg R 1 b_1:=\deg R_1 b 1 := deg R 1 , put u min : = min i u i u_{\min}:=\min_i u_i u m i n := min i u i and v min : = min j v j v_{\min}:=\min_j v_j v m i n := min j v j . Then the entirely numerical bounds
d X ≤ r − q H A ( u min − 1 − a 1 ) , d Y ≤ N − p H A ( v min − 1 − b 1 ) (34) d_X\le r-qH_A(u_{\min}-1-a_1),\qquad d_Y\le N-pH_A(v_{\min}-1-b_1) \tag*{(34)} d X ≤ r − q H A ( u m i n − 1 − a 1 ) , d Y ≤ N − p H A ( v m i n − 1 − b 1 ) (34) hold. On a curve, the global quotient dimensions in Theorem 2.2 satisfy
h ≤ d L + d R ≤ d ( q a 1 + p b 1 ) . h\le d_L+d_R\le d(qa_1+pb_1). h ≤ d L + d R ≤ d ( q a 1 + p b 1 ) .
Proof. Choose complements of X ∩ A 1 × q L 1 X\cap A^{1\times q}L_1 X ∩ A 1 × q L 1 in X X X and of Y ∩ R 1 A p × 1 Y\cap R_1A^{p\times1} Y ∩ R 1 A p × 1 in Y Y Y . The mass vanishes when both arguments belong to the intersections. Its matrix, in these decompositions, has, with the complements listed first, the form
M ν = [ A B C 0 ] = [ A B 0 0 ] + [ 0 0 C 0 ] . M_\nu=
\begin{bmatrix}
A & B\\
C & 0
\end{bmatrix}
=
\begin{bmatrix}
A & B\\
0 & 0
\end{bmatrix}
+
\begin{bmatrix}
0 & 0\\
C & 0
\end{bmatrix}. M ν = [ A C B 0 ] = [ A 0 B 0 ] + [ 0 C 0 0 ] .
The first summand has only d X d_X d X possibly nonzero rows and the second only d Y d_Y d Y possibly nonzero columns. Subadditivity of rank therefore gives rank M ν ≤ d X + d Y \operatorname{rank}M_\nu\le d_X+d_Y rank M ν ≤ d X + d Y . Its dimensions also give rank M ν ≤ min { r , N } \operatorname{rank}M_\nu\le\min\{r,N\} rank M ν ≤ min { r , N } , proving (3.25).
Multiplication by L 1 L_1 L 1 maps ( F u min − 1 − a 1 A ) 1 × q (F_{u_{\min}-1-a_1}A)^{1\times q} ( F u m i n − 1 − a 1 A ) 1 × q injectively into X ∩ A 1 × q L 1 X\cap A^{1\times q}L_1 X ∩ A 1 × q L 1 . Counting its dimension proves the first bound in (3.26); the right bound follows in the same way. On a curve apply this argument to the full spaces ( F j A ) q (F_jA)^q ( F j A ) q . Let Q L : = A 1 × q / A 1 × q L 1 Q_L:=A^{1\times q}/A^{1\times q}L_1 Q L := A 1 × q / A 1 × q L 1 . Its finite dimension means that representatives of a basis all lie in some F j A F_jA F j A . Thus, for all sufficiently large j j j , the map from ( F j A ) q (F_jA)^q ( F j A ) q to Q L Q_L Q L is onto. Its kernel contains the injective image of ( F j − a 1 A ) q (F_{j-a_1}A)^q ( F j − a 1 A ) q under multiplication by L 1 L_1 L 1 . Rank-nullity gives
d L ≤ q ( H A ( j ) − H A ( j − a 1 ) ) = q d a 1 . d_L \le q\bigl(H_A(j)-H_A(j-a_1)\bigr)=q d a_1. d L ≤ q ( H A ( j ) − H A ( j − a 1 ) ) = q d a 1 .
The same proof gives d R ≤ p d b 1 d_R \le p d b_1 d R ≤ p d b 1 . This argument uses regularity and the Hilbert polynomial, and includes singular, reducible and nonreduced curves whenever the four-factor regularity assumptions hold. It also allows isolated zero-dimensional components: their eventual constant contribution to H A H_A H A cancels in H A ( j ) − H A ( j − a 1 ) H_A(j)-H_A(j-a_1) H A ( j ) − H A ( j − a 1 ) . Equivalently, a regular scalar on an isolated finite-dimensional local component is a unit, so its determinant contributes no remainder there. No assumption that all components have dimension one is needed for this argument. □
For a purely additive perturbation one can use E = X E=X E = X , H = Y H=Y H = Y in the square kernel problem, with no extra rational vectors. Its determinant has order h h h , the usual finite-rank update order. For a normalized problem with dim Y − dim X = k \dim Y-\dim X=k dim Y − dim X = k , taking nested original spaces E N E_N E N , H N H_N H N containing X , Y X,Y X , Y gives order k + h k+h k + h in (15) . When the unperturbed normalized problem is already solved, eliminating that normalization block gives the usual order-h h h update instead. Masses arising from division can already be included in spectral data; they must not be added a second time to a count based on consecutive original degrees.
One variable and curves: exact stable counts
For a degree-ordered flag for mixed-type orthogonality, take X = E r X=E_r X = E r and Y = H N Y=H_N Y = H N , with N = r + k N=r+k N = r + k . On a curve of degree d d d write H A ( j ) = d j + χ H_A(j)=dj+\chi H A ( j ) = d j + χ in the stable range. The integer χ \chi χ is the constant term of this Hilbert polynomial. Define the two endpoint counts and two fixed degree shifts by
Δ q : = q H A ( ℓ q ( N ) ) − N , Δ p : = p H A ( ℓ p ( N ) ) − N , \Delta_q := qH_A(\ell_q(N))-N,\qquad\Delta_p := pH_A(\ell_p(N))-N, Δ q := q H A ( ℓ q ( N )) − N , Δ p := p H A ( ℓ p ( N )) − N ,
σ L : = max { e L , γ L + ℓ q ( r ) − ℓ q ( N ) } , σ R : = max { e R , γ R } . \sigma_L := \max\{e_L,\gamma_L+\ell_q(r)-\ell_q(N)\},\qquad\sigma_R := \max\{e_R,\gamma_R\}. σ L := max { e L , γ L + ℓ q ( r ) − ℓ q ( N )} , σ R := max { e R , γ R } .
The endpoint counts measure the unused places in the last degree block, and satisfy 0 ≤ Δ q < q d 0\le\Delta_q<qd 0 ≤ Δ q < q d and 0 ≤ Δ p < p d 0\le\Delta_p<pd 0 ≤ Δ p < p d . For flags the formulas for K L , K R K_L,K_R K L , K R use the largest degrees ℓ q ( r ) \ell_q(r) ℓ q ( r ) , ℓ p ( N ) \ell_p(N) ℓ p ( N ) directly; they do not require a partial block to be a full component degree space.
Corollary 3.8 (Bilateral orders on a curve). Once all the degrees entering the count are in the linear Hilbert range, the degree construction has exact order
t d e g = k + Δ q + Δ p + d ( q σ L + p σ R ) + h . (35) t_{\mathrm{deg}}=k+\Delta_q+\Delta_p+d(q\sigma_L+p\sigma_R)+h. \tag*{(35)} t deg = k + Δ q + Δ p + d ( q σ L + p σ R ) + h . (35) In particular,
t s p a n ≤ t d e g ≤ k + ( q + p ) d − 2 + d ( q max { e L , γ L } + p max { e R , γ R } ) + q a 1 + p b 1 . (36) t_{\mathrm{span}}\le t_{\mathrm{deg}}\le k+(q+p)d-2+d\bigl(q\max\{e_L,\gamma_L\}+p\max\{e_R,\gamma_R\}\bigr)+qa_1+pb_1. \tag*{(36)} t span ≤ t deg ≤ k + ( q + p ) d − 2 + d ( q max { e L , γ L } + p max { e R , γ R } ) + q a 1 + p b 1 . (36) For fixed factors and fixed k k k , this is independent of the level. For a fixed original mass the rank h h h stabilizes along exhaustive nested flags. The degree-construction order is eventually periodic in r r r , with a period dividing lcm ( q d , p d ) \operatorname{lcm}(qd,pd) lcm ( q d , p d ) .
Proof. Write K L = ℓ q ( N ) + σ L K_L=\ell_q(N)+\sigma_L K L = ℓ q ( N ) + σ L and K R = ℓ p ( N ) + σ R K_R=\ell_p(N)+\sigma_R K R = ℓ p ( N ) + σ R . The linear Hilbert formula gives q H A ( K L ) = N + Δ q + q d σ L qH_A(K_L)=N+\Delta_q+qd\sigma_L q H A ( K L ) = N + Δ q + q d σ L and its right counterpart. Substitution in (30) proves (35) . Since r ≤ N r\le N r ≤ N , the monotonicity of ℓ q \ell_q ℓ q bounds σ L \sigma_L σ L by max { e L , γ L } \max\{e_L,\gamma_L\} max { e L , γ L } . The endpoint and mass bounds give (36) . For stabilization of h h h , let the global mass rank be h ∞ h_\infty h ∞ . Choose h ∞ h_\infty h ∞ rows and columns whose mass matrix is invertible; such a minor exists by the definition of rank. Exhaustiveness puts all of them in the two flags at some finite level. Every subsequent restriction has rank at least h ∞ h_\infty h ∞ , and no restriction can exceed the global rank. Hence h = h ∞ h=h_\infty h = h ∞ from that level onward. Finally,
ℓ c ( N ) = ⌈ N − c χ c d ⌉ \ell_c(N)=\left\lceil\frac{N-c\chi}{cd}\right\rceil ℓ c ( N ) = ⌈ c d N − cχ ⌉
in the stable range. Put L : = lcm ( q d , p d ) L:=\operatorname{lcm}(qd,pd) L := lcm ( q d , p d ) . Increasing r r r and N N N by L L L increases ℓ q \ell_q ℓ q by L / ( q d ) L/(qd) L / ( q d ) and ℓ p \ell_p ℓ p by L / ( p d ) L/(pd) L / ( p d ) . Thus Δ q , Δ p \Delta_q,\Delta_p Δ q , Δ p and ℓ q ( r ) − ℓ q ( N ) \ell_q(r)-\ell_q(N) ℓ q ( r ) − ℓ q ( N ) stay unchanged. Every term in (35) then repeats, proving eventual periodicity. □
This gives an explicit answer for the full bilateral mixed-type orthogonality problem on a curve, including all four matrices and any admissible mass. The order may oscillate with the position inside a degree block; it need not be the same integer at consecutive scalar levels. The rank computation (32) detects additional reductions, including cancellations between matrix factors.
On the line d = 1 d=1 d = 1 , χ = 1 \chi=1 χ = 1 and ℓ c ( N ) = ⌈ N / c ⌉ − 1 \ell_c(N)=\lceil N/c\rceil-1 ℓ c ( N ) = ⌈ N / c ⌉ − 1 . Thus an exact formula at every level r ≥ 1 r\ge1 r ≥ 1 is
t d e g = q max { ⌈ r + k q ⌉ + e L , ⌊ r q ⌋ + γ L } + p ⌈ r + k p ⌉ + p max { e R , γ R } − 2 r − k + h . (37) t_{\mathrm{deg}}=q\max\left\{\left\lceil\frac{r+k}{q}\right\rceil+e_L,\left\lfloor\frac{r}{q}\right\rfloor+\gamma_L\right\}+p\left\lceil\frac{r+k}{p}\right\rceil+p\max\{e_R,\gamma_R\}-2r-k+h. \tag*{(37)} t deg = q max { ⌈ q r + k ⌉ + e L , ⌊ q r ⌋ + γ L } + p ⌈ p r + k ⌉ + p max { e R , γ R } − 2 r − k + h . (37)
At r = 0 r=0 r = 0 , the left target space is zero. Omit the second term inside the first maximum, as prescribed in (29) ; the numerator factor then creates no left rational vectors. For example, with q = p = k = 1 q=p=k=1 q = p = k = 1 , L 0 = z a L_0=z^a L 0 = z a and all other factors equal to one, the construction has order one for every a ≥ 0 a\ge0 a ≥ 0 . The normalization alone determines the constant candidate. For r ≥ 1 r\ge1 r ≥ 1 , q = p = 1 q=p=1 q = p = 1 and k = 1 k=1 k = 1 , if the four scalar factors have degrees a 0 , a 1 , b 0 , b 1 a_0,a_1,b_0,b_1 a 0 , a 1 , b 0 , b 1 , this becomes
t d e g = max { 1 + a 1 , a 0 } + max { b 1 , b 0 } + h . t_{\mathrm{deg}}=\max\{1+a_1,a_0\}+\max\{b_1,b_0\}+h. t deg = max { 1 + a 1 , a 0 } + max { b 1 , b 0 } + h .
In particular, four linear factors with no mass give order three, in agreement with the worked example. This count belongs to the reduced formula, and is different from the order a 0 + b 0 + a 1 + b 1 a_0+b_0+a_1+b_1 a 0 + b 0 + a 1 + b 1 when the scalar numerator and denominator have those degrees and the standard spectral formula applies.
For scalar orthogonal blocks of total degree n ≥ 1 n\ge1 n ≥ 1 on any support, take X = F n − 1 A X=F_{n-1}A X = F n − 1 A and Y = F n A Y=F_nA Y = F n A . Thus r = H A ( n − 1 ) r=H_A(n-1) r = H A ( n − 1 ) and k = ρ A ( n ) k=\rho_A(n) k = ρ A ( n ) . If the four scalar factors have filtered degrees a 0 , a 1 , b 0 , b 1 a_0,a_1,b_0,b_1 a 0 , a 1 , b 0 , b 1 , define
u : = max { a 1 , a 0 − 1 } , v : = max { b 1 , b 0 } . u:=\max\{a_1,a_0-1\},\qquad v:=\max\{b_1,b_0\}. u := max { a 1 , a 0 − 1 } , v := max { b 1 , b 0 } .
Theorem 3.6 gives the exact bilateral order
t d e g ( n ) = H A ( n + u ) + H A ( n + v ) − H A ( n ) − H A ( n − 1 ) + h n . (38) t_{\mathrm{deg}}(n)=H_A(n+u)+H_A(n+v)-H_A(n)-H_A(n-1)+h_n. \tag*{(38)} t deg ( n ) = H A ( n + u ) + H A ( n + v ) − H A ( n ) − H A ( n − 1 ) + h n . (38)
This applies whenever the indicated original Gram is normal and the four factors are regular; h n h_n h n is the rank of the actual mass on F n − 1 A × F n A F_{n-1}A\times F_nA F n − 1 A × F n A . On a curve in its stable range,
t d e g ( n ) = d ( max { 1 + a 1 , a 0 } + max { b 1 , b 0 } ) + h n . t_{\mathrm{deg}}(n)=d\left(\max\{1+a_1,a_0\}+\max\{b_1,b_0\}\right)+h_n. t deg ( n ) = d ( max { 1 + a 1 , a 0 } + max { b 1 , b 0 } ) + h n .
For two independent variables the exact formula is
t d e g ( n ) = ( u + v + 1 ) n + u ( u + 3 ) + ( v + 1 ) ( v + 2 ) 2 + h n . t_{\mathrm{deg}}(n)=(u+v+1)n+\frac{u(u+3)+(v+1)(v+2)}{2}+h_n. t deg ( n ) = ( u + v + 1 ) n + 2 u ( u + 3 ) + ( v + 1 ) ( v + 2 ) + h n .
In particular, with four linear factors and zero mass the orders are 3 on the line, 3 n + 5 3n+5 3 n + 5 in two independent variables, and 3 d 3d 3 d on a plane curve once n ≥ max { 1 , d − 1 } n\ge\max\{1,d-1\} n ≥ max { 1 , d − 1 } . These are exact orders of the reduced construction; the consecutive-degree construction has its separate count below. Setting a 1 = b 0 = b 1 = 0 a_1=b_0=b_1=0 a 1 = b 0 = b 1 = 0 and a 0 = a ≥ 1 a_0=a\ge1 a 0 = a ≥ 1 in (38) recovers the usual Christoffel order H A ( n + a − 1 ) − H A ( n − 1 ) H_A(n+a-1)-H_A(n-1) H A ( n + a − 1 ) − H A ( n − 1 ) .
The four entry degrees alone give a convenient bound even before the adjugates:
e L ≤ q a 1 , γ L ≤ ( q − 1 ) a 1 + a 0 , e_L\le qa_1,\qquad\gamma_L\le(q-1)a_1+a_0, e L ≤ q a 1 , γ L ≤ ( q − 1 ) a 1 + a 0 ,
e R ≤ p b 1 , γ R ≤ ( p − 1 ) b 1 + b 0 . e_R\le pb_1,\qquad\gamma_R\le(p-1)b_1+b_0. e R ≤ p b 1 , γ R ≤ ( p − 1 ) b 1 + b 0 .
Inserting these inequalities in (36) gives a bound involving only d , q , p , k d,q,p,k d , q , p , k and the four degrees. This bound refers to the degree-based construction; polynomial rank reduction and further elimination of invertible blocks can lower its order. For example, for a 2 × 3 2\times3 2 × 3 pairing on a conic, four degree-one matrices give t d e g ≤ k + 44 t_{\mathrm{deg}}\le k+44 t deg ≤ k + 44 ; the exact count uses (35) , and the reduced count uses (32) .
Several variables, surfaces and general component degree bounds
For two independent variables, (30) reads
t d e g = q 2 ( K L + 1 ) ( K L + 2 ) + p 2 ( K R + 1 ) ( K R + 2 ) − 2 r − k + h . t_{\mathrm{deg}}=\frac{q}{2}(K_L+1)(K_L+2)+\frac{p}{2}(K_R+1)(K_R+2)-2r-k+h. t deg = 2 q ( K L + 1 ) ( K L + 2 ) + 2 p ( K R + 1 ) ( K R + 2 ) − 2 r − k + h .
Here
ℓ c ( N ) = ⌊ 1 + 8 N / c − 3 2 ⌋ , \ell_c(N)=\left\lfloor\frac{\sqrt{1+8N/c}-3}{2}\right\rfloor, ℓ c ( N ) = ⌊ 2 1 + 8 N / c − 3 ⌋ ,
so the formula is numerical once the factors and component bounds are specified. On a surface in C 3 \mathbb{C}^{3} C 3 defined by one equation of degree d d d , replace each binomial H A ( j ) H_A(j) H A ( j ) by
H A ( j ) = ( j + 3 3 ) − ( j − d + 3 3 ) . H_A(j)=\binom{j+3}{3}-\binom{j-d+3}{3}. H A ( j ) = ( 3 j + 3 ) − ( 3 j − d + 3 ) .
The resulting formula is again exact at every degree, including degrees below the eventual Hilbert polynomial range. Other surfaces and higher-dimensional supports use their own H A H_A H A in (30) ; the polynomial rank formula applies unchanged.
Proposition 3.9. (Growth with the dimension of the support ). Fix the four factors. Assume the original normal flags exist at the levels under consideration. Let X , Y X,Y X , Y be degree-ordered flags, or component degree spaces for which the spread of the bounds among all components on each side stays bounded. Let n n n be the larger of the two largest component degrees. If the support has dimension s ≥ 1 s\geq1 s ≥ 1 and k = O ( n s − 1 ) k=O(n^{s-1}) k = O ( n s − 1 ) as n ⟶ ∞ n\longrightarrow\infty n ⟶ ∞ , then
t s p a n ≤ t d e g = O ( n s − 1 ) ( n ⟶ ∞ ) . t_{\mathrm{span}}\leq t_{\mathrm{deg}}=O(n^{s-1}) \qquad(n\longrightarrow\infty). t span ≤ t deg = O ( n s − 1 ) ( n ⟶ ∞ ) . This includes every admissible balanced or bivariate mass for the fixed denominator factors. Thus the orders are bounded on curves, grow at most linearly on surfaces, and at most quadratically on three-dimensional supports. These upper bounds allow smaller orders for special factors or mass data.
Proof. Write U , V U,V U , V for the largest degrees occurring on the two sides, so n = max { U , V } n=\max\{U,V\} n = max { U , V } , and retain N = r + k N=r+k N = r + k . First compare each target space with its full degree space. For component spaces, choose fixed nonnegative integers D L , D R D_L,D_R D L , D R bounding the two spreads. Then
q H A ( U − D L ) ≤ r ≤ q H A ( U ) , qH_A(U-D_L)\leq r\leq qH_A(U), q H A ( U − D L ) ≤ r ≤ q H A ( U ) ,
p H A ( V − D R ) ≤ N ≤ p H A ( V ) . pH_A(V-D_R)\leq N\leq pH_A(V). p H A ( V − D R ) ≤ N ≤ p H A ( V ) .
Indeed every left component contains F U − D L A F_{U-D_L}A F U − D L A and is contained in F U A F_UA F U A ; the right argument is the same. Define the missing dimensions by
ε X : = q H A ( U ) − r , ε Y : = p H A ( V ) − N . \varepsilon_X:=qH_A(U)-r,\qquad\varepsilon_Y:=pH_A(V)-N. ε X := q H A ( U ) − r , ε Y := p H A ( V ) − N .
The preceding inequalities and the fixed-shift Hilbert difference give
0 ≤ ε X ≤ q ( H A ( U ) − H A ( U − D L ) ) , 0\leq\varepsilon_X\leq q\bigl(H_A(U)-H_A(U-D_L)\bigr), 0 ≤ ε X ≤ q ( H A ( U ) − H A ( U − D L ) ) ,
0 ≤ ε Y ≤ p ( H A ( V ) − H A ( V − D R ) ) . 0\leq\varepsilon_Y\leq p\bigl(H_A(V)-H_A(V-D_R)\bigr). 0 ≤ ε Y ≤ p ( H A ( V ) − H A ( V − D R ) ) .
Both bounds are O ( n s − 1 ) O(n^{s-1}) O ( n s − 1 ) as n ⟶ ∞ n\longrightarrow\infty n ⟶ ∞ . If X , Y X,Y X , Y are flags, all degree blocks below the last one are full, so the corresponding bounds are instead
0 ≤ ε X < q ρ A ( U ) , 0 ≤ ε Y < p ρ A ( V ) . 0\leq\varepsilon_X<q\rho_A(U),\qquad0\leq\varepsilon_Y<p\rho_A(V). 0 ≤ ε X < q ρ A ( U ) , 0 ≤ ε Y < p ρ A ( V ) .
They satisfy the same upper bound. This also proves the assertion when a flag is used on one side and a component space on the other.
These comparisons control the degrees of the original spaces as well. The upper bound is N ≤ p H A ( n ) N\leq pH_A(n) N ≤ p H A ( n ) . If n = U n=U n = U , the left lower bound gives N ≥ r N\geq r N ≥ r of order n s n^s n s ; if n = V n=V n = V , the right lower bound does so directly. The flag lower bounds use the full block immediately below the last one and give the same conclusion. Thus there are positive constants C 1 , C 2 C_1,C_2 C 1 , C 2 such that
C 1 n s ≤ N ≤ C 2 n s for all sufficiently large n . C_1n^s\leq N\leq C_2n^s\qquad\text{for all sufficiently large }n. C 1 n s ≤ N ≤ C 2 n s for all sufficiently large n .
For c ∈ { q , p } c\in\{q,p\} c ∈ { q , p } put b c : = ℓ c ( N ) b_c:=\ell_c(N) b c := ℓ c ( N ) . Its defining minimality gives
c H A ( b c − 1 ) < N ≤ c H A ( b c ) . cH_A(b_c-1)<N\leq cH_A(b_c). c H A ( b c − 1 ) < N ≤ c H A ( b c ) .
Since H A ( j ) H_A(j) H A ( j ) has leading term d j s / s ! dj^s/s! d j s / s ! , these inequalities imply
b c ∼ ( s ! N c d ) 1 / s ( N ⟶ ∞ ) . b_c\sim\left(\frac{s!N}{cd}\right)^{1/s}\qquad(N\longrightarrow\infty). b c ∼ ( c d s ! N ) 1/ s ( N ⟶ ∞ ) .
In particular, each b c b_c b c is bounded above and below by positive constant multiples of n n n . The endpoint remainder η c : = c H A ( b c ) − N \eta_c:=cH_A(b_c)-N η c := c H A ( b c ) − N therefore satisfies
0 ≤ η c < c ρ A ( b c ) = O ( n s − 1 ) ( n ⟶ ∞ ) . 0\leq\eta_c<c\rho_A(b_c)=O(n^{s-1})\qquad(n\longrightarrow\infty). 0 ≤ η c < c ρ A ( b c ) = O ( n s − 1 ) ( n ⟶ ∞ ) .
To compare the two entries in each maximum in (29) , put M L : = max { b q , U } M_L := \max\{b_q,U\} M L := max { b q , U } and M R : = max { b p , V } M_R := \max\{b_p,V\} M R := max { b p , V } . Monotonicity of H A H_A H A gives the exact identities
q H A ( M L ) = max { N + η q , r + ε X } = N + max { η q , ε X − k } , qH_A(M_L)=\max\{N+\eta_q,r+\varepsilon_X\}=N+\max\{\eta_q,\varepsilon_X-k\}, q H A ( M L ) = max { N + η q , r + ε X } = N + max { η q , ε X − k } ,
p H A ( M R ) = max { N + η p , N + ε Y } = N + max { η p , ε Y } . pH_A(M_R)=\max\{N+\eta_p,N+\varepsilon_Y\}=N+\max\{\eta_p,\varepsilon_Y\}. p H A ( M R ) = max { N + η p , N + ε Y } = N + max { η p , ε Y } .
Both excesses over N N N are nonnegative and bounded by a constant times n s − 1 n^{s-1} n s − 1 as n → ∞ n\to\infty n → ∞ . The factor degrees are fixed, and (29) gives
M L ≤ K L ≤ M L + max { e L , γ L } , M_L\le K_L\le M_L+\max\{e_L,\gamma_L\}, M L ≤ K L ≤ M L + max { e L , γ L } ,
M R ≤ K R ≤ M R + max { e R , γ R } . M_R\le K_R\le M_R+\max\{e_R,\gamma_R\}. M R ≤ K R ≤ M R + max { e R , γ R } .
The Hilbert difference for these fixed increments now proves
q H A ( K L ) = N + O ( n s − 1 ) , p H A ( K R ) = N + O ( n s − 1 ) ( n → ∞ ) . qH_A(K_L)=N+O(n^{s-1}),\qquad pH_A(K_R)=N+O(n^{s-1})\qquad(n\to\infty). q H A ( K L ) = N + O ( n s − 1 ) , p H A ( K R ) = N + O ( n s − 1 ) ( n → ∞ ) .
The mass estimate is also a fixed-shift comparison. For component spaces, u min − 1 ≥ U − D L u_{\min}-1\ge U-D_L u m i n − 1 ≥ U − D L , so (34) implies
d X ≤ r − q H A ( U − D L − a 1 ) ≤ q ( H A ( U ) − H A ( U − D L − a 1 ) ) , d_X\le r-qH_A(U-D_L-a_1)\le q\bigl(H_A(U)-H_A(U-D_L-a_1)\bigr), d X ≤ r − q H A ( U − D L − a 1 ) ≤ q ( H A ( U ) − H A ( U − D L − a 1 ) ) ,
d Y ≤ N − p H A ( V − D R − b 1 ) ≤ p ( H A ( V ) − H A ( V − D R − b 1 ) ) . d_Y\le N-pH_A(V-D_R-b_1)\le p\bigl(H_A(V)-H_A(V-D_R-b_1)\bigr). d Y ≤ N − p H A ( V − D R − b 1 ) ≤ p ( H A ( V ) − H A ( V − D R − b 1 ) ) .
For a left flag, ( F U − 1 A ) 1 × q ⊂ X (F_{U-1}A)^{1\times q}\subset X ( F U − 1 A ) 1 × q ⊂ X , and multiplication by L 1 L_1 L 1 maps ( F U − 1 − a 1 A ) 1 × q (F_{U-1-a_1}A)^{1\times q} ( F U − 1 − a 1 A ) 1 × q injectively into X ∩ A 1 × q L 1 X\cap A^{1\times q}L_1 X ∩ A 1 × q L 1 . Hence
d X ≤ q ( H A ( U ) − H A ( U − 1 − a 1 ) ) , d_X\le q\bigl(H_A(U)-H_A(U-1-a_1)\bigr), d X ≤ q ( H A ( U ) − H A ( U − 1 − a 1 ) ) ,
with the analogous estimate on the right. In every case (33) yields h ≤ d X + d Y = O ( n s − 1 ) h\le d_X+d_Y=O(n^{s-1}) h ≤ d X + d Y = O ( n s − 1 ) as n → ∞ n\to\infty n → ∞ . This argument uses the rank of the mass on the chosen finite spaces, so it applies to all the balanced and bivariate masses in the statement.
Substitution in (30) gives
t deg = 2 N − 2 r − k + O ( n s − 1 ) + h = k + O ( n s − 1 ) + h = O ( n s − 1 ) ( n → ∞ ) , t_{\deg}=2N-2r-k+O(n^{s-1})+h=k+O(n^{s-1})+h=O(n^{s-1})\qquad(n\to\infty), t d e g = 2 N − 2 r − k + O ( n s − 1 ) + h = k + O ( n s − 1 ) + h = O ( n s − 1 ) ( n → ∞ ) ,
where the last step uses the assumed bound on k k k . The inequality t s p a n ≤ t deg t_{\mathrm{span}}\le t_{\deg} t span ≤ t d e g follows from Theorem 3.6 .
The comparison is made separately on the two sides and therefore allows q ≠ p q\ne p q = p . For example, on A = C [ z ] A=\mathbb{C}[z] A = C [ z ] take q = 2 q=2 q = 2 , p = 3 p=3 p = 3 , u 1 = u 2 = 3 m u_1=u_2=3m u 1 = u 2 = 3 m and v 1 = v 2 = v 3 = 2 m v_1=v_2=v_3=2m v 1 = v 2 = v 3 = 2 m , with m ≥ 1 m\ge1 m ≥ 1 . Then r = N = 6 m r=N=6m r = N = 6 m , k = 0 k=0 k = 0 , U = b q = 3 m − 1 U=b_q=3m-1 U = b q = 3 m − 1 and V = b p = 2 m − 1 V=b_p=2m-1 V = b p = 2 m − 1 . Each side has zero spread, although the two largest degrees differ by m m m . The estimates above compare each of these degrees with its own flag endpoint. With unrestricted component spreads the exact order formulas still apply, but the bounds on ε X , ε Y , d X , d Y \varepsilon_X,\varepsilon_Y,d_X,d_Y ε X , ε Y , d X , d Y used here need not hold. □
Consecutive degrees and conditions at infinity
To compare with familiar scalar Christoffel formulas, suppose the scalar relation is Q μ ⃗ = P μ Q\vec{\mu}=P\mu Q μ = P μ , where P , Q P,Q P , Q have degrees a , b a,b a , b , and use a spectral construction whose unknown coefficients occupy the full original blocks of degrees from n − b n-b n − b through n + a − 1 n+a-1 n + a − 1 . Its determinant order is
t w i n = H A ( n + a − 1 ) − H A ( n − b − 1 ) , n ≥ b . (39) t_{\mathrm{win}}=H_A(n+a-1)-H_A(n-b-1),\qquad n\ge b. \tag*{(39)} t win = H A ( n + a − 1 ) − H A ( n − b − 1 ) , n ≥ b . (39)
This statement counts that specified interval of degrees. Its use as an invertible spectral system requires the divisibility, degree and normality hypotheses of the chosen spectral construction [7 , 9 ] .
A Hilbert-function calculation alone does not prove those hypotheses on a new support. The general bilateral construction above imposes the full image conditions and uses the image spaces generated by the actual factors.
For one variable (39) is a + b a+b a + b . For two independent variables it is
( a + b ) ( 2 n + a − b + 1 ) 2 . \frac{(a+b)(2n+a-b+1)}{2}. 2 ( a + b ) ( 2 n + a − b + 1 ) .
On a plane curve it counts
∑ j = n − b n + a − 1 min { j + 1 , d } , \sum_{j=n-b}^{n+a-1}\min\{j+1,d\}, j = n − b ∑ n + a − 1 min { j + 1 , d } ,
and equals d ( a + b ) d(a+b) d ( a + b ) when n ≥ b + d − 1 n \ge b+d-1 n ≥ b + d − 1 . In D D D independent variables its exact value is
( n + a + D − 1 D ) − ( n − b + D − 1 D ) \binom{n+a+D-1}{D}-\binom{n-b+D-1}{D} ( D n + a + D − 1 ) − ( D n − b + D − 1 )
These familiar orders and (30) describe different coefficient systems and should be reported separately.
In the one-variable spectral transformation for mixed-type orthogonality μ ^ R = L μ \hat{\mu}R=L\mu μ ^ R = Lμ , the standard compatible step-line construction has order deg det L + deg det R \deg\det L+\deg\det R deg det L + deg det R [5 ] . For invertible leading matrices of degrees a , b a,b a , b , this is q a + p b qa+pb q a + p b . On a curve, simply multiplying this integer by d d d counts complete degree layers but does not by itself construct that spectral determinant. The proved bilateral orders on the curve are (30) , (32) and (35) .
Two examples explain why the extra degree information is needed. On the parabola A = C [ x , y ] / ( y − x 2 ) A=\mathbb{C}[x,y]/(y-x^2) A = C [ x , y ] / ( y − x 2 ) , the scalar polynomial x x x has ambient degree one, but dim A / ( x ) = 1 \dim A/(x)=1 dim A / ( x ) = 1 , whereas a stable total-degree block has size two. The difference comes from a condition on the highest allowed degree, in addition to divisibility at the finite zero. Even on the line, the matrix
L ( z ) = [ 1 z a 0 1 ] , a > 0 , L(z)=\begin{bmatrix}1&z^a\\0&1\end{bmatrix},\qquad a>0, L ( z ) = [ 1 0 z a 1 ] , a > 0 ,
has determinant one and a polynomial inverse. It creates no finite divisibility conditions, but changes component degree bounds. Thus determinant degrees alone cannot cover arbitrary matrix factors with fixed component degree bounds. The rank matrices in (31) retain precisely these degree effects.
Finally, cancelling a common factor from a rational fraction does not automatically cancel it from the functional equation: from z μ ^ = z μ z\hat{\mu}=z\mu z μ ^ = z μ one may still have μ ^ = μ + c δ 0 \hat{\mu}=\mu+c\delta_0 μ ^ = μ + c δ 0 . The remaining mass must be retained when choosing the reduced representation and its order.
3.3.7. How to report the size of a concrete formula. For a given transformation, first state the degree filtration and the component bounds, then compute H A H_A H A , the four quantities in (28) , and the mass matrix rank. Equations (29) and (30) give the exact order of the degree construction, without any further dimension choices. Reducing the two polynomial coefficient matrices gives (32) . If a formula using consecutive degrees or a finite-rank update is used instead, its own order is stated along with its hypotheses. The normalized Gram system always has order N = r + k N=r+k N = r + k and is an additional option at low degrees. For the reduced construction, the polynomial entry determinants have order t + 1 t+1 t + 1 , and the same order-t t t inverse determines the kernel; no new large Gram determinant is needed for that kernel calculation.
For reference, the concrete constructions already displayed in this article have the following orders. Here the symbols r r r , N N N in the last two rows retain their local meaning in those examples.
Table 2.
In the circle example the direct normalized Gram determinant has order four, although the displayed reduced uses order seven. At such a small level the direct system is smaller. In the singular-curve family the update order 4 r 4r 4 r is independent of the section level N N N ; on the two-dimensional torus the update order 2 N + 1 2N+1 2 N + 1 grows linearly. These distinctions prevent a determinant order from being confused with the rank of a weighted kernel difference or with the number of free mass parameters. Balanced and bivariate pairings use the same dimension counts; their different mass freedoms enter through the specified rank h h h . For the Hermitian and symmetric specializations, conjugation and transposition do not change the sizes of the coefficient spaces.
Finite updates of kernels and normalized vectors
The same finite pairings also admit a direct Gram description, useful when a particular rationally transformed pairing has already been specified. The following formulas retain that pairing as one block and add the chosen mass to it. They will be applied to the residue coordinates and spaces of sections in Sections 4 and 5.
Rational solutions with regular denominators
The particular-solution argument in Theorem 2.2 (i) uses only the inverses of the denominator factors. We record this specialization for the balanced pairings used in the geometric constructions, allowing arbitrary numerator factors. In this subsubsection A A A is integral and det L 1 det R 1 ≠ 0 \det L_1\det R_1\ne0 det L 1 det R 1 = 0 . Extend μ \mu μ complex linearly to the rational matrix tests that occur below, and denote the extension by μ ~ \widetilde{\mu} μ . Such an extension exists by completing a vector-space basis of the polynomial tests to one of the rational tests. Its additional values are part of the data: the polynomial moments alone do not determine them. Formula (4) , with μ ~ \widetilde{\mu} μ , defines the corresponding rational pairing β ~ μ \widetilde{\beta}_{\mu} β μ . Set
S L : = L 1 − 1 L 0 , S R : = R 0 R 1 − 1 , β p a r t ( f , g ) : = β ~ μ ( f S L , S R g ) . (40) S_L:=L_1^{-1}L_0,\qquad S_R:=R_0R_1^{-1},\qquad\beta_{\mathrm{part}}(f,g):=\widetilde{\beta}_{\mu}(fS_L,S_Rg). \tag*{(40)} S L := L 1 − 1 L 0 , S R := R 0 R 1 − 1 , β part ( f , g ) := β μ ( f S L , S R g ) . (40)
The inverses exist over K K K by the determinant assumption.
Proposition 3.10 (Particular solution and homogeneous mass). Every balanced solution of {(xref)}(2.1) is uniquely of the form
β ^ = β p a r t + ν , ν ( f L 1 , R 1 g ) = 0 for every f , g , \widehat{\beta}=\beta_{\mathrm{part}}+\nu,\qquad\nu(fL_1,R_1g)=0\quad\text{for every }f,g, β = β part + ν , ν ( f L 1 , R 1 g ) = 0 for every f , g , where ν \nu ν is a balanced bilinear pairing. Uniqueness refers to the fixed rational extension in {(xref)}(3.32). The pairing ν \nu ν is called a homogeneous mass.
Proof. The specialization of Theorem 2.2 (i) follows by the same substitution on the polynomial tests; it is also the calculation needed for the section modules below. Only L 1 , R 1 L_1,R_1 L 1 , R 1 are inverted, and the order of the factors gives ( f L 1 ) S L = f L 0 (fL_1)S_L=fL_0 ( f L 1 ) S L = f L 0 and S R ( R 1 g ) = R 0 g S_R(R_1g)=R_0g S R ( R 1 g ) = R 0 g . Thus β p a r t ( f L 1 , R 1 g ) = β μ ( f L 0 , R 0 g ) \beta_{\mathrm{part}}(fL_1,R_1g)=\beta_{\mu}(fL_0,R_0g) β part ( f L 1 , R 1 g ) = β μ ( f L 0 , R 0 g ) , because the extension agrees with μ \mu μ on polynomial tests. Subtracting this identity from {(xref)}(2.1) shows that ν : = β ^ − β p a r t \nu:=\widehat{\beta}-\beta_{\mathrm{part}} ν := β − β part vanishes on the indicated pairs. Conversely, adding any such ν \nu ν preserves the equation. The difference of balanced pairings is balanced, and the displayed subtraction also proves uniqueness. □
For rows and columns of sections of modules stable under multiplication by A A A , the two cancellation identities in the proof hold in their rational section spaces. Once a pairing on these rational sections is specified and agrees with the original pairing on regular tests, substitution and subtraction give the same particular solution and homogeneous mass. Balancedness is preserved because scalar multiplication commutes with the matrix actions. In Subsection 4.1 the two modules are a line bundle and its dual with values in the canonical differentials of the curve. Their products are differential tests. Subsection 4.2 describes the finite quotient through which the homogeneous masses factor, including at singular points.
From additive masses to a finite reproducing kernel
For scalar polynomial spaces in several variables, additions of one Dirac mass and of finitely many point masses yield the polynomial and kernel formulas of [35 , 36 , 43 ] . The following finite-dimensional calculation gives the kernel and the normalized vectors after an additive change of pairing. It applies to polynomial or rational functions with any prescribed finite bounds. Only the chosen spaces, their pairings and the rational factors enter the calculation. We use the column–row convention of Subsection 2.1.1.
Let E L E_L E L and E R E_R E R be spaces of row and column functions of sizes q q q and p p p , with the same finite complex dimension d d d and bases f 1 , … , f d f_1,\ldots,f_d f 1 , … , f d and g 1 , … , g d g_1,\ldots,g_d g 1 , … , g d . The original pairing β μ \beta_{\mu} β μ is extended to the rational functions occurring below; this extension is denoted β ~ μ \widetilde{\beta}_{\mu} β μ . Let ν \nu ν be an additive pairing on the target spaces. For the rational changes of arguments, set S L : = L 1 − 1 L 0 S_L:=L_1^{-1}L_0 S L := L 1 − 1 L 0 and S R : = R 0 R 1 − 1 S_R:=R_0R_1^{-1} S R := R 0 R 1 − 1 , with the factors of Subsection 2.1, and write
β p a r t ( f , g ) : = β ~ μ ( f S L , S R g ) , β ^ = β p a r t + ν . \beta_{\mathrm{part}}(f,g):=\widetilde{\beta}_{\mu}(fS_L,S_Rg),\qquad\widehat{\beta}=\beta_{\mathrm{part}}+\nu. β part ( f , g ) := β μ ( f S L , S R g ) , β = β part + ν .
Choose bases a 1 , … , a s a_1,\ldots,a_s a 1 , … , a s and b 1 , … , b t b_1,\ldots,b_t b 1 , … , b t of finite rational function spaces containing E L S L E_LS_L E L S L and S R E R S_RE_R S R E R , respectively. Define their coordinate matrices by
f i S L = ∑ α = 1 s ( T L ) i , α a α , S R g j = ∑ β = 1 t b β ( T R ) β , j . (41) f_iS_L=\sum_{\alpha=1}^{s}(T_L)_{i,\alpha}a_{\alpha},\qquad S_Rg_j=\sum_{\beta=1}^{t}b_{\beta}(T_R)_{\beta,j}. \tag*{(41)} f i S L = α = 1 ∑ s ( T L ) i , α a α , S R g j = β = 1 ∑ t b β ( T R ) β , j . (41)
where T L T_L T L is d × s d \times s d × s and T R T_R T R is t × d t \times d t × d . Finally put
G r a t : = [ β ~ μ ( a α , b β ) ] α ∈ { 1 , … , s } , β ∈ { 1 , … , t } , G_{\mathrm{rat}} := [\widetilde{\beta}_{\mu}(a_{\alpha},b_{\beta})]_{\alpha\in\{1,\ldots,s\},\ \beta\in\{1,\ldots,t\}}, G rat := [ β μ ( a α , b β ) ] α ∈ { 1 , … , s } , β ∈ { 1 , … , t } ,
H ν : = [ ν ( f i , g j ) ] i , j ∈ { 1 , … , d } . H_{\nu} := [\nu(f_i,g_j)]_{i,j \in\{1,\ldots,d\}}. H ν := [ ν ( f i , g j ) ] i , j ∈ { 1 , … , d } .
Thus G r a t G_{\mathrm{rat}} G rat contains the rational pairings of the ambient bases used to express f i S L f_iS_L f i S L and S R g j S_Rg_j S R g j , and H ν H_{\nu} H ν contains the mass evaluated on the target basis pairs. When the mass factors through a finite quotient, its entries can be computed from the remainders of these pairs and the prescribed values of the mass on a quotient basis.
Proposition 3.11 (Normality and the finite kernel). With the rational extension, spaces and coordinates above, the transformed Gram is
G ^ = T L G r a t T R + H ν . (42) \widehat{G} = T_LG_{\mathrm{rat}}T_R + H_{\nu}. \tag*{(42)} G = T L G rat T R + H ν . (42) The pairing is normal on E L , E R E_L,E_R E L , E R if and only if det G ^ ≠ 0 \det\widehat{G} \ne0 det G = 0 . Under this condition its reproducing kernel is
K ^ ( P , Q ) = g ( P ) G ^ − 1 f ( Q ) , g ( P ) : = [ g 1 ( P ) ⋯ g d ( P ) ] , f ( Q ) : = [ f i ( Q ) ] i ∈ { 1 , … , d } . (43) \widehat{K}(P,Q) = \boldsymbol{g}(P)\widehat{G}^{-1}\boldsymbol{f}(Q), \qquad\boldsymbol{g}(P) := [g_1(P)\ \cdots\ g_d(P)], \qquad\boldsymbol{f}(Q) := [f_i(Q)]_{i \in\{1,\ldots,d\}}. \tag*{(43)} K ( P , Q ) = g ( P ) G − 1 f ( Q ) , g ( P ) := [ g 1 ( P ) ⋯ g d ( P )] , f ( Q ) := [ f i ( Q ) ] i ∈ { 1 , … , d } . (43) No invertibility of the possibly rectangular matrix G r a t G_{\mathrm{rat}} G rat is assumed.
Suppose in addition that G p a r t : = T L G r a t T R G_{\mathrm{part}} := T_LG_{\mathrm{rat}}T_R G part := T L G rat T R is invertible. Let ρ : = rank H ν \rho:= \operatorname{rank} H_{\nu} ρ := rank H ν and choose a rank factorization H ν = P ν Q ν H_{\nu} = P_{\nu}Q_{\nu} H ν = P ν Q ν , where P ν P_{\nu} P ν is d × ρ d \times\rho d × ρ and Q ν Q_{\nu} Q ν is ρ × d \rho\times d ρ × d , both of rank ρ \rho ρ . Then G ^ \widehat{G} G is invertible if and only if
W ν : = I ρ + Q ν G p a r t − 1 P ν (44) W_{\nu} := I_{\rho} + Q_{\nu}G_{\mathrm{part}}^{-1}P_{\nu} \tag*{(44)} W ν := I ρ + Q ν G part − 1 P ν (44) is invertible. When these conditions hold,
K ^ ( P , Q ) − K p a r t ( P , Q ) = − [ g ( P ) G p a r t − 1 P ν ] W ν − 1 [ Q ν G p a r t − 1 f ( Q ) ] , (45) \widehat{K}(P,Q) - K_{\mathrm{part}}(P,Q) = -[\boldsymbol{g}(P)G_{\mathrm{part}}^{-1}P_{\nu}]W_{\nu}^{-1}[Q_{\nu}G_{\mathrm{part}}^{-1}\boldsymbol{f}(Q)], \tag*{(45)} K ( P , Q ) − K part ( P , Q ) = − [ g ( P ) G part − 1 P ν ] W ν − 1 [ Q ν G part − 1 f ( Q )] , (45) K p a r t ( P , Q ) : = g ( P ) G p a r t − 1 f ( Q ) . K_{\mathrm{part}}(P,Q) := \boldsymbol{g}(P)G_{\mathrm{part}}^{-1}\boldsymbol{f}(Q). K part ( P , Q ) := g ( P ) G part − 1 f ( Q ) . The separation rank of this correction is exactly ρ \rho ρ . For ρ = 0 \rho= 0 ρ = 0 , the correction is zero.
Proof. Expanding both arguments by (3.33) gives [ β p a r t ( f i , g j ) ] = T L G r a t T R [\beta_{\mathrm{part}}(f_i,g_j)] = T_LG_{\mathrm{rat}}T_R [ β part ( f i , g j )] = T L G rat T R . Adding the mass gives (3.34). Normality is precisely invertibility of this finite Gram. If g = ∑ j g j c j g = \sum_j g_jc_j g = ∑ j g j c j , its pairings against the rows f i f_i f i form the column G ^ c \widehat{G}\boldsymbol{c} G c ; multiplication by g ( P ) G ^ − 1 \boldsymbol{g}(P)\widehat{G}^{-1} g ( P ) G − 1 recovers g ( P ) g(P) g ( P ) . The corresponding row calculation proves the other reproduction identity and yields (3.35).
For the additional assertion, form the square matrix of order d + ρ d+\rho d + ρ
D : = [ G p a r t P ν − Q ν I ρ ] . \mathcal{D} := \begin{bmatrix} G_{\mathrm{part}} & P_{\nu} \\ -Q_{\nu} & I_{\rho} \end{bmatrix}. D := [ G part − Q ν P ν I ρ ] . Adding Q ν G p a r t − 1 Q_{\nu}G_{\mathrm{part}}^{-1} Q ν G part − 1 times its upper block row to its lower block row gives diagonal blocks G p a r t G_{\mathrm{part}} G part and I ρ + Q ν G p a r t − 1 P ν = W ν I_{\rho}+Q_{\nu}G_{\mathrm{part}}^{-1}P_{\nu}=W_{\nu} I ρ + Q ν G part − 1 P ν = W ν . Alternatively, subtracting P ν P_{\nu} P ν times its lower block row from its upper block row gives diagonal blocks G p a r t + P ν Q ν = G ^ G_{\mathrm{part}}+P_{\nu}Q_{\nu}=\widehat{G} G part + P ν Q ν = G and I ρ I_{\rho} I ρ . Both operations preserve determinants, proving
det G ^ = det D = det G p a r t det W ν . \det\widehat{G} = \det\mathcal{D} = \det G_{\mathrm{part}}\det W_{\nu}. det G = det D = det G part det W ν . For the inverse formula put B : = G p a r t − 1 P ν B := G_{\mathrm{part}}^{-1}P_{\nu} B := G part − 1 P ν and C : = Q ν G p a r t − 1 C := Q_{\nu}G_{\mathrm{part}}^{-1} C := Q ν G part − 1 . Then G ^ B = P ν ( I ρ + Q ν B ) = P ν W ν \widehat{G}B = P_{\nu}(I_{\rho}+Q_{\nu}B)=P_{\nu}W_{\nu} G B = P ν ( I ρ + Q ν B ) = P ν W ν , so
G ^ ( G p a r t − 1 − B W ν − 1 C ) = I d + P ν C − P ν W ν W ν − 1 C = I d . \widehat{G}(G_{\mathrm{part}}^{-1}-BW_{\nu}^{-1}C) = I_d + P_{\nu}C-P_{\nu}W_{\nu}W_{\nu}^{-1}C = I_d. G ( G part − 1 − B W ν − 1 C ) = I d + P ν C − P ν W ν W ν − 1 C = I d . This proves
G ^ − 1 = G p a r t − 1 − G p a r t − 1 P ν W ν − 1 Q ν G p a r t − 1 . \widehat{G}^{-1} = G_{\mathrm{part}}^{-1}-G_{\mathrm{part}}^{-1}P_{\nu}W_{\nu}^{-1}Q_{\nu}G_{\mathrm{part}}^{-1}. G − 1 = G part − 1 − G part − 1 P ν W ν − 1 Q ν G part − 1 . Multiplying by the two basis matrices proves (3.37). Moreover,
G ^ − 1 − G p a r t − 1 = − G ^ − 1 H ν G p a r t − 1 . \widehat{G}^{-1}-G_{\mathrm{part}}^{-1}=-\widehat{G}^{-1}H_{\nu}G_{\mathrm{part}}^{-1}. G − 1 − G part − 1 = − G − 1 H ν G part − 1 . Its rank is ρ \rho ρ , since both outside factors are invertible. The two chosen bases are complex linearly independent, so this matrix rank is the separation rank of their tensor, as in Subsection 2.1.1. □ \square □
A mass may make G ^ \widehat{G} G invertible even when G p a r t G_{\mathrm{part}} G part is singular; then (3.35) still applies. When both are normal, (3.37) compares two pairings on the same target spaces. The mass values, the rational moments and the choice of finite spaces are the data needed for this computation.
The transformed polynomials and their normalization
The same Gram matrices determine the individual orthogonal vectors. We retain the normalization information by prescribing their classes modulo the lower spaces. For polynomial spaces these classes prescribe the leading coefficients; the argument also applies whenever fixed complements specify the desired new directions.
Write E : = E L E := E_L E := E L and H : = E R H := E_R H := E R for the finite spaces just used, and retain their bases f 1 , … , f d f_1,\ldots,f_d f 1 , … , f d and g 1 , … , g d g_1,\ldots,g_d g 1 , … , g d . These are the spaces against which the next vectors must be orthogonal. Let
β ^ = β p a r t + m , G : = [ β p a r t ( f i , g j ) ] i , j ∈ { 1 , … , d } , G ^ : = [ β ^ ( f i , g j ) ] i , j ∈ { 1 , … , d } . \widehat{\beta}=\beta_{\mathrm{part}}+m,\qquad G:=[\beta_{\mathrm{part}}(f_i,g_j)]_{i,j\in\{1,\ldots,d\}},\qquad\widehat{G}:=[\widehat{\beta}(f_i,g_j)]_{i,j\in\{1,\ldots,d\}}. β = β part + m , G := [ β part ( f i , g j ) ] i , j ∈ { 1 , … , d } , G := [ β ( f i , g j ) ] i , j ∈ { 1 , … , d } .
Here m m m denotes the full additive mass pairing, including any scalar coefficient multiplying it. Thus m = ν m=\nu m = ν in Proposition 3.11, whereas a family written as β ^ = β p a r t + c ν \widehat{\beta}=\beta_{\mathrm{part}}+c\nu β = β part + c ν has m = c ν m=c\nu m = c ν .
Proposition 3.12 (Transformation of normalized orthogonal vectors). Assume that G G G and G ^ \widehat{G} G are invertible. Let Π \Pi Π be a row vector and Σ \Sigma Σ a column vector satisfying
β p a r t ( Π , g j ) = 0 ( j ∈ { 1 , … , d } ) , β p a r t ( f i , Σ ) = 0 ( i ∈ { 1 , … , d } ) . \beta_{\mathrm{part}}(\Pi,g_j)=0\quad(j\in\{1,\ldots,d\}),\qquad\beta_{\mathrm{part}}(f_i,\Sigma)=0\quad(i\in\{1,\ldots,d\}). β part ( Π , g j ) = 0 ( j ∈ { 1 , … , d }) , β part ( f i , Σ ) = 0 ( i ∈ { 1 , … , d }) . Put
f : = [ f i ] i ∈ { 1 , … , d } , g : = [ g 1 ⋯ g d ] , f:=[f_i]_{i\in\{1,\ldots,d\}},\qquad g:=[g_1\ \cdots\ g_d], f := [ f i ] i ∈ { 1 , … , d } , g := [ g 1 ⋯ g d ] , a Π : = [ m ( Π , g j ) ] j ∈ { 1 , … , d } , b Σ : = [ m ( f i , Σ ) ] i ∈ { 1 , … , d } , a_\Pi:=[m(\Pi,g_j)]_{j\in\{1,\ldots,d\}},\qquad b_\Sigma:=[m(f_i,\Sigma)]_{i\in\{1,\ldots,d\}}, a Π := [ m ( Π , g j ) ] j ∈ { 1 , … , d } , b Σ := [ m ( f i , Σ ) ] i ∈ { 1 , … , d } , where a Π a_\Pi a Π is a 1 × d 1\times d 1 × d row and b Σ b_\Sigma b Σ a d × 1 d\times1 d × 1 column. Then the unique transformed vectors with the same classes modulo E E E and H H H , respectively, are
Π ^ : = Π − a Π G ^ − 1 f , Σ ^ : = Σ − g G ^ − 1 b Σ . (46) \widehat{\Pi}:=\Pi-a_\Pi\widehat{G}^{-1}f,\qquad\widehat{\Sigma}:=\Sigma-g\widehat{G}^{-1}b_\Sigma. \tag*{(46)} Π := Π − a Π G − 1 f , Σ := Σ − g G − 1 b Σ . (46) They satisfy
β ^ ( Π ^ , H ) = 0 , β ^ ( E , Σ ^ ) = 0. \widehat{\beta}(\widehat{\Pi},H)=0,\qquad\widehat{\beta}(E,\widehat{\Sigma})=0. β ( Π , H ) = 0 , β ( E , Σ ) = 0. In particular, any prescribed leading coefficients encoded by those quotient classes are preserved.
Proof. All entries of the Gram matrices and of a Π , b Σ a_\Pi,b_\Sigma a Π , b Σ are complex scalars. In particular, the row correction in (3.38) means
a Π G ^ − 1 f = ∑ i , j = 1 d m ( Π , g j ) ( G ^ − 1 ) j , i f i . a_\Pi\widehat{G}^{-1}f=\sum_{i,j=1}^{d}m(\Pi,g_j)(\widehat{G}^{-1})_{j,i}f_i. a Π G − 1 f = i , j = 1 ∑ d m ( Π , g j ) ( G − 1 ) j , i f i .
Pairing this identity with g k g_k g k gives
β ^ ( Π ^ , g k ) = m ( Π , g k ) − ∑ i , j = 1 d m ( Π , g j ) ( G ^ − 1 ) j , i G ^ i , k = 0. \widehat{\beta}(\widehat{\Pi},g_k)=m(\Pi,g_k)-\sum_{i,j=1}^{d}m(\Pi,g_j)(\widehat{G}^{-1})_{j,i}\widehat{G}_{i,k}=0. β ( Π , g k ) = m ( Π , g k ) − i , j = 1 ∑ d m ( Π , g j ) ( G − 1 ) j , i G i , k = 0.
We used the original orthogonality of Π \Pi Π in the first term and G ^ − 1 G ^ = I d \widehat{G}^{-1}\widehat{G}=I_d G − 1 G = I d in the second. For the column correction, its pairing with f k f_k f k is
β ^ ( f k , Σ ^ ) = m ( f k , Σ ) − ∑ i , j = 1 d G ^ k , i ( G ^ − 1 ) i , j m ( f j , Σ ) = 0 , \widehat{\beta}(f_k,\widehat{\Sigma})=m(f_k,\Sigma)-\sum_{i,j=1}^{d}\widehat{G}_{k,i}(\widehat{G}^{-1})_{i,j}m(f_j,\Sigma)=0, β ( f k , Σ ) = m ( f k , Σ ) − i , j = 1 ∑ d G k , i ( G − 1 ) i , j m ( f j , Σ ) = 0 ,
using G ^ G ^ − 1 = I d \widehat{G}\widehat{G}^{-1}=I_d G G − 1 = I d . Thus the order of the inverse factors is verified separately for rows and columns. Both corrections lie in the corresponding lower spaces. They therefore preserve the prescribed classes. If two transformed rows had the same class, their difference would be ∑ i c i f i \sum_i c_i f_i ∑ i c i f i ; its pairings with all g j g_j g j would say [ c 1 ⋯ c d ] G ^ = 0 [c_1\ \cdots\ c_d]\widehat{G}=0 [ c 1 ⋯ c d ] G = 0 . Invertibility forces every c i c_i c i to vanish. For columns the same argument uses G ^ [ c 1 ⋯ c d ] T = 0 \widehat{G}[c_1\ \cdots\ c_d]^{\mathsf{T}}=0 G [ c 1 ⋯ c d ] T = 0 . This proves uniqueness. □
To obtain normalized blocks, take larger spaces E + ⊃ E E^+\supset E E + ⊃ E and H + ⊃ H H^+\supset H H + ⊃ H with equal dimension increments. Choose fixed complements with row matrix Π 0 \Pi^0 Π 0 and column matrix Σ 0 \Sigma^0 Σ 0 . If the common increment is a a a , these matrices have sizes a × q a\times q a × q and p × a p\times a p × a . Define
Π : = Π 0 − β p a r t ( Π 0 , g ) G − 1 f , \Pi:=\Pi^0-\beta_{\mathrm{part}}(\Pi^0,g)G^{-1}f, Π := Π 0 − β part ( Π 0 , g ) G − 1 f ,
Σ : = Σ 0 − g G − 1 β p a r t ( f , Σ 0 ) . \Sigma:=\Sigma^0-gG^{-1}\beta_{\mathrm{part}}(f,\Sigma^0). Σ := Σ 0 − g G − 1 β part ( f , Σ 0 ) .
The pairing tables before G − 1 G^{-1} G − 1 and after it have sizes a × d a\times d a × d and d × a d\times a d × a , respectively. Pairing the first equation with g g g subtracts its own table multiplied by G − 1 G = I d G^{-1}G=I_d G − 1 G = I d ; pairing the second with f f f uses G G − 1 = I d GG^{-1}=I_d G G − 1 = I d . Thus every row of Π \Pi Π and column of Σ \Sigma Σ satisfies the orthogonality hypotheses of Proposition 3.12, while the subtracted terms lie in the lower spaces and preserve the chosen classes. Applying its formulas to each of them gives blocks Π ^ \widehat{\Pi} Π and Σ ^ \widehat{\Sigma} Σ with the chosen leading classes. If the transformed pairing is also normal on E + × H + E^{+} \times H^{+} E + × H + , then
S ^ : = β ^ ( Π ^ , Σ ^ ) \widehat{S}:=\widehat{\beta}(\widehat{\Pi},\widehat{\Sigma}) S := β ( Π , Σ )
is invertible. Indeed, in the completed bases the Gram matrix is
[ β ^ ( f , g ) β ^ ( f , Σ ^ ) β ^ ( Π ^ , g ) β ^ ( Π ^ , Σ ^ ) ] = [ G ^ 0 0 S ^ ] . \begin{bmatrix}
\widehat{\beta}(f,g) & \widehat{\beta}(f,\widehat{\Sigma}) \\
\widehat{\beta}(\widehat{\Pi},g) & \widehat{\beta}(\widehat{\Pi},\widehat{\Sigma})
\end{bmatrix}
=
\begin{bmatrix}
\widehat{G} & 0 \\
0 & \widehat{S}
\end{bmatrix}. [ β ( f , g ) β ( Π , g ) β ( f , Σ ) β ( Π , Σ ) ] = [ G 0 0 S ] .
The two off-diagonal blocks vanish by the orthogonality just proved. Its determinant is det G ^ det S ^ \det\widehat{G}\det\widehat{S} det G det S ; normality on the larger spaces and det G ^ ≠ 0 \det\widehat{G} \ne0 det G = 0 force det S ^ ≠ 0 \det\widehat{S} \ne0 det S = 0 . Keeping Π ^ \widehat{\Pi} Π and replacing Σ ^ \widehat{\Sigma} Σ by Σ ^ S ^ − 1 \widehat{\Sigma}\widehat{S}^{-1} Σ S − 1 produces dual blocks whose mutual pairing is the identity. This last normalization changes the leading classes of the column block; it is a separate choice from prescribing both leading blocks. Iterating along nested normal cuts gives biorthogonal block families: every new row or column block is orthogonal to all preceding blocks on the opposite side.
The computation uses the same small matrix as the kernel correction. If H m : = [ m ( f i , g j ) ] H_m := [m(f_i,g_j)] H m := [ m ( f i , g j )] has a rank factorization H m = P m Q m H_m = P_mQ_m H m = P m Q m of rank ρ \rho ρ , substitution of the inverse formula in the proof of Proposition 3.11 into (3.38) uses I ρ + Q m G − 1 P m I_{\rho} + Q_mG^{-1}P_m I ρ + Q m G − 1 P m , exactly as in (3.36). When the mass has finite rank on the whole space, a Π a_{\Pi} a Π and b Σ b_{\Sigma} b Σ are obtained from its same finite remainder data, even as the polynomial spaces increase. This does not mean that only finitely many orthogonal polynomials change.
Finally, these formulas control parameter dependence. Suppose the spaces, bases and prescribed leading classes are fixed, while the moments of both pairings on the relevant enlarged spaces are holomorphic in a parameter near a given value. If the needed Gram determinants are nonzero at that value, their inverses are holomorphic in a neighbourhood, by the adjugate formula. The particular projections, the transformed vectors in (3.38), and the dual normalization by S ^ − 1 \widehat{S}^{-1} S − 1 are consequently holomorphic there. The same argument gives rational coefficients when the moments are rational functions of the parameter. The curve family in Section 5 applies this observation at a collision of the mass support with a singular point.
Residues and matrix division on curves and surfaces
The finite formulas of Section 3 require linear conditions that recognize polynomial images and values of masses on the remaining classes. This section constructs those data by residues. A quotient records the remainder after division. A perfect pairing on that quotient identifies every linear functional on its classes with one dual class. Choosing bases then turns the pairing into a finite matrix, whose entries are the residues given below. Inserting these entries in the finite construction produces the normalized polynomials as well as their kernel.
The curve construction first supplies its geometric Cauchy kernel and then its matrix quotient pairing. The surface construction uses a second equation to make the quotient finite and proves its residue formula directly. The finite-dimensional update formulas apply to both.
The underlying finite linear algebra can be expressed in terms of remainder classes and their dual functionals. A class [ f ] [f] [ f ] in a quotient V / T V V/TV V / T V is the remainder information that survives adding T a Ta T a to f f f . A pairing b : Q × Q d → C b: Q \times Q^{\mathrm{d}} \to\mathbb{C} b : Q × Q d → C is perfect when every linear functional on either space is pairing with a unique element of the other. If G : = [ b ( e i , η j ) ] i , j ∈ { 1 , … , m } G := [b(e_i,\eta_j)]_{i,j\in\{1,\ldots,m\}} G := [ b ( e i , η j ) ] i , j ∈ { 1 , … , m } is its matrix in two bases, this means that G G G is square and invertible. The coevaluation tensor is the corresponding inverse written as
C : = ∑ i , j = 1 m ( G − T ) i , j e i ⊗ η j . \mathcal{C}:=\sum_{i,j=1}^{m}(G^{-\mathsf{T}})_{i,j}e_i\otimes\eta_j. C := i , j = 1 ∑ m ( G − T ) i , j e i ⊗ η j .
Indeed, contracting the second factor with e k e_k e k gives ∑ i , j ( G − T ) i , j G k , j e i = e k \sum_{i,j}(G^{-\mathsf{T}})_{i,j}G_{k,j}e_i=e_k ∑ i , j ( G − T ) i , j G k , j e i = e k . This calculation fixes our index order: the first tensor factor labels the rows, and the second labels the columns. The residue formulas below construct b b b and C \mathcal{C} C without first inverting G G G . Their value for the perturbation problem is that they turn divisibility and the remaining mass values into finite computations. The quotient is chosen for the equation under consideration: numerator-image conditions use the corresponding division map, whereas homogeneous masses use the denominator test map. The residue method applies separately to these choices.
The global rank of a mass pairing ν \nu ν is the dimension of the image of g ↦ [ f ↦ ν ( f , g ) ] g \mapsto[f \mapsto\nu(f,g)] g ↦ [ f ↦ ν ( f , g )] . Its kernel is the right radical, consisting of the columns pairing to zero with every row. Thus this global rank is the dimension of the column module modulo that radical. Every finite matrix of mass pairings has rank at most the global rank. The singular families below identify the finite cuts on which the full rank is attained.
A global Cauchy kernel on an integral curve
On a curve, division by a regular denominator leaves a finite-dimensional space of remainder classes. Residue duality represents their linear functionals and hence the values of a mass on the chosen polynomial or section bases. Adding these values to the particular pairing gives the Gram matrix used in Subsection 3.4 to compute the kernel and normalized polynomials. The construction below establishes this duality also when the support of the mass meets a singular point.
The geometric kernel S F S_F S F has a prescribed diagonal pole. Multiplication by T ( P ) − T ( Q ) T(P)-T(Q) T ( P ) − T ( Q ) produces a regular kernel C T \mathcal{C}_T C T whose class represents the inverse of a finite residue pairing. The finite reproducing kernel K β K_\beta K β then uses the full orthogonality pairing and its Gram inverse. The first two objects encode division and its remainder duality; the third reproduces the chosen polynomial or section space. Subsection 4.2.2 assembles these data explicitly. We then give a noncommutative family whose normality, mass rank and normalized polynomial blocks persist through confluence at a singular point, including non-Gorenstein examples.
The global construction of S F S_F S F ensures that multiplying it by a function vanishing on the diagonal produces a regular tensor even at a singular pair. This property will justify the matrix divided difference in Theorem 4.4 ; its residue normalization will give the reconstruction identity. A finite projection then supplies an expression in trace-dual frames and identifies the restriction on rational factors arising from a change of geometric data.
For concrete calculations, polynomial reduction provides representatives of the quotient classes; it may be carried out by module Gröbner bases or by a division adapted to the equations. The residue pairing provides the dual coordinates of those classes, and S F S_F S F gives their inverse tensor through a matrix divided difference. These descriptions can be used together. In Section 5, coefficient extraction computes the moments and Grams directly, while the explicit geometric kernel identifies the tensor reproducing the mass quotient.
The geometric kernel and the finite reproducing kernel
The finite reproducing kernel of Subsection 2.1.1 is determined by a pairing on two finite-dimensional spaces. Its coefficients contain the inverse Gram matrix. Changing the weight or the masses therefore changes this kernel. Where the basis sections are regular, the finite kernel is regular on the diagonal as well.
The geometric Szegő kernel has different defining data: a curve, a line bundle and a prescribed diagonal residue. It provides the analogue of the Cauchy kernel used to form transforms and residues. Once these geometric data are fixed, changing the orthogonality functional does not change this kernel. Bertola uses both objects: his Szegő kernel is normalized in [12 (2.22)] , whereas the correspondence in [12 Section 3.1] concerns finite Christoffel–Darboux kernels. The classical theta-function description on smooth complex curves appears in Fay [46 ] . A bundle formulation, including bundles of higher rank, and its residue interpretation are given in [18 Section 3.4 and Proposition 5.1.1] . For smooth projective curves over a finite field with a fixed point at infinity, Han and Yu [13 Theorems 6.2 and 6.3] identify the finite residual Casimir tensor with the intrinsic principal part of the normalized Szegő kernel and derive residue reproduction, after extension to their coefficient field. Thus the scalar smooth residue construction is an established antecedent of the result below. The present construction extends this quotient formula to integral singular curves, retains the dualizing module at non-Gorenstein points, and specifies the ordered matrix operations that enter the bilateral transformation of mixed-type orthogonality.
A kernel defined along the entire diagonal
Let C C C be an integral projective curve over C \mathbb{C} C : a compact algebraic curve with one irreducible component and no nilpotent functions in its structure sheaf O C \mathcal{O}_C O C . We allow singular points. For an open set U U U , O C ( U ) \mathcal{O}_C(U) O C ( U ) denotes the functions regular on U U U ; the local ring O C , s \mathcal{O}_{C,s} O C , s consists of their germs near s s s . Its dualizing sheaf ω C \omega_C ω C replaces the sheaf of regular differentials in Serre duality. It agrees with regular differentials on the smooth part of C C C , and it may fail to be a line bundle at a singular point. The normalization is the smooth curve mapping finitely to C C C that separates the local branches at its singularities. Locally the dualizing sheaf can be described there: a meromorphic differential η \eta η represents a regular dualizing differential at s s s if ∑ a ↦ s Res a ( g η ) = 0 \sum_{a\mapsto s}\operatorname{Res}_{a}(g\eta)=0 ∑ a ↦ s Res a ( g η ) = 0 for every g ∈ O C , s g\in\mathcal{O}_{C,s} g ∈ O C , s . The sum runs over the branches above s s s . This is Rosenlicht’s definition of regular differentials on a singular curve; see [45 Chapter IV] . Thus regularity includes the relations among the branches. Its identification with the dualizing sheaf is part of the local and global duality framework in [39 Introduction and Sections 5, 7–10] . A residue on a branch means the coefficient of t − 1 d t t^{-1}\mathrm{d}t t − 1 d t in a local parameter t t t . A differential that is regular in this dualizing sense may have poles on individual branches. For example, at a node the pair ( d u / u , − d v / v ) (\mathrm{d}u/u,-\mathrm{d}v/v) ( d u / u , − d v / v ) has this property: a regular function has the same value at the origins of the two branches, so the two residues of its product with this pair cancel. The point is called Gorenstein when the dualizing module has one local generator; the argument below also treats points where more than one generator is needed.
The global residue theorem in this convention follows from the usual theorem on the normalization. Indeed, let ϖ : C ~ → C \varpi:\widetilde{C}\to C ϖ : C → C be the normalization and let η \eta η be a meromorphic differential, regarded also as a differential on the smooth projective curve C ~ \widetilde{C} C . By the definition of the residue at a point of C C C ,
∑ s ∈ C Res s η = ∑ s ∈ C ∑ a ∈ ϖ − 1 ( s ) Res a η = ∑ a ∈ C ~ Res a η = 0. \sum_{s\in C}\operatorname{Res}_{s}\eta=\sum_{s\in C}\sum_{a\in\varpi^{-1}(s)}\operatorname{Res}_{a}\eta=\sum_{a\in\widetilde{C}}\operatorname{Res}_{a}\eta=0. s ∈ C ∑ Res s η = s ∈ C ∑ a ∈ ϖ − 1 ( s ) ∑ Res a η = a ∈ C ∑ Res a η = 0.
Only finitely many terms can be nonzero. At a point where η \eta η is regular in Rosenlicht’s sense, its contribution vanishes by taking g = 1 g=1 g = 1 in the local condition. This is the global residue identity used in Theorem 4.4 .
A line bundle has one local generator on each sufficiently small open set; a section is a regular function times that generator, with the prescribed changes of generator on overlaps. For a line bundle F F F , write
F D : = H o m O C ( F , ω C ) = F − 1 ⊗ ω C . F^{\mathrm{D}}:=\mathcal{H}om_{\mathcal{O}_{C}}(F,\omega_{C})=F^{-1}\otimes\omega_{C}. F D := H o m O C ( F , ω C ) = F − 1 ⊗ ω C .
The product of a section of F F F and a section of F D F^{\mathrm{D}} F D is thus a dualizing differential. We assume that F F F is acyclic , meaning
H 0 ( C , F ) = H 1 ( C , F ) = 0. (47) H^{0}(C,F)=H^{1}(C,F)=0. \tag*{(47)} H 0 ( C , F ) = H 1 ( C , F ) = 0. (47)
The first vanishing excludes nonzero global regular sections of F F F ; the second will remove the obstruction to prescribing the residue. Serre duality identifies H 0 ( C , F D ) H^{0}(C,F^{\mathrm{D}}) H 0 ( C , F D ) with H 1 ( C , F ) ∗ H^{1}(C,F)^{*} H 1 ( C , F ) ∗ and H 1 ( C , F D ) H^{1}(C,F^{\mathrm{D}}) H 1 ( C , F D ) with H 0 ( C , F ) ∗ H^{0}(C,F)^{*} H 0 ( C , F ) ∗ . Thus both groups for F D F^{\mathrm{D}} F D vanish as well.
Let p 1 , p 2 : C × C → C p_{1},p_{2}:C\times C\to C p 1 , p 2 : C × C → C be the two projections and let i : C → C × C i:C\to C\times C i : C → C × C be the diagonal map, i ( P ) = ( P , P ) i(P)=(P,P) i ( P ) = ( P , P ) . Set
J : = J Δ , M F : = F ⊠ F D = p 1 ∗ F ⊗ p 2 ∗ F D . J:=\mathcal{J}_{\Delta},\qquad\mathcal{M}_{F}:=F\boxtimes F^{\mathrm{D}}=p_{1}^{*}F\otimes p_{2}^{*}F^{\mathrm{D}}. J := J Δ , M F := F ⊠ F D = p 1 ∗ F ⊗ p 2 ∗ F D .
The diagonal ideal J J J consists of functions vanishing when the two arguments coincide. On an affine product it is generated by the differences a ( P ) − a ( Q ) a(P)-a(Q) a ( P ) − a ( Q ) of regular functions on the curve.
On a smooth curve, a Cauchy kernel is a section of M F \mathcal{M}_{F} M F with a simple pole on Δ \Delta Δ . At a singular point the diagonal ideal need not have one generator, so we use the sheaf H o m ( J , M F ) \mathcal{H}om(J,\mathcal{M}_{F}) H o m ( J , M F ) . Its sections are locally linear maps from the diagonal ideal to M F \mathcal{M}_{F} M F . Equivalently, one may view such a section as a meromorphic kernel S S S for which j S jS j S is regular for every j ∈ J j\in J j ∈ J . Away from the diagonal, J = O C × C J=\mathcal{O}_{C\times C} J = O C × C , so this condition makes S S S regular there. On the smooth part it is precisely the usual simple-pole condition. The next exact sequence describes this pole condition in elementary terms: its kernel consists of regular kernels, and its last term records the one freely prescribed residue along the diagonal.
Proposition 4.1 (The diagonal residue sequence). There is an exact sequence
0 ⟶ M F ⟶ H o m ( J , M F ) → res Δ O Δ ⟶ 0. (48) 0\longrightarrow\mathcal{M}_{F}\longrightarrow\mathcal{H}om(J,\mathcal{M}_{F})\xrightarrow{\operatorname{res}_{\Delta}}\mathcal{O}_{\Delta}\longrightarrow0. \tag*{(48)} 0 ⟶ M F ⟶ H o m ( J , M F ) res Δ O Δ ⟶ 0. (48)
Under (47) , there is a unique global section S F S_{F} S F with res Δ S F = 1 \operatorname{res}_{\Delta}S_{F}=1 res Δ S F = 1 . The sign of the residue map is chosen so that, at a smooth point, in a local coordinate z z z and a frame $e of F F F ,
S F ( P , Q ) = e ( P ) e ( Q ) − 1 d z ( Q ) z ( P ) − z ( Q ) + a a section regular near the diagonal . (49) S_{F}(P,Q)=\frac{e(P)e(Q)^{-1}\mathrm{d}z(Q)}{z(P)-z(Q)}+a\text{ a section regular near the diagonal}. \tag*{(49)} S F ( P , Q ) = z ( P ) − z ( Q ) e ( P ) e ( Q ) − 1 d z ( Q ) + a a section regular near the diagonal . (49)
Thus Res Q = P S F ( P , Q ) f ( Q ) = − f ( P ) \operatorname{Res}_{Q=P}S_{F}(P,Q)f(Q)=-f(P) Res Q = P S F ( P , Q ) f ( Q ) = − f ( P ) for a local regular section f f f of F F F . Proof. An integral noetherian curve is Cohen–Macaulay: its local rings have dimension at most one, and a nonzero element in the maximal ideal of a one-dimensional local domain is a nonzerodivisor. This permits the use of its dualizing sheaf also at a non-Gorenstein point. We first compute the local quotient that records the diagonal residue. We then check that its identification with functions on the diagonal is independent of the coordinates. The final cohomology calculation constructs the global kernel.
Choose an affine neighbourhood U = Spec A U = \operatorname{Spec} A U = Spec A on which F F F is trivial; A A A is its algebra of regular functions. A Noether normalization chooses a nonconstant function x x x such that A A A is a finite module over C [ x ] \mathbb{C}[x] C [ x ] . Geometrically this is a finite map from U U U to the affine line. It is torsion free over this polynomial ring, since A A A is a domain, and hence it is a finite free module. Denote the image of x x x in A A A by a a a , and put
R : = A ⊗ C A , B : = A ⊗ C C [ x ] = A [ x ] . R := A \otimes_{\mathbb{C}} A,\qquad B := A \otimes_{\mathbb{C}} \mathbb{C}[x] = A[x]. R := A ⊗ C A , B := A ⊗ C C [ x ] = A [ x ] .
The second factor C [ x ] \mathbb{C}[x] C [ x ] is included in the second copy of A A A . The diagonal is the R R R -module A A A obtained by multiplying the two factors. As a B B B -module, it is B / ( x − a ) B/(x-a) B / ( x − a ) . Consequently
0 ⟶ B → x − a B ⟶ A ⟶ 0 (50) 0 \longrightarrow B \xrightarrow{x-a} B \longrightarrow A \longrightarrow0 \tag*{(50)} 0 ⟶ B x − a B ⟶ A ⟶ 0 (50)
is exact. Injectivity follows by comparing the highest coefficient of a polynomial multiplied by the monic polynomial x − a x-a x − a .
Here is the finite-duality identification used in the computation, including its normalization. Let K A K_A K A be the fraction field of A A A . A meromorphic differential can be written u d x u\,\mathrm{d}x u d x with u ∈ K A u \in K_A u ∈ K A . The description of regular dualizing differentials by the trace condition is
ω A = { u d x : Tr K A / C ( x ) ( b u ) ∈ C [ x ] for every b ∈ A } . \omega_A = \{u\,\mathrm{d}x : \operatorname{Tr}_{K_A/\mathbb{C}(x)}(bu) \in\mathbb{C}[x]\ \text{for every } b \in A\}. ω A = { u d x : Tr K A / C ( x ) ( b u ) ∈ C [ x ] for every b ∈ A } .
It identifies the differential u d x u\,\mathrm{d}x u d x with the functional b ↦ Tr K A / C ( x ) ( b u ) d x b \mapsto\operatorname{Tr}_{K_A/\mathbb{C}(x)}(bu)\,\mathrm{d}x b ↦ Tr K A / C ( x ) ( b u ) d x in Hom C [ x ] ( A , C [ x ] ) d x \operatorname{Hom}_{\mathbb{C}[x]}(A,\mathbb{C}[x])\mathrm{d}x Hom C [ x ] ( A , C [ x ]) d x ; see [39 Definition 2.1, Example 2.1.2 and Lemma 2.2] . The trace pairing on the finite separable field extension is nondegenerate. Therefore every functional on A A A extends to the fraction fields and has a unique expression of this form. Its values lie in C [ x ] d x \mathbb{C}[x]\mathrm{d}x C [ x ] d x precisely when the displayed condition holds. This also fixes the identification on the smooth part: for an unramified local branch the differential trace is the ordinary trace of the branch coefficients.
After using d x \mathrm{d}x d x as a basis of the differentials on the base, finite freeness gives
A ⊗ C ω A ≃ Hom B ( R , B ) . A \otimes_{\mathbb{C}} \omega_A \simeq\operatorname{Hom}_B(R,B). A ⊗ C ω A ≃ Hom B ( R , B ) .
For clarity, the scalar-extension map sends c ⊗ ℓ c \otimes\ell c ⊗ ℓ to the B B B -linear map c ′ ⊗ b ↦ c c ′ ℓ ( b ) c' \otimes b \mapsto cc'\ell(b) c ′ ⊗ b ↦ c c ′ ℓ ( b ) ; it is an isomorphism because a basis of the second copy of A A A over C [ x ] \mathbb{C}[x] C [ x ] and its dual basis remain dual after extending coefficients to the first copy.
The module R R R is finite free over B B B . For an R R R -module P P P , the adjunction needed here is the explicit isomorphism
Hom R ( P , Hom B ( R , B ) ) ⟶ Hom B ( P , B ) , Φ ⟼ [ p ↦ Φ ( p ) ( 1 ) ] . \operatorname{Hom}_R(P,\operatorname{Hom}_B(R,B)) \longrightarrow\operatorname{Hom}_B(P,B),\qquad\Phi\longmapsto[p \mapsto\Phi(p)(1)]. Hom R ( P , Hom B ( R , B )) ⟶ Hom B ( P , B ) , Φ ⟼ [ p ↦ Φ ( p ) ( 1 )] .
Its inverse sends λ \lambda λ to the map p ↦ [ r ↦ λ ( r p ) ] p \mapsto[r \mapsto\lambda(rp)] p ↦ [ r ↦ λ ( r p )] . The R R R -action on Hom B ( R , B ) \operatorname{Hom}_B(R,B) Hom B ( R , B ) is ( r ℓ ) ( s ) = ℓ ( r s ) (r\ell)(s) = \ell(rs) ( r ℓ ) ( s ) = ℓ ( r s ) , so both formulas are module maps and their two compositions are the identity. A projective resolution over R R R is also projective over B B B : each term is a direct summand of a free R R R -module, and R R R is free over B B B . Applying these inverse maps to the terms and differentials of the resolution therefore identifies the corresponding Ext groups. Applied to the diagonal module and (4.4), it yields
Ext R 1 ( A , A ⊗ C ω A ) ≃ A , Ext R j ( A , A ⊗ C ω A ) = 0 ( j ≠ 1 ) . \operatorname{Ext}^1_R(A,A \otimes_{\mathbb{C}} \omega_A) \simeq A,\qquad\operatorname{Ext}^j_R(A,A \otimes_{\mathbb{C}} \omega_A) = 0\quad(j \ne1). Ext R 1 ( A , A ⊗ C ω A ) ≃ A , Ext R j ( A , A ⊗ C ω A ) = 0 ( j = 1 ) .
More explicitly, applying Hom B ( − , B ) \operatorname{Hom}_B(-,B) Hom B ( − , B ) to the two free terms of (4.4) gives the complex B → x − a B B \xrightarrow{x-a} B B x − a B . Its degree-zero kernel is zero and its degree-one quotient is B / ( x − a ) = A B/(x-a)=A B / ( x − a ) = A ; there are no further terms. Thus Ext 1 \operatorname{Ext}^1 Ext 1 records exactly one copy of the functions on the diagonal. In the residue sequence it measures the failure of a map on the diagonal ideal to extend to a map on the whole ring. The element x − a x-a x − a belongs to the auxiliary ring B B B ; this calculation does not require the diagonal ideal in R R R to have one generator.
We next verify compatibility on overlapping neighbourhoods. Use the preceding trace normalization and fix the sign of the Ext identification so that, wherever x x x is a local coordinate, the class of d x ( Q ) / ( x ( P ) − x ( Q ) ) \mathrm{d}x(Q)/(x(P)-x(Q)) d x ( Q ) / ( x ( P ) − x ( Q )) has value 1. This assertion can be checked directly on the resolution. At a smooth point where the finite map is unramified, the completed local rings of the branch and the line are both C [ [ x − x 0 ] ] \mathbb{C}[[x-x_0]] C [[ x − x 0 ]] . In the completion along the corresponding pair of points the trace on that branch is the identity, so the adjunction above reduces to the two-term calculation for the principal diagonal equation x ( Q ) − x ( P ) x(Q)-x(P) x ( Q ) − x ( P ) . Its connecting map takes the class of b ( P , Q ) d x ( Q ) / ( x ( Q ) − x ( P ) ) b(P,Q)\,dx(Q)/(x(Q)-x(P)) b ( P , Q ) d x ( Q ) / ( x ( Q ) − x ( P )) to b ( P , P ) b(P,P) b ( P , P ) . Taking the negative of this map gives precisely the stated normalization with denominator x ( P ) − x ( Q ) x(P)-x(Q) x ( P ) − x ( Q ) .
This normalized value is unchanged if another local coordinate z z z is used. Indeed,
x ( P ) − x ( Q ) = u ( P , Q ) ( z ( P ) − z ( Q ) ) , u ( P , P ) = d x d z ( P ) ≠ 0. x(P)-x(Q)=u(P,Q)(z(P)-z(Q)),\qquad u(P,P)=\frac{dx}{dz}(P)\ne0. x ( P ) − x ( Q ) = u ( P , Q ) ( z ( P ) − z ( Q )) , u ( P , P ) = d z d x ( P ) = 0.
The numerator d x ( Q ) dx(Q) d x ( Q ) contributes the same factor d x / d z dx/dz d x / d z on the diagonal, so it cancels the factor u ( P , P ) u(P,P) u ( P , P ) in the denominator. Changing a frame e e e of F F F to v e ve v e likewise multiplies the kernel by v ( P ) / v ( Q ) v(P)/v(Q) v ( P ) / v ( Q ) , whose restriction to the diagonal is 1 1 1 .
For two choices of finite coordinate, these calculations apply on a common dense open subset of their overlap: in characteristic zero each finite map is unramified away from finitely many points, and the smooth points form a dense open subset. The local Ext calculation makes the degree-one Ext sheaf locally free of rank one as a sheaf on the diagonal. The two normalized maps from that sheaf to O Δ \mathcal{O}_{\Delta} O Δ agree on this dense open subset. After a local trivialization their difference is multiplication by a regular function vanishing on a dense open subset of the integral curve. That function is zero. Hence the maps agree on the whole overlap, including its singular points, and they glue.
The glued local calculations can also be recorded as
R H o m C × C ( O Δ , p 2 ∗ ω C ) ≃ i ∗ O C [ − 1 ] . \mathop{\mathcal{R}\!Hom}_{C\times C}(\mathcal{O}_{\Delta},p_2^*\omega_C)\simeq i_*\mathcal{O}_C[-1]. R H o m C × C ( O Δ , p 2 ∗ ω C ) ≃ i ∗ O C [ − 1 ] .
Here derived Hom records all the Ext sheaves at once, and the shift [ − 1 ] [-1] [ − 1 ] places its sole nonzero sheaf in degree one. Tensoring by the line bundle p 1 ∗ F ⊗ p 2 ∗ F − 1 p_1^*F\otimes p_2^*F^{-1} p 1 ∗ F ⊗ p 2 ∗ F − 1 preserves this formula: its restriction to the diagonal is canonically trivial. We obtain, in particular,
H o m ( O Δ , M F ) = 0 , E x t 1 ( O Δ , M F ) ≃ O Δ . \mathop{\mathcal{H}\!om}(\mathcal{O}_{\Delta},\mathcal{M}_F)=0,\qquad\mathop{\mathcal{E}\!xt}\nolimits^1(\mathcal{O}_{\Delta},\mathcal{M}_F)\simeq\mathcal{O}_{\Delta}. H o m ( O Δ , M F ) = 0 , E x t 1 ( O Δ , M F ) ≃ O Δ .
Thus the residue identification is intrinsic. The coordinate calculation and its compatibility establish here the diagonal case of relative duality directly.
Apply H o m ( − , M F ) \mathop{\mathcal{H}\!om}(-,\mathcal{M}_F) H o m ( − , M F ) to 0 → J → O C × C → O Δ → 0 0\to J\to\mathcal{O}_{C\times C}\to\mathcal{O}_{\Delta}\to0 0 → J → O C × C → O Δ → 0 . The two displayed identifications, together with E x t 1 ( O C × C , M F ) = 0 \mathop{\mathcal{E}\!xt}\nolimits^1(\mathcal{O}_{C\times C},\mathcal{M}_F)=0 E x t 1 ( O C × C , M F ) = 0 , give (4.2). Its first map sends a regular section m m m to the multiplication map j ↦ j m j\mapsto jm j ↦ j m . On a smooth coordinate neighbourhood, J J J is generated by z ( P ) − z ( Q ) z(P)-z(Q) z ( P ) − z ( Q ) . Our choice of sign gives
res Δ ( b ( P , Q ) e ( P ) e ( Q ) − 1 d z ( Q ) z ( P ) − z ( Q ) ) = b ( P , P ) . \operatorname{res}_{\Delta}\left(\frac{b(P,Q)e(P)e(Q)^{-1}dz(Q)}{z(P)-z(Q)}\right)=b(P,P). res Δ ( z ( P ) − z ( Q ) b ( P , Q ) e ( P ) e ( Q ) − 1 d z ( Q ) ) = b ( P , P ) .
This is the normalization in (4.3).
Finally, Künneth and (4.1) imply H 0 ( C × C , M F ) = H 1 ( C × C , M F ) = 0 H^0(C\times C,\mathcal{M}_F)=H^1(C\times C,\mathcal{M}_F)=0 H 0 ( C × C , M F ) = H 1 ( C × C , M F ) = 0 : the first group is H 0 ( F ) ⊗ H 0 ( F D ) H^0(F)\otimes H^0(F^D) H 0 ( F ) ⊗ H 0 ( F D ) , and the second is [ H 1 ( F ) ⊗ H 0 ( F D ) ] ⊕ [ H 0 ( F ) ⊗ H 1 ( F D ) ] [H^1(F)\otimes H^0(F^D)]\oplus[H^0(F)\otimes H^1(F^D)] [ H 1 ( F ) ⊗ H 0 ( F D )] ⊕ [ H 0 ( F ) ⊗ H 1 ( F D )] . Every summand is zero. The cohomology sequence of (4.2) now makes the residue map an isomorphism
H 0 ( C × C , H o m ( J , M F ) ) ⟶ H 0 ( C , O C ) = C . H^0(C\times C,\mathop{\mathcal{H}\!om}(J,\mathcal{M}_F))\longrightarrow H^0(C,\mathcal{O}_C)=\mathbb{C}. H 0 ( C × C , H o m ( J , M F )) ⟶ H 0 ( C , O C ) = C .
The preimage of 1 1 1 is the required section S F S_F S F .
The proposition constructs S F S_F S F also at ( s , s ) (s,s) ( s , s ) when s s s is singular. Its residue there is the quotient map in (4.2). In particular, the construction retains ω C \omega_C ω C and the full ideal J J J at that point. On a smooth curve it recovers the usual Szegő kernel by uniqueness. The proof is the residue construction of that kernel, with relative duality providing its formulation on the singular curve.
An expression through a finite projection. Let π : C → P 1 \pi:C\to\mathbb{P}^1 π : C → P 1 be finite of degree b b b . The direct image E : = π ∗ F E:=\pi_*F E := π ∗ F assigns to an open set V V V of the base the sections F ( π − 1 ( V ) ) F(\pi^{-1}(V)) F ( π − 1 ( V )) . It is a vector bundle of rank b b b : locally on the smooth base it is a finite torsion-free module over a discrete valuation ring and is therefore free. Finite duality gives
π ∗ F D ≃ H o m P 1 ( E , ω P 1 ) . \pi_*F^D\simeq\mathop{\mathcal{H}\!om}_{\mathbb{P}^1}(E,\omega_{\mathbb{P}^1}). π ∗ F D ≃ H o m P 1 ( E , ω P 1 ) .
The evaluation associated with this duality is denoted T π : π ∗ ω C → ω P 1 T_\pi:\pi_*\omega_C\to\omega_{\mathbb{P}^1} T π : π ∗ ω C → ω P 1 . It is the trace on dualizing differentials. In local module terms it sends a functional ℓ ∈ Hom C [ x ] ( A , C [ x ] ) d x \ell\in\operatorname{Hom}_{\mathbb{C}[x]}(A,\mathbb{C}[x])dx ℓ ∈ Hom C [ x ] ( A , C [ x ]) d x to ℓ ( 1 ) \ell(1) ℓ ( 1 ) . This formulation remains meaningful when several points of a fibre meet or a fibre is ramified.
Since π \pi π is finite, the cohomology of E E E is the cohomology of F F F . Write its splitting on P 1 \mathbb{P}^{1} P 1 as E ≃ ⨁ α = 1 b O ( a α ) E \simeq\bigoplus_{\alpha=1}^{b} \mathcal{O}(a_{\alpha}) E ≃ ⨁ α = 1 b O ( a α ) . The vanishing of H 0 ( E ) H^{0}(E) H 0 ( E ) forces a α ≤ − 1 a_{\alpha} \leq-1 a α ≤ − 1 , whereas the vanishing of H 1 ( E ) H^{1}(E) H 1 ( E ) forces a α ≥ − 1 a_{\alpha} \geq-1 a α ≥ − 1 . Hence
E ≃ O ( − 1 ) b , π ∗ F D ≃ O ( − 1 ) b . (51) E \simeq\mathcal{O}(-1)^{b}, \qquad\pi_{*}F^{\mathrm{D}} \simeq\mathcal{O}(-1)^{b}. \tag*{(51)} E ≃ O ( − 1 ) b , π ∗ F D ≃ O ( − 1 ) b . (51)
Here O ( a ) \mathcal{O}(a) O ( a ) is the line bundle of degree a a a on the projective line. The numerical conditions follow from dim H 0 ( O ( a ) ) = max { a + 1 , 0 } \dim H^{0}(\mathcal{O}(a)) = \max\{a+1,0\} dim H 0 ( O ( a )) = max { a + 1 , 0 } and dim H 1 ( O ( a ) ) = max { − a − 1 , 0 } \dim H^{1}(\mathcal{O}(a)) = \max\{-a-1,0\} dim H 1 ( O ( a )) = max { − a − 1 , 0 } . Fix such a splitting. On the affine base with coordinate x x x , it gives a frame e 1 , … , e b e_{1},\ldots,e_{b} e 1 , … , e b of E E E and a trace-dual frame η 1 , … , η b \eta_{1},\ldots,\eta_{b} η 1 , … , η b of π ∗ F D \pi_{*}F^{\mathrm{D}} π ∗ F D characterized by
Tr π ( e α η β ) = δ α , β d x . (52) \operatorname{Tr}_{\pi}(e_{\alpha}\eta_{\beta})=\delta_{\alpha,\beta}\,dx. \tag*{(52)} Tr π ( e α η β ) = δ α , β d x . (52)
Thus the e α e_{\alpha} e α are sections of F F F above this affine chart, while the η α \eta_{\alpha} η α are sections of F D F^{\mathrm{D}} F D .
Proposition 4.2 (Projection formula for the geometric kernel). With frames compatible with (4.5), the kernel of Proposition 4.1 is
S F ( P , Q ) = ∑ α = 1 b e α ( P ) η α ( Q ) x ( P ) − x ( Q ) . (53) S_{F}(P,Q)=\frac{\sum_{\alpha=1}^{b}e_{\alpha}(P)\eta_{\alpha}(Q)}{x(P)-x(Q)}. \tag*{(53)} S F ( P , Q ) = x ( P ) − x ( Q ) ∑ α = 1 b e α ( P ) η α ( Q ) . (53) The expression defines a section of H o m ( J , M F ) \mathcal{H}om(J,\mathcal{M}_{F}) H o m ( J , M F ) through ramified fibres and at singular points. It has no poles away from the diagonal.
Proof. Write d : = x ( P ) − x ( Q ) d:=x(P)-x(Q) d := x ( P ) − x ( Q ) and C π ( P , Q ) : = ∑ α e α ( P ) η α ( Q ) \mathcal{C}_{\pi}(P,Q):=\sum_{\alpha}e_{\alpha}(P)\eta_{\alpha}(Q) C π ( P , Q ) := ∑ α e α ( P ) η α ( Q ) . Modulo d d d , this numerator is the coevaluation tensor for (4.6): it represents the identity map of the finite free module E E E over C [ x ] \mathbb{C}[x] C [ x ] . For every regular function a a a on π − 1 ( A 1 ) \pi^{-1}(A^{1}) π − 1 ( A 1 ) , multiplication by a a a is self-adjoint for the evaluation pairing between E E E and its dual. The identity tensor consequently satisfies
( a ( P ) − a ( Q ) ) C π ( P , Q ) ∈ d Γ ( π − 1 ( A 1 ) 2 , M F ) . (a(P)-a(Q))\mathcal{C}_{\pi}(P,Q)\in d\,\Gamma(\pi^{-1}(A^{1})^{2},\mathcal{M}_{F}). ( a ( P ) − a ( Q )) C π ( P , Q ) ∈ d Γ ( π − 1 ( A 1 ) 2 , M F ) .
One can verify this statement directly: applying multiplication by a a a to the first tensor factor represents the map m ↦ a m m\mapsto am m ↦ am ; applying its dual to the second represents the same map. In coordinates, if a e α = ∑ β A β , α e β ae_{\alpha}=\sum_{\beta}A_{\beta,\alpha}e_{\beta} a e α = ∑ β A β , α e β , trace duality gives a η β = ∑ α A β , α η α a\eta_{\beta}=\sum_{\alpha}A_{\beta,\alpha}\eta_{\alpha} a η β = ∑ α A β , α η α . Thus both actions on ∑ α e α ⊗ η α \sum_{\alpha}e_{\alpha}\otimes\eta_{\alpha} ∑ α e α ⊗ η α give ∑ α , β A β , α e β ⊗ η α \sum_{\alpha,\beta}A_{\beta,\alpha}e_{\beta}\otimes\eta_{\alpha} ∑ α , β A β , α e β ⊗ η α after identifying the two base coordinates. This is the asserted divisibility by d d d .
The differences a ( P ) − a ( Q ) a(P)-a(Q) a ( P ) − a ( Q ) generate J J J . It follows that j C π / d j\mathcal{C}_{\pi}/d j C π / d is regular for every j ∈ J j\in J j ∈ J . Multiplication by d d d is injective on M F \mathcal{M}_{F} M F on this affine product. For example, its module is finite free over C [ x ( P ) , x ( Q ) ] \mathbb{C}[x(P),x(Q)] C [ x ( P ) , x ( Q )] by (4.5). Thus the quotient determines an unambiguous map J → M F J\to\mathcal{M}_{F} J → M F . This proves the required membership in the Hom sheaf. It also explains the cancellation at distinct points with x ( P ) = x ( Q ) x(P)=x(Q) x ( P ) = x ( Q ) : off the diagonal, J J J is the unit ideal, so the kernel itself is regular. The argument uses the whole finite module, and does not require a fibre to consist of distinct points.
On a smooth unramified fibre, trace duality is ordinary duality on the direct sum of the one-dimensional fibres. It gives C π ( P , P ) = d x ( P ) \mathcal{C}_{\pi}(P,P)=dx(P) C π ( P , P ) = d x ( P ) , with the F F F factors evaluated. The quotient in (4.7) therefore has diagonal residue 1 1 1 . Its residue is a regular function on the integral diagonal, so this equality extends to the ramified and singular points as well.
It remains to check the points above infinity. Write e 0 , α , η 0 , α e_{0,\alpha},\eta_{0,\alpha} e 0 , α , η 0 , α for the affine frames already used, and set w = 1 / x w=1/x w = 1/ x . For compatible frames on the two charts, the transition of O ( − 1 ) \mathcal{O}(-1) O ( − 1 ) and the equality d x = − x 2 d w dx=-x^{2}dw d x = − x 2 d w give
e 0 , α = x e ∞ , α , η 0 , α = − x η ∞ , α . e_{0,\alpha}=xe_{\infty,\alpha},\qquad\eta_{0,\alpha}=-x\eta_{\infty,\alpha}. e 0 , α = x e ∞ , α , η 0 , α = − x η ∞ , α .
The numerator and denominator in (4.7) both acquire the factor − x ( P ) x ( Q ) -x(P)x(Q) − x ( P ) x ( Q ) . Their quotient is therefore C π , ∞ ( P , Q ) / ( w ( P ) − w ( Q ) ) \mathcal{C}_{\pi,\infty}(P,Q)/(w(P)-w(Q)) C π , ∞ ( P , Q ) / ( w ( P ) − w ( Q )) on the other chart. On the two crossed charts it takes the forms
∑ α e 0 , α ( P ) η ∞ , α ( Q ) 1 − x ( P ) w ( Q ) , ∑ α e ∞ , α ( P ) η 0 , α ( Q ) 1 − w ( P ) x ( Q ) . \frac{\sum_{\alpha}e_{0,\alpha}(P)\eta_{\infty,\alpha}(Q)}{1-x(P)w(Q)},\qquad
\frac{\sum_{\alpha}e_{\infty,\alpha}(P)\eta_{0,\alpha}(Q)}{1-w(P)x(Q)}. 1 − x ( P ) w ( Q ) ∑ α e 0 , α ( P ) η ∞ , α ( Q ) , 1 − w ( P ) x ( Q ) ∑ α e ∞ , α ( P ) η 0 , α ( Q ) .
These expressions are regular near the corresponding points at infinity. When both points lie above infinity, the same ideal argument applies with w w w in place of x x x . We have obtained a global section with residue 1 1 1 , which equals S F S_{F} S F by uniqueness. □ \square □
The trace-dual frames used here come from the global splitting (51) . Changes between such frames are constant matrices, and the numerator in (53) is invariant under their dual changes. An arbitrary nonconstant change of affine frame also changes the behaviour at infinity and cannot be substituted into this assertion. Independence of the chosen projection follows from the global uniqueness of S F S_F S F .
Geometric changes and bilateral changes of the weight.
For the matrix division below, fix an affine open U ⊂ C U\subset C U ⊂ C and use the notation
A : = O C ( U ) , M : = F ( U ) , N : = F D ( U ) . (54) A:=\mathcal{O}_C(U),\qquad M:=F(U),\qquad N:=F^D(U). \tag*{(54)} A := O C ( U ) , M := F ( U ) , N := F D ( U ) . (54)
Here M M M is a projective module of rank one and N N N is its dual with values in ω A \omega_A ω A . Their evaluation takes values in ω A \omega_A ω A , not necessarily in A A A . The kernel S F S_F S F remains fixed. Multiplying it by differences of matrix entries will produce regular tensors for finite quotient pairings, while the orthogonality functional determines the separate finite reproducing kernel.
Matrix quotient duality and bilateral masses
Retain the curve, line bundle and affine modules of (54) , and write ω : = ω C ( U ) \omega:=\omega_C(U) ω := ω C ( U ) . We use the normalization (49) , so that Res Q = P S F ( P , Q ) f ( Q ) = − f ( P ) \operatorname{Res}_{Q=P}S_F(P,Q)f(Q)=-f(P) Res Q = P S F ( P , Q ) f ( Q ) = − f ( P ) for a local section f f f of F F F at a smooth point. Recall that N = Hom A ( M , ω ) N=\operatorname{Hom}_A(M,\omega) N = Hom A ( M , ω ) ; its elements already carry their canonical differential factor.
For a scalar denominator h h h , multiplying S F ( P , Q ) S_F(P,Q) S F ( P , Q ) by h ( P ) − h ( Q ) h(P)-h(Q) h ( P ) − h ( Q ) removes its diagonal singularity. The resulting tensor encodes the remainder classes modulo h h h . The same construction works for a matrix denominator, provided that the inverse in the residue pairing and the transpose in the dual quotient are placed in the correct order.
Theorem 4.4 (Residue kernel of a matrix quotient). Let T ∈ Mat n ( A ) T\in\operatorname{Mat}_n(A) T ∈ Mat n ( A ) have nonzero determinant, and let Z ⊂ U Z\subset U Z ⊂ U be the finite set of zeros of det T \det T det T . Define
Q T : = M n / T M n , Q T D : = N n / T N n . Q_T:=M^n/TM^n,\qquad Q_T^D:=N^n/TN^n. Q T := M n / T M n , Q T D := N n / T N n . These are finite-dimensional complex vector spaces. The formula
b T ( [ f ] , [ η ] ) : = ∑ a ∈ Z Res a ( η T T − 1 f ) (55) b_T([f],[\eta]):=\sum_{a\in Z}\operatorname{Res}_a\bigl(\eta^T T^{-1}f\bigr) \tag*{(55)} b T ([ f ] , [ η ]) := a ∈ Z ∑ Res a ( η T T − 1 f ) (55)
defines a perfect bilinear pairing between them. At a singular point, Res a \operatorname{Res}_a Res a denotes the sum of the residues on the branches of the normalization above a a a .
The matrix of sections
C T ( P , Q ) : = [ T ( P ) − T ( Q ) ] S F ( P , Q ) (56) \mathcal{C}_T(P,Q):=[T(P)-T(Q)]S_F(P,Q) \tag*{(56)} C T ( P , Q ) := [ T ( P ) − T ( Q )] S F ( P , Q ) (56)
is regular on U × U U \times U U × U . Its row index belongs to M n M^n M n and its column index to N n N^n N n . Its class in Q T ⊗ C Q T D Q_T \otimes_{\mathbb{C}} Q_T^D Q T ⊗ C Q T D is the coevaluation tensor of (55) : contracting its second factor against any [ f ] ∈ Q T [f] \in Q_T [ f ] ∈ Q T returns [ f ] [f] [ f ] .
Proof. The adjugate identity shows that det T \det T det T annihilates both quotients. Since A A A is the coordinate ring of an affine integral curve and det T ≠ 0 \det T \ne0 det T = 0 , their support is finite and their complex dimensions are finite. Replacing f f f by f + T a f+Ta f + T a changes the differential in (55) by η T a \eta^T a η T a . Replacing η \eta η by η + T T b \eta+T^T b η + T T b changes it by b T f b^T f b T f . Both are regular dualizing differentials near Z Z Z , so their residues there vanish. At a singular point it is the sum over its branches that vanishes; individual branch residues need not vanish. Thus the formula descends to the two stated quotients.
We verify perfection by reducing the displayed residue pairing to polynomial division in one variable. Choose a Noether normalization B = C [ x ] ⊂ A B=\mathbb{C}[x]\subset A B = C [ x ] ⊂ A . The module A A A is finite and torsion free over B B B , hence free. Since M M M is a finite projective A A A -module, it is a direct summand of a finite free A A A -module and therefore a finite free B B B -module as well. Write its B B B -rank as b b b . The trace description of the canonical module gives
ω A ≃ Hom B ( A , B ) d x , N = Hom A ( M , ω A ) ≃ Hom B ( M , B ) d x ; (57) \omega_A \simeq\operatorname{Hom}_B(A,B)\,dx,\qquad N=\operatorname{Hom}_A(M,\omega_A)\simeq\operatorname{Hom}_B(M,B)\,dx; \tag*{(57)} ω A ≃ Hom B ( A , B ) d x , N = Hom A ( M , ω A ) ≃ Hom B ( M , B ) d x ; (57)
the first identification is the finite-map case of [39 Theorem 5.1 and Remark 5.2] . In this description, a differential η \eta η corresponds to the rule a ↦ Tr π ( a η ) a\mapsto\operatorname{Tr}_{\pi}(a\eta) a ↦ Tr π ( a η ) , with values in B d x Bdx B d x . Thus the trace of η \eta η itself is the value of this rule at 1 1 1 .
The second identification in (57) can be checked directly. Suppress the common factor d x dx d x for this calculation, and give Hom B ( A , B ) \operatorname{Hom}_B(A,B) Hom B ( A , B ) the A A A -action ( a ℓ ) ( c ) : = ℓ ( a c ) (a\ell)(c):=\ell(ac) ( a ℓ ) ( c ) := ℓ ( a c ) . An A A A -linear map ψ : M → Hom B ( A , B ) \psi:M\to\operatorname{Hom}_B(A,B) ψ : M → Hom B ( A , B ) determines the B B B -linear functional λ ψ ( m ) : = ψ ( m ) ( 1 ) \lambda_\psi(m):=\psi(m)(1) λ ψ ( m ) := ψ ( m ) ( 1 ) . Conversely, a B B B -linear functional λ : M → B \lambda:M\to B λ : M → B determines
ψ λ ( m ) ( a ) : = λ ( a m ) . \psi_\lambda(m)(a):=\lambda(am). ψ λ ( m ) ( a ) := λ ( am ) .
This rule is A A A -linear in m m m , since ψ λ ( c m ) ( a ) = λ ( a c m ) = ( c ψ λ ( m ) ) ( a ) \psi_\lambda(cm)(a)=\lambda(acm)=(c\psi_\lambda(m))(a) ψ λ ( c m ) ( a ) = λ ( a c m ) = ( c ψ λ ( m )) ( a ) . The two constructions are inverse: evaluation of ψ λ ( m ) \psi_\lambda(m) ψ λ ( m ) at 1 1 1 returns λ ( m ) \lambda(m) λ ( m ) , while λ ψ ( a m ) = ψ ( a m ) ( 1 ) = ψ ( m ) ( a ) \lambda_\psi(am)=\psi(am)(1)=\psi(m)(a) λ ψ ( am ) = ψ ( am ) ( 1 ) = ψ ( m ) ( a ) . This proves the asserted adjunction with its normalization. In particular, N N N is free over B B B , even when it is not free over A A A , because it is the B B B -dual of the finite free module M M M . Choose a B B B -basis of M n M^n M n and its trace-dual basis of N n N^n N n ; the trace of the evaluation of a basis element against its dual is d x dx d x , and against any other dual basis element it is zero. The adjugate identity makes T T T injective on the torsion-free module M n M^n M n . Thus its matrix D ( x ) ∈ Mat b ( B ) D(x)\in\operatorname{Mat}_b(B) D ( x ) ∈ Mat b ( B ) in these bases has nonzero determinant, and its dual action T T T^T T T is represented by D ( x ) T D(x)^T D ( x ) T . For the coordinate columns v , w v,w v , w of f , η f,\eta f , η , respectively, finite duality gives the identity of meromorphic differentials on the base
Tr π ( η T T − 1 f ) = w T D ( x ) − 1 v d x , π : Spec A ⟶ A x 1 . \operatorname{Tr}_{\pi}(\eta^T T^{-1}f)=w^T D(x)^{-1}v\,dx,\qquad\pi:\operatorname{Spec} A\longrightarrow A^1_x. Tr π ( η T T − 1 f ) = w T D ( x ) − 1 v d x , π : Spec A ⟶ A x 1 .
The trace preserves the sum of residues over each fibre:
Res x = c Tr π ξ = ∑ a ↦ c Res a ξ . (58) \operatorname{Res}_{x=c}\operatorname{Tr}_{\pi}\xi=\sum_{a\mapsto c}\operatorname{Res}_a\xi. \tag*{(58)} Res x = c Tr π ξ = a ↦ c ∑ Res a ξ . (58)
Here residues at singular points mean the sums on their normalization branches. To check the formula, pass to the normalization of the curve. In the completed local ring of a branch above c c c , choose a parameter so that x − c = z e x-c=z^e x − c = z e . Since d x = e z e − 1 d z dx=ez^{e-1}dz d x = e z e − 1 d z , summing over the e e e conjugate parameters gives
Tr ( z k d z ) = { ( x − c ) ( k + 1 ) / e − 1 d x , e ∣ k + 1 , 0 , e ∤ k + 1. \operatorname{Tr}(z^k dz)=
\begin{cases}
(x-c)^{(k+1)/e-1}dx, & e\mid k+1,\\
0, & e\nmid k+1.
\end{cases} Tr ( z k d z ) = { ( x − c ) ( k + 1 ) / e − 1 d x , 0 , e ∣ k + 1 , e ∤ k + 1.
For a Laurent differential ∑ k a k z k d z \sum_k a_kz^k dz ∑ k a k z k d z , the coefficient of d x / ( x − c ) dx/(x-c) d x / ( x − c ) is therefore a − 1 a_{-1} a − 1 . Summing this calculation over the branches proves (58) . Away from Z Z Z , η T T − 1 f \eta^T T^{-1}f η T T − 1 f is a regular dualizing differential and has zero branch-sum residue. Thus (55) is exactly the sum of the finite residues of $w^T D^{-1}v,dx`.
It remains to check this pairing for the polynomial matrix D D D . Take a Smith form U D V = diag ( d 1 , … , d n b ) UDV=\operatorname{diag}(d_1,\ldots,d_{nb}) U D V = diag ( d 1 , … , d nb ) , where U , V U,V U , V are invertible over B B B and every d j d_j d j is nonzero. The changes v ↦ U v v\mapsto Uv v ↦ U v and w ↦ V T w w\mapsto V^{T}w w ↦ V T w identify the two quotient modules and the pairing with the direct sum of
B / ( d j ) × B / ( d j ) ⟶ C , ( [ a ] , [ b ] ) ⟼ ∑ d j ( c ) = 0 Res x = c a ( x ) b ( x ) d j ( x ) d x . B/(d_j)\times B/(d_j)\longrightarrow\mathbb{C},\qquad([a],[b])\longmapsto\sum_{d_j(c)=0}\operatorname{Res}_{x=c}\frac{a(x)b(x)}{d_j(x)}\,\mathrm{d}x. B / ( d j ) × B / ( d j ) ⟶ C , ([ a ] , [ b ]) ⟼ d j ( c ) = 0 ∑ Res x = c d j ( x ) a ( x ) b ( x ) d x .
Each scalar pairing is perfect. Indeed, near a zero c c c of order k k k , write z : = x − c z:=x-c z := x − c and d j = z k u ( z ) d_j=z^ku(z) d j = z k u ( z ) with u ( 0 ) ≠ 0 u(0)\ne0 u ( 0 ) = 0 . The residue is the coefficient of z k − 1 z^{k-1} z k − 1 in a b / u ab/u ab / u modulo z k z^k z k . Multiplication by the unit u − 1 u^{-1} u − 1 is invertible, and the coefficient pairing on C [ z ] / ( z k ) \mathbb{C}[z]/(z^k) C [ z ] / ( z k ) has an invertible anti-diagonal Gram in the monomial basis. Explicitly, the classes u ( z ) z k − 1 − j u(z)z^{k-1-j} u ( z ) z k − 1 − j are dual to z j z^j z j , for j ∈ { 0 , … , k − 1 } j\in\{0,\ldots,k-1\} j ∈ { 0 , … , k − 1 } : the coefficient of z k − 1 z^{k-1} z k − 1 in z i u ( z ) z k − 1 − j / u ( z ) z^iu(z)z^{k-1-j}/u(z) z i u ( z ) z k − 1 − j / u ( z ) is δ i , j \delta_{i,j} δ i , j . The Chinese remainder decomposition combines these local pairings. A constant d j d_j d j contributes a zero quotient. This proves perfection of the original pairing and derives its residue normalization explicitly. The dual quotient contains T T T^{T} T T because it is the adjoint action for the evaluation into ω A \omega_A ω A , not because ω A \omega_A ω A is identified with A A A .
Each entry T i , j ( P ) − T i , j ( Q ) T_{i,j}(P)-T_{i,j}(Q) T i , j ( P ) − T i , j ( Q ) belongs to the diagonal ideal. The defining property of S F ∈ Hom ( I Δ , F ⊠ F D ) S_F\in\operatorname{Hom}(\mathcal{I}_{\Delta},F\boxtimes F^D) S F ∈ Hom ( I Δ , F ⊠ F D ) therefore makes (56) regular, including at pairs of singular points. Since U U U is affine, its entries are finite sums of tensors from M ⊗ C N M\otimes_{\mathbb{C}}N M ⊗ C N .
It remains to verify that this regular tensor is the inverse of the residue pairing. For f ∈ M n f\in M^n f ∈ M n , its contraction is
I ( P ) : = ∑ a ∈ Z Res Q = a ( C T ( P , Q ) T ( Q ) − 1 f ( Q ) ) . I(P):=\sum_{a\in Z}\operatorname{Res}_{Q=a}\left(\mathcal{C}_T(P,Q)T(Q)^{-1}f(Q)\right). I ( P ) := a ∈ Z ∑ Res Q = a ( C T ( P , Q ) T ( Q ) − 1 f ( Q ) ) .
The inverse is on the right of C T \mathcal{C}_T C T . Keeping that order, the differential inside the sum is
T ( P ) S F ( P , Q ) T ( Q ) − 1 f ( Q ) − S F ( P , Q ) f ( Q ) . T(P)S_F(P,Q)T(Q)^{-1}f(Q)-S_F(P,Q)f(Q). T ( P ) S F ( P , Q ) T ( Q ) − 1 f ( Q ) − S F ( P , Q ) f ( Q ) .
For a general smooth P ∈ U ∖ Z P\in U\setminus Z P ∈ U ∖ Z , the two terms together have no pole at Q = P Q=P Q = P : this also follows directly from the regularity of C T \mathcal{C}_T C T . Their remaining poles lie in Z Z Z or in the finite complement C ∖ U \mathbb{C}\setminus U C ∖ U . The global residue theorem gives
I ( P ) = − T ( P ) A f ( P ) + B f ( P ) , I(P)=-T(P)A_f(P)+B_f(P), I ( P ) = − T ( P ) A f ( P ) + B f ( P ) ,
where
A f ( P ) : = ∑ a ∈ C ∖ U Res Q = a ( S F ( P , Q ) T ( Q ) − 1 f ( Q ) ) , A_f(P):=\sum_{a\in\mathbb{C}\setminus U}\operatorname{Res}_{Q=a}\left(S_F(P,Q)T(Q)^{-1}f(Q)\right), A f ( P ) := a ∈ C ∖ U ∑ Res Q = a ( S F ( P , Q ) T ( Q ) − 1 f ( Q ) ) ,
B f ( P ) : = ∑ a ∈ C ∖ U Res Q = a ( S F ( P , Q ) f ( Q ) ) . B_f(P):=\sum_{a\in\mathbb{C}\setminus U}\operatorname{Res}_{Q=a}\left(S_F(P,Q)f(Q)\right). B f ( P ) := a ∈ C ∖ U ∑ Res Q = a ( S F ( P , Q ) f ( Q ) ) .
These are regular sections on U U U . Indeed, the residues are taken at fixed points outside U U U , so they do not meet the diagonal as P P P varies in U U U ; taking their finite Laurent coefficients preserves regularity in P P P . In particular A f ∈ M n A_f\in M^n A f ∈ M n .
Apply the global residue theorem once more to S F ( P , Q ) f ( Q ) S_F(P,Q)f(Q) S F ( P , Q ) f ( Q ) . Inside U U U its only nonzero residue is the diagonal residue − f ( P ) -f(P) − f ( P ) . Consequently B f ( P ) = f ( P ) B_f(P)=f(P) B f ( P ) = f ( P ) , and hence
I ( P ) = f ( P ) − T ( P ) A f ( P ) . I(P)=f(P)-T(P)A_f(P). I ( P ) = f ( P ) − T ( P ) A f ( P ) .
Both sides are regular on U U U , so the identity obtained for general P P P holds throughout U U U . It gives [ I ] = [ f ] [I]=[f] [ I ] = [ f ] in Q T Q_T Q T . Because the pairing is perfect, this reproduction identity determines its coevaluation tensor uniquely and proves the theorem. □ \square □
The theorem supplies image-membership tests for any chosen matrix T T T . If [ e 1 ] , … , [ e m ] [e_1],\ldots,[e_m] [ e 1 ] , … , [ e m ] and [ e 1 D ] , … , [ e m D ] [e_1^D],\ldots,[e_m^D] [ e 1 D ] , … , [ e m D ] are residue-dual quotient bases, put
ℓ j ( f ) : = b T ( [ f ] , [ e j D ] ) , j ∈ { 1 , … , m } . \ell_j(f):=b_T([f],[e_j^D]),\qquad j\in\{1,\ldots,m\}. ℓ j ( f ) := b T ([ f ] , [ e j D ]) , j ∈ { 1 , … , m } .
Then [ f ] = ∑ j ℓ j ( f ) [ e j ] [f]=\sum_j\ell_j(f)[e_j] [ f ] = ∑ j ℓ j ( f ) [ e j ] , so f ∈ T M n f\in TM^n f ∈ T M n if and only if all ℓ j ( f ) \ell_j(f) ℓ j ( f ) vanish. Prescribed mass values c j c_j c j on [ e j ] [e_j] [ e j ] give the functional ν [ f ] = ∑ j c j ℓ j ( f ) = b T ( [ f ] , ∑ j c j [ e j D ] ) \nu[f]=\sum_j c_j\ell_j(f)=b_T([f],\sum_jc_j[e_j^D]) ν [ f ] = ∑ j c j ℓ j ( f ) = b T ([ f ] , ∑ j c j [ e j D ]) . Thus the residue calculation tests division by T T T and describes every functional vanishing on its image. In a bilateral problem, apply it separately to the division map being tested and to the denominator map defining the homogeneous mass.
A small singular quotient makes the differential module concrete. Take A = C [ t 3 , t 4 , t 5 ] A=\mathbb{C}[t^3,t^4,t^5] A = C [ t 3 , t 4 , t 5 ] , whose only singular point is t = 0 t=0 t = 0 , and use the scalar denominator T = t 3 T=t^3 T = t 3 . The allowed exponents of functions in A A A are 0 , 3 , 4 , 5 , … 0,3,4,5,\ldots 0 , 3 , 4 , 5 , … . A Laurent differential t k d t t^k\mathrm{d}t t k d t belongs to ω A \omega_A ω A precisely when multiplying by any such function gives zero residue at 0 0 0 . Thus its exponent cannot be − 1 , − 4 , − 5 , … -1,-4,-5,\ldots − 1 , − 4 , − 5 , … , and
ω A = A t − 3 d t + A t − 2 d t . \omega_A=A\,t^{-3}\mathrm{d}t+A\,t^{-2}\mathrm{d}t. ω A = A t − 3 d t + A t − 2 d t .
Both generators are needed at 0 0 0 . Division by t 3 t^3 t 3 leaves the following bases of the two three-dimensional quotients:
A / ( t 3 ) : 1 , t 4 , t 5 , ω A / t 3 ω A : t − 3 d t , t − 2 d t , t 2 d t . A/(t^3):\quad1,t^4,t^5,\qquad\omega_A/t^3\omega_A:\quad t^{-3}\mathrm{d}t,t^{-2}\mathrm{d}t,t^2\mathrm{d}t. A / ( t 3 ) : 1 , t 4 , t 5 , ω A / t 3 ω A : t − 3 d t , t − 2 d t , t 2 d t .
For example, the first list consists of the allowed exponents that are not 3 3 3 plus another allowed exponent. The same subtraction gives the second list from − 3 , − 2 , 0 , 1 , … -3,-2,0,1,\ldots − 3 , − 2 , 0 , 1 , … . The pairing is Res t = 0 ( f η / t 3 ) \operatorname{Res}_{t=0}(f\eta/t^3) Res t = 0 ( f η / t 3 ) , so its entry is 1 1 1 exactly when the two displayed exponents sum to 2 2 2 . Its Gram matrix is
( 0 0 1 0 1 0 1 0 0 ) . \begin{pmatrix}
0 & 0 & 1\\
0 & 1 & 0\\
1 & 0 & 0
\end{pmatrix}. 0 0 1 0 1 0 1 0 0 .
This verifies perfection at a non-Gorenstein point and shows explicitly why the dual remainder classes are differentials.
The construction has an ordered product rule. If T a , T b T_a,T_b T a , T b are two matrices of the same size with entries in A A A , then
C T a T b ( P , Q ) = T a ( P ) C T b ( P , Q ) + C T a ( P , Q ) T b ( Q ) . \mathcal{C}_{T_aT_b}(P,Q)=T_a(P)\mathcal{C}_{T_b}(P,Q)+\mathcal{C}_{T_a}(P,Q)T_b(Q). C T a T b ( P , Q ) = T a ( P ) C T b ( P , Q ) + C T a ( P , Q ) T b ( Q ) .
Indeed, after removing the common scalar kernel, the right-hand side is
T a ( P ) [ T b ( P ) − T b ( Q ) ] + [ T a ( P ) − T a ( Q ) ] T b ( Q ) = T a ( P ) T b ( P ) − T a ( Q ) T b ( Q ) . T_a(P)\left[T_b(P)-T_b(Q)\right]+\left[T_a(P)-T_a(Q)\right]T_b(Q)=T_a(P)T_b(P)-T_a(Q)T_b(Q). T a ( P ) [ T b ( P ) − T b ( Q ) ] + [ T a ( P ) − T a ( Q ) ] T b ( Q ) = T a ( P ) T b ( P ) − T a ( Q ) T b ( Q ) .
This proof retains the order of the factors and requires no commutativity between the two matrices.
Remark 4.5 (Row–column convention). The arguments of b T b_T b T are columns. If instead f f f is a row with entries in M M M , g g g is a column with entries in N N N , and
ν ( f , g ) = c ∑ a ∈ Z Res a ( f T − 1 g ) , c ≠ 0 , \nu(f,g)=c\sum_{a\in Z}\operatorname{Res}_a(fT^{-1}g),\qquad c\ne0, ν ( f , g ) = c a ∈ Z ∑ Res a ( f T − 1 g ) , c = 0 ,
then ν ( f , g ) = c b T T ( f T , g ) \nu(f,g)=c\,b_{T^\mathsf{T}}(f^\mathsf{T},g) ν ( f , g ) = c b T T ( f T , g ) . Thus its quotient pair is M n / T M n M^n/TM^n M n / T M n and N n / T N n N^n/TN^n N n / T N n , and its coevaluation in that order is c − 1 C T T c^{-1}\mathcal{C}_{T^\mathsf{T}} c − 1 C T T . This transpose changes the matrix indices, without exchanging the two points. The column–row kernel convention of (7) subsequently exchanges the two tensor factors.
The bilateral test quotient
We now connect the quotient theorem with the four factors of Subsection 2.1. For a balanced q × p q\times p q × p test matrix, write
D i ( H ) : = L i T H R i T , T i : = R i ⊗ L i T , i ∈ { 0 , 1 } . D_i(H):=L_i^\mathsf{T}HR_i^\mathsf{T},\qquad T_i:=R_i\otimes L_i^\mathsf{T},\qquad i\in\{0,1\}. D i ( H ) := L i T H R i T , T i := R i ⊗ L i T , i ∈ { 0 , 1 } .
Vectorization by columns gives
vec ( D i ( H ) ) = T i vec ( H ) . (59) \operatorname{vec}(D_i(H))=T_i\operatorname{vec}(H). \tag*{(59)} vec ( D i ( H )) = T i vec ( H ) . (59)
For clarity, the entry in row a a a and column b b b of D i ( H ) D_i(H) D i ( H ) is ∑ c , d ( L i ) c , a H c , d ( R i ) b , d \sum_{c,d}(L_i)_{c,a}H_{c,d}(R_i)_{b,d} ∑ c , d ( L i ) c , a H c , d ( R i ) b , d . Placing H c , d H_{c,d} H c , d in position c + q ( d − 1 ) c+q(d-1) c + q ( d − 1 ) gives precisely the Kronecker-product entry ( R i ) b , d ( L i ) c , a (R_i)_{b,d}(L_i)_{c,a} ( R i ) b , d ( L i ) c , a in T i T_i T i . Here T i T_i T i has size q p qp q p , regardless of whether a particular mass also admits a representation using a smaller matrix quotient. Assume det L 1 det R 1 ≠ 0 \det L_1\det R_1\ne0 det L 1 det R 1 = 0 . The numerator factors are independent of the denominators and no matrix commutation is imposed.
Let E \mathcal{E} E be the module of scalar test entries; it is A A A for function tests and ω \omega ω for tests obtained by evaluating F ⊗ F D → ω C F\otimes F^D\to\omega_C F ⊗ F D → ω C . Put
V E : = C q p , Q E : = V E / T 1 V E . V_{\mathcal{E}}:=\mathbb{C}^{qp},\qquad Q_{\mathcal{E}}:=V_{\mathcal{E}}/T_1V_{\mathcal{E}}. V E := C q p , Q E := V E / T 1 V E .
Choose a complex linear extension μ ~ \widetilde{\mu} μ of μ \mu μ to the rational tests needed below. Applying the subtraction argument of Proposition 3.10 to these test modules gives every solution in the form
μ ~ [ v ] = μ ~ [ T 0 T 1 − 1 v ] + ν [ v ] , ν ∈ Q E ∗ . (60) \widetilde{\mu}[v]=\widetilde{\mu}[T_0T_1^{-1}v]+\nu[v],\qquad\nu\in Q_{\mathcal{E}}^*. \tag*{(60)} μ [ v ] = μ [ T 0 T 1 − 1 v ] + ν [ v ] , ν ∈ Q E ∗ . (60)
Indeed, evaluating the rational term at T 1 v T_1v T 1 v returns μ [ T 0 v ] \mu[T_0v] μ [ T 0 v ] , and the difference of any two solutions vanishes on T 1 V E T_1V_{\mathcal{E}} T 1 V E . The ordered product is
T 0 T 1 − 1 = ( R 0 R 1 − 1 ) ⊗ ( L 1 − 1 L 0 ) T . T_0T_1^{-1}=(R_0R_1^{-1})\otimes(L_1^{-1}L_0)^{\mathsf{T}}. T 0 T 1 − 1 = ( R 0 R 1 − 1 ) ⊗ ( L 1 − 1 L 0 ) T .
Reversing T 0 T_0 T 0 and T 1 − 1 T_1^{-1} T 1 − 1 would generally fail to solve the functional equation. The choice of rational extension is part of the data; changing it changes the particular solution and the remaining mass in (4.14).
The two test modules have different canonical dual quotients:
A q p / T 1 A q p pairs with ω q p / T 1 T ω q p . (61) A^{qp}/T_1A^{qp}\quad\text{pairs with}\quad\omega^{qp}/T_1^{\mathsf{T}}\omega^{qp}. \tag*{(61)} A q p / T 1 A q p pairs with ω q p / T 1 T ω q p . (61)
ω q p / T 1 ω q p pairs with A q p / T 1 T A q p . (62) \omega^{qp}/T_1\omega^{qp}\quad\text{pairs with}\quad A^{qp}/T_1^{\mathsf{T}}A^{qp}. \tag*{(62)} ω q p / T 1 ω q p pairs with A q p / T 1 T A q p . (62)
In both lines the pairing is the sum of residues of η T T 1 − 1 v \eta^{\mathsf{T}}T_1^{-1}v η T T 1 − 1 v . The representative η \eta η carries the differential in the first line; v v v carries it in the second.
Here is the precise link with Theorem 4.4 . The quotients are supported on finitely many points, and the line F F F has a frame e e e on a neighbourhood of that finite set. For (4.15), use v ↦ e v v\mapsto ev v ↦ e v and η ↦ e − 1 η \eta\mapsto e^{-1}\eta η ↦ e − 1 η . These identify the two quotients with M q p / T 1 M q p M^{qp}/T_1M^{qp} M q p / T 1 M q p and N q p / T 1 T N q p N^{qp}/T_1^{\mathsf{T}}N^{qp} N q p / T 1 T N q p near their support, which suffices to identify the quotients themselves. For (4.16), use v ↦ e − 1 v v\mapsto e^{-1}v v ↦ e − 1 v and η ↦ e η \eta\mapsto e\eta η ↦ eη . The resulting pairing is the one for T 1 T T_1^{\mathsf{T}} T 1 T with its arguments exchanged, since
( e − 1 v ) T ( T 1 T ) − 1 ( e η ) = η T T 1 − 1 v . (e^{-1}v)^{\mathsf{T}}(T_1^{\mathsf{T}})^{-1}(e\eta)=\eta^{\mathsf{T}}T_1^{-1}v. ( e − 1 v ) T ( T 1 T ) − 1 ( eη ) = η T T 1 − 1 v .
Its coevaluation is therefore obtained by exchanging the tensor factors of C T 1 T \mathcal{C}_{T_1^{\mathsf{T}}} C T 1 T and transporting them by the chosen frame. A change of frame changes these representatives and their dual coordinates together, leaving the functional unchanged. This local choice does not require F F F to be trivial on all of U U U .
This distinction matters when C C C is not Gorenstein: the differential module ω \omega ω need not be free of rank one. Writing a differential in fractional coordinates does not turn its coefficient module into A A A . Likewise, the present balanced test quotient retains its A A A -module relations; it does not impose those relations on a bivariate pairing that originally had independent arguments.
Computing the finite pairing from the quotient
The residue coordinates enter the finite construction in two places. For polynomial division, under the regularity assumptions of Subsection 3.2, take the lifts P L , P R P_L,P_R P L , P R of (3.2). A row candidate b b b satisfies polynomial divisibility precisely when the class of ( a L b ) T (a_Lb)^{\mathsf{T}} ( a L b ) T modulo P L T A q P_L^{\mathsf{T}}A^q P L T A q vanishes; for a column candidate c c c , the corresponding condition is that the class of a R c a_Rc a R c modulo P R A p P_RA^p P R A p vanishes. Theorem 4.4, with a local frame at the finite support as above, supplies residue functionals detecting these classes. Evaluating the functionals on the chosen ambient basis vectors and selecting independent rows gives the divisibility rows of Γ L \Gamma_L Γ L and Γ R \Gamma_R Γ R . The remaining rows test membership of the divided vectors in the prescribed finite spaces, as in Proposition 3.1. Thus residue tests and the degree or section bounds together give the full image conditions used in Theorems 3.3 and 3.5.
For masses, the relevant quotient is instead the one defined by the denominator action D 1 D_1 D 1 . Its residue coordinates give the mass entries in those same finite formulas. They also give a direct Gram computation by Subsection 3.4, as follows. Choose d d d row sections f i ∈ M 1 × q f_i\in M^{1\times q} f i ∈ M 1 × q and d d d column sections g j ∈ N p × 1 g_j\in N^{p\times1} g j ∈ N p × 1 spanning the prescribed spaces. Their products give differential tests H ( f i , g j ) : = [ ( f i ) a ( g j ) b ] a ∈ { 1 , … , q } , b ∈ { 1 , … , p } H(f_i,g_j):=[(f_i)a(g_j)b]_{a\in\{1,\ldots,q\},\,b\in\{1,\ldots,p\}} H ( f i , g j ) := [( f i ) a ( g j ) b ] a ∈ { 1 , … , q } , b ∈ { 1 , … , p } , so the homogeneous mass is an element of the dual of Mat q × p ( ω ) / D 1 Mat q × p ( ω ) \operatorname{Mat}_{q\times p}(\omega)/D_1\operatorname{Mat}_{q\times p}(\omega) Mat q × p ( ω ) / D 1 Mat q × p ( ω ) . Choose a basis ζ 1 , … , ζ m \zeta_1,\ldots,\zeta_m ζ 1 , … , ζ m of this quotient and reduce each test H ( f i , g j ) H(f_i,g_j) H ( f i , g j ) modulo D 1 D_1 D 1 :
[ H ( f i , g j ) ] = ∑ a = 1 m c i , j , a ζ a , ( H ν ) i , j = ∑ a = 1 m c i , j , a ν [ ζ a ] . [H(f_i,g_j)]=\sum_{a=1}^{m}c_{i,j,a}\zeta_a,\qquad
(H_\nu)_{i,j}=\sum_{a=1}^{m}c_{i,j,a}\nu[\zeta_a]. [ H ( f i , g j )] = a = 1 ∑ m c i , j , a ζ a , ( H ν ) i , j = a = 1 ∑ m c i , j , a ν [ ζ a ] .
If ζ a D \zeta_a^D ζ a D denotes the residue-dual basis, the reduction coefficients are the pairings of [ H ( f i , g j ) ] [H(f_i,g_j)] [ H ( f i , g j )] , after column vectorization, with ζ a D \zeta_a^D ζ a D under the duality (4.16). This computes every c i , j , a c_{i,j,a} c i , j , a by a residue, and the displayed finite sum then inserts the prescribed mass parameters ν [ ζ a ] \nu[\zeta_a] ν [ ζ a ] . Reduction takes place in the actual differential module, including its relations at singular points.
Next express f i S L f_iS_L f i S L and S R g j S_Rg_j S R g j in finite rational bases as in (3.33). Their rational pairings form G r a t G_{\mathrm{rat}} G rat , and their coordinate matrices are T L , T R T_L,T_R T L , T R . The complete target Gram is G ^ = T L G r a t T R + H ν \widehat{G}=T_LG_{\mathrm{rat}}T_R+H_\nu G = T L G rat T R + H ν . If it is invertible, (3.35) gives the reproducing kernel, and Proposition 3.12 gives the next orthogonal rows and columns with their prescribed leading classes. This is the calculation to be performed after the geometric quotient is known.
The three pairings have separate roles. The original pairing is β μ \beta_\mu β μ . The chosen rational extension gives β p a r t ( f , g ) = β ~ μ ( f S L , S R g ) \beta_{\mathrm{part}}(f,g)=\widetilde{\beta}_\mu(fS_L,S_Rg) β part ( f , g ) = β μ ( f S L , S R g ) , and the target pairing is β ^ = β p a r t + ν \widehat{\beta}=\beta_{\mathrm{part}}+\nu β = β part + ν . Formula (45) compares the target and particular kernels on the same target spaces when both Grams are invertible. These finite Grams belong to the orthogonality problem; the geometric kernel S F S_F S F is determined by the curve and the line bundle.
Finite residue corrections on surfaces
On a curve, a regular denominator has a finite zero set, and the quotients in Theorem 4.4 are finite dimensional. On a surface, its zero set is usually a curve. For example, C [ x , y ] / ( x ) = C [ y ] \mathbb{C}[x,y]/(x)=\mathbb{C}[y] C [ x , y ] / ( x ) = C [ y ] has infinite complex dimension. To obtain a finite residue problem, we impose a second equation: a scalar polynomial h h h that meets the determinant divisor in finitely many points. The resulting quotient retains the multiplicities at these points. Its dual describes those masses that satisfy both the matrix equation and the additional scalar equation.
We first work in the plane, where polynomial divided differences give an explicit tensor reproducing this quotient. Scalar complete-intersection residue duality and its Bézoutian tensor are classical; we use the formulation in [20 Proposition 4.3 and Lemma 4.4] . Residue methods for polynomial matrix division are developed in [21 ] . The calculation below specifies the quotient, the transposed dual and the order of the matrix factors needed in the present perturbation problem. Its output is a finite residue tensor. A global geometric Szegő kernel on a surface would require a different construction.
A matrix equation and a scalar cut
Let A : = C [ x , y ] A:=\mathbb{C}[x,y] A := C [ x , y ] , let T ∈ Mat n ( A ) T\in\operatorname{Mat}_n(A) T ∈ Mat n ( A ) and put d : = det T d:=\det T d := det T . Assume that d ≠ 0 d\ne0 d = 0 and that h ∈ A h\in A h ∈ A satisfies
0 < dim C A / ( d , h ) < ∞ . (63) 0<\dim_{\mathbb{C}} A/(d,h)<\infty. \tag*{(63)} 0 < dim C A / ( d , h ) < ∞. (63)
Thus d d d and h h h have no common curve component, and their common zero set Z Z Z is finite. In particular, h h h is a nonzerodivisor in A A A , and multiplication by d d d is injective in A / ( h ) A/(h) A / ( h ) . Indeed, d d d and h h h have no common irreducible factor, so h ∣ d a h\mid da h ∣ d a implies h ∣ a h\mid a h ∣ a . These are the regularity properties used below. They do not require the points of Z Z Z to be simple.
Fix the volume form d x ∧ d y \mathrm{d}x\wedge\mathrm{d}y d x ∧ d y and the order ( d , h ) (d,h) ( d , h ) of the two equations. Define the scalar residue functional by
λ d , h ( a ) : = ∑ ξ ∈ Z Res ξ a ( x , y ) d x ∧ d y d ( x , y ) h ( x , y ) . (64) \lambda_{d,h}(a):=\sum_{\xi\in Z}\operatorname{Res}_{\xi}\frac{a(x,y)\,\mathrm{d}x\wedge\mathrm{d}y}{d(x,y)h(x,y)}. \tag*{(64)} λ d , h ( a ) := ξ ∈ Z ∑ Res ξ d ( x , y ) h ( x , y ) a ( x , y ) d x ∧ d y . (64)
Here Res ξ \operatorname{Res}_{\xi} Res ξ is the local Grothendieck residue with these ordered denominators. Analytically it is the integral
1 ( 2 π i ) 2 ∫ Γ ξ a ( x , y ) d x ∧ d y d ( x , y ) h ( x , y ) , Γ ξ : ∣ d ∣ = r 1 , ∣ h ∣ = r 2 , \frac{1}{(2\pi i)^2}\int_{\Gamma_\xi}\frac{a(x,y)\,\mathrm{d}x\wedge\mathrm{d}y}{d(x,y)h(x,y)},\qquad\Gamma_\xi:\ |d|=r_1,\quad|h|=r_2, ( 2 π i ) 2 1 ∫ Γ ξ d ( x , y ) h ( x , y ) a ( x , y ) d x ∧ d y , Γ ξ : ∣ d ∣ = r 1 , ∣ h ∣ = r 2 ,
on a sufficiently small cycle enclosing only ξ \xi ξ , with orientation d arg d ∧ d arg h > 0 \mathrm{d}\arg d\wedge\mathrm{d}\arg h>0 d arg d ∧ d arg h > 0 . The algebraic residue is the same functional and is computed by finite polynomial reduction below. For denominators x r , y s x^r,y^s x r , y s at the origin it extracts [ x r − 1 y s − 1 ] a [x^{r-1}y^{s-1}]a [ x r − 1 y s − 1 ] a , by applying the one-variable residue formula twice. Thus repeated zeros retain derivative information. This convention gives λ x , y ( 1 ) = 1 \lambda_{x,y}(1)=1 λ x , y ( 1 ) = 1 . If every intersection point is simple, the same functional is
λ d , h ( a ) = ∑ ξ ∈ Z a ( ξ ) ( d x h y − d y h x ) ( ξ ) . \lambda_{d,h}(a)=\sum_{\xi\in Z}\frac{a(\xi)}{(d_xh_y-d_yh_x)(\xi)}. λ d , h ( a ) = ξ ∈ Z ∑ ( d x h y − d y h x ) ( ξ ) a ( ξ ) .
The residue definition in (4.18) also applies when this Jacobian vanishes. It then retains the corresponding local multiplicities. It annihilates ( d , h ) (d,h) ( d , h ) , and ( [ a ] , [ b ] ) ↦ λ d , h ( a b ) ([a],[b])\mapsto\lambda_{d,h}(ab) ([ a ] , [ b ]) ↦ λ d , h ( ab ) is a perfect pairing on A / ( d , h ) A/(d,h) A / ( d , h ) . We give the scalar reconstruction calculation that proves this perfection immediately before Theorem 4.8 .
We use columns for both arguments of the matrix residue pairing. Define
Q T , h : = A n T A n + h A n , Q T , h D : = A n T ‾ A n + h A n . (65) Q_{T,h}:=\frac{A^n}{TA^n+hA^n},\qquad Q^D_{T,h}:=\frac{A^n}{\overline{T}A^n+hA^n}. \tag*{(65)} Q T , h := T A n + h A n A n , Q T , h D := T A n + h A n A n . (65)
The superscript D D D records the dual role of the second quotient. Its entries are understood to multiply the fixed volume form. This identification is available here because the canonical module of the plane is freely generated by d x ∧ d y \mathrm{d}x \wedge\mathrm{d}y d x ∧ d y .
Theorem 4.6 (Duality for a finite matrix quotient). Under (4.17), both quotients in (4.19) are finite dimensional, and
b T , h ( [ f ] , [ η ] ) : = λ d , h ( η T adj ( T ) f ) (66) b_{T,h}([f],[\eta]) := \lambda_{d,h}(\eta^{\mathsf{T}}\operatorname{adj}(T)f) \tag*{(66)} b T , h ([ f ] , [ η ]) := λ d , h ( η T adj ( T ) f ) (66) *is a perfect bilinear pairing between them. Consequently every complex linear functional on Q T , h Q_{T,h} Q T , h has a unique expression [ f ] ↦ b T , h ( [ f ] , [ η ] ) [f] \mapsto b_{T,h}([f],[\eta]) [ f ] ↦ b T , h ([ f ] , [ η ]) with [ η ] ∈ Q T , h D [\eta] \in Q_{T,h}^{D} [ η ] ∈ Q T , h D .
Proof. Write J : = adj ( T ) J := \operatorname{adj}(T) J := adj ( T ) . The identities T J = J T = d I n TJ = JT = dI_n T J = J T = d I n show that d d d annihilates both quotients. Since h h h also annihilates them, they are quotients of ( A / ( d , h ) ) n (A/(d,h))^n ( A / ( d , h ) ) n and are finite dimensional.
First check that the formula is independent of the representatives. Replacing f f f by f + T a f+Ta f + T a adds d η T a d\eta^{\mathsf{T}}a d η T a to its numerator. Replacing η \eta η by η + T T b \eta+T^{\mathsf{T}}b η + T T b adds d b T f db^{\mathsf{T}}f d b T f . Adding a multiple of h h h to either argument adds a numerator divisible by h h h . All these terms have zero residue. In particular, the transpose in the second quotient is exactly what makes this second substitution valid.
We next show that a class pairing to zero with every opposite class must itself be zero. Suppose first that b T , h ( [ f ] , [ η ] ) = 0 b_{T,h}([f],[\eta])=0 b T , h ([ f ] , [ η ]) = 0 for every η ∈ A n \eta\in A^n η ∈ A n . Taking η = a e i \eta=ae_i η = a e i , with arbitrary a ∈ A a\in A a ∈ A , shows that each component of J f Jf J f pairs to zero with all of A / ( d , h ) A/(d,h) A / ( d , h ) under the scalar residue pairing. Its perfection implies
J f ∈ ( d , h ) A n . Jf \in(d,h)A^n. J f ∈ ( d , h ) A n .
Work now in A / ( h ) A/(h) A / ( h ) , denoting classes by bars. The last inclusion means that J ‾ f ‾ = d ‾ a ‾ \overline{J}\,\overline{f}=\overline{d}\,\overline{a} J f = d a for some a ‾ ∈ ( A / ( h ) ) n \overline{a}\in(A/(h))^n a ∈ ( A / ( h ) ) n . Multiplication by T ‾ \overline{T} T gives
d ‾ f ‾ = d ‾ T ‾ a ‾ . \overline{d}\,\overline{f}=\overline{d}\,\overline{T}\,\overline{a}. d f = d T a .
Multiplication by d ‾ \overline{d} d is injective, so cancellation yields f ‾ = T ‾ a ‾ \overline{f}=\overline{T}\,\overline{a} f = T a . Lifting this equality to A A A gives f ∈ T A n + h A n f\in TA^n+hA^n f ∈ T A n + h A n . Thus [ f ] = 0 [f]=0 [ f ] = 0 in Q T , h Q_{T,h} Q T , h .
For the other argument, write the same scalar as λ d , h ( f T J T η ) \lambda_{d,h}(f^{\mathsf{T}}J^{\mathsf{T}}\eta) λ d , h ( f T J T η ) . If it vanishes for every f f f , scalar duality gives J T η ∈ ( d , h ) A n J^{\mathsf{T}}\eta\in(d,h)A^n J T η ∈ ( d , h ) A n . Apply the preceding cancellation argument with T T T^{\mathsf{T}} T T and J T J^{\mathsf{T}} J T to obtain η ∈ T T A n + h A n \eta\in T^{\mathsf{T}}A^n+hA^n η ∈ T T A n + h A n . Hence the pairing has zero radical on either side. Since the two spaces are finite dimensional, the induced maps into their complex duals are injective in both directions and therefore are isomorphisms. This proves perfection and the asserted uniqueness.
The sign agrees with our normalization: for T = x I n T=xI_n T = x I n and h = y h=y h = y , the formula uses d = x n d=x^n d = x n and J = x n − 1 I n J=x^{n-1}I_n J = x n − 1 I n and gives the ordinary dot product on C n \mathbb{C}^n C n . □
Canonical modules on a singular surface
Remark 4.7 (Canonical modules on a singular surface). The duality argument has the same form on a Cohen–Macaulay affine surface with coordinate ring R R R and canonical module ω R \omega_R ω R . Take T ∈ Mat n ( R ) T\in\operatorname{Mat}_n(R) T ∈ Mat n ( R ) and write d : = det T d:=\det T d := det T . If h h h is regular on R R R , d d d is regular modulo h h h , and ( d , h ) (d,h) ( d , h ) has finite support, the two quotients are
R n T R n + h R n , ω R n T T ω R n + h ω R n . \frac{R^n}{TR^n+hR^n}, \qquad\frac{\omega_R^n}{T^{\mathsf{T}}\omega_R^n+h\omega_R^n}. T R n + h R n R n , T T ω R n + h ω R n ω R n .
The second quotient uses the canonical module, which need not have one generator at a singular point. Put R d , h : = R / ( d , h ) R_{d,h}:=R/(d,h) R d , h := R / ( d , h ) and Ω d , h : = ω R / ( d , h ) ω R \Omega_{d,h}:=\omega_R/(d,h)\omega_R Ω d , h := ω R / ( d , h ) ω R . The scalar duality needed for the matrix argument is the R d , h R_{d,h} R d , h -linear identification
Θ : Ω d , h → ≃ Hom C ( R d , h , C ) , ( a b ) ( b ) : = ℓ ( a b ) . \Theta:\Omega_{d,h}\xrightarrow{\ \simeq\ }\operatorname{Hom}_{\mathbb{C}}(R_{d,h},\mathbb{C}), \qquad(ab)(b):=\ell(ab). Θ : Ω d , h ≃ Hom C ( R d , h , C ) , ( ab ) ( b ) := ℓ ( ab ) .
Here is a local description of this identification and of the regularity needed for cancellation. At a point of the finite support, complete the local ring and denote the result by R ^ \widehat{R} R . The ordered pair ( d , h ) (d,h) ( d , h ) is a system of parameters. In a Cohen–Macaulay local ring every system of parameters is a regular sequence, so the order can be reversed from ( h , d ) (h,d) ( h , d ) to ( d , h ) (d,h) ( d , h ) here. The power-series ring S : = C [ [ u , v ] ] S:=\mathbb{C}[[u,v]] S := C [[ u , v ]] acts by u ↦ d u\mapsto d u ↦ d and v ↦ h v\mapsto h v ↦ h . To see that R ^ \widehat{R} R is finite over S S S , lift a complex basis of R ^ / ( d , h ) R ^ \widehat{R}/(d,h)\widehat{R} R / ( d , h ) R . Reduction modulo ( d , h ) (d,h) ( d , h ) writes every element as a linear combination of these lifts plus terms divisible by d d d or h h h . Repeating this reduction on the remaining coefficients gives power series in d , h d,h d , h multiplying the same finite set of lifts. These series converge in R ^ \widehat{R} R , because the ( d , h ) (d,h) ( d , h ) -adic topology agrees with the maximal-ideal topology when the quotient has finite length. The lifts therefore generate R ^ \widehat{R} R over S S S . It has depth two as an S S S -module, because the two parameters form a regular sequence. The Auslander–Buchsbaum formula therefore gives projective dimension zero over the regular local ring S S S , so R ^ \widehat{R} R is finite free over S S S . The canonical-module construction for a finite algebra over a regular ring identifies the completed canonical module with
ω ^ R ≃ Hom S ( R ^ , S ) d u ∧ d v ; \widehat{\omega}_R \simeq\operatorname{Hom}_S(\widehat{R},S)\,\mathrm{d}u \wedge\mathrm{d}v; ω R ≃ Hom S ( R , S ) d u ∧ d v ;
This is [14 Theorem 21.15] in equal dimensions; Corollaries 21.17 and 21.18 there give compatibility with localization and completion. Choosing an S S S -basis of R ^ \widehat{R} R and reducing its dual basis modulo ( u , v ) (u,v) ( u , v ) gives
ω ^ R ( d , h ) ω ^ R ≃ Hom C ( R ^ / ( d , h ) R ^ , C ) . \frac{\widehat{\omega}_R}{(d,h)\widehat{\omega}_R} \simeq\operatorname{Hom}_{\mathbb{C}}\left(\widehat{R}/(d,h)\widehat{R},\mathbb{C}\right). ( d , h ) ω R ω R ≃ Hom C ( R / ( d , h ) R , C ) .
Both sides are unchanged by completion, since they have finite length. Combining these identifications over the finitely many support points gives Θ \Theta Θ . The same free-module description shows that h , d h,d h , d is a regular sequence on ω R \omega_R ω R near those points: multiplication by v v v , and then by u u u modulo v v v , is injective on the free S S S -module Hom S ( R ^ , S ) \operatorname{Hom}_S(\widehat{R},S) Hom S ( R , S ) .
The residue trace is now a concrete linear functional:
λ d , h ( [ η ] ) : = Θ ( [ η ] ) ( 1 ) , λ d , h ( a [ η ] ) = Θ ( [ η ] ) ( a ) . \lambda_{d,h}([\eta]) := \Theta([\eta])(1), \qquad\lambda_{d,h}(a[\eta]) = \Theta([\eta])(a). λ d , h ([ η ]) := Θ ([ η ]) ( 1 ) , λ d , h ( a [ η ]) = Θ ([ η ]) ( a ) .
Consequently the pairing ( a , [ η ] ) ↦ λ d , h ( a [ η ] ) (a,[\eta]) \mapsto\lambda_{d,h}(a[\eta]) ( a , [ η ]) ↦ λ d , h ( a [ η ]) is perfect: every functional on R d , h R_{d,h} R d , h is Θ ( [ η ] ) \Theta([\eta]) Θ ([ η ]) for a unique class [ η ] [\eta] [ η ] , and a nonzero element of R d , h R_{d,h} R d , h is detected by a linear functional. With the ordered parameters ( d , h ) (d,h) ( d , h ) , this trace is the local-duality residue of the generalized fraction with numerator η \eta η and denominator ( d , h ) (d,h) ( d , h ) . In the completed description it is the constant coefficient of the S S S -linear functional corresponding to η \eta η , evaluated at 1 1 1 ; it is summed over the support points. This is the coefficient realization of the residue trace; compare [39 Sections 7 and 9] .
To obtain the matrix pairing, put J : = adj ( T ) J:=\operatorname{adj}(T) J := adj ( T ) and use λ d , h ( η T J f ) \lambda_{d,h}(\eta^T J f) λ d , h ( η T J f ) . The identities T J = J T = d I n TJ=JT=dI_n T J = J T = d I n show that this expression vanishes on each of the submodules divided out above. If it vanishes for every η \eta η , scalar perfection gives J f ∈ ( d , h ) R n Jf\in(d,h)R^n J f ∈ ( d , h ) R n . Modulo h h h , write J f = d a Jf=da J f = d a and multiply by T T T . Cancellation of the regular element d d d gives f = T a f=Ta f = T a . For the other radical, scalar perfection gives J T η ∈ ( d , h ) ω R n J^T\eta\in(d,h)\omega_R^n J T η ∈ ( d , h ) ω R n . Multiplication by T T T^T T T and cancellation of d d d on ω R / h ω R \omega_R/h\omega_R ω R / h ω R give η ∈ T T ω R n + h ω R n \eta\in T^T\omega_R^n+h\omega_R^n η ∈ T T ω R n + h ω R n . These arguments apply at every support point; outside that set the quotients are zero. Thus both radicals vanish and the finite matrix pairing is perfect. The explicit divided differences in the next subsection are stated in the planar coordinates x , y x,y x , y .
The polynomial tensor that reproduces the quotient
For a perfect pairing of two finite-dimensional spaces, its coevaluation tensor is the inverse pairing written without choosing coordinates. More explicitly, if e 1 , … , e s e_1,\ldots,e_s e 1 , … , e s is a basis of the first space and e 1 D , … , e s D e^D_1,\ldots,e^D_s e 1 D , … , e s D is its dual basis in the second, the tensor is ∑ i e i ⊗ e i D \sum_i e_i\otimes e^D_i ∑ i e i ⊗ e i D . Contracting the second factor against f f f gives ∑ i e i b ( f , e i D ) = f \sum_i e_i b(f,e^D_i)=f ∑ i e i b ( f , e i D ) = f . We now construct this tensor for b T , h b_{T,h} b T , h by polynomial operations on T T T and h h h .
Use two points P = ( x , y ) P=(x,y) P = ( x , y ) and Q = ( u , v ) Q=(u,v) Q = ( u , v ) , with the intermediate point U = ( u , y ) U=(u,y) U = ( u , y ) . For a scalar polynomial or a polynomial matrix F F F , define the entrywise divided differences
δ x F : = F ( P ) − F ( U ) x − u , δ y F : = F ( U ) − F ( Q ) y − v . \delta_xF:=\frac{F(P)-F(U)}{x-u}, \qquad\delta_yF:=\frac{F(U)-F(Q)}{y-v}. δ x F := x − u F ( P ) − F ( U ) , δ y F := y − v F ( U ) − F ( Q ) .
Both are polynomial: the respective numerator vanishes when x = u x=u x = u or y = v y=v y = v . Thus the displayed quotients introduce no poles, including on the diagonal P = Q P=Q P = Q .
For the scalar complete intersection, put
B d , h ( P , Q ) : = δ x d δ y h − δ y d δ x h . B_{d,h}(P,Q):=\delta_x d\,\delta_y h-\delta_y d\,\delta_x h. B d , h ( P , Q ) := δ x d δ y h − δ y d δ x h .
The scalar Bézoutian identity states that
λ d , h ( Q ) ( B d , h ( P , Q ) a ( Q ) ) = a ( P ) ( m o d ( d ( P ) , h ( P ) ) ) ( a ∈ A ) . (67) \lambda_{d,h}^{(Q)}\left(B_{d,h}(P,Q)a(Q)\right)=a(P)\pmod{(d(P),h(P))}\qquad(a\in A). \tag*{(67)} λ d , h ( Q ) ( B d , h ( P , Q ) a ( Q ) ) = a ( P ) ( mod ( d ( P ) , h ( P ))) ( a ∈ A ) . (67)
The superscript means that the residue acts in u , v u,v u , v , leaving x , y x,y x , y as polynomial coefficients. This is the coevaluation identity for the scalar residue pairing, valid for the entire quotient, including its nilpotent elements [20 Proposition 4.3 and Lemma 4.4] . For d = x d=x d = x and h = y h=y h = y , both the Bézoutian and the residue of 1 1 1 are 1 1 1 , fixing the same orientation as (64) .
Here is a derivation that also explains the multiplicities. The residue change-of-equations rule says that, if ordered denominator columns satisfy g = A f g=Af g = A f , then
Res g ( det ( A ) a ) = Res f ( a ) \operatorname{Res}_{g}\left(\det(A)a\right)=\operatorname{Res}_{f}(a) Res g ( det ( A ) a ) = Res f ( a )
whenever the corresponding quotients are finite and the sequences are regular; it applies with finite coefficient algebras as well [19 Section 2, especially Proposition 2.1] . Work over the coefficient algebra R : = C [ x , y ] / ( d ( P ) , h ( P ) ) R:=\mathbb{C}[x,y]/(d(P),h(P)) R := C [ x , y ] / ( d ( P ) , h ( P )) . The divided differences give in R [ u , v ] R[u,v] R [ u , v ] the exact identity
[ d ( Q ) h ( Q ) ] = [ δ x d δ y d δ x h δ y h ] [ u − x v − y ] . \begin{bmatrix} d(Q) \\ h(Q) \end{bmatrix}
=
\begin{bmatrix} \delta_xd & \delta_yd \\ \delta_xh & \delta_yh \end{bmatrix}
\begin{bmatrix} u-x \\ v-y \end{bmatrix}. [ d ( Q ) h ( Q ) ] = [ δ x d δ x h δ y d δ y h ] [ u − x v − y ] .
The determinant is B d , h B_{d,h} B d , h . Both denominator sequences are regular over R R R , and their quotients are finite free R R R -modules: respectively R ⊗ C A / ( d , h ) R\otimes_{\mathbb{C}} A/(d,h) R ⊗ C A / ( d , h ) and R R R . The change-of-equations rule therefore gives
λ d , h ( Q ) ( B d , h a ( Q ) ) = Res u − x , v − y ( a ( u , v ) ) = a ( x , y ) in R . \lambda_{d,h}^{(Q)}\left(B_{d,h}a(Q)\right)
=
\operatorname{Res}_{u-x,v-y}\left(a(u,v)\right)
=
a(x,y)
\quad\text{in }R. λ d , h ( Q ) ( B d , h a ( Q ) ) = Res u − x , v − y ( a ( u , v ) ) = a ( x , y ) in R .
The last equality is the two successive Cauchy coefficient formulas. It remains valid for nilpotent x , y x,y x , y : on a local component write x = c + n x=c+n x = c + n with n N = 0 n^N=0 n N = 0 , and expand 1 / ( u − x ) = ∑ j = 0 N − 1 n j / ( u − c ) j + 1 1/(u-x)=\sum_{j=0}^{N-1}n^j/(u-c)^{j+1} 1/ ( u − x ) = ∑ j = 0 N − 1 n j / ( u − c ) j + 1 ; the residue is the finite Taylor expansion at c c c . Do the same for v − y v-y v − y . This proves (67) on all remainder classes. Finally, if λ d , h ( a b ) = 0 \lambda_{d,h}(ab)=0 λ d , h ( ab ) = 0 for every b b b , expand B d , h = ∑ ℓ p ℓ ( P ) q ℓ ( Q ) B_{d,h}=\sum_{\ell}p_{\ell}(P)q_{\ell}(Q) B d , h = ∑ ℓ p ℓ ( P ) q ℓ ( Q ) in finite tensors. The reconstruction identity gives [ a ] = ∑ ℓ [ p ℓ ] λ d , h ( q ℓ a ) = 0 [a]=\sum_{\ell}[p_{\ell}]\lambda_{d,h}(q_{\ell}a)=0 [ a ] = ∑ ℓ [ p ℓ ] λ d , h ( q ℓ a ) = 0 . Thus the scalar pairing has zero radical and is perfect.
Theorem 4.8 (Matrix coevaluation for a general scalar cut). Under (63) , define
C T , h ( P , Q ) : = δ x T δ y h − δ y T δ x h . (68) \mathcal{C}_{T,h}(P,Q):=\delta_xT\,\delta_yh-\delta_yT\,\delta_xh. \tag*{(68)} C T , h ( P , Q ) := δ x T δ y h − δ y T δ x h . (68) Its class in Q T , h ⊗ C Q T , h D Q_{T,h}\otimes_{\mathbb{C}}Q_{T,h}^{D} Q T , h ⊗ C Q T , h D is the coevaluation tensor of (66) . The row index of C T , h \mathcal{C}_{T,h} C T , h belongs to the first quotient and its column index to the second. Equivalently, for every f ∈ A n f\in A^n f ∈ A n ,
λ d , h ( Q ) ( C T , h ( P , Q ) adj ( T ( Q ) ) f ( Q ) ) ≡ f ( P ) ( m o d T ( P ) A n n + h ( P ) A n n ) . (69) \lambda_{d,h}^{(Q)}\left(\mathcal{C}_{T,h}(P,Q)\operatorname{adj}(T(Q))f(Q)\right)
\equiv f(P)\pmod{T(P)A_n^n+h(P)A_n^n}. \tag*{(69)} λ d , h ( Q ) ( C T , h ( P , Q ) adj ( T ( Q )) f ( Q ) ) ≡ f ( P ) ( mod T ( P ) A n n + h ( P ) A n n ) . (69) The residue on a vector is taken entrywise. This statement permits arbitrary intersection multiplicities and imposes no restriction that h h h depend on only one coordinate.
Proof. Put J : = adj ( T ) J:=\operatorname{adj}(T) J := adj ( T ) and form C J , h \mathcal{C}_{J,h} C J , h by the same definition (68) . We first establish an identity of polynomial matrices before taking any residue or quotient:
B d , h ( P , Q ) I n = C T , h ( P , Q ) J ( Q ) + T ( P ) C J , h ( P , Q ) + [ h ( P ) − h ( Q ) ] δ x T δ y J . (70) B_{d,h}(P,Q)I_n
=
\mathcal{C}_{T,h}(P,Q)J(Q)+T(P)\mathcal{C}_{J,h}(P,Q)
+\left[h(P)-h(Q)\right]\delta_xT\,\delta_yJ. \tag*{(70)} B d , h ( P , Q ) I n = C T , h ( P , Q ) J ( Q ) + T ( P ) C J , h ( P , Q ) + [ h ( P ) − h ( Q ) ] δ x T δ y J . (70)
To derive it, take the two consecutive differences of T J = d I n TJ=dI_n T J = d I n . The product rule, with the factors kept in their original order, gives
δ x d I n = δ x T J ( U ) + T ( P ) δ x J , \delta_xd\,I_n=\delta_xT\,J(U)+T(P)\delta_xJ, δ x d I n = δ x T J ( U ) + T ( P ) δ x J ,
δ y d I n = δ y T J ( Q ) + T ( U ) δ y J . \delta_yd\,I_n=\delta_yT\,J(Q)+T(U)\delta_yJ. δ y d I n = δ y T J ( Q ) + T ( U ) δ y J .
Multiply the first equality by δ y h \delta_yh δ y h and subtract the second multiplied by δ x h \delta_xh δ x h . After subtracting C T , h ( P , Q ) J ( Q ) \mathcal{C}_{T,h}(P,Q)J(Q) C T , h ( P , Q ) J ( Q ) , the remaining terms are
δ x T [ J ( U ) − J ( Q ) ] δ y h + T ( P ) δ x J δ y h − T ( U ) δ y J δ x h . \delta_xT\left[J(U)-J(Q)\right]\delta_yh
+
T(P)\delta_xJ\,\delta_yh
-
T(U)\delta_yJ\,\delta_xh. δ x T [ J ( U ) − J ( Q ) ] δ y h + T ( P ) δ x J δ y h − T ( U ) δ y J δ x h .
Use J ( U ) − J ( Q ) = ( y − v ) δ y J J(U)-J(Q)=(y-v)\delta_yJ J ( U ) − J ( Q ) = ( y − v ) δ y J and the scalar identity
( y − v ) δ y h = h ( P ) − h ( Q ) − ( x − u ) δ x h . (y-v)\delta_yh=h(P)-h(Q)-(x-u)\delta_xh. ( y − v ) δ y h = h ( P ) − h ( Q ) − ( x − u ) δ x h .
Since ( x − u ) δ x T = T ( P ) − T ( U ) (x-u)\delta_xT=T(P)-T(U) ( x − u ) δ x T = T ( P ) − T ( U ) , the expression becomes
[ h ( P ) − h ( Q ) ] δ x T δ y J − [ T ( P ) − T ( U ) ] δ y J δ x h + T ( P ) δ x J δ y h − T ( U ) δ y J δ x h . \left[h(P)-h(Q)\right]\delta_xT\,\delta_yJ
-
\left[T(P)-T(U)\right]\delta_yJ\,\delta_xh
+
T(P)\delta_xJ\,\delta_yh
-
T(U)\delta_yJ\,\delta_xh. [ h ( P ) − h ( Q ) ] δ x T δ y J − [ T ( P ) − T ( U ) ] δ y J δ x h + T ( P ) δ x J δ y h − T ( U ) δ y J δ x h .
The two terms containing T ( U ) T(U) T ( U ) cancel. The other terms containing T ( P ) T(P) T ( P ) give T ( P ) C J , h T(P)\mathcal{C}_{J,h} T ( P ) C J , h , proving (70) . Only the divided differences of the scalar h h h have been moved past matrices.
Now multiply (70) on the right by f ( Q ) f(Q) f ( Q ) and apply λ d , h ( Q ) \lambda_{d,h}^{(Q)} λ d , h ( Q ) . By (67) , the left-hand side is f ( P ) f(P) f ( P ) modulo ( d ( P ) , h ( P ) ) A n (d(P),h(P))A^n ( d ( P ) , h ( P )) A n . The second term on the right belongs to T ( P ) A n T(P)A^n T ( P ) A n . In the last term, the contribution with h ( P ) h(P) h ( P ) belongs to h ( P ) A n h(P)A^n h ( P ) A n , while the contribution with h ( Q ) h(Q) h ( Q ) has zero residue. Finally,
d ( P ) A n ⊆ T ( P ) A n , d ( P ) I n = T ( P ) J ( P ) . d(P)A^n \subseteq T(P)A^n,\qquad d(P)I_n=T(P)J(P). d ( P ) A n ⊆ T ( P ) A n , d ( P ) I n = T ( P ) J ( P ) .
Passing to Q T , h Q_{T,h} Q T , h therefore leaves precisely (69) .
The contraction in that formula pairs the second tensor factor with [ f ] [f] [ f ] by (66) ; this explains the order C T , h J ( Q ) f ( Q ) \mathcal{C}_{T,h}J(Q)f(Q) C T , h J ( Q ) f ( Q ) . Theorem 4.6 makes this pairing perfect. Hence a tensor whose contraction is the identity is its unique coevaluation tensor, as asserted. □
For a cut h = k ( y ) h=k(y) h = k ( y ) , the second summand in (68) vanishes, and the formula reduces to
C T , k ( P , Q ) = T ( x , y ) − T ( u , y ) x − u k ( y ) − k ( v ) y − v . \mathcal{C}_{T,k}(P,Q)=\frac{T(x,y)-T(u,y)}{x-u}\frac{k(y)-k(v)}{y-v}. C T , k ( P , Q ) = x − u T ( x , y ) − T ( u , y ) y − v k ( y ) − k ( v ) .
Repeated roots of k k k are allowed. For a general cut, both summands are needed, as the following small example shows. Take
T = [ x y 0 x ] , h = y 2 − x . T=\begin{bmatrix}x&y\\0&x\end{bmatrix},\qquad h=y^2-x. T = [ x 0 y x ] , h = y 2 − x .
Then d = x 2 d=x^2 d = x 2 and A / ( d , h ) ≃ C [ y ] / ( y 4 ) A/(d,h)\simeq\mathbb{C}[y]/(y^4) A / ( d , h ) ≃ C [ y ] / ( y 4 ) by substituting x = y 2 x=y^2 x = y 2 . The scalar residue is the coefficient of y 3 y^3 y 3 in this remainder. Its normalization agrees with λ d , h ( d x h y − d y h x ) = λ d , h ( 4 x y ) = 4 \lambda_{d,h}(d_xh_y-d_yh_x)=\lambda_{d,h}(4xy)=4 λ d , h ( d x h y − d y h x ) = λ d , h ( 4 x y ) = 4 . The four divided differences give
δ x T = I 2 , δ y T = [ 0 1 0 0 ] , δ x h = − 1 , δ y h = y + v . \delta_xT=I_2,\qquad\delta_yT=\begin{bmatrix}0&1\\0&0\end{bmatrix},\qquad\delta_xh=-1,\qquad\delta_yh=y+v. δ x T = I 2 , δ y T = [ 0 0 1 0 ] , δ x h = − 1 , δ y h = y + v .
Consequently
C T , h ( P , Q ) = [ y + v 1 0 y + v ] . \mathcal{C}_{T,h}(P,Q)=\begin{bmatrix}y+v&1\\0&y+v\end{bmatrix}. C T , h ( P , Q ) = [ y + v 0 1 y + v ] .
The off-diagonal entry comes from the second summand of (68) . In the matrix quotient the relations are y 2 e 1 = 0 y^2e_1=0 y 2 e 1 = 0 and y e 1 + y 2 e 2 = 0 ye_1+y^2e_2=0 y e 1 + y 2 e 2 = 0 , after eliminating x x x . They give the basis e 1 , e 2 , y e 2 , y 2 e 2 e_1,e_2,ye_2,y^2e_2 e 1 , e 2 , y e 2 , y 2 e 2 , on which we now check the claimed reproduction. In the residue variable, put u = v 2 u=v^2 u = v 2 and take the coefficient of v 3 v^3 v 3 . Multiplication by the adjugate gives
[ y + v 1 0 y + v ] [ v 2 − v 0 v 2 ] = [ y v 2 + v 3 − y v 0 y v 2 + v 3 ] . \begin{bmatrix}y+v&1\\0&y+v\end{bmatrix}
\begin{bmatrix}v^2&-v\\0&v^2\end{bmatrix}
=
\begin{bmatrix}yv^2+v^3&-yv\\0&yv^2+v^3\end{bmatrix}. [ y + v 0 1 y + v ] [ v 2 0 − v v 2 ] = [ y v 2 + v 3 0 − y v y v 2 + v 3 ] .
Consequently contraction sends the four basis vectors to e 1 , e 2 , y e 2 , − y e 1 e_1,e_2,ye_2,-ye_1 e 1 , e 2 , y e 2 , − y e 1 , respectively. The last image equals y 2 e 2 y^2e_2 y 2 e 2 in the quotient by its defining relation, so all four classes are reproduced. The basis also follows directly from the relations: y 2 e 1 = 0 y^2e_1=0 y 2 e 1 = 0 and y e 1 = − y 2 e 2 ye_1=-y^2e_2 y e 1 = − y 2 e 2 imply y 3 e 2 = 0 y^3e_2=0 y 3 e 2 = 0 ; the four remaining classes are independent, as the residue Gram against e 2 , e 1 , y e 1 , y 2 e 1 e_2,e_1,ye_1,y^2e_1 e 2 , e 1 , y e 1 , y 2 e 1 in the transposed quotient is
[ 0 0 1 0 0 0 0 − 1 1 0 − 1 0 0 − 1 0 0 ] , det = 1. \begin{bmatrix}
0&0&1&0\\
0&0&0&-1\\
1&0&-1&0\\
0&-1&0&0
\end{bmatrix},
\qquad\det=1. 0 0 1 0 0 0 0 − 1 1 0 − 1 0 0 − 1 0 0 , det = 1.
From a finite residue to a bilateral correction.
The constructions above allow T T T to be an ordered product of noncommuting polynomial matrices. The second equation is the scalar h h h , so it commutes with all the matrix operations in the proofs. For row–column pairings, the orientation is the same as in Remark 4.5. Namely, if f f f is a polynomial row, g g g a polynomial column and c ≠ 0 c\ne0 c = 0 , then
ν ( f , g ) : = c λ d , h ( f adj ( T ) g ) = c b T T , h ( f T , g ) . \nu(f,g):=c\lambda_{d,h}(f\operatorname{adj}(T)g)=cb_{T^\mathsf{T},h}(f^\mathsf{T},g). ν ( f , g ) := c λ d , h ( f adj ( T ) g ) = c b T T , h ( f T , g ) .
Its quotient pair is therefore Q T T , h , Q T , h Q_{T^\mathsf{T},h},Q_{T,h} Q T T , h , Q T , h , and its coevaluation in this order is c − 1 C T , h c^{-1}\mathcal{C}_{T,h} c − 1 C T , h . The transpose changes matrix indices without interchanging P P P and Q Q Q . Passing to the column–row convention of the finite reproducing kernel then exchanges the tensor factors. More generally, return to the polynomial test matrix H H H of the bilateral functional equation. As in (59) , column vectorization gives
vec ( L 1 T H R 1 T ) = T 1 vec ( H ) , T 1 : = R 1 ⊗ L 1 T . \operatorname{vec}(L_1^{\mathsf{T}} H R_1^{\mathsf{T}})=T_1\operatorname{vec}(H),\qquad T_1:=R_1\otimes L_1^{\mathsf{T}}. vec ( L 1 T H R 1 T ) = T 1 vec ( H ) , T 1 := R 1 ⊗ L 1 T .
Assume that T 1 T_1 T 1 and a chosen scalar h h h satisfy (63) . The masses which annihilate both the image of T 1 T_1 T 1 and multiplication by h h h are exactly the functionals on Q T 1 , h Q_{T_1,h} Q T 1 , h . Theorem 4.6 writes every such mass uniquely as
ν [ H ] = λ det T 1 , h ( η T adj ( T 1 ) vec ( H ) ) , [ η ] ∈ Q T 1 , h D . \nu[H]=\lambda_{\det T_1,h}\left(\eta^{\mathsf{T}}\operatorname{adj}(T_1)\operatorname{vec}(H)\right),\qquad[\eta]\in Q_{T_1,h}^{D}. ν [ H ] = λ d e t T 1 , h ( η T adj ( T 1 ) vec ( H ) ) , [ η ] ∈ Q T 1 , h D .
The coevaluation of Theorem 4.8 recovers these finite remainder classes. The scalar equation specifies the finite sector: the full dual of coker T 1 \operatorname{coker}T_1 coker T 1 may also contain masses that do not annihilate h h h .
To use such a mass in an orthogonality problem, choose the prescribed finite row and column spaces and evaluate its entries H ν = [ ν ( f i , g j ) ] i , j H_{\nu}=[\nu(f_i,g_j)]_{i,j} H ν = [ ν ( f i , g j ) ] i , j . The finite-coordinate calculation of Proposition 3.11 then gives G ^ = T L G r a t T R + H ν \widehat{G}=T_LG_{\mathrm{rat}}T_R+H_{\nu} G = T L G rat T R + H ν . Its invertibility is exactly the normality condition on those spaces. If both G ^ \widehat{G} G and G p a r t = T L G r a t T R G_{\mathrm{part}}=T_LG_{\mathrm{rat}}T_R G part = T L G rat T R are invertible, (45) gives the mass correction to their finite reproducing kernels, of separation rank rank H ν \operatorname{rank}H_{\nu} rank H ν . This step uses the residue data together with the rational moments and the chosen polynomial spaces; perfection of the residue quotient alone does not imply normality of the complete problem.
A polynomial residue tensor on the singular cone
The finite residue construction also gives an explicit reproducing tensor on a singular surface. In this example the calculation reduces to the plane by writing each function on the surface in two components. The resulting tensor reproduces classes in a finite polynomial quotient. Its construction requires no orthogonality measure or choice of polynomial degree bounds.
Take the complex coordinate algebras
A : = C [ x , y , z ] / ( z 2 − x 2 − y 2 ) , B : = C [ x , y ] . A:=\mathbb{C}[x,y,z]/(z^2-x^2-y^2),\qquad B:=\mathbb{C}[x,y]. A := C [ x , y , z ] / ( z 2 − x 2 − y 2 ) , B := C [ x , y ] .
The polynomial defining the surface is irreducible: its quadratic form has rank three and therefore cannot be a product of two linear forms. The surface has its unique singular point at the origin, where the three partial derivatives vanish. Every class in A A A has a unique representative p 0 ( x , y ) + z p 1 ( x , y ) p_0(x,y)+zp_1(x,y) p 0 ( x , y ) + z p 1 ( x , y ) , because division by the monic polynomial z 2 − x 2 − y 2 z^2-x^2-y^2 z 2 − x 2 − y 2 leaves degree at most one in z z z . Thus A A A is a free B B B -module with basis 1 , z 1,z 1 , z . Define
φ [ p 0 + z p 1 ] : = p 1 . (71) \varphi[p_0+zp_1]:=p_1. \tag*{(71)} φ [ p 0 + z p 1 ] := p 1 . (71)
This map takes the coefficient of z z z after using the surface equation to reduce higher powers. The matrix of the B B B -valued pairing ( p , q ) ↦ φ [ p q ] (p,q)\mapsto\varphi[pq] ( p , q ) ↦ φ [ pq ] in the basis 1 , z 1,z 1 , z is
( 0 1 1 0 ) . \begin{pmatrix}
0 & 1\\
1 & 0
\end{pmatrix}. ( 0 1 1 0 ) .
It is invertible over B B B , including above the singular point. Consequently every B B B -linear functional on A A A is uniquely of the form p ↦ φ [ c p ] p\mapsto\varphi[cp] p ↦ φ [ c p ] for some c ∈ A c\in A c ∈ A . In the canonical-module description associated with the finite projection to the ( x , y ) (x,y) ( x , y ) -plane, this says that φ d x ∧ d y \varphi\,\mathrm{d}x\wedge\mathrm{d}y φ d x ∧ d y is a generator. The reproduction calculation below uses the coefficient map (71) directly.
Fix a real parameter ε \varepsilon ε and put
a : = 1 + ε 2 , h : = a x 2 + ε 2 y 2 , k : = y 2 + ε 2 . a:=1+\varepsilon^2,\qquad h:=ax^2+\varepsilon^2y^2,\qquad k:=y^2+\varepsilon^2. a := 1 + ε 2 , h := a x 2 + ε 2 y 2 , k := y 2 + ε 2 .
The parameter is real, while x , y , z x,y,z x , y , z are coordinate classes in the complex algebra A A A . In particular, the finite quotients below are taken over C \mathbb{C} C ; their supporting points need not be real. The matrix needed for the residue calculation is
T : = [ h 2 x y 0 2 h ] , adj T : = [ 2 h − 2 x y 0 h ] , det T = 2 h 2 . T:=\begin{bmatrix} h & 2xy\\ 0 & 2h \end{bmatrix},\qquad
\operatorname{adj}T:=\begin{bmatrix} 2h & -2xy\\ 0 & h \end{bmatrix},\qquad
\det T=2h^2. T := [ h 0 2 x y 2 h ] , adj T := [ 2 h 0 − 2 x y h ] , det T = 2 h 2 .
This matrix also has a bilateral denominator interpretation. Put ℓ : = x − i ε z \ell:=x-\mathrm{i}\varepsilon z ℓ := x − i ε z and r : = x + i ε z r:=x+\mathrm{i}\varepsilon z r := x + i ε z , and choose
L 1 = [ ℓ y 0 ℓ ] , R 1 = [ r y 0 2 r ] . L_1=\begin{bmatrix} \ell& y\\ 0 & \ell\end{bmatrix},\qquad
R_1=\begin{bmatrix} r & y\\ 0 & 2r \end{bmatrix}. L 1 = [ ℓ 0 y ℓ ] , R 1 = [ r 0 y 2 r ] .
The surface equation gives ℓ r = h \ell r=h ℓ r = h and ℓ + r = 2 x \ell+r=2x ℓ + r = 2 x . Consequently R 1 L 1 = T R_1L_1=T R 1 L 1 = T , although the factors themselves involve the surface coordinate z z z and do not commute in general. All entries of their product belong to B B B . The polynomial h h h is monic up to the nonzero factor a a a as a polynomial in x x x , and k k k is monic in y y y . Their simultaneous quotients can therefore be computed by successive polynomial division, including at ε = 0 \varepsilon=0 ε = 0 .
For j ∈ { 1 , 2 } j\in\{1,2\} j ∈ { 1 , 2 } and p ∈ B p\in B p ∈ B , divide first in x x x by ( h / a ) j (h/a)^j ( h / a ) j and then divide the coefficients of the remainder in y y y by k k k . Denote the final remainder by rem j ( p ) \operatorname{rem}_j(p) rem j ( p ) . It has degree less than 2 j 2j 2 j in x x x and less than two in y y y . Define
ρ j [ p ] : = a − j [ x 2 j − 1 y ] rem j ( p ) , Λ j : = ρ j ∘ φ . (72) \rho_j[p]:=a^{-j}[x^{2j-1}y]\operatorname{rem}_j(p),\qquad\Lambda_j:=\rho_j\circ\varphi. \tag*{(72)} ρ j [ p ] := a − j [ x 2 j − 1 y ] rem j ( p ) , Λ j := ρ j ∘ φ . (72)
Here [ x u y v ] [x^u y^v] [ x u y v ] extracts the indicated coefficient. The factor a − j a^{-j} a − j compensates for the leading coefficient a j a^j a j of h j h^j h j in x x x . With the residue convention of (64) , ρ j \rho_j ρ j is the scalar residue for the ordered denominators ( h j , k ) (h^j,k) ( h j , k ) and differential d x ∧ d y \mathrm{d}x\wedge\mathrm{d}y d x ∧ d y . Indeed, summing the residues in x x x of a reduced numerator over h j h^j h j picks its x 2 j − 1 x^{2j-1} x 2 j − 1 coefficient divided by a j a^j a j ; terms of lower x x x degree have no x − 1 x^{-1} x − 1 term at infinity. The subsequent sum of the y y y residues over the monic quadratic k k k picks the coefficient of y y y . This gives exactly (72) , including when either polynomial has multiple roots.
These functionals give perfect finite pairings. To see this without assuming that the supporting points are simple, first work over R : = C [ y ] / ( k ) R:=\mathbb{C}[y]/(k) R := C [ y ] / ( k ) . Division by the monic polynomial ( h / a ) j (h/a)^j ( h / a ) j gives the free R R R -basis 1 , x , … , x 2 j − 1 1,x,\ldots,x^{2j-1} 1 , x , … , x 2 j − 1 . The pairing that takes the coefficient of x 2 j − 1 x^{2j-1} x 2 j − 1 in a product has zeros whenever the sum of the two basis exponents is less than 2 j − 1 2j-1 2 j − 1 , and has ones when that sum equals 2 j − 1 2j-1 2 j − 1 . Reversing the columns of its matrix therefore gives a triangular matrix with diagonal entries one. Its determinant is a unit in R R R .
Next, the coefficient of y y y gives an invertible pairing on R R R : its matrix in the basis 1 , y 1,y 1 , y is
( 0 1 1 0 ) , \begin{pmatrix}0&1\\1&0\end{pmatrix}, ( 0 1 1 0 ) ,
since y 2 = − ε 2 y^2=-\varepsilon^2 y 2 = − ε 2 . Composing the two coefficient pairings and multiplying by a − j a^{-j} a − j proves that ( [ p ] , [ q ] ) ↦ ρ j [ p q ] ([p],[q])\mapsto\rho_j[pq] ([ p ] , [ q ]) ↦ ρ j [ pq ] is perfect on B / ( h j , k ) B/(h^j,k) B / ( h j , k ) . Composing once more with (71) proves that ( [ p ] , [ q ] ) ↦ Λ j [ p q ] ([p],[q])\mapsto\Lambda_j[pq] ([ p ] , [ q ]) ↦ Λ j [ pq ] is perfect on A / ( h j , k ) A/(h^j,k) A / ( h j , k ) . This last quotient has the 8 j 8j 8 j basis classes x u y v z w x^u y^v z^w x u y v z w , with u ∈ { 0 , … , 2 j − 1 } u\in\{0,\ldots,2j-1\} u ∈ { 0 , … , 2 j − 1 } and v , w ∈ { 0 , 1 } v,w\in\{0,1\} v , w ∈ { 0 , 1 } .
Polynomial multiplication of a functional means ( b Λ ) [ p ] : = Λ [ b p ] (b\Lambda)[p]:=\Lambda[bp] ( b Λ ) [ p ] := Λ [ b p ] . The definitions imply
h Λ 2 = Λ 1 , h Λ 1 = 0 , h 2 Λ 2 = 0 , k Λ 1 = k Λ 2 = 0. (73) h\Lambda_2=\Lambda_1,\qquad h\Lambda_1=0,\qquad h^2\Lambda_2=0,\qquad k\Lambda_1=k\Lambda_2=0. \tag*{(73)} h Λ 2 = Λ 1 , h Λ 1 = 0 , h 2 Λ 2 = 0 , k Λ 1 = k Λ 2 = 0. (73)
For the first identity, divide a polynomial in B B B by h h h and write its remainder as c 0 ( y ) + x c 1 ( y ) c_0(y)+xc_1(y) c 0 ( y ) + x c 1 ( y ) . Modulo h 2 h^2 h 2 , its product with h h h has x 3 x^3 x 3 coefficient a c 1 ( y ) ac_1(y) a c 1 ( y ) . The factors a − 2 a^{-2} a − 2 and a − 1 a^{-1} a − 1 in (4.26) give h ρ 2 = ρ 1 h\rho_2=\rho_1 h ρ 2 = ρ 1 . Since h h h belongs to B B B , it commutes with φ \varphi φ , giving the stated identity for Λ j \Lambda_j Λ j . The remaining identities follow because the remainders annihilate the respective defining ideals.
Define the matrix functional
ν : = ( adj T ) Λ 2 = ( 2 Λ 1 − 2 x y Λ 2 0 Λ 1 ) . (74) \nu:=(\operatorname{adj}T)\Lambda_2=
\begin{pmatrix}
2\Lambda_1&-2xy\Lambda_2\\
0&\Lambda_1
\end{pmatrix}. \tag*{(74)} ν := ( adj T ) Λ 2 = ( 2 Λ 1 0 − 2 x y Λ 2 Λ 1 ) . (74)
For a row f f f and a column g g g , it acts as ν ( f , g ) = Λ 2 [ f adj ( T ) g ] \nu(f,g)=\Lambda_2[f\operatorname{adj}(T)g] ν ( f , g ) = Λ 2 [ f adj ( T ) g ] . It is a homogeneous mass for the denominator pair just defined. Indeed, T = R 1 L 1 T=R_1L_1 T = R 1 L 1 and the adjugate product rule give
L 1 adj ( T ) R 1 = L 1 adj ( L 1 ) adj ( R 1 ) R 1 = det ( L 1 ) det ( R 1 ) I 2 = 2 h 2 I 2 . \begin{aligned}
L_1\operatorname{adj}(T)R_1=L_1\operatorname{adj}(L_1)\operatorname{adj}(R_1)R_1 \\
&=\det(L_1)\det(R_1)I_2=2h^2I_2.
\end{aligned} L 1 adj ( T ) R 1 = L 1 adj ( L 1 ) adj ( R 1 ) R 1 = det ( L 1 ) det ( R 1 ) I 2 = 2 h 2 I 2 .
Therefore
ν ( f L 1 , R 1 g ) = 2 Λ 2 [ h 2 f g ] = 0 , k ν = 0 , \nu(fL_1,R_1g)=2\Lambda_2[h^2fg]=0,\qquad k\nu=0, ν ( f L 1 , R 1 g ) = 2 Λ 2 [ h 2 f g ] = 0 , k ν = 0 ,
by (73) . Thus both defining annihilation equations hold without choosing an underlying measure.
The adjugate identities and (73) show that this pairing depends only on the classes of f T f^T f T and g g g in
Q L : = A 2 / ( T T A 2 + k A 2 ) , Q R : = A 2 / ( T A 2 + k A 2 ) , Q_L:=A^2/(T^TA^2+kA^2),\qquad Q_R:=A^2/(TA^2+kA^2), Q L := A 2 / ( T T A 2 + k A 2 ) , Q R := A 2 / ( T A 2 + k A 2 ) ,
respectively. For example, adding T b Tb T b to g g g contributes 2 Λ 2 [ h 2 f b ] = 0 2\Lambda_2[h^2fb]=0 2 Λ 2 [ h 2 f b ] = 0 . Adding a multiple of k k k also contributes zero. These quotients are finite dimensional, because both h 2 h^2 h 2 and k k k annihilate them. More precisely, each has the sixteen basis classes
x u y v z w e i , u , v , w ∈ { 0 , 1 } , i ∈ { 1 , 2 } . x^u y^v z^w e_i,\qquad u,v,w\in\{0,1\},\quad i\in\{1,2\}. x u y v z w e i , u , v , w ∈ { 0 , 1 } , i ∈ { 1 , 2 } .
To obtain this basis, first separate the coefficients of 1 , z 1,z 1 , z . Over C [ y ] / ( k ) \mathbb{C}[y]/(k) C [ y ] / ( k ) , both T T T and T T T^{\mathsf T} T T are quadratic matrix polynomials in x x x with invertible leading matrix diag ( a , 2 a ) \operatorname{diag}(a,2a) diag ( a , 2 a ) . Polynomial division therefore leaves a unique vector remainder of degree less than two in x x x . The two coefficients in y y y and the two coefficients in z z z give the displayed basis. This also proves its independence.
The pairing is perfect on these quotients. There is a direct determinant check: order the eight scalar monomials x u y v z w x^{u}y^{v}z^{w} x u y v z w lexicographically by ( u , v , w ) (u,v,w) ( u , v , w ) . The matrix H H H of Λ 1 [ p q ] \Lambda_1[pq] Λ 1 [ pq ] is a − 1 a^{-1} a − 1 times the reversal matrix. Only complementary exponent triples contribute the coefficient x y z xyz x y z . Thus det H = a − 8 \det H=a^{-8} det H = a − 8 . The matrix of ν \nu ν , listing component 1 before component 2, has the block form ( 2 H ∗ 0 H ) \begin{pmatrix}2H&*\\0&H\end{pmatrix} ( 2 H 0 ∗ H ) , and its determinant is
2 8 ( det H ) 2 = 256 ( 1 + ε 2 ) 16 . 2^{8}(\det H)^2=\frac{256}{(1+\varepsilon^2)^{16}}. 2 8 ( det H ) 2 = ( 1 + ε 2 ) 16 256 .
It remains nonzero at ε = 0 \varepsilon=0 ε = 0 , where the support reaches the vertex. A coevaluation tensor for this pairing reproduces a quotient class by contraction with the opposite argument, as in the following explicit identity.
Proposition 4.9 (Polynomial coevaluation on the cone). For the independent arguments P = ( x , y , z ) P=(x,y,z) P = ( x , y , z ) and Q = ( u , v , w ) Q=(u,v,w) Q = ( u , v , w ) , the coevaluation of (4.28) has the representative
C ν ( P , Q ) : = ( z + w ) ( y + v ) 2 ( a ( x + u ) 0 2 y 2 a ( x + u ) ) (75) \mathcal{C}_{\nu}(P,Q):=\frac{(z+w)(y+v)}{2}
\begin{pmatrix}
a(x+u)&0\\
2y&2a(x+u)
\end{pmatrix} \tag*{(75)} C ν ( P , Q ) := 2 ( z + w ) ( y + v ) ( a ( x + u ) 2 y 0 2 a ( x + u ) ) (75) It represents a tensor in Q L ⊗ C Q R Q_L\otimes_{\mathbb{C}}Q_R Q L ⊗ C Q R , with the first factor indexed by its rows and the second by its columns.
For every column f ∈ A 2 f\in A^2 f ∈ A 2 ,
Λ 2 , Q [ C ν ( P , Q ) adj ( T ( Q ) ) T f ( Q ) ] = f ( P ) ( m o d T ( P ) T A 2 + k ( P ) A 2 ) . (76) \Lambda_{2,Q}\left[\mathcal{C}_{\nu}(P,Q)\operatorname{adj}(T(Q))^{\mathsf T}f(Q)\right]
=f(P)\pmod{T(P)^{\mathsf T}A^2+k(P)A^2}. \tag*{(76)} Λ 2 , Q [ C ν ( P , Q ) adj ( T ( Q ) ) T f ( Q ) ] = f ( P ) ( mod T ( P ) T A 2 + k ( P ) A 2 ) . (76) Here Λ 2 , Q \Lambda_{2,Q} Λ 2 , Q acts entry by entry in the Q Q Q variables. The transposed identity, with P , Q P,Q P , Q exchanged, reproduces the classes of Q R Q_R Q R . The displayed representative is polynomial in ε \varepsilon ε and remains defined at the singular vertex.
Proof. First apply Theorem 4.8 on the plane to T T T^{\mathsf T} T T and k k k . Their divided differences are
T ( x , y ) T − T ( u , y ) T x − u = ( a ( x + u ) 0 2 y 2 a ( x + u ) ) , k ( y ) − k ( v ) y − v = y + v . \frac{T(x,y)^{\mathsf T}-T(u,y)^{\mathsf T}}{x-u}
=
\begin{pmatrix}
a(x+u)&0\\
2y&2a(x+u)
\end{pmatrix},
\qquad
\frac{k(y)-k(v)}{y-v}=y+v. x − u T ( x , y ) T − T ( u , y ) T = ( a ( x + u ) 2 y 0 2 a ( x + u ) ) , y − v k ( y ) − k ( v ) = y + v .
Let B T ( x , y ; u , v ) B_T(x,y;u,v) B T ( x , y ; u , v ) be the product of these two displayed expressions. The theorem pairs this tensor with the scalar residue for the denominators ( det T , k ) = ( 2 h 2 , k ) (\det T,k)=(2h^2,k) ( det T , k ) = ( 2 h 2 , k ) . That scalar residue is ρ 2 / 2 \rho_2/2 ρ 2 /2 . Consequently its reproduction identity for a polynomial column b ∈ B 2 b\in B^2 b ∈ B 2 reads
1 2 ρ 2 , Q [ B T ( x , y ; u , v ) adj ( T ( u , v ) ) T b ( u , v ) ] = b ( x , y ) ( m o d T ( x , y ) T B 2 + k ( x , y ) B 2 ) . \frac{1}{2}\rho_{2,Q}\left[B_T(x,y;u,v)\operatorname{adj}(T(u,v))^{\mathsf T}b(u,v)\right]
=b(x,y)\pmod{T(x,y)^{\mathsf T}B^2+k(x,y)B^2}. 2 1 ρ 2 , Q [ B T ( x , y ; u , v ) adj ( T ( u , v ) ) T b ( u , v ) ] = b ( x , y ) ( mod T ( x , y ) T B 2 + k ( x , y ) B 2 ) .
Thus B T / 2 B_T/2 B T /2 is the tensor appropriate to the functional ρ 2 \rho_2 ρ 2 used in (4.26). This explains the factor 1 / 2 1/2 1/2 in (4.29).
It remains to pass from the plane to the surface. Write f ( Q ) = f 0 ( u , v ) + w f 1 ( u , v ) f(Q)=f_0(u,v)+wf_1(u,v) f ( Q ) = f 0 ( u , v ) + w f 1 ( u , v ) with f 0 , f 1 ∈ B 2 f_0,f_1\in B^2 f 0 , f 1 ∈ B 2 . Reduction by w 2 = u 2 + v 2 w^2=u^2+v^2 w 2 = u 2 + v 2 gives
( z + w ) ( f 0 + w f 1 ) = z f 0 + ( u 2 + v 2 ) f 1 + w ( f 0 + z f 1 ) . (z+w)(f_0+wf_1)=zf_0+(u^2+v^2)f_1+w(f_0+zf_1). ( z + w ) ( f 0 + w f 1 ) = z f 0 + ( u 2 + v 2 ) f 1 + w ( f 0 + z f 1 ) .
Taking its w w w coefficient therefore returns f 0 + z f 1 f_0+zf_1 f 0 + z f 1 . The other matrices in the reproduction identity have entries in B B B , so they commute with this coefficient extraction. Applying the planar identity separately to f 0 f_0 f 0 and f 1 f_1 f 1 now gives f 0 ( x , y ) + z f 1 ( x , y ) f_0(x,y)+zf_1(x,y) f 0 ( x , y ) + z f 1 ( x , y ) modulo T T A 2 + k A 2 T^{\mathsf T}A^2+kA^2 T T A 2 + k A 2 , which is (4.30). The identity on the other quotient follows from the corresponding transposed planar identity and the same coefficient extraction. This proof uses no division by z z z or by a branch discriminant, so it also applies at the vertex. □ \square □
This polynomial residue tensor reproduces the finite quotient at every parameter value, including at the cone vertex.
An explicit bilateral family on singular curves
The family constructed in this section has explicit geometric and algebraic data: the curve, the two spaces of sections, the four polynomial matrix factors, the added mass and the resulting polynomial blocks. Normality is established for every real parameter, including the value at which the denominator zeros meet the singular point. This collision of zeros is the confluence considered here; the underlying curve remains fixed.
The construction uses different left and right section modules whose products are canonical differentials. Their scalar Gram matrices are shown to be invertible, providing the initial normal pairings. Residue duality on the corresponding quotients supplies the masses, with the quotient pairings remaining nondegenerate when their support reaches the singular point. The resulting matrix mass has global rank 4 r 4r 4 r , and an explicit constant minor shows that this rank is attained on every section space with N ≥ 2 r N \ge2r N ≥ 2 r , including at confluence. These scalar pairings determine upper triangular 2 × 2 2 \times2 2 × 2 matrix pairings. The common triangular structure makes their Gram determinants products of the two diagonal block determinants. These blocks control normality, meaning invertibility of the Gram matrix, while the nonzero off-diagonal entries retain the order dependence of the matrix factors. An explicit complex coefficient is chosen so that the same mass coupling applies to every real parameter and every finite space in the family. The proof also shows that almost every complex coupling has this simultaneous property. Once these nonvanishing determinants are established, the Gram formulas give the normalized polynomials and their regular limits by evaluation at the confluent parameter.
The finite spaces have the same scalar bound N ≥ r N \ge r N ≥ r in both components. Explicit bases and the nonvanishing of all required determinants establish the existence of normalized polynomials at precisely these bounds. The matrix character of the denominators is verified separately from triangularity: their entry ideals prevent either denominator from becoming a scalar matrix under invertible polynomial changes of rows and columns.
The curve and two explicit modules of sections
An acyclic line bundle, with no global sections and no first cohomology, provides the geometric data for the construction. These vanishing conditions give matching dimensions for the initial section spaces and determine the geometric residue kernel. Pairing with the canonical dual then supplies the differentials needed for the residue calculation.
For an integer r ≥ 2 r \ge2 r ≥ 2 , consider the coordinate algebra
A r : = C + t r C [ t ] = C [ t r , t r + 1 , … , t 2 r − 1 ] . A_r := \mathbb{C} + t^r\mathbb{C}[t] = \mathbb{C}[t^r,t^{r+1},\ldots,t^{2r-1}]. A r := C + t r C [ t ] = C [ t r , t r + 1 , … , t 2 r − 1 ] .
Let C r C_r C r be the curve obtained by gluing Spec A r \operatorname{Spec} A_r Spec A r and Spec C [ t − 1 ] \operatorname{Spec}\mathbb{C}[t^{-1}] Spec C [ t − 1 ] along Spec C [ t , t − 1 ] \operatorname{Spec}\mathbb{C}[t,t^{-1}] Spec C [ t , t − 1 ] . Its normalization is P 1 \mathbb{P}^1 P 1 , and the point ∞ \infty ∞ is smooth. For any integer b ≥ r b \ge r b ≥ r , the function t b t^b t b defines a finite map π : C r → P 1 \pi:C_r\to\mathbb{P}^1 π : C r → P 1 , so C r C_r C r is projective. The quotient C [ t ] / A r \mathbb{C}[t]/A_r C [ t ] / A r has basis t , … , t r − 1 t,\ldots,t^{r-1} t , … , t r − 1 . Consequently the arithmetic genus is g = r − 1 g=r-1 g = r − 1 . The canonical module at t = 0 t=0 t = 0 will have r − 1 r-1 r − 1 minimal generators. For r ≥ 3 r \ge3 r ≥ 3 it is therefore not free of rank one: this is the non-Gorenstein singularity used below.
The affine section modules and the coefficients of their canonical differentials are described by the following polynomial spaces:
ℓ 0 : = 1 + t r − 1 , M r : = C ℓ 0 + t r C [ t ] , \ell_0 := 1+t^{r-1}, \qquad M_r := \mathbb{C}\ell_0+t^r\mathbb{C}[t], ℓ 0 := 1 + t r − 1 , M r := C ℓ 0 + t r C [ t ] ,
K r : = span C { 1 , t , … , t r − 2 } + t r C [ t ] , N r : = { k ∈ C [ t ] : [ t r − 1 ] k = − [ t 0 ] k } . K_r := \operatorname{span}_{\mathbb{C}}\{1,t,\ldots,t^{r-2}\}+t^r\mathbb{C}[t], \qquad N_r := \{k\in\mathbb{C}[t]:[t^{r-1}]k=-[t^0]k\}. K r := span C { 1 , t , … , t r − 2 } + t r C [ t ] , N r := { k ∈ C [ t ] : [ t r − 1 ] k = − [ t 0 ] k } .
Here [ t j ] k [t^j]k [ t j ] k denotes the coefficient of t j t^j t j in k k k . In particular,
N r = C ( 1 − t r − 1 ) + span C { t , … , t r − 2 } + t r C [ t ] , N_r=\mathbb{C}(1-t^{r-1})+\operatorname{span}_{\mathbb{C}}\{t,\ldots,t^{r-2}\}+t^r\mathbb{C}[t], N r = C ( 1 − t r − 1 ) + span C { t , … , t r − 2 } + t r C [ t ] ,
where the middle span is zero if r = 2 r=2 r = 2 . Each of these spaces is an A r A_r A r -module. The polynomial factor of a differential in the canonical module belongs to K r K_r K r :
(5.1)
ω A r = K r d t t r . (77) \omega_{A_r}=K_r\frac{dt}{t^r}. \tag*{(77)} ω A r = K r t r d t . (77)
Indeed, the residue condition Res 0 ( a η ) = 0 \operatorname{Res}_0(a\eta)=0 Res 0 ( a η ) = 0 for every a ∈ A r a\in A_r a ∈ A r excludes t − 1 d t t^{-1}dt t − 1 d t and every t j d t t^jdt t j d t with j < − r j<-r j < − r , and allows t − r d t , … , t − 2 d t t^{-r}dt,\ldots,t^{-2}dt t − r d t , … , t − 2 d t and all polynomial differentials. These are exactly the terms in (5.1). More explicitly, if η = ∑ j c j t j d t \eta=\sum_j c_jt^jdt η = ∑ j c j t j d t , the test a = 1 a=1 a = 1 forces c − 1 = 0 c_{-1}=0 c − 1 = 0 , and the tests a = t m a=t^m a = t m , m ≥ r m\ge r m ≥ r , force c − m − 1 = 0 c_{-m-1}=0 c − m − 1 = 0 . No other negative coefficient is constrained. The polynomial factors 1 , t , … , t r − 2 1,t,\ldots,t^{r-2} 1 , t , … , t r − 2 generate K r K_r K r over A r A_r A r ; their classes are independent modulo the maximal ideal times K r K_r K r . Indeed every product by a nonconstant element of A r A_r A r has degree at least r r r . This also verifies the stated number of generators.
Proposition 5.1 (The acyclic line bundle). The module M r M_r M r , glued to O \mathcal{O} O at infinity by its inclusion in C [ t , t − 1 ] \mathbb{C}[t,t^{-1}] C [ t , t − 1 ] , defines a line bundle L r \mathcal{L}_r L r of degree zero on C r C_r C r . The line bundle
F r : = L r ( ( r − 2 ) ∞ ) F_r := \mathcal{L}_r((r-2)\infty) F r := L r (( r − 2 ) ∞ ) has H 0 ( C r , F r ) = H 1 ( C r , F r ) = 0 H^0(C_r,F_r)=H^1(C_r,F_r)=0 H 0 ( C r , F r ) = H 1 ( C r , F r ) = 0 . Its canonical dual F r D : = Hom ( F r , ω C r ) F_r^D:=\operatorname{Hom}(F_r,\omega_{C_r}) F r D := Hom ( F r , ω C r ) has affine module N r d t / t r N_r\,\mathrm{d}t/t^r N r d t / t r , and M r N r ⊂ K r M_rN_r\subset K_r M r N r ⊂ K r .
Proof. Away from t = 0 t=0 t = 0 , the element t r t^r t r is a unit and M r M_r M r is the structure sheaf, whereas at the origin M r M_r M r is generated by ℓ 0 \ell_0 ℓ 0 . The latter description follows in the local ring of the singular curve from the identity
t r 1 + t r − 1 = t r − t 2 r − 1 1 − t 2 r − 2 ∈ ( A r ) 0 . \frac{t^r}{1+t^{r-1}}=\frac{t^r-t^{2r-1}}{1-t^{2r-2}}\in(A_r)_0. 1 + t r − 1 t r = 1 − t 2 r − 2 t r − t 2 r − 1 ∈ ( A r ) 0 .
The denominator has constant term one and belongs to A r A_r A r . Multiplying the numerator by an arbitrary polynomial still leaves all its exponents at least r r r . Thus every element of t r C [ t ] t^r\mathbb{C}[t] t r C [ t ] is a local multiple of ℓ 0 \ell_0 ℓ 0 , and the module is locally free of rank one.
The quotient C [ t ] / M r \mathbb{C}[t]/M_r C [ t ] / M r has dimension r − 1 r-1 r − 1 , the same as C [ t ] / A r \mathbb{C}[t]/A_r C [ t ] / A r . The normalization sequences for these two modules therefore give χ ( L r ) = χ ( O C r ) \chi(\mathcal{L}_r)=\chi(\mathcal{O}_{C_r}) χ ( L r ) = χ ( O C r ) , hence deg L r = 0 \deg\mathcal{L}_r=0 deg L r = 0 . A global section of F r F_r F r is an element of M r M_r M r whose degree is at most r − 2 r-2 r − 2 . Such an element has the form c ( 1 + t r − 1 ) + t r p ( t ) c(1+t^{r-1})+t^r p(t) c ( 1 + t r − 1 ) + t r p ( t ) . Its degree bound first forces c = 0 c=0 c = 0 and then p = 0 p=0 p = 0 . Hence H 0 ( C r , F r ) = 0 H^0(C_r,F_r)=0 H 0 ( C r , F r ) = 0 . Since deg F r = r − 2 = g − 1 \deg F_r=r-2=g-1 deg F r = r − 2 = g − 1 , Riemann–Roch gives χ ( F r ) = 0 \chi(F_r)=0 χ ( F r ) = 0 , and H 1 ( C r , F r ) = 0 H^1(C_r,F_r)=0 H 1 ( C r , F r ) = 0 follows.
The canonical dual is determined by those k ∈ C ( t ) k\in\mathbb{C}(t) k ∈ C ( t ) satisfying k M r ⊂ K r kM_r\subset K_r k M r ⊂ K r . For such an element, both k ℓ 0 k\ell_0 k ℓ 0 and k t r kt^r k t r are polynomials. Since ℓ 0 \ell_0 ℓ 0 and t r t^r t r are relatively prime, k k k must itself be a polynomial. Its product with t r C [ t ] t^r\mathbb{C}[t] t r C [ t ] already belongs to K r K_r K r , so the condition on the other generator is exactly
[ t r − 1 ] ( k ℓ 0 ) = [ t r − 1 ] k + [ t 0 ] k = 0. [t^{r-1}](k\ell_0)=[t^{r-1}]k+[t^0]k=0. [ t r − 1 ] ( k ℓ 0 ) = [ t r − 1 ] k + [ t 0 ] k = 0.
Thus ( K r : M r ) = N r (K_r:M_r)=N_r ( K r : M r ) = N r , proving the assertion about F r D F_r^D F r D and the inclusion M r N r ⊂ K r M_rN_r\subset K_r M r N r ⊂ K r . □
For N ≥ r N\ge r N ≥ r , the complete spaces of sections
U N : = H 0 ( C r , F r ( N ∞ ) ) , V N : = H 0 ( C r , F r D ( N ∞ ) ) U_N:=H^0(C_r,F_r(N\infty)),\qquad V_N:=H^0(C_r,F_r^D(N\infty)) U N := H 0 ( C r , F r ( N ∞ )) , V N := H 0 ( C r , F r D ( N ∞ ))
have the following ordered bases:
f N = [ 1 + t r − 1 , t r , t r + 1 , … , t N + r − 2 ] , (78) \mathfrak{f}_N=[1+t^{r-1},t^r,t^{r+1},\ldots,t^{N+r-2}], \tag*{(78)} f N = [ 1 + t r − 1 , t r , t r + 1 , … , t N + r − 2 ] , (78)
g N = [ 1 − t r − 1 , t , … , t r − 2 , t r , t r + 1 , … , t N ] . (79) \mathfrak{g}_N=[1-t^{r-1},t,\ldots,t^{r-2},t^r,t^{r+1},\ldots,t^N]. \tag*{(79)} g N = [ 1 − t r − 1 , t , … , t r − 2 , t r , t r + 1 , … , t N ] . (79)
The entries of the second list are multiplied by d t / t r \mathrm{d}t/t^r d t / t r when regarded as sections; products are expressed through their polynomial parts, with this differential factor understood. The degree bound N + r − 2 N+r-2 N + r − 2 in the first list comes from the twist of the original line bundle. The bound in the second list is determined by its global dual,
F r D = L r − 1 ⊗ ω C r ( − ( r − 2 ) ∞ ) . F_r^D=\mathcal{L}_r^{-1}\otimes\omega_{C_r}(-(r-2)\infty). F r D = L r − 1 ⊗ ω C r ( − ( r − 2 ) ∞ ) .
The order at infinity of k d t / t r k\,\mathrm{d}t/t^r k d t / t r is r − 2 − deg k r-2-\deg k r − 2 − deg k . Membership in F r D ( N ∞ ) F_r^D(N\infty) F r D ( N ∞ ) requires that order to be at least r − 2 − N r-2-N r − 2 − N , giving the bound deg k ≤ N \deg k\le N deg k ≤ N . The lists have 1 + ( N − 1 ) = N 1+(N-1)=N 1 + ( N − 1 ) = N and 1 + ( r − 2 ) + ( N − r + 1 ) = N 1+(r-2)+(N-r+1)=N 1 + ( r − 2 ) + ( N − r + 1 ) = N elements, respectively. They are therefore bases of two spaces of the same dimension.
A finite projection relates these section spaces to vector polynomials on the line. For π = t b \pi=t^b π = t b with b ≥ r b\ge r b ≥ r , acyclicity and the splitting of vector bundles on P 1 \mathbb{P}^1 P 1 give
π ∗ F r = O ( − 1 ) b , π ∗ F r D = O ( − 1 ) b . \pi_*F_r=\mathcal{O}(-1)^b,\qquad\pi_*F_r^D=\mathcal{O}(-1)^b. π ∗ F r = O ( − 1 ) b , π ∗ F r D = O ( − 1 ) b .
The pushforwards are torsion-free of rank b b b over the smooth base P 1 \mathbb{P}^1 P 1 , so they are vector bundles. A finite map has no higher direct images, and hence preserves the cohomology of these sheaves. In the splitting into line bundles on P 1 \mathbb{P}^1 P 1 , a summand O ( a ) \mathcal{O}(a) O ( a ) has both cohomology groups zero exactly when a = − 1 a=-1 a = − 1 . Serre duality gives the same cohomology vanishing for F r D F_r^D F r D . Thus the bounds N = b k N=bk N = bk correspond to polynomial components of degree at most k − 1 k-1 k − 1 in both projected bundles. The modules on the curve remain M r M_r M r and N r d t / t r N_r\mathrm{d}t/t^r N r d t / t r ; projection does not replace the latter by a free A r A_r A r -module.
The same module relations determine an explicit geometric kernel:
S r ( t , s ) = [ t r − 1 s r − 1 ( t − s ) − 1 s r ] d s = [ s t r − 1 t − s − 1 ] d s s r . (80) S_r(t,s)=\left[\frac{t^{r-1}}{s^{r-1}(t-s)}-\frac{1}{s^r}\right]\mathrm{d}s=\left[\frac{s t^{r-1}}{t-s}-1\right]\frac{\mathrm{d}s}{s^r}. \tag*{(80)} S r ( t , s ) = [ s r − 1 ( t − s ) t r − 1 − s r 1 ] d s = [ t − s s t r − 1 − 1 ] s r d s . (80)
Its normalization and compatibility with the two modules follow from its local expansions. At a fixed smooth point t t t , the factor t r − 1 / s r − 1 t^{r-1}/s^{r-1} t r − 1 / s r − 1 is one at s = t s=t s = t , so the polar term is d s / ( t − s ) \mathrm{d}s/(t-s) d s / ( t − s ) , whose residue in the second variable is − 1 -1 − 1 , the convention used for the geometric kernel. In the first variable, the coefficients of t 0 t^0 t 0 and t r − 1 t^{r-1} t r − 1 are both − s − r -s^{-r} − s − r and the intervening coefficients are zero, as required by M r M_r M r . These first-variable coefficients are obtained by expanding 1 / ( t − s ) = − ∑ j ≥ 0 t j s − j − 1 1/(t-s)=-\sum_{j\geq0}t^j s^{-j-1} 1/ ( t − s ) = − ∑ j ≥ 0 t j s − j − 1 at t = 0 t=0 t = 0 . The corresponding expansion in the second variable is s t r − 1 / ( t − s ) = ∑ j ≥ 1 s r − j t j − 1 s t^{r-1}/(t-s)=\sum_{j\geq1}s^{r-j}t^{j-1} s t r − 1 / ( t − s ) = ∑ j ≥ 1 s r − j t j − 1 at s = 0 s=0 s = 0 . After extracting d s / s r \mathrm{d}s/s^r d s / s r , the constant coefficient is − 1 -1 − 1 and its coefficient of s r − 1 s^{r-1} s r − 1 is 1 1 1 , as required by N r N_r N r . The behavior at both points at infinity is expressed in the coordinates w = 1 / t w=1/t w = 1/ t and v = 1 / s v=1/s v = 1/ s , in which
S r = w − ( r − 2 ) v r − 2 [ 1 w − v + w r − 2 ] d v . S_r=w^{-(r-2)}v^{r-2}\left[\frac{1}{w-v}+w^{r-2}\right]\mathrm{d}v. S r = w − ( r − 2 ) v r − 2 [ w − v 1 + w r − 2 ] d v .
The factors w − ( r − 2 ) w^{-(r-2)} w − ( r − 2 ) and v r − 2 d v v^{r-2}\mathrm{d}v v r − 2 d v are the local frames of F r F_r F r and F r D F_r^D F r D at infinity. The remaining expression is regular off the smooth diagonal and has its prescribed simple pole there.
At the singular pair, extension of the kernel is determined by multiplication by the ideal of the diagonal. On the affine product this ideal is generated by t m − s m t^m-s^m t m − s m , r ≤ m ≤ 2 r − 1 r\leq m\leq2r-1 r ≤ m ≤ 2 r − 1 . The elementary identity
( t m − s m ) [ s t r − 1 t − s − 1 ] = ∑ j = 0 m − 1 t m + r − 2 − j s j + 1 − t m + s m (t^m-s^m)\left[\frac{s t^{r-1}}{t-s}-1\right]=\sum_{j=0}^{m-1}t^{m+r-2-j}s^{j+1}-t^m+s^m ( t m − s m ) [ t − s s t r − 1 − 1 ] = j = 0 ∑ m − 1 t m + r − 2 − j s j + 1 − t m + s m
shows that multiplication of (5.4) by such a generator gives a polynomial times d s / s r \mathrm{d}s/s^r d s / s r . Its coefficients of t 0 t^0 t 0 and t r − 1 t^{r-1} t r − 1 are both s m s^m s m , and its intervening t t t -coefficients vanish. Its coefficients of s 0 s^0 s 0 and s r − 1 s^{r-1} s r − 1 are − t m -t^m − t m and t m t^m t m . Consequently it lies in M r ⊗ C N r d s / s r M_r\otimes_{\mathbb{C}}N_r\mathrm{d}s/s^r M r ⊗ C N r d s / s r . The acyclic residue construction therefore identifies (5.4) with the normalized global section of Hom ( I Δ , F r ⊠ F r D ) \operatorname{Hom}(I_\Delta,F_r\boxtimes F_r^D) Hom ( I Δ , F r ⊠ F r D ) .
A scalar pairing with invertible Gram matrices at every stated bound
The scalar pairing used in the diagonal matrix blocks is obtained from the following integral functional on K r K_r K r :
Γ [ p ] : = ∫ 1 / 4 1 / 2 p ( t ) t − r d t . \Gamma[p]:=\int_{1/4}^{1/2}p(t)t^{-r}\,\mathrm{d}t. Γ [ p ] := ∫ 1/4 1/2 p ( t ) t − r d t .
It pairs f ∈ M r f\in M_r f ∈ M r with g d t / t r ∈ F r D g\mathrm{d}t/t^r\in F_r^D g d t / t r ∈ F r D by Γ [ f g ] \Gamma[fg] Γ [ f g ] . The different bases in (5.2)–(5.3) need not produce a symmetric Gram matrix. Its invertibility follows instead from the signs of the two evaluation determinants and their integral relation to the Gram determinant.
Lemma 5.2 (Invertibility of the scalar Gram matrices). Let N ≥ r N\geq r N ≥ r and let w w w be a continuous positive function on [ 1 / 4 , 1 / 2 ] [1/4,1/2] [ 1/4 , 1/2 ] . Then
det [ ∫ 1 / 4 1 / 2 f i ( t ) g j ( t ) w ( t ) t − r d t ] i , j ∈ { 1 , … , N } > 0 , \det\left[\int_{1/4}^{1/2}f_i(t)g_j(t)w(t)t^{-r}\,\mathrm{d}t\right]_{i,j\in\{1,\ldots,N\}}>0, det [ ∫ 1/4 1/2 f i ( t ) g j ( t ) w ( t ) t − r d t ] i , j ∈ { 1 , … , N } > 0 , where f i , g j f_i,g_j f i , g j are the lists (5.2)–(5.3). For polynomial weights in A r A_r A r with rational coefficients, all these entries are rational numbers.
Proof. Take 1 / 4 < t 1 < ⋯ < t N < 1 / 2 1/4<t_1<\cdots<t_N<1/2 1/4 < t 1 < ⋯ < t N < 1/2 and put V ( t 1 , … , t N ) : = ∏ i < j ( t j − t i ) > 0 V(t_1,\ldots,t_N):=\prod_{i<j}(t_j-t_i)>0 V ( t 1 , … , t N ) := ∏ i < j ( t j − t i ) > 0 . The determinant det [ f i ( t j ) ] \det[f_i(t_j)] det [ f i ( t j )] is the sum of the two alternants whose exponent lists are
( 0 , r , r + 1 , … , N + r − 2 ) , ( r − 1 , r , r + 1 , … , N + r − 2 ) . (0,r,r+1,\ldots,N+r-2),\qquad(r-1,r,r+1,\ldots,N+r-2). ( 0 , r , r + 1 , … , N + r − 2 ) , ( r − 1 , r , r + 1 , … , N + r − 2 ) .
Both lists are strictly increasing. For exponents 0 ≤ a 1 < ⋯ < a N 0\leq a_1<\cdots<a_N 0 ≤ a 1 < ⋯ < a N , the alternant formula gives det [ t j a i ] = V s λ ( t ) \det[t_j^{a_i}]=V s_\lambda(t) det [ t j a i ] = V s λ ( t ) , with λ i = a N − i + 1 − ( N − i ) \lambda_i=a_{N-i+1}-(N-i) λ i = a N − i + 1 − ( N − i ) . Here λ \lambda λ is a decreasing list of nonnegative integers and s λ s_\lambda s λ is its Schur polynomial. Its tableau expansion is a nonempty sum of monomials with nonnegative integer coefficients. Consequently it is strictly positive when every t j > 0 t_j>0 t j > 0 . Applying this formula to each of the two displayed exponent lists proves positivity of the left evaluation determinant.
For the right determinant, the constant term in its first row gives the exponent list 0 , 1 , … , r − 2 , r , … , N 0,1,\ldots,r-2,r,\ldots,N 0 , 1 , … , r − 2 , r , … , N . The corresponding alternant is V e N − r + 1 ( t 1 , … , t N ) V e_{N-r+1}(t_1,\ldots,t_N) V e N − r + 1 ( t 1 , … , t N ) , where e j e_j e j denotes the elementary element symmetric polynomial of degree j j j . This identity follows by expanding V ( t 1 , … , t N , z ) = V ( t ) ∏ j ( z − t j ) V(t_1,\ldots,t_N,z)=V(t)\prod_j(z-t_j) V ( t 1 , … , t N , z ) = V ( t ) ∏ j ( z − t j ) along its last column and comparing coefficients of z r − 1 z^{r-1} z r − 1 . To put the other term, − t − r − 1 -t^{-r-1} − t − r − 1 , in increasing exponent order requires r − 2 r-2 r − 2 row interchanges. Its contribution is therefore ( − 1 ) r − 1 V ∏ j t j (-1)^{r-1}V\prod_j t_j ( − 1 ) r − 1 V ∏ j t j . The resulting decomposition is
det [ g i ( t j ) ] = V [ e N − r + 1 ( t ) + ( − 1 ) r − 1 ∏ j t j ] = V ∏ j t j [ e r − 1 ( t 1 − 1 , … , t N − 1 ) + ( − 1 ) r − 1 ] > 0. \begin{aligned}
\det[g_i(t_j)] = V\left[e_{N-r+1}(t)+(-1)^{r-1}\prod_j t_j\right] \\
&=V\prod_j t_j\left[e_{r-1}(t_1^{-1},\ldots,t_N^{-1})+(-1)^{r-1}\right]>0.
\end{aligned} det [ g i ( t j )] = V [ e N − r + 1 ( t ) + ( − 1 ) r − 1 j ∏ t j ] = V j ∏ t j [ e r − 1 ( t 1 − 1 , … , t N − 1 ) + ( − 1 ) r − 1 ] > 0.
The last inequality follows because each t j − 1 > 2 t_j^{-1}>2 t j − 1 > 2 and e r − 1 ( t 1 − 1 , … , t N − 1 ) > 1 e_{r-1}(t_1^{-1},\ldots,t_N^{-1})>1 e r − 1 ( t 1 − 1 , … , t N − 1 ) > 1 . The signs of the evaluation determinants pass to the Gram determinant through an integral identity. With d σ ( t ) : = w ( t ) t − r d t d\sigma(t):=w(t)t^{-r}\,dt d σ ( t ) := w ( t ) t − r d t , expansion of det [ ∫ f i g j d σ ] \det[\int f_i g_j\,d\sigma] det [ ∫ f i g j d σ ] by permutations expresses each product of N N N integrals as an integral on the product interval. Antisymmetrization of the f f f factors then gives
det G w = 1 N ! ∫ [ 1 / 4 , 1 / 2 ] N det [ f i ( t j ) ] det [ g i ( t j ) ] ∏ j = 1 N d σ ( t j ) . \det G_w=\frac{1}{N!}\int_{[1/4,1/2]^N}\det[f_i(t_j)]\det[g_i(t_j)]\prod_{j=1}^{N}d\sigma(t_j). det G w = N ! 1 ∫ [ 1/4 , 1/2 ] N det [ f i ( t j )] det [ g i ( t j )] j = 1 ∏ N d σ ( t j ) .
A permutation of the t j t_j t j changes the sign of each determinant, so their product is invariant. The diagonals have measure zero, and the N ! N! N ! ordered regions have equal integrals. Canceling the factor 1 / N ! 1/N! 1/ N ! gives
det G w = ∫ 1 / 4 < t 1 < ⋯ < t N < 1 / 2 det [ f i ( t j ) ] det [ g i ( t j ) ] ∏ j = 1 N w ( t j ) t j − r d t 1 ⋯ d t N . \det G_w=\int_{1/4<t_1<\cdots<t_N<1/2}\det[f_i(t_j)]\det[g_i(t_j)]\prod_{j=1}^{N}w(t_j)t_j^{-r}\,dt_1\cdots dt_N. det G w = ∫ 1/4 < t 1 < ⋯ < t N < 1/2 det [ f i ( t j )] det [ g i ( t j )] j = 1 ∏ N w ( t j ) t j − r d t 1 ⋯ d t N .
Every factor is positive in the interior of this simplex, so the integral is strictly positive.
The rationality of the moments follows from the missing coefficient in f i g j ∈ K r f_i g_j\in K_r f i g j ∈ K r : its coefficient of t r − 1 t^{r-1} t r − 1 is zero, and multiplication by a polynomial in A r A_r A r preserves this property. Division by t r t^r t r therefore produces no t − 1 t^{-1} t − 1 term, so every monomial integral has the form
∫ 1 / 4 1 / 2 t m − r d t = ( 1 / 2 ) m − r + 1 − ( 1 / 4 ) m − r + 1 m − r + 1 , m ≠ r − 1 , \int_{1/4}^{1/2}t^{m-r}\,dt=\frac{(1/2)^{m-r+1}-(1/4)^{m-r+1}}{m-r+1},\qquad m\ne r-1, ∫ 1/4 1/2 t m − r d t = m − r + 1 ( 1/2 ) m − r + 1 − ( 1/4 ) m − r + 1 , m = r − 1 ,
which is rational. This proves the assertion about moments. □
Two residue functionals and their perfect quotients
For the residue masses, the coordinates x : = t r x:=t^r x := t r , y : = t r + 1 y:=t^{r+1} y := t r + 1 give the denominator family
h u : = x 2 + u = t 2 r + u . h_u:=x^2+u=t^{2r}+u. h u := x 2 + u = t 2 r + u .
For a polynomial p ∈ K r p\in K_r p ∈ K r , define
ρ 1 , u [ p ] : = ∑ j ≥ 0 ( − u ) j [ t 3 r − 1 + 2 r j ] p , (81) \rho_{1,u}[p]:=\sum_{j\geq0}(-u)^j[t^{3r-1+2rj}]p, \tag*{(81)} ρ 1 , u [ p ] := j ≥ 0 ∑ ( − u ) j [ t 3 r − 1 + 2 r j ] p , (81)
ρ 2 , u [ p ] : = ∑ j ≥ 0 ( j + 1 ) ( − u ) j [ t 5 r − 1 + 2 r j ] p . (82) \rho_{2,u}[p]:=\sum_{j\geq0}(j+1)(-u)^j[t^{5r-1+2rj}]p. \tag*{(82)} ρ 2 , u [ p ] := j ≥ 0 ∑ ( j + 1 ) ( − u ) j [ t 5 r − 1 + 2 r j ] p . (82)
Only finitely many terms occur for each test. These are the sums of the finite residues of p d t / ( t r h u ) p\,dt/(t^r h_u) p d t / ( t r h u ) and p d t / ( t r h u 2 ) p\,dt/(t^r h_u^2) p d t / ( t r h u 2 ) , respectively. The coefficient formulas also define them at u = 0 u=0 u = 0 , when all the zeros of h u h_u h u meet the singular point. Their behavior under multiplication by the denominator is expressed by
ρ 2 , u [ h u p ] = ρ 1 , u [ p ] , ρ 1 , u [ h u p ] = 0 , ρ 2 , u [ h u 2 p ] = 0. (83) \rho_{2,u}[h_u p]=\rho_{1,u}[p],\qquad\rho_{1,u}[h_u p]=0,\qquad\rho_{2,u}[h_u^2p]=0. \tag*{(83)} ρ 2 , u [ h u p ] = ρ 1 , u [ p ] , ρ 1 , u [ h u p ] = 0 , ρ 2 , u [ h u 2 p ] = 0. (83)
Lemma 5.3 (Perfect residue pairings). For every u ∈ C u\in\mathbb{C} u ∈ C , the two multiplication pairings induced by ρ 1 , u \rho_{1,u} ρ 1 , u and ρ 2 , u \rho_{2,u} ρ 2 , u are perfect, meaning nondegenerate in both arguments, on
M r h u M r × N r h u N r , M r h u 2 M r × N r h u 2 N r . (84) \frac{M_r}{h_uM_r}\times\frac{N_r}{h_uN_r},\qquad
\frac{M_r}{h_u^2M_r}\times\frac{N_r}{h_u^2N_r}. \tag*{(84)} h u M r M r × h u N r N r , h u 2 M r M r × h u 2 N r N r . (84) The dimensions of each pair of arguments are 2 r 2r 2 r and 4 r 4r 4 r , respectively. On polynomial representatives with rational coefficients, their Gram entries are polynomials in u u u with rational coefficients.
Proof. Set v = t 2 r v=t^{2r} v = t 2 r and B = C [ v ] B=\mathbb{C}[v] B = C [ v ] . The map
Ψ [ p ] : = ∑ j ≥ 0 [ t 3 r − 1 + 2 r j ] p v j , p ∈ K r , \Psi[p]:=\sum_{j\geq0}[t^{3r-1+2rj}]p\,v^j,\qquad p\in K_r, Ψ [ p ] := j ≥ 0 ∑ [ t 3 r − 1 + 2 r j ] p v j , p ∈ K r ,
is B B B -linear. In fact, Ψ [ v p ] = v Ψ [ p ] + [ t r − 1 ] p = v Ψ [ p ] \Psi[vp]=v\Psi[p]+[t^{r-1}]p=v\Psi[p] Ψ [ v p ] = v Ψ [ p ] + [ t r − 1 ] p = v Ψ [ p ] . The missing coefficient is thus exactly the condition needed for linearity over this polynomial ring; it is supplied by the canonical module used in the pairing.
Bases adapted to this polynomial-valued pairing are indexed by the following set of 2 r 2r 2 r exponents:
E r : = { 0 } ∪ { r , … , 2 r − 1 } ∪ { 2 r + 1 , … , 3 r − 1 } . E_r:=\{0\}\cup\{r,\ldots,2r-1\}\cup\{2r+1,\ldots,3r-1\}. E r := { 0 } ∪ { r , … , 2 r − 1 } ∪ { 2 r + 1 , … , 3 r − 1 } .
For e ∈ E r e\in E_r e ∈ E r , let m e = t e m_e=t^e m e = t e , except that m 0 = 1 + t r − 1 m_0=1+t^{r-1} m 0 = 1 + t r − 1 . Let n e = t 3 r − 1 − e n_e=t^{3r-1-e} n e = t 3 r − 1 − e , except that n 3 r − 1 = 1 − t r − 1 n_{3r-1}=1-t^{r-1} n 3 r − 1 = 1 − t r − 1 . These are B B B -bases of M r M_r M r and N r N_r N r . Their spanning property includes the polynomial t 2 r t^{2r} t 2 r absent from the monomial lists, since
v ( 1 + t r − 1 ) = t 2 r + t 3 r − 1 , v ( 1 − t r − 1 ) = t 2 r − t 3 r − 1 . v(1+t^{r-1})=t^{2r}+t^{3r-1},\qquad v(1-t^{r-1})=t^{2r}-t^{3r-1}. v ( 1 + t r − 1 ) = t 2 r + t 3 r − 1 , v ( 1 − t r − 1 ) = t 2 r − t 3 r − 1 .
The other monomials are generated by powers of v v v . For independence, the constant residue class modulo 2 r 2r 2 r first forces the coefficient of the modified generator to vanish; the remaining exponents occupy distinct residue classes.
These two bases are dual for Ψ \Psi Ψ :
Ψ [ m e n e ′ ] = δ e , e ′ . (85) \Psi[m_en_{e'}]=\delta_{e,e'}. \tag*{(85)} Ψ [ m e n e ′ ] = δ e , e ′ . (85)
The set E r E_r E r has exactly one representative of each residue class modulo 2 r 2r 2 r : its three parts give the classes 0 , r , … , 2 r − 1 0,r,\ldots,2r-1 0 , r , … , 2 r − 1 , and 1 , … , r − 1 1,\ldots,r-1 1 , … , r − 1 , respectively. The map Ψ \Psi Ψ retains the exponents 3 r − 1 + 2 r j 3r-1+2rj 3 r − 1 + 2 r j with j ≥ 0 j\geq0 j ≥ 0 . The products can therefore be checked in four cases.
If e ≠ 0 e\ne0 e = 0 and e ′ ≠ 3 r − 1 e'\ne3r-1 e ′ = 3 r − 1 , both factors are monomials and
m e n e ′ = t 3 r − 1 + e − e ′ . m_en_{e'}=t^{3r-1+e-e'}. m e n e ′ = t 3 r − 1 + e − e ′ .
A retained exponent requires e − e ′ = 2 r j e-e'=2rj e − e ′ = 2 r j . Distinct elements of E r E_r E r have distinct residue classes, so this forces e = e ′ e=e' e = e ′ and j = 0 j=0 j = 0 . Its contribution to Ψ \Psi Ψ is 1 1 1 .
If e = 0 e=0 e = 0 and e ′ ≠ 3 r − 1 e'\ne3r-1 e ′ = 3 r − 1 , then
m 0 n e ′ = t 3 r − 1 − e ′ + t 4 r − 2 − e ′ . m_0n_{e'}=t^{3r-1-e'}+t^{4r-2-e'}. m 0 n e ′ = t 3 r − 1 − e ′ + t 4 r − 2 − e ′ .
The first exponent is retained precisely when e ′ = 0 e'=0 e ′ = 0 ; the equation − e ′ = 2 r j -e'=2rj − e ′ = 2 r j has no other solution with e ′ ≥ 0 e'\geq0 e ′ ≥ 0 and j ≥ 0 j\geq0 j ≥ 0 . For the second exponent, the required equation is e ′ = r − 1 − 2 r j e'=r-1-2rj e ′ = r − 1 − 2 r j . When j = 0 j=0 j = 0 , the value r − 1 r-1 r − 1 is absent from E r E_r E r , and for j ≥ 1 j\geq1 j ≥ 1 the value is negative. Thus this product contributes δ 0 , e ′ \delta_{0,e'} δ 0 , e ′ .
If e ≠ 0 e\ne0 e = 0 and e ′ = 3 r − 1 e'=3r-1 e ′ = 3 r − 1 , then
m e n 3 r − 1 = t e − t e + r − 1 . m_en_{3r-1}=t^e-t^{e+r-1}. m e n 3 r − 1 = t e − t e + r − 1 .
Since e ≤ 3 r − 1 e\leq3r-1 e ≤ 3 r − 1 , the first term contributes only at e = 3 r − 1 e=3r-1 e = 3 r − 1 , with value 1 1 1 . The second exponent would require e = 2 r + 2 r j e=2r+2rj e = 2 r + 2 r j . The value 2 r 2r 2 r is absent from E r E_r E r , and all values with j ≥ 1 j\geq1 j ≥ 1 exceed its largest element. This case therefore contributes δ e , 3 r − 1 \delta_{e,3r-1} δ e , 3 r − 1 .
Finally, if e = 0 e=0 e = 0 and e ′ = 3 r − 1 e'=3r-1 e ′ = 3 r − 1 , both generators are modified:
m 0 n 3 r − 1 = ( 1 + t r − 1 ) ( 1 − t r − 1 ) = 1 − t 2 r − 2 . m_0n_{3r-1}=(1+t^{r-1})(1-t^{r-1})=1-t^{2r-2}. m 0 n 3 r − 1 = ( 1 + t r − 1 ) ( 1 − t r − 1 ) = 1 − t 2 r − 2 .
Both exponents are smaller than 3 r − 1 3r-1 3 r − 1 , so the contribution is zero, as required by δ 0 , 3 r − 1 = 0 \delta_{0,3r-1}=0 δ 0 , 3 r − 1 = 0 . These four computations prove (5.9) for every pair of basis elements.
Evaluation and differentiation of this polynomial-valued pairing recover the two residue functionals:
ρ 1 , u [ p ] = Ψ [ p ] ( − u ) , ρ 2 , u [ p ] = d d v Ψ [ p ] ( v ) ∣ v = − u . \rho_{1,u}[p]=\Psi[p](-u),\qquad\rho_{2,u}[p]=\left.\frac{\mathrm{d}}{\mathrm{d}v}\Psi[p](v)\right|_{v=-u}. ρ 1 , u [ p ] = Ψ [ p ] ( − u ) , ρ 2 , u [ p ] = d v d Ψ [ p ] ( v ) v = − u .
In particular, for b , c ∈ B b,c\in B b , c ∈ B , B B B -bilinearity and (5.9) give ρ 1 , u [ b m e n e ′ ] = δ e , e ′ b ( − u ) c ( − u ) \rho_{1,u}[bm_en_{e'}]=\delta_{e,e'}b(-u)c(-u) ρ 1 , u [ b m e n e ′ ] = δ e , e ′ b ( − u ) c ( − u ) . Thus the 2 r × 2 r 2r\times2r 2 r × 2 r Gram on the quotient by h u = v + u h_u=v+u h u = v + u is the identity in these bases. For the quotient by h u 2 h_u^2 h u 2 , write b = b 0 + b 1 ( v + u ) b=b_0+b_1(v+u) b = b 0 + b 1 ( v + u ) and c = c 0 + c 1 ( v + u ) c=c_0+c_1(v+u) c = c 0 + c 1 ( v + u ) . Then ( b c ) ′ ( − u ) = b 0 c 1 + b 1 c 0 (bc)'(-u)=b_0c_1+b_1c_0 ( b c ) ′ ( − u ) = b 0 c 1 + b 1 c 0 . The scalar Gram in the basis 1 , v + u 1,v+u 1 , v + u is
[ 0 1 1 0 ] \begin{bmatrix}
0 & 1\\
1 & 0
\end{bmatrix} [ 0 1 1 0 ]
and has determinant − 1 -1 − 1 . The full 4 r × 4 r 4r \times4r 4 r × 4 r Gram is a direct sum of 2 r 2r 2 r copies of this matrix. This proves both nondegeneracy assertions and both dimensions without requiring that the roots of h u h_u h u be distinct. The same evaluation and derivative formulas show that ( ( v + u ) b ′ ) ∣ v = − u = b ( − u ) \left.((v+u)b')\right|_{v=-u}=b(-u) (( v + u ) b ′ ) ∣ v = − u = b ( − u ) , ( ( v + u ) b ) ∣ v = − u = 0 \left.((v+u)b)\right|_{v=-u}=0 (( v + u ) b ) ∣ v = − u = 0 and ( ( ( v + u ) 2 b ) ′ ) ∣ v = − u = 0 \left.(((v+u)^2b)')\right|_{v=-u}=0 ((( v + u ) 2 b ) ′ ) v = − u = 0 give (5.7). Polynomial dependence on u u u follows directly from (5.5)–(5.6). □
The perfect pairings in (5.8) relate the section module to its canonical dual. Their persistence at u = 0 u=0 u = 0 is therefore compatible with the non-Gorenstein nature of A r A_r A r for r ≥ 3 r\geq3 r ≥ 3 : the second argument retains the distinct canonical module, rather than replacing it by a second copy of A r / ( h u ) A_r/(h_u) A r / ( h u ) . Perfection is established for this dual pair of modules.
A fixed coefficient giving normality for every real parameter
The matrix construction below reduces normality to the nonvanishing of determinants that are polynomial in the parameter and in the mass coefficient. A consequence of algebraic independence supplies one coefficient for which all such determinants remain nonzero. This fixes the coupling for the whole family before any finite bound is chosen.
Lemma 5.4 (A uniform coefficient). Let P ( u , z ) ∈ Q [ u , z ] P(u,z)\in\mathbb{Q}[u,z] P ( u , z ) ∈ Q [ u , z ] and suppose that P ( u , 0 ) ≠ 0 P(u,0)\ne0 P ( u , 0 ) = 0 for every u ∈ R u\in\mathbb{R} u ∈ R . Then
P ( u , κ ) ≠ 0 for every u ∈ R , κ : = e − 1 + i e − 2 . P(u,\kappa)\ne0 \quad\text{for every }u\in\mathbb{R},\qquad\kappa:=e^{-1}+ie^{-\sqrt{2}}. P ( u , κ ) = 0 for every u ∈ R , κ := e − 1 + i e − 2 . The same κ \kappa κ works for any collection of polynomials satisfying these hypotheses.
Proof. Write a = e − 1 a=e^{-1} a = e − 1 and b = e − 2 b=e^{-\sqrt{2}} b = e − 2 . These two real numbers are algebraically independent over Q \mathbb{Q} Q . Indeed, a polynomial relation would be a linear relation over Q \mathbb{Q} Q among the numbers exp ( − j − k 2 ) \exp(-j-k\sqrt{2}) exp ( − j − k 2 ) for distinct pairs of nonnegative integers ( j , k ) (j,k) ( j , k ) . Their exponents are distinct algebraic numbers. The Lindemann–Weierstrass theorem states that such exponentials are linearly independent over the algebraic numbers; see [40 Chapter 1, Section 3] and [22 ] .
Suppose now that P ( u , a + i b ) = 0 P(u,a+ib)=0 P ( u , a + ib ) = 0 for a real u u u . The specialization P ( u , z ) P(u,z) P ( u , z ) is a nonzero polynomial, because P ( u , 0 ) ≠ 0 P(u,0)\ne0 P ( u , 0 ) = 0 . It follows that a + i b a+ib a + ib is algebraic over Q ( u ) \mathbb{Q}(u) Q ( u ) . Its conjugate is a root of the same polynomial, since the coefficients are real. Therefore a a a and b b b both belong to an algebraic extension of Q ( u ) \mathbb{Q}(u) Q ( u ) . Adjoining algebraic elements does not increase transcendence degree, whereas a field generated by the single real number u u u has transcendence degree at most one over Q \mathbb{Q} Q . The field Q ( a , b ) \mathbb{Q}(a,b) Q ( a , b ) would consequently have transcendence degree at most one. Their algebraic independence gives degree two, a contradiction. Since this argument is independent of the degree of P P P and requires no choice of exceptional values of u u u , the single coefficient κ \kappa κ has the stated simultaneous property. □
Almost every complex coupling has the same uniform property. More precisely, there is one set E ⊂ C \mathcal{E}\subset\mathbb{C} E ⊂ C of planar Lebesgue measure zero such that, for every κ ∉ E \kappa\notin\mathcal{E} κ ∈ / E ,
P ( u , κ ) ≠ 0 ( u ∈ R ) P(u,\kappa)\ne0 \qquad(u\in\mathbb{R}) P ( u , κ ) = 0 ( u ∈ R )
simultaneously for every P ∈ Q [ u , z ] P\in\mathbb{Q}[u,z] P ∈ Q [ u , z ] satisfying P ( u , 0 ) ≠ 0 P(u,0)\ne0 P ( u , 0 ) = 0 on the real line. To prove this assertion, define
E : = ⋃ 0 ≠ Q ∈ Q [ X , Y ] { a + i b : a , b ∈ R , Q ( a , b ) = 0 } . \mathcal{E}:=\bigcup_{0\ne Q\in\mathbb{Q}[X,Y]}\{a+ib:a,b\in\mathbb{R},\ Q(a,b)=0\}. E := 0 = Q ∈ Q [ X , Y ] ⋃ { a + ib : a , b ∈ R , Q ( a , b ) = 0 } .
Suppose that P ( u 0 , κ ) = 0 P(u_0,\kappa)=0 P ( u 0 , κ ) = 0 for some real u 0 u_0 u 0 , and write κ = a + i b \kappa=a+ib κ = a + ib . The condition P ( u 0 , 0 ) ≠ 0 P(u_0,0)\ne0 P ( u 0 , 0 ) = 0 makes P ( u 0 , z ) P(u_0,z) P ( u 0 , z ) a nonzero polynomial. Hence κ \kappa κ is algebraic over Q ( u 0 ) \mathbb{Q}(u_0) Q ( u 0 ) , and so is its conjugate, since this polynomial has real coefficients. The equalities
a = κ + κ ‾ 2 , b = κ − κ ‾ 2 i a=\frac{\kappa+\overline{\kappa}}{2},\qquad b=\frac{\kappa-\overline{\kappa}}{2i} a = 2 κ + κ , b = 2 i κ − κ
show that both numbers are algebraic over Q ( u 0 , i ) \mathbb{Q}(u_0,i) Q ( u 0 , i ) . Consequently trdeg Q Q ( a , b ) ≤ 1 \operatorname{trdeg}_{\mathbb{Q}}\mathbb{Q}(a,b)\leq1 trdeg Q Q ( a , b ) ≤ 1 . They satisfy a nonzero polynomial relation over Q \mathbb{Q} Q , so κ ∈ E \kappa\in\mathcal{E} κ ∈ E . This proves the assertion for every coupling outside the single set E \mathcal{E} E .
There are countably many rational polynomials Q Q Q . Each nonzero one has a zero set of planar measure zero: write Q ( X , Y ) = ∑ j = 0 m q j ( X ) Y j Q(X,Y)=\sum_{j=0}^{m}q_j(X)Y^j Q ( X , Y ) = ∑ j = 0 m q j ( X ) Y j and choose one nonzero coefficient q j q_j q j . Except at its finitely many real roots, a vertical section of the zero set consists of finitely many roots of the nonzero polynomial Q ( a , Y ) Q(a,Y) Q ( a , Y ) . The exceptional vertical lines also have planar measure zero. Fubini’s theorem on each bounded rectangle, followed by a countable union of rectangles, therefore gives measure zero for the whole zero set. The countable union defining E \mathcal{E} E has measure zero as well. This argument controls all real parameters together. The explicit coefficient in Lemma 5.4 belongs to its complement by the algebraic independence already proved.
The four factors and the normal matrix pairings
For ε ∈ R \varepsilon\in\mathbb{R} ε ∈ R , set
u = ε 2 , ℓ ε : = x − i ε , a ε : = x + i ε , h : = x 2 + ε 2 . u=\varepsilon^2,\qquad\ell_{\varepsilon}:=x-i\varepsilon,\qquad a_{\varepsilon}:=x+i\varepsilon,\qquad h:=x^2+\varepsilon^2. u = ε 2 , ℓ ε := x − i ε , a ε := x + i ε , h := x 2 + ε 2 .
The four factors are
L 1 = [ ℓ ε y 0 ℓ ε ] , R 1 = [ a ε y 0 2 a ε ] , L 0 = [ 1 0 0 x ] , R 0 = [ x 0 0 1 ] . L_1=\begin{bmatrix}\ell_{\varepsilon} & y \\ 0 & \ell_{\varepsilon}\end{bmatrix},\qquad
R_1=\begin{bmatrix}a_{\varepsilon} & y \\ 0 & 2a_{\varepsilon}\end{bmatrix},\qquad
L_0=\begin{bmatrix}1 & 0 \\ 0 & x\end{bmatrix},\qquad
R_0=\begin{bmatrix}x & 0 \\ 0 & 1\end{bmatrix}. L 1 = [ ℓ ε 0 y ℓ ε ] , R 1 = [ a ε 0 y 2 a ε ] , L 0 = [ 1 0 0 x ] , R 0 = [ x 0 0 1 ] .
Their determinants are respectively ℓ ε 2 \ell_{\varepsilon}^2 ℓ ε 2 , 2 a ε 2 2a_{\varepsilon}^2 2 a ε 2 , x x x and x x x , all nonzero in the domain A r A_r A r . The order of the two denominator matrices matters:
T : = R 1 L 1 = [ h 2 x y 0 2 h ] , L 1 R 1 = [ h ( 3 x + i ε ) y 0 2 h ] . (86) T:=R_1L_1=\begin{bmatrix}h & 2xy \\ 0 & 2h\end{bmatrix},\qquad
L_1R_1=\begin{bmatrix}h & (3x+i\varepsilon)y \\ 0 & 2h\end{bmatrix}. \tag*{(86)} T := R 1 L 1 = [ h 0 2 x y 2 h ] , L 1 R 1 = [ h 0 ( 3 x + i ε ) y 2 h ] . (86)
adj ( T ) = adj ( L 1 ) adj ( R 1 ) = [ 2 h − 2 x y 0 h ] , det T = 2 h 2 . \operatorname{adj}(T)=\operatorname{adj}(L_1)\operatorname{adj}(R_1)=
\begin{bmatrix}2h & -2xy \\ 0 & h\end{bmatrix},\qquad
\det T=2h^2. adj ( T ) = adj ( L 1 ) adj ( R 1 ) = [ 2 h 0 − 2 x y h ] , det T = 2 h 2 .
The integral and residue parts have distinct supports. The functional Γ \Gamma Γ integrates over a real interval, while for real ε ≠ 0 \varepsilon\ne0 ε = 0 the zeros of h = t 2 r + ε 2 h=t^{2r}+\varepsilon^2 h = t 2 r + ε 2 are complex points of the curve outside that interval. These zeros support the residue masses. The resulting complex matrix pairings are normal on the stated finite spaces, as proved below.
For a matrix W W W with entries in A r A_r A r , the notation Γ W \Gamma W Γ W means the matrix of functionals with entries ( Γ W ) a , b [ p ] = Γ [ W a , b p ] (\Gamma W)_{a,b}[p]=\Gamma[W_{a,b}p] ( Γ W ) a , b [ p ] = Γ [ W a , b p ] . With this convention, the original pairing, the particular solution and the mass correction are specified by
μ ε : = Γ [ h x ( 3 x + i ε ) y 0 2 h ] , μ p a r t : = x Γ I 2 , (87) \mu_{\varepsilon}:=\Gamma\begin{bmatrix}h & x(3x+i\varepsilon)y \\ 0 & 2h\end{bmatrix},\qquad
\mu_{\mathrm{part}}:=x\Gamma I_2, \tag*{(87)} μ ε := Γ [ h 0 x ( 3 x + i ε ) y 2 h ] , μ part := x Γ I 2 , (87)
ν u : = adj ( T ) ρ 2 , u = [ 2 ρ 1 , u − 2 x y ρ 2 , u 0 ρ 1 , u ] , μ ^ ε : = μ p a r t + κ ν u . (88) \nu_u:=\operatorname{adj}(T)\rho_{2,u}=
\begin{bmatrix}2\rho_{1,u} & -2xy\rho_{2,u} \\ 0 & \rho_{1,u}\end{bmatrix},\qquad
\widehat{\mu}_{\varepsilon}:=\mu_{\mathrm{part}}+\kappa\nu_u. \tag*{(88)} ν u := adj ( T ) ρ 2 , u = [ 2 ρ 1 , u 0 − 2 x y ρ 2 , u ρ 1 , u ] , μ ε := μ part + κ ν u . (88)
For example, ( x y ρ 2 , u ) [ p ] = ρ 2 , u [ x y p ] (xy\rho_{2,u})[p]=\rho_{2,u}[xyp] ( x y ρ 2 , u ) [ p ] = ρ 2 , u [ x y p ] . For a row f ∈ M r 1 × 2 \mathbf{f}\in M_r^{1\times2} f ∈ M r 1 × 2 and a column g ∈ N r 2 × 1 \mathbf{g}\in N_r^{2\times1} g ∈ N r 2 × 1 , a matrix of functionals defines the pairing β μ ( f , g ) = ∑ a , b μ a , b [ f a g b ] \beta_{\mu}(\mathbf{f},\mathbf{g})=\sum_{a,b}\mu_{a,b}[f_ag_b] β μ ( f , g ) = ∑ a , b μ a , b [ f a g b ] . The differential in the second argument is understood. With this convention,
β L μ R ( f , g ) = β μ ( f L , R g ) . \beta_{L\mu R}(\mathbf{f},\mathbf{g})=\beta_{\mu}(\mathbf{f}L,R\mathbf{g}). β Lμ R ( f , g ) = β μ ( f L , R g ) .
Theorem 5.5 (A normal bilateral family). For every integer r ≥ 2 r\ge2 r ≥ 2 and every ε ∈ R \varepsilon\in\mathbb{R} ε ∈ R , the functionals (5.11)–(5.12) satisfy
L 1 μ ^ ε R 1 = L 0 μ ε R 0 . (89) L_1\widehat{\mu}_{\varepsilon}R_1=L_0\mu_{\varepsilon}R_0. \tag*{(89)} L 1 μ ε R 1 = L 0 μ ε R 0 . (89) Both pairings are normal on U N 1 × 2 × V N 2 × 1 U_N^{1\times2}\times V_N^{2\times1} U N 1 × 2 × V N 2 × 1 for every N ≥ r N\ge r N ≥ r , using the same scalar bound in each component. The homogeneous mass ν u \nu_u ν u has global rank 4 r 4r 4 r . Its restriction to these finite spaces has rank 4 r 4r 4 r if and only if N ≥ 2 r N\ge2r N ≥ 2 r , for every value of the parameter, including ε = 0 \varepsilon=0 ε = 0 .
Proof. The functional equation follows from the particular solution and the homogeneous mass separately. The polynomial densities in (5.11) give
L 1 ( x I 2 ) R 1 = x [ h ( 3 x + i ε ) y 0 2 h ] = L 0 [ h x ( 3 x + i ε ) y 0 2 h ] R 0 . L_1(xI_2)R_1=x\begin{bmatrix}h & (3x+i\varepsilon)y \\ 0 & 2h\end{bmatrix}
=L_0\begin{bmatrix}h & x(3x+i\varepsilon)y \\ 0 & 2h\end{bmatrix}R_0. L 1 ( x I 2 ) R 1 = x [ h 0 ( 3 x + i ε ) y 2 h ] = L 0 [ h 0 x ( 3 x + i ε ) y 2 h ] R 0 .
Thus μ p a r t \mu_{\mathrm{part}} μ part is a particular solution. For the mass, preserving the adjugate order gives
L 1 ν u R 1 = L 1 adj ( L 1 ) adj ( R 1 ) R 1 ρ 2 , u = 2 h 2 ρ 2 , u I 2 = 0 \begin{aligned}
L_1\nu_uR_1=L_1\operatorname{adj}(L_1)\operatorname{adj}(R_1)R_1\rho_{2,u} \\
&=2h^2\rho_{2,u}I_2=0
\end{aligned} L 1 ν u R 1 = L 1 adj ( L 1 ) adj ( R 1 ) R 1 ρ 2 , u = 2 h 2 ρ 2 , u I 2 = 0
by (5.7). This proves (5.13).
When all N N N basis vectors of the first component precede those of the second, the Gram matrices inherit the triangular form of the matrix pairings. Their blocks are expressed through
G w : = [ Γ [ w f i g j ] ] i , j ∈ { 1 , … , N } , H 1 ( u ) : = [ ρ 1 , u [ f i g j ] ] i , j ∈ { 1 , … , N } , H 12 ( u ) : = [ − 2 ρ 2 , u [ x y f i g j ] ] i , j ∈ { 1 , … , N } . G_w := [\Gamma[wf_i g_j]]_{i,j\in\{1,\ldots,N\}},
\qquad
H_1(u) := [\rho_{1,u}[f_i g_j]]_{i,j\in\{1,\ldots,N\}},
\qquad
H_{12}(u) := [-2\rho_{2,u}[xyf_i g_j]]_{i,j\in\{1,\ldots,N\}}. G w := [ Γ [ w f i g j ] ] i , j ∈ { 1 , … , N } , H 1 ( u ) := [ ρ 1 , u [ f i g j ] ] i , j ∈ { 1 , … , N } , H 12 ( u ) := [ − 2 ρ 2 , u [ x y f i g j ] ] i , j ∈ { 1 , … , N } .
The original Gram has the block form
G o r i g = [ G h G x ( 3 x + i ε ) y 0 2 G h ] , det G o r i g = 2 N ( det G h ) 2 . G_{\mathrm{orig}} =
\begin{bmatrix}
G_h & G_x(3x+i\varepsilon)y \\
0 & 2G_h
\end{bmatrix},
\qquad
\det G_{\mathrm{orig}} = 2^N(\det G_h)^2. G orig = [ G h 0 G x ( 3 x + i ε ) y 2 G h ] , det G orig = 2 N ( det G h ) 2 .
On the integration interval, h ( t ) > 0 h(t)>0 h ( t ) > 0 for every real ε \varepsilon ε . Lemma 5.2 therefore proves that this determinant is nonzero. The determinant depends only on the diagonal blocks, so positivity of the scalar pairing is sufficient even when the matrix density is not positive.
The same ordering gives the target Gram in triangular form:
G ^ N ( u ) = [ G x + 2 κ H 1 ( u ) κ H 12 ( u ) 0 G x + κ H 1 ( u ) ] . (90) \widehat{G}_N(u)=
\begin{bmatrix}
G_x+2\kappa H_1(u) & \kappa H_{12}(u) \\
0 & G_x+\kappa H_1(u)
\end{bmatrix}. \tag*{(90)} G N ( u ) = [ G x + 2 κ H 1 ( u ) 0 κ H 12 ( u ) G x + κ H 1 ( u ) ] . (90)
The entries of G x G_x G x are rational and det G x > 0 \det G_x>0 det G x > 0 , because x ( t ) > 0 x(t)>0 x ( t ) > 0 on the same interval. The entries of H 1 ( u ) H_1(u) H 1 ( u ) are polynomials in u u u with rational coefficients. The two determinant polynomials relevant to Lemma 5.4 are
det ( G x + 2 z H 1 ( u ) ) , det ( G x + z H 1 ( u ) ) . \det(G_x+2zH_1(u)),\qquad\det(G_x+zH_1(u)). det ( G x + 2 z H 1 ( u )) , det ( G x + z H 1 ( u )) .
They belong to Q [ u , z ] \mathbb{Q}[u,z] Q [ u , z ] , because each determinant is a finite signed sum of products of the specified entries. Their constant term in z z z is det G x \det G_x det G x , independent of u u u and nonzero. Both diagonal blocks in (5.14) are therefore invertible for every real u u u , simultaneously for all r r r and N N N . This proves target normality for u = ε 2 u=\varepsilon^2 u = ε 2 .
The global mass rank is determined by the perfect pairing ρ 2 , u \rho_{2,u} ρ 2 , u from Lemma 5.3, applied in each component. Composition in its second argument with adj ( T ) \operatorname{adj}(T) adj ( T ) gives the mass. Perfection therefore characterizes its right radical, for v ∈ N r 2 v\in N_r^2 v ∈ N r 2 , by
adj ( T ) v ∈ h 2 N r 2 if and only if v ∈ T N r 2 . \operatorname{adj}(T)v\in h^2N_r^2
\quad\text{if and only if}\quad
v\in TN_r^2. adj ( T ) v ∈ h 2 N r 2 if and only if v ∈ T N r 2 .
One implication follows from adj ( T ) T = 2 h 2 I 2 \operatorname{adj}(T)T=2h^2I_2 adj ( T ) T = 2 h 2 I 2 . Conversely, if adj ( T ) v = h 2 w \operatorname{adj}(T)v=h^2w adj ( T ) v = h 2 w , multiplication by T T T gives 2 h 2 v = h 2 T w 2h^2v=h^2Tw 2 h 2 v = h 2 T w . The module N r N_r N r is torsion-free and h ≠ 0 h\ne0 h = 0 , so cancellation gives v = T ( w / 2 ) v=T(w/2) v = T ( w /2 ) . Hence
rank ν u = dim C ( N r 2 / T N r 2 ) . (91) \operatorname{rank}\nu_u=\dim_{\mathbb{C}}(N_r^2/TN_r^2). \tag*{(91)} rank ν u = dim C ( N r 2 / T N r 2 ) . (91)
The triangular matrix in (5.10) gives an exact sequence
0 ⟶ N r / h N r ⟶ N r 2 / T N r 2 ⟶ N r / h N r ⟶ 0. 0\longrightarrow N_r/hN_r\longrightarrow N_r^2/TN_r^2\longrightarrow N_r/hN_r\longrightarrow0. 0 ⟶ N r / h N r ⟶ N r 2 / T N r 2 ⟶ N r / h N r ⟶ 0.
The maps send [ a ] [a] [ a ] to [ a 0 ] T \left[\begin{smallmatrix}a&0\end{smallmatrix}\right]^{\mathsf T} [ a 0 ] T and [ a b ] T \left[\begin{smallmatrix}a&b\end{smallmatrix}\right]^{\mathsf T} [ a b ] T to [ b ] [b] [ b ] . If [ a 0 ] T = T [ c d ] T \left[\begin{smallmatrix}a&0\end{smallmatrix}\right]^{\mathsf T}=T\left[\begin{smallmatrix}c&d\end{smallmatrix}\right]^{\mathsf T} [ a 0 ] T = T [ c d ] T , then 2 h d = 0 2hd=0 2 h d = 0 , so d = 0 d=0 d = 0 and a = h c a=hc a = h c ; this proves injectivity. If b b b is a multiple of h h h , subtracting a suitable T [ 0 d ] T T\left[\begin{smallmatrix}0&d\end{smallmatrix}\right]^{\mathsf T} T [ 0 d ] T leaves a representative in the first component, proving exactness in the middle. Both end terms have dimension 2 r 2r 2 r , so (5.15) gives rank ν u = 4 r \operatorname{rank}\nu_u=4r rank ν u = 4 r .
On the finite section spaces, this global rank is attained at N = 2 r N=2r N = 2 r , where the lists (5.2)–(5.3) represent bases of M r / h M r M_r/hM_r M r / h M r and N r / h N r N_r/hN_r N r / h N r . Relative to the dual bases used in Lemma 5.3, their only replacement is the high monomial t 2 r − 1 t^{2r-1} t 2 r − 1 by t 2 r t^{2r} t 2 r . The relations are
t 2 r = − u ( 1 + t r − 1 ) − t 3 r − 1 ( m o d h M r ) , t 2 r = t 3 r − 1 − u ( 1 − t r − 1 ) ( m o d h N r ) . t^{2r}=-u(1+t^{r-1})-t^{3r-1}\pmod{hM_r},
\qquad
t^{2r}=t^{3r-1}-u(1-t^{r-1})\pmod{hN_r}. t 2 r = − u ( 1 + t r − 1 ) − t 3 r − 1 ( mod h M r ) , t 2 r = t 3 r − 1 − u ( 1 − t r − 1 ) ( mod h N r ) .
In each replacement the coefficient of the removed monomial is − 1 -1 − 1 or 1 1 1 ; the other term is a multiple of the retained first basis vector. Therefore each change of basis has determinant + 1 +1 + 1 or − 1 -1 − 1 , independently of u u u . The quotient pairing had identity Gram in the dual B B B -bases, so its Gram in the displayed section bases satisfies det H 1 ( u ) = ± 1 \det H_1(u)=\pm1 det H 1 ( u ) = ± 1 at this bound, and the matrix of the full mass has determinant
det [ 2 H 1 ( u ) H 12 ( u ) 0 H 1 ( u ) ] = 2 2 r ( det H 1 ( u ) ) 2 = 2 2 r . (92) \det
\begin{bmatrix}
2H_1(u) & H_{12}(u) \\
0 & H_1(u)
\end{bmatrix}
=2^{2r}(\det H_1(u))^2=2^{2r}. \tag*{(92)} det [ 2 H 1 ( u ) 0 H 12 ( u ) H 1 ( u ) ] = 2 2 r ( det H 1 ( u ) ) 2 = 2 2 r . (92)
Every larger section space contains this minor. Together with the global upper bound 4 r 4r 4 r , this proves the claimed rank for N ≥ 2 r N\ge2r N ≥ 2 r . If N < 2 r N<2r N < 2 r , the mass Gram has size 2 N < 4 r 2N<4r 2 N < 4 r and cannot attain that rank. ∙იდ
The target Gram G N ( u ) G_N(u) G N ( u ) remains nonsingular simultaneously for every r ≥ 2 r \ge2 r ≥ 2 , N ≥ r N \ge r N ≥ r and real u u u for any coupling κ ∉ E \kappa\notin\mathcal{E} κ ∈ / E . Indeed, each diagonal determinant is one of the rational-coefficient polynomials covered above, with nonzero constant term det G x \det G_x det G x . The set E \mathcal{E} E is independent of these indices and of u u u , so almost every complex coupling works for the entire collection of target Grams. At u = ε 2 u=\varepsilon^2 u = ε 2 this gives target normality for every real parameter of the bilateral family. Such a coupling is nonzero, because 0 ∈ E 0 \in\mathcal{E} 0 ∈ E ; multiplication of the mass by it preserves the rank 4 r 4r 4 r on the stated cuts. The fixed explicit coefficient will continue to be used in the formulas below.
Theorem 4.4 gives a geometric description of this particular mass through the matrix T = R 1 L 1 T=R_1L_1 T = R 1 L 1 in (86) . This 2 × 2 2 \times2 2 × 2 product defines the quotient relevant to the chosen mass, whereas the 4 × 4 4 \times4 4 × 4 vectorized denominator operator R 1 ⊗ L 1 T R_1 \otimes L_1^{\mathsf{T}} R 1 ⊗ L 1 T acts on all matrix tests. For the mass constructed here, the residue formula is
ν u ( f , g ) = 2 ∑ a ∈ { h = 0 } Res a ( f T − 1 g d t t r ) = 2 b T T ( f T , g ) . \nu_u(\mathbf{f},\mathbf{g})=2\sum_{a\in\{h=0\}}\operatorname{Res}_a\left(\mathbf{f}T^{-1}\mathbf{g}\frac{\mathrm{d}t}{t^r}\right)=2b_{T^{\mathsf{T}}}(\mathbf{f}^{\mathsf{T}},\mathbf{g}). ν u ( f , g ) = 2 a ∈ { h = 0 } ∑ Res a ( f T − 1 g t r d t ) = 2 b T T ( f T , g ) .
At a singular support point the residues are interpreted through the canonical module, or equivalently by (81) –(82) . The transpose in the last expression is necessary because the first argument of b T T b_{T^{\mathsf{T}}} b T T is a column. Consequently the two quotients are M r 2 / T T M r 2 M_r^2/T^{\mathsf{T}}M_r^2 M r 2 / T T M r 2 and N r 2 / T N r 2 N_r^2/TN_r^2 N r 2 / T N r 2 , and the tensor of dual basis elements (the coevaluation tensor) of ν u \nu_u ν u is
1 2 C T T ( t , s ) = 1 2 ( T T ( t ) − T T ( s ) ) S r ( t , s ) (93) \frac{1}{2}C_{T^{\mathsf{T}}}(t,s)=\frac{1}{2}\left(T^{\mathsf{T}}(t)-T^{\mathsf{T}}(s)\right)S_r(t,s) \tag*{(93)} 2 1 C T T ( t , s ) = 2 1 ( T T ( t ) − T T ( s ) ) S r ( t , s ) (93)
in their tensor product. Both the transpose and the factor 1 / 2 1/2 1/2 follow from the pairing convention: det T = 2 h 2 \det T=2h^2 det T = 2 h 2 and ν u = 2 b T T \nu_u=2b_{T^{\mathsf{T}}} ν u = 2 b T T . Formula (93) links the global kernel with the finite mass, including the confluence value u = 0 u=0 u = 0 .
The finite kernel correction and the noncommuting factors
The constant full-rank minor obtained above gives a polynomial factorization of the mass Gram that persists through confluence. This factorization reduces the kernel correction to a matrix whose size is the global mass rank, independently of the section bound.
Corollary 5.6 (A uniform correction of size 4 r 4r 4 r ). Let N ≥ 2 r N \ge2r N ≥ 2 r . Relative to the particular pairing μ p a r t = x Γ I 2 \mu_{\mathrm{part}}=x\Gamma I_2 μ part = x Γ I 2 , the reproducing kernel of μ ~ ε \widetilde{\mu}_{\varepsilon} μ ε has a correction of rank exactly 4 r 4r 4 r . It can be computed by inverting a 4 r × 4 r 4r \times4r 4 r × 4 r matrix that is nonsingular for every real ε \varepsilon ε . Its factorization can be chosen polynomially in u = ε 2 u=\varepsilon^2 u = ε 2 .
Proof. Let H N ( u ) H_N(u) H N ( u ) be the 2 N × 2 N 2N \times2N 2 N × 2 N mass Gram and put G p a r t : = diag ( G x , G x ) G_{\mathrm{part}}:=\operatorname{diag}(G_x,G_x) G part := diag ( G x , G x ) . Choose row and column index sets I , J I,J I , J corresponding to the bound N = 2 r N=2r N = 2 r in the two components. By (92) , H ∗ ( u ) : = H N ( u ) [ I , J ] H_*(u):=H_N(u)[I,J] H ∗ ( u ) := H N ( u ) [ I , J ] has constant determinant 2 2 r 2^{2r} 2 2 r . Let E I , E J E_I,E_J E I , E J have as their columns the standard coordinate vectors indexed by I , J I,J I , J , respectively, so that H ∗ = E I T H N E J H_*=E_I^{\mathsf{T}}H_NE_J H ∗ = E I T H N E J . Define
P N ( u ) : = H N ( u ) E J , Q N ( u ) : = H ∗ ( u ) − 1 E I T H N ( u ) . P_N(u):=H_N(u)E_J,\qquad Q_N(u):=H_*(u)^{-1}E_I^{\mathsf{T}}H_N(u). P N ( u ) := H N ( u ) E J , Q N ( u ) := H ∗ ( u ) − 1 E I T H N ( u ) .
Both matrices have polynomial entries, since the denominator of H ∗ − 1 H_*^{-1} H ∗ − 1 is constant. The identity H N = P N Q N H_N=P_NQ_N H N = P N Q N follows by moving the selected rows and columns to the front, where the block decomposition is H N = [ H ∗ A ; B D ] H_N=[H_* \ A; B\ D] H N = [ H ∗ A ; B D ] . The first 4 r 4r 4 r columns are independent, and every remaining column is their linear combination because rank H N = 4 r \operatorname{rank}H_N=4r rank H N = 4 r . Its first 4 r 4r 4 r entries force the coefficients to be H ∗ − 1 A H_*^{-1}A H ∗ − 1 A , so D = B H ∗ − 1 A D=BH_*^{-1}A D = B H ∗ − 1 A . Thus
H N = [ H ∗ B ] H ∗ − 1 [ H ∗ A ] H_N=
\begin{bmatrix}
H_*\\
B
\end{bmatrix}
H_*^{-1}
\begin{bmatrix}
H_* & A
\end{bmatrix} H N = [ H ∗ B ] H ∗ − 1 [ H ∗ A ]
in the permuted coordinates. Returning to the original order gives exactly P N Q N P_NQ_N P N Q N . The equality first holds over C ( u ) \mathbb{C}(u) C ( u ) ; since all entries are polynomials, it is a polynomial identity. Thus the factorization is fixed uniformly through the confluence.
The matrix governing the inverse correction is
Z N ( u ) : = I 4 r + κ Q N ( u ) G p a r t − 1 P N ( u ) . Z_N(u):=I_{4r}+\kappa Q_N(u)G_{\mathrm{part}}^{-1}P_N(u). Z N ( u ) := I 4 r + κ Q N ( u ) G part − 1 P N ( u ) .
The determinant identity
det Z N ( u ) = det G ^ N ( u ) det G p a r t \det Z_N(u)=\frac{\det\widehat{G}_N(u)}{\det G_{\mathrm{part}}} det Z N ( u ) = det G part det G N ( u )
shows that Z N ( u ) Z_N(u) Z N ( u ) is invertible for every real u u u . The relation to the full inverse follows by multiplication. For G = G p a r t G=G_{\mathrm{part}} G = G part , U = κ P N U=\kappa P_N U = κ P N and V = Q N V=Q_N V = Q N , the defining equality Z N = I + V G − 1 U Z_N=I+VG^{-1}U Z N = I + V G − 1 U gives Z N − 1 + Z N − 1 V G − 1 U = I Z_N^{-1}+Z_N^{-1}VG^{-1}U=I Z N − 1 + Z N − 1 V G − 1 U = I , which cancels the additional terms in ( G + U V ) ( G − 1 − G − 1 U Z N − 1 V G − 1 ) (G+UV)(G^{-1}-G^{-1}UZ_N^{-1}VG^{-1}) ( G + U V ) ( G − 1 − G − 1 U Z N − 1 V G − 1 ) . The resulting inverse identity is the Woodbury formula
G ^ N ( u ) − 1 = G p a r t − 1 − κ G p a r t − 1 P N ( u ) Z N ( u ) − 1 Q N ( u ) G p a r t − 1 . (94) \widehat{G}_N(u)^{-1}=G_{\mathrm{part}}^{-1}-\kappa G_{\mathrm{part}}^{-1}P_N(u)Z_N(u)^{-1}Q_N(u)G_{\mathrm{part}}^{-1}. \tag*{(94)} G N ( u ) − 1 = G part − 1 − κ G part − 1 P N ( u ) Z N ( u ) − 1 Q N ( u ) G part − 1 . (94)
This factorization introduces no parameter poles in P N , Q N P_N,Q_N P N , Q N .
To pass from the Gram inverse to the reproducing kernel, let F N ( Q ) F_N(Q) F N ( Q ) be the 2 N × 2 2N\times2 2 N × 2 matrix whose rows are the evaluations of the row-section basis, and let G N ( P ) G_N(P) G N ( P ) be the 2 × 2 N 2\times2N 2 × 2 N matrix whose columns are the evaluations of the column-section basis. In the column–row convention of the manuscript, the reproducing kernel is
K ^ N ( P , Q ) = G N ( P ) G ^ N ( u ) − 1 F N ( Q ) . \widehat{K}_N(P,Q)=G_N(P)\widehat{G}_N(u)^{-1}F_N(Q). K N ( P , Q ) = G N ( P ) G N ( u ) − 1 F N ( Q ) .
Substitution of (5.18) gives
K ^ N ( P , Q ) − K p a r t , N ( P , Q ) = − κ [ G N ( P ) G p a r t − 1 P N ( u ) ] Z N ( u ) − 1 [ Q N ( u ) G p a r t − 1 F N ( Q ) ] . (95) \widehat{K}_N(P,Q)-K_{\mathrm{part},N}(P,Q)=-\kappa\left[G_N(P)G_{\mathrm{part}}^{-1}P_N(u)\right]Z_N(u)^{-1}\left[Q_N(u)G_{\mathrm{part}}^{-1}F_N(Q)\right]. \tag*{(95)} K N ( P , Q ) − K part , N ( P , Q ) = − κ [ G N ( P ) G part − 1 P N ( u ) ] Z N ( u ) − 1 [ Q N ( u ) G part − 1 F N ( Q ) ] . (95)
Proposition 3.11 gives separation rank rank ( κ H N ) = 4 r \operatorname{rank}(\kappa H_N)=4r rank ( κ H N ) = 4 r . This compares the target kernel with the particular kernel on the same section spaces. The original functional μ ε \mu_\varepsilon μ ε has a different Gram and is not the comparison pairing in (5.19). □
The same Gram inverse determines normalized polynomial sections. Their leading classes are fixed by the passage to the next section space, and orthogonality determines the remaining coefficients. The construction requires normality of consecutive section spaces and therefore applies from the initial bound of the family. The additional full-rank condition is needed only for the fixed-size factorization of the correction.
Corollary 5.7 (Normalized polynomial sections through confluence). Fix r ≥ 2 r\geq2 r ≥ 2 and N ≥ r N\geq r N ≥ r . Use the lower spaces U N 1 × 2 U_N^{1\times2} U N 1 × 2 and V N 2 × 1 V_N^{2\times1} V N 2 × 1 , and the upper spaces with N N N replaced by N + 1 N+1 N + 1 , in the bases (5.2)–(5.3). There are unique row and column blocks for μ ^ ε \widehat{\mu}_\varepsilon μ ε orthogonal to the respective lower spaces and with prescribed leading classes
P ^ N ( t ) ≡ t N + r − 1 I 2 ( m o d U N 1 × 2 ) , Q ^ N ( t ) ≡ t N + 1 I 2 ( m o d V N 2 × 1 ) . \widehat{P}_N(t)\equiv t^{N+r-1}I_2\pmod{U_N^{1\times2}},\qquad\widehat{Q}_N(t)\equiv t^{N+1}I_2\pmod{V_N^{2\times1}}. P N ( t ) ≡ t N + r − 1 I 2 ( mod U N 1 × 2 ) , Q N ( t ) ≡ t N + 1 I 2 ( mod V N 2 × 1 ) . The congruences are interpreted row by row and column by column. The rows are polynomial representatives in M r M_r M r , and the columns represent sections in N r d t / t r N_r\,dt/t^r N r d t / t r , with the differential understood. Their mutual 2 × 2 2\times2 2 × 2 pairing matrix is invertible.
Let P N , Q N \mathcal{P}_N,\mathcal{Q}_N P N , Q N be the blocks with these same leading classes for the particular pairing x Γ I 2 x\Gamma I_2 x Γ I 2 . With the basis matrices used in (5.19), the transformed blocks are explicitly
P ^ N = P N − κ ν u ( P N , G N ) G ^ N ( u ) − 1 F N , \widehat{\mathcal{P}}_N=\mathcal{P}_N-\kappa\nu_u(\mathcal{P}_N,G_N)\widehat{G}_N(u)^{-1}F_N, P N = P N − κ ν u ( P N , G N ) G N ( u ) − 1 F N , Q ^ N = Q N − κ G N G ^ N ( u ) − 1 ν u ( F N , Q N ) . (96) \widehat{\mathcal{Q}}_N=\mathcal{Q}_N-\kappa G_N\widehat{G}_N(u)^{-1}\nu_u(F_N,\mathcal{Q}_N). \tag*{(96)} Q N = Q N − κ G N G N ( u ) − 1 ν u ( F N , Q N ) . (96) Pairings of blocks are taken entry by entry. Their coefficients are rational functions of u = ε 2 u=\varepsilon^2 u = ε 2 , with no poles for u ≥ 0 u\geq0 u ≥ 0 . As functions of ε \varepsilon ε , the blocks are holomorphic near every real parameter, and their limit at zero is the normalized block computed directly from μ ^ 0 \widehat{\mu}_0 μ 0 . If N ≥ 2 r N\geq2r N ≥ 2 r , their correction can use the same 4 r × 4 r 4r\times4r 4 r × 4 r matrix Z N ( u ) Z_N(u) Z N ( u ) as the kernel correction.
Proof. The new scalar basis elements at level N + 1 N+1 N + 1 are exactly t N + r − 1 t^{N+r-1} t N + r − 1 on the left and t N + 1 t^{N+1} t N + 1 on the right. These leading monomials determine the particular normalized blocks after subtraction of their
projections onto the lower spaces. For A N ( t ) : = t N + r − 1 I 2 A_N(t) := t^{N+r-1}I_2 A N ( t ) := t N + r − 1 I 2 and B N ( t ) : = t N + 1 I 2 B_N(t) := t^{N+1}I_2 B N ( t ) := t N + 1 I 2 , with β p a r t \beta_{\mathrm{part}} β part denoting the particular pairing, this gives
P N = A N − β p a r t ( A N , G N ) G p a r t − 1 F N , \mathcal{P}_N=A_N-\beta_{\mathrm{part}}(A_N,G_N)G_{\mathrm{part}}^{-1}F_N, P N = A N − β part ( A N , G N ) G part − 1 F N ,
Q N = B N − G N G p a r t − 1 β p a r t ( F N , B N ) . \mathcal{Q}_N=B_N-G_NG_{\mathrm{part}}^{-1}\beta_{\mathrm{part}}(F_N,B_N). Q N = B N − G N G part − 1 β part ( F N , B N ) .
The two moment blocks have sizes 2 × 2 N 2\times2N 2 × 2 N and 2 N × 2 2N\times2 2 N × 2 , respectively, so both resulting blocks are 2 × 2 2\times2 2 × 2 . The leading monomials specify fixed complements to the lower spaces; subtraction of their particular-pairing projections constructs P N , Q N \mathcal{P}_N,\mathcal{Q}_N P N , Q N while preserving these leading classes. The particular lower Gram is invertible by Lemma 5.2 ; Theorem 5.5 gives invertibility of the transformed Grams at both N N N and N + 1 N+1 N + 1 . For the row correction, the target pairing of P N \mathcal{P}_N P N with the lower column basis is κ ν u ( P N , G N ) \kappa\nu_u(\mathcal{P}_N,G_N) κ ν u ( P N , G N ) , since its particular pairing with that basis vanishes. A correction C F N C F_N C F N has target pairing C G ^ N C\widehat{G}_N C G N . The equation for orthogonality is therefore C G ^ N = κ ν u ( P N , G N ) C\widehat{G}_N=\kappa\nu_u(\mathcal{P}_N,G_N) C G N = κ ν u ( P N , G N ) , whose unique solution is the first correction in (96) . The column equation is G ^ N D = κ ν u ( F N , Q N ) \widehat{G}_N D=\kappa\nu_u(F_N,\mathcal{Q}_N) G N D = κ ν u ( F N , Q N ) and gives the second correction. This is the explicit specialization of Proposition 3.12 with m = κ ν u m=\kappa\nu_u m = κ ν u . In the completed bases, these subtractions are invertible unit triangular changes. The upper Gram becomes block diagonal with the lower Gram and the mutual block pairing. Its invertibility proves that the latter is nondegenerate. The fixed bases at N = r N=r N = r provide an initial block of 2 r 2r 2 r rows and columns with invertible Gram G ^ r \widehat{G}_r G r . Appending the successive blocks above gives biorthogonal block families exhausting the two section modules: each later block is orthogonal to the lower spaces containing all earlier blocks. The families are therefore constructed entirely from the section spaces N ≥ r N\ge r N ≥ r .
The bases, complements and leading classes are independent of the parameter. The particular moments are constant, and the mass moments are polynomial in u u u . Writing G ^ N − 1 = adj ( G ^ N ) / det G ^ N \widehat{G}_N^{-1}=\operatorname{adj}(\widehat{G}_N)/\det\widehat{G}_N G N − 1 = adj ( G N ) / det G N makes each resulting coefficient a rational function of u u u , with the fixed number κ \kappa κ in its coefficient field. Its denominator is nonzero for every real u u u by the normality proof, in particular for u ≥ 0 u\ge0 u ≥ 0 . For each real ε 0 \varepsilon_0 ε 0 , the denominator composed with u = ε 2 u=\varepsilon^2 u = ε 2 is nonzero at ε 0 \varepsilon_0 ε 0 and hence in a complex neighborhood of that point. The coefficients are holomorphic there. At zero the formulas can consequently be evaluated by direct substitution. Their leading classes and orthogonality remain the same, so uniqueness identifies the limit with the normalized blocks for μ ^ 0 \widehat{\mu}_0 μ 0 . This also gives the parameter conclusion of Subsubsection 3.4.3 in explicit coordinates. For N ≥ 2 r N\ge2r N ≥ 2 r , (94) expresses the same computation through Z N − 1 Z_N^{-1} Z N − 1 . The analogous blocks for the original pairing are rational in ε \varepsilon ε and regular near the real axis; its density in (87) also contains a term linear in ε \varepsilon ε . The rational dependence on u u u asserted above concerns the particular and transformed blocks.
The Gram calculation establishes normality and confluence, while the factor identities describe the algebraic character of the transformation. If E 1 , 2 E_{1,2} E 1 , 2 denotes the elementary matrix with its only nonzero entry in position ( 1 , 2 ) (1,2) ( 1 , 2 ) , then
[ L 0 , L 1 ] = ( 1 − x ) y E 1 , 2 , [ R 0 , R 1 ] = ( x − 1 ) y E 1 , 2 , [ L 1 , R 1 ] = a ε y E 1 , 2 . (97) [L_0,L_1]=(1-x)yE_{1,2},\qquad[R_0,R_1]=(x-1)yE_{1,2},\qquad[L_1,R_1]=a_\varepsilon yE_{1,2}. \tag*{(97)} [ L 0 , L 1 ] = ( 1 − x ) y E 1 , 2 , [ R 0 , R 1 ] = ( x − 1 ) y E 1 , 2 , [ L 1 , R 1 ] = a ε y E 1 , 2 . (97)
Thus the order matters between the numerator and denominator on each side, and also in the two matrix products displayed in (86) . The actions on separate arguments of a pairing still commute as left and right actions; this is distinct from the matrix commutators in (97) .
The entry ideals give a further obstruction to reducing the denominators to scalar matrices by unimodular left and right multiplications over A r A_r A r . Such multiplications preserve the ideal generated by all matrix entries. For L 1 L_1 L 1 this ideal is ( ℓ ε , y ) (\ell_\varepsilon,y) ( ℓ ε , y ) , whereas the ideal of ℓ ε I 2 \ell_\varepsilon I_2 ℓ ε I 2 is ( ℓ ε ) (\ell_\varepsilon) ( ℓ ε ) . At ε = 0 \varepsilon=0 ε = 0 , their inequality follows from y / x = t ∉ A r y/x=t\notin A_r y / x = t ∈ / A r . At ε ≠ 0 \varepsilon\ne0 ε = 0 , y y y is not divisible by ℓ ε \ell_\varepsilon ℓ ε even in C [ t ] \mathbb{C}[t] C [ t ] , since the roots of ℓ ε \ell_\varepsilon ℓ ε are nonzero and y y y does not vanish there. Comparing determinants shows that any scalar matrix equivalent to L 1 L_1 L 1 would have to be a nonzero constant multiple of ℓ ε I 2 \ell_\varepsilon I_2 ℓ ε I 2 : the only units of A r A_r A r are constants. This excludes every scalar possibility. The same argument applies to R 1 R_1 R 1 with a ε a_\varepsilon a ε . The numerators have determinant x x x , also a nonunit, and are not polynomial changes of basis.
The associated rational changes of arguments make their relation to dual changes of basis explicit:
S L = L 1 − 1 L 0 = [ 1 / ℓ ε − x y / ℓ ε 2 0 x / ℓ ε ] , S R = R 0 R 1 − 1 = [ x / a ε − x y / ( 2 a ε 2 ) 0 1 / ( 2 a ε ) ] . S_L=L_1^{-1}L_0=
\begin{bmatrix}
1/\ell_\varepsilon& -xy/\ell_\varepsilon^2\\
0 & x/\ell_\varepsilon
\end{bmatrix},
\qquad
S_R=R_0R_1^{-1}=
\begin{bmatrix}
x/a_\varepsilon& -xy/(2a_\varepsilon^2)\\
0 & 1/(2a_\varepsilon)
\end{bmatrix}. S L = L 1 − 1 L 0 = [ 1/ ℓ ε 0 − x y / ℓ ε 2 x / ℓ ε ] , S R = R 0 R 1 − 1 = [ x / a ε 0 − x y / ( 2 a ε 2 ) 1/ ( 2 a ε ) ] .
Their determinants satisfy det S L det S R = x 2 / ( 2 h 2 ) ≠ 1 \det S_L\det S_R=x^2/(2h^2)\ne1 det S L det S R = x 2 / ( 2 h 2 ) = 1 , which shows that the two rational transformations do not form inverse changes of dual bases. The common triangular filtration preserved by all four matrices provides the block structure that makes the diagonal Gram calculations sufficient for normality. The entry-ideal obstruction preserves the distinction between these denominators and scalar matrices over A r A_r A r . The resulting normality statement applies to the complete section spaces (78) –(79) , repeated in the two components with the same scalar bound; it does not extend the assertion to arbitrary component degree bounds.
Standard and Laurent specializations
The general finite system acquires more explicit forms when the coordinate algebra and the transformation factors are specified. On the line, partial fractions describe the possible poles, while Taylor coefficients determine whether the inverse polynomial matrix product has a pole at any of these points. Once these finite conditions hold, the coefficients at infinity impose the prescribed degree bounds. Together, the two sets of conditions determine the constraint matrices associated with the four factors.
The Christoffel and Geronimus cases simplify different parts of this construction. Christoffel multiplication introduces neither denominator poles nor homogeneous masses, so only the polynomial-image and normalization equations remain. Geronimus division instead fixes a Gram block on the denominator images, which can be eliminated before the remaining equations are solved. The Laurent specialization retains both features but also requires the adjoint convention and the degree filtration to be specified. The concluding torus example calculates their effect on determinants, polynomial blocks and kernel ranks.
Rational bilateral mixed-type orthogonality in one variable
In one variable, principal parts and Jordan chains describe the divisibility conditions, and Cauchy transforms give the rational moments needed in the same finite system. Let A = C [ z ] A=\mathbb{C}[z] A = C [ z ] and let the nonnegative multiindices m ∈ N q \boldsymbol{m}\in\mathbb{N}^{q} m ∈ N q , n ∈ N p \boldsymbol{n}\in\mathbb{N}^{p} n ∈ N p be arbitrary. The polynomial row and column spaces are X m : = { f ∈ A 1 × q : deg f i < m i } X_{\boldsymbol{m}}:=\{f\in A^{1\times q}:\deg f_i<m_i\} X m := { f ∈ A 1 × q : deg f i < m i } and Y n : = { g ∈ A p × 1 : deg g j < n j } Y_{\boldsymbol{n}}:=\{g\in A^{p\times1}:\deg g_j<n_j\} Y n := { g ∈ A p × 1 : deg g j < n j } , with ∣ m ∣ : = ∑ i m i = r |\boldsymbol{m}|:=\sum_i m_i=r ∣ m ∣ := ∑ i m i = r and ∣ n ∣ : = ∑ j n j = r + k |\boldsymbol{n}|:=\sum_j n_j=r+k ∣ n ∣ := ∑ j n j = r + k . A zero bound makes the corresponding component zero. The four factors are regular matrix polynomials, so their rational inverses exist, while their order remains relevant because factors on the same side need not commute. The construction also includes shared poles and confluent masses.
Let d L = deg det L 1 d_L=\deg\det L_1 d L = deg det L 1 , d R = deg det R 1 d_R=\deg\det R_1 d R = deg det R 1 . For every zero λ \lambda λ of det L 1 \det L_1 det L 1 , of multiplicity a λ a_\lambda a λ , use the row pole functions e i t ( z − λ ) a \dfrac{e_i^t}{(z-\lambda)^a} ( z − λ ) a e i t , i ∈ { 1 , … , q } i\in\{1,\ldots,q\} i ∈ { 1 , … , q } , a ∈ { 1 , … , a λ } a\in\{1,\ldots,a_\lambda\} a ∈ { 1 , … , a λ } . Similarly use e j ( z − ξ ) b \dfrac{e_j}{(z-\xi)^b} ( z − ξ ) b e j at the zeros of det R 1 \det R_1 det R 1 with their determinant multiplicities. The sum of the root multiplicities is d L d_L d L on the left and d R d_R d R on the right, so these families have sizes s L = q d L s_L=qd_L s L = q d L and s R = p d R s_R=pd_R s R = p d R . The determinant multiplicities bound the pole order of every entry of the inverse matrix. The displayed families therefore contain all required pole terms, although individual entries may require fewer of them. Let an original normal pair of polynomial spaces E , H E,H E , H of common dimension N N N containing the polynomial parts of X m S L X_{\boldsymbol{m}}S_L X m S L and S R Y n S_RY_{\boldsymbol{n}} S R Y n be fixed. Adjoining the pole families to its dual bases gives the required ambient bases. Retaining only the pole directions needed by the rational images can reduce the order of the remaining system.
A rational function splits into a polynomial part and pole terms. For example, ( z 2 + 1 ) / ( z − 1 ) = z + 1 + 2 / ( z − 1 ) (z^2+1)/(z-1)=z+1+2/(z-1) ( z 2 + 1 ) / ( z − 1 ) = z + 1 + 2/ ( z − 1 ) : the last term is its principal part at 1 1 1 . In general this is the sum of negative Laurent powers at the chosen point. The symbol [ z a ] ∞ [z^a]_\infty [ z a ] ∞ extracts the coefficient of z a z^a z a from the polynomial part.
The distinction between the principal part and the polynomial part separates the two requirements on an inverse product. Vanishing of all finite principal parts makes it polynomial; vanishing of its forbidden coefficients at infinity then places it in the prescribed polynomial space. Both requirements are linear in the coordinates of the chosen ambient basis. The following theorem identifies their common kernel and applies the general finite construction to it.
Theorem 6.1 (One-variable bilateral rational formulas). Consider on these ambient spaces the linear conditions
every finite principal part of b L 0 − 1 L 1 vanishes , [ z a ] ∞ ( b L 0 − 1 L 1 ) i = 0 ( a ≥ m i ) , every finite principal part of R 1 R 0 − 1 c vanishes , [ z b ] ∞ ( R 1 R 0 − 1 c ) j = 0 ( b ≥ n j ) . \begin{aligned}
&\text{every finite principal part of }bL_0^{-1}L_1\text{ vanishes}, && [z^a]_\infty(bL_0^{-1}L_1)_i=0 \quad(a\ge m_i),\\
&\text{every finite principal part of }R_1R_0^{-1}c\text{ vanishes}, && [z^b]_\infty(R_1R_0^{-1}c)_j=0 \quad(b\ge n_j).
\end{aligned} every finite principal part of b L 0 − 1 L 1 vanishes , every finite principal part of R 1 R 0 − 1 c vanishes , [ z a ] ∞ ( b L 0 − 1 L 1 ) i = 0 ( a ≥ m i ) , [ z b ] ∞ ( R 1 R 0 − 1 c ) j = 0 ( b ≥ n j ) .
A choice of independent rows forms Γ L , Γ R \Gamma_L,\Gamma_R Γ L , Γ R . Their kernels are exactly the two rational images, and their row counts are N + q d L − r N+qd_L-r N + q d L − r and N + p d R − r − k N+pd_R-r-k N + p d R − r − k .
The allowed masses are those of Theorem 2.2 , under either convention. The balanced mass space satisfies
dim { ν balanced : L 1 ν R 1 = 0 } = p d L + q d R . (98) \dim\{\nu\ \textit{balanced}:L_1\nu R_1=0\}=pd_L+qd_R. \tag*{(98)} dim { ν balanced : L 1 ν R 1 = 0 } = p d L + q d R . (98)
The bivariate masses are (6) ; each mass has rank h ≤ d L + d R h\le d_L+d_R h ≤ d L + d R in either case. Using (12) and the complete constraints in (15) , the remaining square system, after eliminating the original dual block, has order
t = N − r + q d L + p d R + h . (99) t=N-r+qd_L+pd_R+h. \tag*{(99)} t = N − r + q d L + p d R + h . (99)
Its determinant tests final normality and its inverse gives the normalized right block (18) , its left counterpart, and the kernel (22) . All finite principal parts in the recovered polynomial vectors cancel, and the prescribed component degree bounds hold. Neither a denominator-only nor a numerator-only intermediate pairing is required to be normal.
Proof. The adjugate expression for L 1 − 1 L_1^{-1} L 1 − 1 shows that each component of f S L fS_L f S L has denominator dividing det L 1 \det L_1 det L 1 . Partial fractions therefore express it as a polynomial plus the stated row pole functions. On the right, S R g = R 0 adj ( R 1 ) g / det R 1 S_Rg=R_0\operatorname{adj}(R_1)g/\det R_1 S R g = R 0 adj ( R 1 ) g / det R 1 . Thus its scalar denominators divide det R 1 \det R_1 det R 1 , and partial fractions give the stated column pole functions, with the same bound on each pole order. Since the original spaces contain the polynomial parts, adjoining these pole functions expresses every required rational product in the enlarged bases.
The image conditions follow from the uniqueness of the inverse rational product: an ambient row b b b has only the possible preimage f = b L 0 − 1 L 1 − 1 f=bL_0^{-1}L_1^{-1} f = b L 0 − 1 L 1 − 1 . A rational function of one variable is polynomial if and only if all its finite principal parts vanish. After this condition is satisfied, the coefficient tests at infinity impose exactly deg f i < m i \deg f_i<m_i deg f i < m i in each component. Thus the common kernel of the left tests is X m S L X_mS_L X m S L . The right inverse product is R 1 R 0 − 1 c R_1R_0^{-1}c R 1 R 0 − 1 c , in that order, and the same argument proves the right kernel assertion. Because all factors are invertible over C ( z ) \mathbb{C}(z) C ( z ) , multiplying by S L S_L S L or S R S_R S R is injective. The two image dimensions are therefore r r r and r + k r+k r + k . Polynomial and proper pole terms have zero intersection, so the ambient dimensions are exactly N + q d L N+qd_L N + q d L and N + p d R N+pd_R N + p d R . The ranks of the full lists of tests are consequently N + q d L − r N+qd_L-r N + q d L − r and N + p d R − r − k N+pd_R-r-k N + p d R − r − k . Selecting independent rows preserves their common kernels and gives the asserted row counts. The lists are finite: each ambient basis vector has finitely many poles of finite order and a polynomial part of finite degree.
The same decomposition supplies explicit extensions of π L , π R \pi_L,\pi_R π L , π R on the whole ambient space: each extension takes the allowed component powers of the polynomial part of the inverse rational product. On a constrained image the inverse product is already an allowed polynomial, so the extension agrees there with the actual inverse. These are the recovery maps required by Theorems 3.3 and 3.5 . In the reduced system of Theorems 3.3 and 3.5 , the column blocks have sizes p d R pd_R p d R , h h h and N + q d L − r N+qd_L-r N + q d L − r , in the order of (15) . Their sum is N − r + q d L + p d R + h N-r+qd_L+pd_R+h N − r + q d L + p d R + h , proving (99) . The right image conditions, remaining pole equations, mass equations and k k k normalization equations give the same total number of rows. The resulting square system is therefore precisely the specialization of those theorems, with their determinant criterion and polynomial formulas.
The mass dimensions follow independently from Smith normal form over C [ z ] \mathbb{C}[z] C [ z ] , which splits each polynomial matrix quotient into scalar quotients by its invariant factors. If an invariant factor is d ( z ) d(z) d ( z ) of degree a a a , division by d d d gives the basis 1 , z , … , z a − 1 1,z,\ldots,z^{a-1} 1 , z , … , z a − 1 of C [ z ] / ( d ) \mathbb{C}[z]/(d) C [ z ] / ( d ) ; a constant factor gives the zero quotient. Hence its dimension is a a a . The product of the invariant factors is a nonzero constant times the determinant, so their degrees add to the determinant degree. Thus the quotient dimensions are d L , d R d_L,d_R d L , d R , and Theorem 2.2 gives both (98) and the mass-rank bound. □
Canonical Jordan chains and polynomial quotients
Divisibility by ( z − λ ) k (z-\lambda)^k ( z − λ ) k requires the first k k k Taylor coefficients to vanish. A matrix division has the same structure after an invertible local change of components, but those component directions can depend on z z z . A Jordan chain here is the finite list of Taylor coefficients of one such direction. The partial multiplicity is the power of z − λ z-\lambda z − λ by which the corresponding component must be divisible. For the polynomial lifts P L , P R P_L,P_R P L , P R from (10) , these directions are obtained at each zero λ \lambda λ from a local Smith normal form, P L = U λ diag ( ( z − λ ) θ i λ ) V λ P_L=U_\lambda\operatorname{diag}((z-\lambda)^{\theta_i^\lambda})V_\lambda P L = U λ diag (( z − λ ) θ i λ ) V λ . For every positive partial multiplicity, the Taylor coefficients of V λ − 1 e i V_{\lambda}^{-1}e_i V λ − 1 e i through order θ λ , i L − 1 \theta_{\lambda,i}^{L}-1 θ λ , i L − 1 give the complete right chain [ v λ , i , ℓ ] ℓ ∈ { 0 , … , θ λ , i L − 1 } [v_{\lambda,i,\ell}]_{\ell\in\{0,\ldots,\theta_{\lambda,i}^{L}-1\}} [ v λ , i , ℓ ] ℓ ∈ { 0 , … , θ λ , i L − 1 } . For P R P_R P R , the rows of the inverse left Smith factor similarly give complete left chains [ w ξ , j , ℓ ] ℓ ∈ { 0 , … , θ ξ , j R − 1 } [w_{\xi,j,\ell}]_{\ell\in\{0,\ldots,\theta_{\xi,j}^{R}-1\}} [ w ξ , j , ℓ ] ℓ ∈ { 0 , … , θ ξ , j R − 1 } . Their equations are
∑ a = 0 ℓ P L ( a ) ( λ ) a ! v λ , i , ℓ − a = 0 , ∑ a = 0 ℓ w ξ , j , ℓ − a P R ( a ) ( ξ ) a ! = 0 , (100) \sum_{a=0}^{\ell}\frac{P_L^{(a)}(\lambda)}{a!}v_{\lambda,i,\ell-a}=0,\qquad\sum_{a=0}^{\ell}w_{\xi,j,\ell-a}\frac{P_R^{(a)}(\xi)}{a!}=0, \tag*{(100)} a = 0 ∑ ℓ a ! P L ( a ) ( λ ) v λ , i , ℓ − a = 0 , a = 0 ∑ ℓ w ξ , j , ℓ − a a ! P R ( a ) ( ξ ) = 0 , (100)
in the respective chain ranges. The chain equations follow by expanding P L ( V λ − 1 e i ) = ( z − λ ) θ λ , i L U λ e i P_L(V_{\lambda}^{-1}e_i)=(z-\lambda)^{\theta_{\lambda,i}^{L}}U_{\lambda}e_i P L ( V λ − 1 e i ) = ( z − λ ) θ λ , i L U λ e i , and its right counterpart, below the indicated power. One chain of full length for each positive Smith exponent records every component of the local quotient. This completeness is needed for division, since a shorter collection of solutions of (6.3) can leave a pole undetected. These classical Smith and maximal root-polynomial constructions [15 ] enter matrix Christoffel quasideterminants in [2 Theorem 2] and rational transformations in [1 ] .
Proposition 6.2. (The Jordan entries of the constraint matrices.) Put p L : = a L b p_L:=a_Lb p L := a L b and p R : = a R c p_R:=a_Rc p R := a R c . Complete finite division tests are
γ λ , i , ℓ ( b ) : = ∑ a = 0 ℓ p L ( a ) ( λ ) a ! v λ , i , ℓ − a , ℓ ∈ { 0 , … , θ λ , i L − 1 } , \gamma_{\lambda,i,\ell}(b):=\sum_{a=0}^{\ell}\frac{p_L^{(a)}(\lambda)}{a!}v_{\lambda,i,\ell-a},\qquad\ell\in\{0,\ldots,\theta_{\lambda,i}^{L}-1\}, γ λ , i , ℓ ( b ) := a = 0 ∑ ℓ a ! p L ( a ) ( λ ) v λ , i , ℓ − a , ℓ ∈ { 0 , … , θ λ , i L − 1 } , δ ξ , j , ℓ ( c ) : = ∑ a = 0 ℓ w ξ , j , ℓ − a p R ( a ) ( ξ ) a ! , ℓ ∈ { 0 , … , θ ξ , j R − 1 } . (101) \delta_{\xi,j,\ell}(c):=\sum_{a=0}^{\ell}w_{\xi,j,\ell-a}\frac{p_R^{(a)}(\xi)}{a!},\qquad\ell\in\{0,\ldots,\theta_{\xi,j}^{R}-1\}. \tag*{(101)} δ ξ , j , ℓ ( c ) := a = 0 ∑ ℓ w ξ , j , ℓ − a a ! p R ( a ) ( ξ ) , ℓ ∈ { 0 , … , θ ξ , j R − 1 } . (101) Their vanishing is equivalent to p L P L − 1 p_LP_L^{-1} p L P L − 1 and P R − 1 p R P_R^{-1}p_R P R − 1 p R being polynomial. The component bounds impose the additional conditions
[ z a ] ∞ ( p L P L − 1 ) i = 0 ( a ≥ m i ) , [ z b ] ∞ ( P R − 1 p R ) j = 0 ( b ≥ n j ) . (102) [z^{a}]_{\infty}(p_LP_L^{-1})_i=0\ (a\ge m_i),\qquad[z^{b}]_{\infty}(P_R^{-1}p_R)_j=0\ (b\ge n_j). \tag*{(102)} [ z a ] ∞ ( p L P L − 1 ) i = 0 ( a ≥ m i ) , [ z b ] ∞ ( P R − 1 p R ) j = 0 ( b ≥ n j ) . (102) Evaluating these tests on the ambient basis and selecting independent rows gives exactly the matrices (3.1).
Proof. Near λ \lambda λ , write the right Smith direction as V λ − 1 e i = ∑ j ≥ 0 v λ , i , j ( z − λ ) j V_{\lambda}^{-1}e_i=\sum_{j\ge0}v_{\lambda,i,j}(z-\lambda)^j V λ − 1 e i = ∑ j ≥ 0 v λ , i , j ( z − λ ) j . Multiplying this Taylor expansion by that of p L p_L p L shows that the coefficient of ( z − λ ) ℓ (z-\lambda)^{\ell} ( z − λ ) ℓ in ( p L V λ − 1 ) i (p_LV_{\lambda}^{-1})_i ( p L V λ − 1 ) i is exactly the first expression in (6.4).
The inverse product in Smith coordinates is
p L P L − 1 = p L V λ − 1 diag ( ( z − λ ) − θ λ , i L ) U λ − 1 . p_LP_L^{-1}=p_LV_{\lambda}^{-1}\operatorname{diag}\left((z-\lambda)^{-\theta_{\lambda,i}^{L}}\right)U_{\lambda}^{-1}. p L P L − 1 = p L V λ − 1 diag ( ( z − λ ) − θ λ , i L ) U λ − 1 .
Both U λ U_{\lambda} U λ and its inverse are regular at the point. Thus the product has no pole there if and only if the i i i th component of p L V λ − 1 p_LV_{\lambda}^{-1} p L V λ − 1 vanishes to order at least θ λ , i L \theta_{\lambda,i}^{L} θ λ , i L for every i i i . These are precisely the listed Taylor-coefficient conditions. Applying them at all roots removes every finite pole and proves polynomiality. On the right write P R = U ξ D ξ V ξ P_R=U_{\xi}D_{\xi}V_{\xi} P R = U ξ D ξ V ξ . Then P R − 1 p R = V ξ − 1 D ξ − 1 U ξ − 1 p R P_R^{-1}p_R=V_{\xi}^{-1}D_{\xi}^{-1}U_{\xi}^{-1}p_R P R − 1 p R = V ξ − 1 D ξ − 1 U ξ − 1 p R . The j j j th row of U ξ − 1 U_{\xi}^{-1} U ξ − 1 has Taylor coefficients w ξ , j , ℓ w_{\xi,j,\ell} w ξ , j , ℓ . Multiplying it by the Taylor series of p R p_R p R gives the second sum in (6.4). Its first θ ξ , j R \theta_{\xi,j}^{R} θ ξ , j R coefficients must vanish before multiplication by ( z − ξ ) − θ ξ , j R (z-\xi)^{-\theta_{\xi,j}^{R}} ( z − ξ ) − θ ξ , j R . The factor V ξ − 1 V_{\xi}^{-1} V ξ − 1 is invertible and regular, so these conditions are also sufficient.
Once the finite poles have been removed, the quotient bounds remain to be imposed. The polynomial quotients are b L 0 − 1 L 1 bL_0^{-1}L_1 b L 0 − 1 L 1 and R 1 R 0 − 1 c R_1R_0^{-1}c R 1 R 0 − 1 c . The coefficient conditions in (6.5) therefore select exactly the prescribed polynomial spaces. Only finitely many such coefficients can occur: each quotient of an ambient basis vector has a finite polynomial part, and there are finitely many basis vectors. Evaluating all tests on that basis and selecting independent rows consequently gives finite complete constraint matrices. □ \square □
For P L = [ z 2 0 − z 1 ] P_L=\begin{bmatrix}z^2&0\\-z&1\end{bmatrix} P L = [ z 2 − z 0 1 ] and a row p L = [ p 1 p 2 ] p_L=[p_1\ p_2] p L = [ p 1 p 2 ] , the chain v 0 = e 1 v_0=e_1 v 0 = e 1 , v 1 = e 2 v_1=e_2 v 1 = e 2 gives p 1 ( 0 ) = 0 p_1(0)=0 p 1 ( 0 ) = 0 and p 1 ′ ( 0 ) + p 2 ( 0 ) = 0 p_1'(0)+p_2(0)=0 p 1 ′ ( 0 ) + p 2 ( 0 ) = 0 . The second condition couples the derivative of the first component to the value of the second because the chain direction itself varies with the variable. The inverse product exhibits this coupling:
P L − 1 = [ z − 2 0 z − 1 1 ] , p L P L − 1 = [ p 1 + z p 2 z 2 p 2 ] . P_L^{-1}=\begin{bmatrix}z^{-2} & 0 \\ z^{-1} & 1\end{bmatrix},\qquad
p_LP_L^{-1}=\begin{bmatrix}\dfrac{p_1+zp_2}{z^2} & p_2\end{bmatrix}. P L − 1 = [ z − 2 z − 1 0 1 ] , p L P L − 1 = [ z 2 p 1 + z p 2 p 2 ] .
The numerator in the first component is divisible by z 2 z^2 z 2 exactly when its constant and linear coefficients vanish, giving the two tests. The row [ − z 1 ] [-z\ 1] [ − z 1 ] has quotient [ 0 1 ] [0\ 1] [ 0 1 ] , whereas [ z 0 ] [z\ 0] [ z 0 ] has quotient [ z − 1 0 ] [z^{-1}\ 0] [ z − 1 0 ] : it passes the point-value test but retains a pole. The degree conditions have an independent role, as is already apparent for a unimodular factor. Its finite chain list can be empty while its quotient violates the prescribed component bounds: for P L = [ 1 z 0 1 ] P_L=\begin{bmatrix}1&z\\0&1\end{bmatrix} P L = [ 1 0 z 1 ] , the row [ 1 0 ] [1\ 0] [ 1 0 ] has quotient [ 1 − z ] [1\ -z] [ 1 − z ] , outside the space of constant rows.
Cauchy data and confluence
With the divisibility conditions determined, the remaining input consists of the rational moments. For either convention, (3.4) gives the polynomial–pole entries β ~ μ ( b ℓ , e j / ( z − ξ ) b ) \widetilde{\beta}_{\mu}(b_\ell,e_j/(z-\xi)^b) β μ ( b ℓ , e j / ( z − ξ ) b ) and β ~ μ ( e i / ( z − λ ) a , a ℓ ) \widetilde{\beta}_{\mu}(e_i/(z-\lambda)^a,a_\ell) β μ ( e i / ( z − λ ) a , a ℓ ) in the original dual bases b ℓ , a ℓ b_\ell,a_\ell b ℓ , a ℓ . For a finite bivariate measure on a compact support avoiding both poles, put G ~ i , j ( λ , ξ ) : = ∬ d μ i , j ( x , y ) ( x − λ ) ( y − ξ ) \widetilde{\mathcal{G}}_{i,j}(\lambda,\xi):=\iint\frac{d\mu_{i,j}(x,y)}{(x-\lambda)(y-\xi)} G i , j ( λ , ξ ) := ∬ ( x − λ ) ( y − ξ ) d μ i , j ( x , y ) . Then
( E 0 ) ( i , λ , a ) , ( j , ξ , b ) = ∂ λ a − 1 ∂ ξ b − 1 G ~ i , j ( λ , ξ ) ( a − 1 ) ! ( b − 1 ) ! . (103) (E_0)_{(i,\lambda,a),(j,\xi,b)}
=\frac{\partial_\lambda^{a-1}\partial_\xi^{b-1}\widetilde{\mathcal{G}}_{i,j}(\lambda,\xi)}{(a-1)!(b-1)!}. \tag*{(103)} ( E 0 ) ( i , λ , a ) , ( j , ξ , b ) = ( a − 1 )! ( b − 1 )! ∂ λ a − 1 ∂ ξ b − 1 G i , j ( λ , ξ ) . (103)
Because the two integrations use independent arguments, that independence persists when λ = ξ \lambda=\xi λ = ξ . For a balanced measure, F i , j ( w ) : = ∫ ( z − w ) − 1 d μ i , j ( z ) F_{i,j}(w):=\int(z-w)^{-1}d\mu_{i,j}(z) F i , j ( w ) := ∫ ( z − w ) − 1 d μ i , j ( z ) gives G ~ i , j ( λ , ξ ) = F i , j ( λ ) − F i , j ( ξ ) λ − ξ \widetilde{\mathcal{G}}_{i,j}(\lambda,\xi)=\frac{F_{i,j}(\lambda)-F_{i,j}(\xi)}{\lambda-\xi} G i , j ( λ , ξ ) = λ − ξ F i , j ( λ ) − F i , j ( ξ ) . The sign follows from
1 ( z − λ ) ( z − ξ ) = 1 λ − ξ ( 1 z − λ − 1 z − ξ ) . \frac{1}{(z-\lambda)(z-\xi)}
=\frac{1}{\lambda-\xi}\left(\frac{1}{z-\lambda}-\frac{1}{z-\xi}\right). ( z − λ ) ( z − ξ ) 1 = λ − ξ 1 ( z − λ 1 − z − ξ 1 ) .
Differentiation under the integral is legitimate because the compact support stays a positive distance from the poles. Since ∂ w k ( z − w ) − 1 = k ! ( z − w ) − k − 1 \partial_w^k(z-w)^{-1}=k!(z-w)^{-k-1} ∂ w k ( z − w ) − 1 = k ! ( z − w ) − k − 1 , when the poles coincide the product is ( z − λ ) − a − b (z-\lambda)^{-a-b} ( z − λ ) − a − b and the entry is F i , j ( a + b − 1 ) ( λ ) / ( a + b − 1 ) ! F_{i,j}^{(a+b-1)}(\lambda)/(a+b-1)! F i , j ( a + b − 1 ) ( λ ) / ( a + b − 1 )! . For independent arguments the same differentiation applies to each denominator separately, giving (6.6) with the same positive sign. When no measure representation is specified, the rational moments (3.4) provide the same blocks as algebraic data.
These Cauchy entries are the ones used in Subsection 3.1, where they determine the three-by-three reduced matrix from the original polynomials. Subsubsection 3.2.6 extends that calculation to matrix factors and a bilateral mass on a curve. In the present one-variable setting, the corresponding image constraints are the Jordan tests of Proposition 6.2, so the moments and constraints together specify the reduced system.
Bilateral masses and their parameters
The distinction between the rank of a mass and the dimension of its parameter space already occurs for q = 2 q=2 q = 2 , p = 3 p=3 p = 3 , L 1 = diag ( z , 1 ) L_1=\operatorname{diag}(z,1) L 1 = diag ( z , 1 ) and R 1 = diag ( z , 1 , 1 ) R_1=\operatorname{diag}(z,1,1) R 1 = diag ( z , 1 , 1 ) . For these factors, the balanced mass ν ( f , g ) = f 1 ′ ( 0 ) g 1 ( 0 ) + f 1 ( 0 ) g 1 ′ ( 0 ) \nu(f,g)=f_1'(0)g_1(0)+f_1(0)g_1'(0) ν ( f , g ) = f 1 ′ ( 0 ) g 1 ( 0 ) + f 1 ( 0 ) g 1 ′ ( 0 ) is the first derivative of the product f 1 g 1 f_1g_1 f 1 g 1 at zero. On the two value–derivative coordinates its Gram is [ 0 1 ; 1 0 ] [0\ 1;\ 1\ 0] [ 0 1 ; 1 0 ] , so its rank is two. If both arguments are multiplied by their denominator factors, their first components contain z z z and the product contains z 2 z^2 z 2 ; its derivative at zero vanishes. This proves L 1 R 1 = 0 L_1R_1=0 L 1 R 1 = 0 . The unilateral actions do not annihilate this mass: with the left denominator alone, ν ( [ z 0 ] , e 1 ) = 1 \nu([z\ 0],e_1)=1 ν ([ z 0 ] , e 1 ) = 1 , while the right denominator alone gives ν ( e 1 T , [ z 0 ] T ) = 1 \nu(e_1^T,[z\ 0]^T)=1 ν ( e 1 T , [ z 0 ] T ) = 1 . Neither unilateral interference holds. The determinant degrees are d L = d R = 1 d_L=d_R=1 d L = d R = 1 , giving balanced mass dimension p d L + q d R = 3 + 2 = 5 p d_L+q d_R=3+2=5 p d L + q d R = 3 + 2 = 5 . The bivariate space is infinite dimensional: the rank-one masses ν j ( f , g ) : = f 1 ( 0 ) [ z j ] g 1 \nu_j(f,g):=f_1(0)[z^j]g_1 ν j ( f , g ) := f 1 ( 0 ) [ z j ] g 1 , j ∈ N j\in\mathbb{N} j ∈ N , are homogeneous and independent, since ν j ( e 1 T , z k e 1 ) = δ j , k \nu_j(e_1^T,z^k e_1)=\delta_{j,k} ν j ( e 1 T , z k e 1 ) = δ j , k . Thus finite rank does not mean finitely many scalar mass parameters.
An explicit order bound for the one-variable reduced system
The pole counts also yield a numerical bound on the reduced order once the original spaces have been chosen by degree. For a nonzero matrix polynomial M M M , let deg M \deg M deg M be the largest degree of an entry, and put
b L : = max { 0 , deg adj L 1 + deg L 0 − d L } , b_L:=\max\{0,\deg\operatorname{adj}L_1+\deg L_0-d_L\}, b L := max { 0 , deg adj L 1 + deg L 0 − d L } ,
b R : = max { 0 , deg R 0 + deg adj R 1 − d R } . b_R:=\max\{0,\deg R_0+\deg\operatorname{adj}R_1-d_R\}. b R := max { 0 , deg R 0 + deg adj R 1 − d R } .
The adjugate formulas show that the polynomial parts of f S L f_{SL} f S L and S R G S_{RG} S R G have degrees below max i m i + b L \max_i m_i+b_L max i m i + b L and max j n j + b R \max_j n_j+b_R max j n j + b R , respectively. For example, on the left the numerator degree is at most max i m i − 1 + \max_i m_i-1+ max i m i − 1 + deg adj L 1 L_1 L 1 + deg L 0 L_0 L 0 ; polynomial division subtracts d L d_L d L from this bound whenever the polynomial part is nonzero.
For c c c components, order the column monomials as e 1 , … , e c , z e 1 , … , z e c , z 2 e 1 , … e_1,\ldots,e_c,ze_1,\ldots,ze_c,z^2e_1,\ldots e 1 , … , e c , z e 1 , … , z e c , z 2 e 1 , … , and let V c , N V_{c,N} V c , N be the span of the first N N N terms. Assume the original pairing normal on every V q , N × V p , N V_{q,N}\times V_{p,N} V q , N × V p , N . A sufficient common dimension is
N = max { q ( max i m i + b L ) , p ( max j n j + b R ) } . N=\max\left\{q\left(\max_i m_i+b_L\right),p\left(\max_j n_j+b_R\right)\right\}. N = max { q ( i max m i + b L ) , p ( j max n j + b R ) } .
Indeed the first q M qM q M row monomials contain all components of degree below M M M , and the same assertion holds on the column side. Thus these original spaces contain both required polynomial parts.
If the components of each multiindex differ by at most one, then q max i m i ≤ r + q − 1 q\max_i m_i\le r+q-1 q max i m i ≤ r + q − 1 and p max j n j ≤ r + k + p − 1 p\max_j n_j\le r+k+p-1 p max j n j ≤ r + k + p − 1 . Subtracting r r r from the displayed choice of N N N and using (6.2) gives
t ≤ max { q b L + q − 1 , p b R + p − 1 + k } + q d L + p d R + h . t\le\max\{qb_L+q-1,pb_R+p-1+k\}+qd_L+pd_R+h. t ≤ max { q b L + q − 1 , p b R + p − 1 + k } + q d L + p d R + h .
For fixed factors and fixed k k k , every term on the right is bounded independently of the level, since h ≤ d L + d R h\le d_L+d_R h ≤ d L + d R . More generally, a fixed bound on the spread of each multiindex gives the same conclusion with the corresponding larger constants. For unrestricted multiindices the exact order remains (6.2). This choice can be compared with the construction in Subsection 3.3, which fixes its ambient spaces directly from the four factors and component bounds. Its one-variable order is (3.29); (3.24) calculates a smaller order by removing dependent polynomial coefficient columns. The different counts reflect the ambient spaces used in the same finite construction.
Christoffel and Geronimus reductions
The roles of the numerator and denominator factors become separate in the Christoffel and Geronimus cases. In the former, the transformed arguments are polynomial images, so the construction reduces to their membership conditions. In the latter, the functional equation fixes the pairing on denominator images, providing a nondegenerate block that can be eliminated. Both reductions retain independent factors and arbitrary finite polynomial row and column spaces.
Bilateral Christoffel multiplication
For identity denominators set L 1 = I q L_1=I_q L 1 = I q , R 1 = I p R_1=I_p R 1 = I p , L : = L 0 L:=L_0 L := L 0 , R : = R 0 R:=R_0 R := R 0 , and assume that det L \det L det L and det R \det R det R are regular. The homogeneous mass vanishes and β ^ ( f , g ) = β μ ( f L , R g ) \widehat{\beta}(f,g)=\beta_\mu(fL,Rg) β ( f , g ) = β μ ( f L , R g ) . Choose finite target spaces X ⊂ A 1 × q X\subset A^{1\times q} X ⊂ A 1 × q and Y ⊂ A p × 1 Y\subset A^{p\times1} Y ⊂ A p × 1 , and an original normal pair E , H E,H E , H of dimension N N N containing X L XL X L and R Y RY R Y . In its dual bases B , A \mathcal{B},\mathcal{A} B , A , choose matrices Γ L , Γ R \Gamma_L,\Gamma_R Γ L , Γ R with independent rows whose kernels are the coefficient spaces of X L XL X L and R Y RY R Y . Thus their equations test whether a polynomial equals f L fL f L or R g Rg R g with f ∈ X f\in X f ∈ X , g ∈ Y g\in Y g ∈ Y . For the dimensions in the proposition, Γ L \Gamma_L Γ L has size ( N − r ) × N (N-r)\times N ( N − r ) × N and Γ R \Gamma_R Γ R has size ( N − r − k ) × N (N-r-k)\times N ( N − r − k ) × N .
Since these polynomial spaces already contain the images, no additional rational vectors are required. Orthogonality restricts the unknown coefficients to the directions described by the left constraints. The resulting system determines the multiplied polynomial R a ^ \widehat{R a} R a ; the image conditions then guarantee that division by R R R gives a polynomial in Y Y Y .
Proposition 6.3 (Bilateral Christoffel formulas). Suppose dim X = r \dim X=r dim X = r , dim Y = r + k \dim Y=r+k dim Y = r + k and λ : Y → C k \lambda:Y\to\mathbb{C}^k λ : Y → C k is the prescribed right normalization. Extend R g ↦ λ ( g ) Rg\mapsto\lambda(g) R g ↦ λ ( g ) from R Y RY R Y to H H H , with matrix η ∈ C k × N \eta\in\mathbb{C}^{k\times N} η ∈ C k × N in the basis A \mathcal{A} A . Put
T C , k : = [ Γ R Γ L T η Γ L T ] , J k = [ 0 ( N − r − k ) × k I k ] . T_{C,k}:=\begin{bmatrix}\Gamma_R\Gamma_L^{\mathsf T}\\ \eta\Gamma_L^{\mathsf T}\end{bmatrix},
\qquad
J_k=\begin{bmatrix}0_{(N-r-k)\times k}\\ I_k\end{bmatrix}. T C , k := [ Γ R Γ L T η Γ L T ] , J k = [ 0 ( N − r − k ) × k I k ] . The matrix has order N − r N-r N − r . For k > 0 k>0 k > 0 the normalized problem has a unique solution if and only if T C , k T_{C,k} T C , k is invertible, and then
a ^ = R − 1 A Γ L T T C , k − 1 J k . \widehat{a}=R^{-1}\mathcal{A}\Gamma_L^{\mathsf T}T_{C,k}^{-1}J_k. a = R − 1 A Γ L T T C , k − 1 J k . For a square pair, let M C : = Γ R Γ L T M_C:=\Gamma_R\Gamma_L^{\mathsf T} M C := Γ R Γ L T . Normality is equivalent to det M C ≠ 0 \det M_C\ne0 det M C = 0 , and
R ( P ) K ^ ( P , Q ) L ( Q ) = A ( P ) [ I N − Γ L T M C − 1 Γ R ] B ( Q ) . (104) R(P)\widehat{K}(P,Q)L(Q)=\mathcal{A}(P)\left[I_N-\Gamma_L^{\mathsf T}M_C^{-1}\Gamma_R\right]\mathcal{B}(Q). \tag*{(104)} R ( P ) K ( P , Q ) L ( Q ) = A ( P ) [ I N − Γ L T M C − 1 Γ R ] B ( Q ) . (104) The correction in (6.7) has tensor rank exactly N − r N-r N − r .
Proof. Original duality identifies the pairing with the ordinary product of coefficient rows and columns. Orthogonality to X L XL X L therefore means that the column z z z annihilates ker Γ L \ker\Gamma_L ker Γ L , and consequently z = Γ L T u z=\Gamma_L^{\mathsf T}u z = Γ L T u , uniquely, because Γ L \Gamma_L Γ L has independent rows. Membership in R Y RY R Y and normalization become respectively Γ R Γ L T u = 0 \Gamma_R\Gamma_L^{\mathsf T}u=0 Γ R Γ L T u = 0 and η Γ L T u = I k \eta\Gamma_L^{\mathsf T}u=I_k η Γ L T u = I k . This is the square system T C , k L u = J k T^L_{C,k}u=J_k T C , k L u = J k , proving the criterion and the formula. Its image condition also ensures that division by R R R gives a polynomial in Y Y Y .
For a square pair, a coefficient column z ∈ ker Γ R z\in\ker\Gamma_R z ∈ ker Γ R is in the right radical precisely when every coefficient row of X L XL X L annihilates it. The latter condition is z ∈ im Γ L T z\in\operatorname{im}\Gamma_L^{\mathsf T} z ∈ im Γ L T . Thus normality is equivalent to ker Γ R ∩ im Γ L T = 0 \ker\Gamma_R\cap\operatorname{im}\Gamma_L^{\mathsf T}=0 ker Γ R ∩ im Γ L T = 0 . This is equivalent to invertibility of M C M_C M C . Then P C : = I N − Γ L T M C − 1 Γ R P_C:=I_N-\Gamma_L^{\mathsf T}M_C^{-1}\Gamma_R P C := I N − Γ L T M C − 1 Γ R is the projection onto ker Γ R \ker\Gamma_R ker Γ R along im Γ L T \operatorname{im}\Gamma_L^{\mathsf T} im Γ L T : it is the identity on the first space and zero on the second. In coordinates, Γ R P C = 0 \Gamma_RP_C=0 Γ R P C = 0 and P C z = z P_Cz=z P C z = z whenever Γ R z = 0 \Gamma_Rz=0 Γ R z = 0 . Also x T P C = x T x^{\mathsf T}P_C=x^{\mathsf T} x T P C = x T for every x ∈ ker Γ L x\in\ker\Gamma_L x ∈ ker Γ L , because x T Γ L T = 0 x^{\mathsf T}\Gamma_L^{\mathsf T}=0 x T Γ L T = 0 . Consequently A ( P ) P C B ( Q ) \mathcal{A}(P)P_C\mathcal{B}(Q) A ( P ) P C B ( Q ) has the required column and row images and the reproducing property, giving (6.7). Its difference from A ( P ) B ( Q ) \mathcal{A}(P)\mathcal{B}(Q) A ( P ) B ( Q ) has coefficient matrix − Γ L T M C − 1 Γ R -\Gamma_L^{\mathsf T}M_C^{-1}\Gamma_R − Γ L T M C − 1 Γ R . The first and last factors have full column and row rank N − r N-r N − r , respectively, and the middle one is invertible. The coefficient matrix therefore has rank N − r N-r N − r . Independence of the ambient basis functions identifies this matrix rank with the tensor rank of the kernel correction. □
If dim X = r + k \dim X=r+k dim X = r + k , dim Y = r \dim Y=r dim Y = r and ζ \zeta ζ extends the left normalization through f ↦ f L f\mapsto fL f ↦ f L , put T C , k L : = [ Γ L Γ R T ; ζ Γ R T ] T^L_{C,k}:=[\Gamma_L\Gamma_R^{\mathsf T};\ \zeta\Gamma_R^{\mathsf T}] T C , k L := [ Γ L Γ R T ; ζ Γ R T ] . Here T C , k L T^L_{C,k} T C , k L has order N − r N-r N − r and the lower identity in J k J_k J k has k k k rows, with the zero block of height N − r − k N-r-k N − r − k . Indeed left orthogonality writes the row coefficients as u T Γ R u^{\mathsf T}\Gamma_R u T Γ R ; left image membership and normalization are Γ L Γ R T u = 0 \Gamma_L\Gamma_R^{\mathsf T}u=0 Γ L Γ R T u = 0 and ζ Γ R T u = I k \zeta\Gamma_R^{\mathsf T}u=I_k ζ Γ R T u = I k . Thus the normalized row block is b ^ = ( ( T C , k L ) − 1 J k ) T Γ R B L − 1 \widehat b=((T^L_{C,k})^{-1}J_k)^{\mathsf T}\Gamma_R\mathcal{B}L^{-1} b = (( T C , k L ) − 1 J k ) T Γ R B L − 1 . The image equations ensure that its displayed division is polynomial. Both constraint matrices include quotient bounds at infinity, which finite-zero conditions alone do not impose.
For the matrix specialization of [2 Theorem 2] , take A = C [ z ] A=\mathbb{C}[z] A = C [ z ] , q = p q=p q = p , L = W L=W L = W monic of degree d d d , and R = I p R=I_p R = I p . Write P j P_j P j for the original monic left matrix polynomials and π j \pi_j π j for their complete Jordan-jet blocks, as in Definition 13 of that reference. Put Π n , d : = [ π n ⋯ π n + d − 1 ] T \Pi_{n,d}:=[\pi_n\ \cdots\ \pi_{n+d-1}]^{\mathsf T} Π n , d := [ π n ⋯ π n + d − 1 ] T and P n , d : = [ P n ⋯ P n + d − 1 ] T \mathcal{P}_{n,d}:=[P_n\ \cdots\ P_{n+d-1}]^{\mathsf T} P n , d := [ P n ⋯ P n + d − 1 ] T . The complete finite conditions have p d pd p d scalar columns, because deg det W = p d \deg\det W=pd deg det W = p d . Thus each π j \pi_j π j has size p × p d p\times pd p × p d , Π n , d \Pi_{n,d} Π n , d has size p d × p d pd\times pd p d × p d , and P n , d \mathcal{P}_{n,d} P n , d has size p d × p pd\times p p d × p . The relation with the matrix formula follows by expanding P ^ n W \widehat P_nW P n W in the monic original basis. Its leading coefficient and degree give the term P n + d P_{n+d} P n + d . Pairing with the original right blocks of degrees below n n n forces all coefficients at P j P_j P j , j < n j<n j < n , to vanish. The remaining coefficient row is C ∈ C p × p d C\in\mathbb{C}^{p\times pd} C ∈ C p × p d , so P ^ n W = P n + d + C P n , d \widehat P_nW=P_{n+d}+C\mathcal{P}_{n,d} P n W = P n + d + C P n , d . Applying the complete division tests to this identity gives π n + d + C Π n , d = 0 \pi_{n+d}+C\Pi_{n,d}=0 π n + d + C Π n , d = 0 . Solving for C C C yields
P ^ n W = P n + d − π n + d Π n , d − 1 P n , d . \widehat P_nW=P_{n+d}-\pi_{n+d}\Pi_{n,d}^{-1}\mathcal{P}_{n,d}. P n W = P n + d − π n + d Π n , d − 1 P n , d .
the Schur complement in formula (10) of that theorem. Conversely, complete jets give divisibility and a monic quotient, so det Π n , d ≠ 0 \det\Pi_{n,d}\ne0 det Π n , d = 0 is precisely the invertibility condition for the remaining finite system. When the leading coefficient is singular, the same distinction between finite divisibility and component degrees persists: the quotient bounds supply conditions not detected by the finite jets.
Remark 6.4 (Normality with vanishing auxiliary adjacent minors). On A = C [ x , y ] / ( x 2 + y 2 − 1 ) A=\mathbb{C}[x,y]/(x^2+y^2-1) A = C [ x , y ] / ( x 2 + y 2 − 1 ) put μ k [ f ] : = 1 2 ∫ 1 / 4 1 / 2 x k [ f ( x , 1 − x 2 ) + f ( x , − 1 − x 2 ) ] d x \mu_k[f]:=\frac{1}{2}\int_{1/4}^{1/2}x^k[f(x,\sqrt{1-x^2})+f(x,-\sqrt{1-x^2})]\,dx μ k [ f ] := 2 1 ∫ 1/4 1/2 x k [ f ( x , 1 − x 2 ) + f ( x , − 1 − x 2 )] d x , k ≥ 0 k\ge0 k ≥ 0 . For every nonzero real class f = a ( x ) + y b ( x ) f=a(x)+yb(x) f = a ( x ) + y b ( x ) , μ k [ f 2 ] = ∫ 1 / 4 1 / 2 x k [ a ( x ) 2 + ( 1 − x 2 ) b ( x ) 2 ] d x > 0 \mu_k[f^2]=\int_{1/4}^{1/2}x^k[a(x)^2+(1-x^2)b(x)^2]\,dx>0 μ k [ f 2 ] = ∫ 1/4 1/2 x k [ a ( x ) 2 + ( 1 − x 2 ) b ( x ) 2 ] d x > 0 . All principal Gram matrices in 1 , x , y , x 2 , x y , … 1,x,y,x^2,xy,\ldots 1 , x , y , x 2 , x y , … are therefore positive definite. For n , k ≥ 1 n,k\ge1 n , k ≥ 1 , the auxiliary minor Θ 2 n ( k ) \Theta_{2n}^{(k)} Θ 2 n ( k ) of [23 Section 2.2, Remark 2.6] has rows x k + j , x k + j − 1 y x^{k+j},x^{k+j-1}y x k + j , x k + j − 1 y , j ∈ { 0 , … , n − 1 } j\in\{0,\ldots,n-1\} j ∈ { 0 , … , n − 1 } , and the first 2 n 2n 2 n columns of that basis: n n n rows of each y y y -parity, but n + 1 n+1 n + 1 even and n − 1 n-1 n − 1 odd columns. The even–odd cross moments vanish because the two branches are averaged with equal weights. After grouping rows and columns by parity, the only nonzero rectangles have sizes n × ( n + 1 ) n\times(n+1) n × ( n + 1 ) and n × ( n − 1 ) n\times(n-1) n × ( n − 1 ) . Their total rank is at most n + ( n − 1 ) = 2 n − 1 n+(n-1)=2n-1 n + ( n − 1 ) = 2 n − 1 , so Θ 2 n ( k ) = 0 \Theta_{2n}^{(k)}=0 Θ 2 n ( k ) = 0 . These minors occur as denominators in [23 (2.12), (2.13)] and the coefficients of (2.14)], while Proposition 6.3 applies to the normal transformation μ k + 1 = x μ k \mu_{k+1}=x\mu_k μ k + 1 = x μ k .
Polynomial elimination and Geronimus division
For factors satisfying (2.3), fix a solution β ^ \widehat\beta β . Its moments on U : = X ∩ A 1 × q L 1 U:=X\cap A^{1\times q}L_1 U := X ∩ A 1 × q L 1 and V : = Y ∩ R 1 A p × 1 V:=Y\cap R_1A^{p\times1} V := Y ∩ R 1 A p × 1 are fixed by β ^ ( f L 1 , R 1 g ) = β μ ( f L 0 , R 0 g ) \widehat\beta(fL_1,R_1g)=\beta_\mu(fL_0,R_0g) β ( f L 1 , R 1 g ) = β μ ( f L 0 , R 0 g ) , independently of preimages. A nonsingular order-d d d minor gives families P i : = f i L 1 P_i := f_iL_1 P i := f i L 1 , Q i : = R 1 g i Q_i := R_1g_i Q i := R 1 g i with β μ ( f i L 0 , R 0 g j ) = δ i , j \beta_\mu(f_iL_0,R_0g_j)=\delta_{i,j} β μ ( f i L 0 , R 0 g j ) = δ i , j after changing bases. Collect them as rows of P P P and columns of Q Q Q , and complete to bases [ P ; Φ ] [P;\Phi] [ P ; Φ ] , [ Q Ψ ] [Q\ \Psi] [ Q Ψ ] of X , Y X,Y X , Y . Put C : = β ^ ( Φ , Q ) C := \hat{\beta}(\Phi,Q) C := β ^ ( Φ , Q ) , D : = β ^ ( P , Ψ ) D := \hat{\beta}(P,\Psi) D := β ^ ( P , Ψ ) , E : = β ^ ( Φ , Ψ ) E := \hat{\beta}(\Phi,\Psi) E := β ^ ( Φ , Ψ ) , Z : = Φ − C P Z := \Phi-CP Z := Φ − C P , W : = Ψ − Q D W := \Psi-QD W := Ψ − Q D and S G : = E − C D S_G := E-CD S G := E − C D . These subtractions make the remaining families orthogonal to the known block P , Q P,Q P , Q . Their Gram matrix is therefore the reduced matrix S G S_G S G , the Schur complement of that block.
The known block comes from moments fixed by the functional equation, so its construction does not require rational inversion. Subtracting the pairings with this block leaves polynomial families orthogonal to it, and the remaining calculation involves only finite numerical inverses. This is the reason that the reduction also applies to compatible nonregular factors.
Proposition 6.5 (Bilateral formulas without polynomial division). The target pairing has rank d + rank S G d+\operatorname{rank}S_G d + rank S G . For dim X = r \dim X=r dim X = r , dim Y = r + k \dim Y=r+k dim Y = r + k , let M ^ \hat{M} M ^ be its Gram matrix in these bases with the k k k normalization rows appended. Then
T G , k : = [ S G λ ( W ) ] , det M ^ = det T G , k , a ^ = W T G , k − 1 [ 0 ( r − d ) × k I k ] . (105) T_{G,k}:=\begin{bmatrix}S_G\\\lambda(W)\end{bmatrix},\qquad\det\hat{M}=\det T_{G,k},\qquad\hat{a}=WT_{G,k}^{-1}\begin{bmatrix}0_{(r-d)\times k}\\I_k\end{bmatrix}. \tag*{(105)} T G , k := [ S G λ ( W ) ] , det M ^ = det T G , k , a ^ = W T G , k − 1 [ 0 ( r − d ) × k I k ] . (105) For k > 0 k>0 k > 0 , the normalized problem has a unique solution exactly when det T G , k ≠ 0 \det T_{G,k}\ne0 det T G , k = 0 . For a square pair, normality is equivalent to det S G ≠ 0 \det S_G\ne0 det S G = 0 , and
K ^ ( P 0 , Q 0 ) = ∑ i = 1 d Q i ( P 0 ) P i ( Q 0 ) + W ( P 0 ) S G − 1 Z ( Q 0 ) . (106) \hat{K}(P_0,Q_0)=\sum_{i=1}^{d}Q_i(P_0)P_i(Q_0)+W(P_0)S_G^{-1}Z(Q_0). \tag*{(106)} K ^ ( P 0 , Q 0 ) = i = 1 ∑ d Q i ( P 0 ) P i ( Q 0 ) + W ( P 0 ) S G − 1 Z ( Q 0 ) . (106) The last term has tensor rank r − d r-d r − d .
Proof. With r r r rows and r + k r+k r + k columns, the block sizes are C : ( r − d ) × d C:(r-d)\times d C : ( r − d ) × d , D : d × ( r + k − d ) D:d\times(r+k-d) D : d × ( r + k − d ) and E : ( r − d ) × ( r + k − d ) E:(r-d)\times(r+k-d) E : ( r − d ) × ( r + k − d ) . Direct multiplication gives
[ I d 0 − C I r − d ] [ I d D C E ] [ I d − D 0 I r + k − d ] = [ I d 0 0 S G ] . \begin{bmatrix}I_d&0\\-C&I_{r-d}\end{bmatrix}
\begin{bmatrix}I_d&D\\C&E\end{bmatrix}
\begin{bmatrix}I_d&-D\\0&I_{r+k-d}\end{bmatrix}
=
\begin{bmatrix}I_d&0\\0&S_G\end{bmatrix}. [ I d − C 0 I r − d ] [ I d C D E ] [ I d 0 − D I r + k − d ] = [ I d 0 0 S G ] . The two invertible changes of basis have determinant one when square and implement the subtractions Z = Φ − C P Z=\Phi-CP Z = Φ − C P and W = Ψ − Q D W=\Psi-QD W = Ψ − Q D . The diagonal form consequently gives the rank d + rank S G d+\operatorname{rank}S_G d + rank S G of the original Gram. Orthogonality to P P P forces a candidate into W c W^c W c ; its remaining equations are S G c = 0 S_Gc=0 S G c = 0 , λ ( W ) c = I k \lambda(W)c=I_k λ ( W ) c = I k . In the augmented Gram the column change replaces the normalization block [ λ ( Q ) λ ( Ψ ) ] [\lambda(Q)\ \lambda(\Psi)] [ λ ( Q ) λ ( Ψ )] by [ λ ( Q ) λ ( W ) ] [\lambda(Q)\ \lambda(W)] [ λ ( Q ) λ ( W )] . Subtracting λ ( Q ) \lambda(Q) λ ( Q ) times the first d d d rows then makes its first block zero. These unit triangular operations give a block diagonal square matrix with blocks I d I_d I d and T G , k = [ S G ; λ ( W ) ] T_{G,k}=[S_G;\lambda(W)] T G , k = [ S G ; λ ( W )] , of respective orders d d d and r + k − d r+k-d r + k − d . Hence det M ^ = det T G , k \det\hat{M}=\det T_{G,k} det M ^ = det T G , k . The system T G , k c = [ 0 ; I k ] T_{G,k}c=[0;I_k] T G , k c = [ 0 ; I k ] proves both the formula and the uniqueness criterion in (6.8). For a square pair, the inverse diagonal Gram gives (6.9). The independent families Z , W Z,W Z , W have size r − d r-d r − d , and S G − 1 S_G^{-1} S G − 1 is invertible, so the last tensor has rank r − d r-d r − d . □
When the dimension excess occurs on the left, transposition of the construction gives [ S G T , λ ( Z ) ] [S_G^{\mathsf T},\lambda(Z)] [ S G T , λ ( Z )] . It allows d = 0 d=0 d = 0 and inverts only numerical matrices.
For Geronimus division set L 0 = I q L_0=I_q L 0 = I q , R 0 = I p R_0=I_p R 0 = I p , L : = L 1 L:=L_1 L := L 1 , R : = R 1 R:=R_1 R := R 1 with regular determinants. Formula (2.4) computes C , D , E C,D,E C , D , E ; for example, C α , i = β ~ μ ( ϕ α L − 1 , g i ) + ν ( ϕ α , R g i ) C_{\alpha,i}=\tilde{\beta}_\mu(\phi_\alpha L^{-1},g_i)+\nu(\phi_\alpha,Rg_i) C α , i = β ~ μ ( ϕ α L − 1 , g i ) + ν ( ϕ α , R g i ) . The first term of (6.9) is R ( P 0 ) K d μ ( P 0 , Q 0 ) L ( Q 0 ) R(P_0)\mathcal{K}^{\mu}_d(P_0,Q_0)L(Q_0) R ( P 0 ) K d μ ( P 0 , Q 0 ) L ( Q 0 ) , with K d μ : = ∑ i g i f i \mathcal{K}^{\mu}_d:=\sum_i g_if_i K d μ := ∑ i g i f i . Similarly D i , α = β ~ μ ( f i , R − 1 ψ α ) + ν ( f i L , ψ α ) D_{i,\alpha}=\tilde{\beta}_\mu(f_i,R^{-1}\psi_\alpha)+\nu(f_iL,\psi_\alpha) D i , α = β ~ μ ( f i , R − 1 ψ α ) + ν ( f i L , ψ α ) and E α , β = β ~ μ ( ϕ α L − 1 , R − 1 ψ β ) + ν ( ϕ α , ψ β ) E_{\alpha,\beta}=\tilde{\beta}_\mu(\phi_\alpha L^{-1},R^{-1}\psi_\beta)+\nu(\phi_\alpha,\psi_\beta) E α , β = β ~ μ ( ϕ α L − 1 , R − 1 ψ β ) + ν ( ϕ α , ψ β ) . The mass terms remain in these three Gram blocks because ν L V R = 0 \nu LVR=0 ν L V R = 0 removes the mass only when both arguments belong to their denominator images. Consequently the terms in C C C and D D D are part of the Geronimus data even though the known block has no mass contribution.
Additive and scalar specialization For β ^ = β μ + ν \hat{\beta}=\beta_\mu+\nu β ^ = β μ + ν , any bilinear mass has finite rank h h h on prescribed finite spaces. Setting both rational factors equal to the identity in Subsection 3.2 gives, for an original normal square pair, T = − I h − H 0 F 0 \mathbb{T}=-I_h-H_0F_0 T = − I h − H 0 F 0 and K ^ = A [ I N − F 0 ( I h + H 0 F 0 ) − 1 H 0 ] B \hat{K}=\mathcal{A}[I_N-F_0(I_h+H_0F_0)^{-1}H_0]\mathcal{B} K ^ = A [ I N − F 0 ( I h + H 0 F 0 ) − 1 H 0 ] B . Here F 0 ∈ C N × h F_0\in\mathbb{C}^{N\times h} F 0 ∈ C N × h and H 0 ∈ C h × N H_0\in\mathbb{C}^{h\times N} H 0 ∈ C h × N factor the mass matrix in the original dual bases, as in (3.5). For scalar polynomials in several variables and masses at finitely many points, this factorization is obtained from the evaluation matrices of the original polynomial bases. The displayed formula then gives the multivariate
Uvarov kernel update studied in [35 , 36 , 43 ] , together with the corresponding polynomial formulas and the invertibility criterion for the modified moment functional. The finite-space construction here also permits component degree bounds and general bilinear masses.
This Uvarov addition is the four-factor case L 1 = L 0 = a I q L_1=L_0=aI_q L 1 = L 0 = a I q , R 1 = R 0 = I p R_1=R_0=I_p R 1 = R 0 = I p when a ∈ A a\in A a ∈ A is regular and ν ( a f , g ) = 0 \nu(af,g)=0 ν ( a f , g ) = 0 for all f , g f,g f , g . For a balanced mass at finitely many points of an integral positive-dimensional support, a nonzero function vanishing at those points exists. A sufficiently high power annihilates all its value and derivative terms and remains regular, providing a a a even for confluent masses. Taking q = p = 1 q=p=1 q = p = 1 , L 1 = b L_1=b L 1 = b , L 0 = a L_0=a L 0 = a , and R 0 = R 1 = 1 R_0=R_1=1 R 0 = R 1 = 1 gives b μ ^ = a μ b\hat{\mu}=a\mu b μ ^ = a μ . Proposition 2.1 and Subsubsection 6.2.2 also cover nonregular a , b a,b a , b . For example, on A = C [ x , y ] / ( x y ) A=\mathbb{C}[x,y]/(xy) A = C [ x , y ] / ( x y ) take a = x ( x + 2 ) a=x(x+2) a = x ( x + 2 ) and b = x ( x + 1 ) b=x(x+1) b = x ( x + 1 ) . Both annihilators are ( y ) (y) ( y ) . Choose u ( f ) = ∫ 1 2 ( t + 1 ) f ( t , 0 ) d t + ∫ 1 2 f ( 0 , t ) d t u(f)=\int_1^2(t+1)f(t,0)\,\mathrm{d}t+\int_1^2f(0,t)\,\mathrm{d}t u ( f ) = ∫ 1 2 ( t + 1 ) f ( t , 0 ) d t + ∫ 1 2 f ( 0 , t ) d t and v ( f ) = ∫ 1 2 ( t + 2 ) f ( t , 0 ) d t + 3 ∫ 1 2 f ( 0 , t ) d t v(f)=\int_1^2(t+2)f(t,0)\,\mathrm{d}t+3\int_1^2f(0,t)\,\mathrm{d}t v ( f ) = ∫ 1 2 ( t + 2 ) f ( t , 0 ) d t + 3 ∫ 1 2 f ( 0 , t ) d t . Then b v = a u bv=au b v = a u , but cancellation of x x x loses the free component mass: [ ( x + 1 ) v − ( x + 2 ) u ] ( f ) = ∫ 1 2 f ( 0 , t ) d t [(x+1)v-(x+2)u](f)=\int_1^2f(0,t)\,\mathrm{d}t [( x + 1 ) v − ( x + 2 ) u ] ( f ) = ∫ 1 2 f ( 0 , t ) d t . Both functionals are positive definite: a class has the unique form c + x p ( x ) + y q ( y ) c+xp(x)+yq(y) c + x p ( x ) + y q ( y ) , and vanishing on an interval of each axis forces both one-variable restrictions, hence the class, to vanish. The same example also exhibits the finite elimination: the degree-two space has basis 1 , x , y , x 2 , y 2 1,x,y,x^2,y^2 1 , x , y , x 2 , y 2 and dimension five. Take P = b = x ( x + 1 ) P=b=x(x+1) P = b = x ( x + 1 ) and Q = 4 / 55 Q=4/55 Q = 4/55 . Since
v ( b ) = ∫ 1 2 t ( t + 1 ) ( t + 2 ) d t = 55 4 , v(b)=\int_1^2t(t+1)(t+2)\,\mathrm{d}t=\frac{55}{4}, v ( b ) = ∫ 1 2 t ( t + 1 ) ( t + 2 ) d t = 4 55 ,
the known block is v ( P Q ) = 1 v(PQ)=1 v ( P Q ) = 1 . Thus d = 1 d=1 d = 1 and the remaining Gram in (106) has order 5 − 1 = 4 5-1=4 5 − 1 = 4 .
Laurent supports and a bilateral perturbation on a torus
For Laurent polynomials, the finite construction depends on the degree filtration induced by the chosen embedding and on the adjoint used in the second argument. Expressing conjugation in the bilinear notation preserves the determinant formulas, while the filtration determines the spaces to which they apply. A mass distributed along a circle in a torus provides an explicit calculation of both effects.
The product of circles T r : = { z : ∣ z j ∣ = 1 , j ∈ { 1 , … , r } } \mathbb{T}^r:=\{z:|z_j|=1,\ j\in\{1,\ldots,r\}\} T r := { z : ∣ z j ∣ = 1 , j ∈ { 1 , … , r }} has Laurent coordinate algebra A : = C [ z 1 ± 1 , … , z r ± 1 ] / J A:=\mathbb{C}[z_1^{\pm1},\ldots,z_r^{\pm1}]/J A := C [ z 1 ± 1 , … , z r ± 1 ] / J after imposing relations J J J . Indeed z j : = x j + i y j z_j:=x_j+\mathrm{i}y_j z j := x j + i y j and z j − 1 = x j − i y j z_j^{-1}=x_j-\mathrm{i}y_j z j − 1 = x j − i y j identify the full Laurent algebra with C [ x 1 , y 1 , … , x r , y r ] / ( x j 2 + y j 2 − 1 : j ∈ { 1 , … , r } ) \mathbb{C}[x_1,y_1,\ldots,x_r,y_r]/(x_j^2+y_j^2-1:j\in\{1,\ldots,r\}) C [ x 1 , y 1 , … , x r , y r ] / ( x j 2 + y j 2 − 1 : j ∈ { 1 , … , r }) . Laurent polynomials allow positive and negative integer powers. Real-coordinate degree becomes Laurent length ∣ α ∣ 1 : = ∑ j ∣ α j ∣ |\alpha|_1:=\sum_j|\alpha_j| ∣ α ∣ 1 := ∑ j ∣ α j ∣ . The space F N A F_NA F N A consists of classes admitting a representative with all monomials of length at most N N N . The full-torus Christoffel construction is developed in [24 ] . Mixed-type Laurent orthogonality on the unit circle, including diagonal Christoffel and Geronimus perturbations, is treated in [41 ] .
The Laurent algebra also encodes conjugation of values through the operation ( ∑ α c α z α ) ⋆ : = ∑ α c α ‾ z − α (\sum_\alpha c_\alpha z^\alpha)^\star:=\sum_\alpha\overline{c_\alpha}z^{-\alpha} ( ∑ α c α z α ) ⋆ := ∑ α c α z − α . Assume J ⋆ = J J^\star=J J ⋆ = J , so this operation is well defined on classes; the spaces F N A F_NA F N A are then ⋆ \star ⋆ -stable. For a matrix let B † = ( B ⋆ ) T B^\dagger=(B^\star)^\mathsf{T} B † = ( B ⋆ ) T . For rows f , h f,h f , h of lengths g , p g,p g , p , define ⟨ f , h ⟩ μ : = β μ ( f , h † ) \langle f,h\rangle_\mu:=\beta_\mu(f,h^\dagger) ⟨ f , h ⟩ μ := β μ ( f , h † ) . It is linear in f f f and conjugate-linear in h h h ; Hermitian symmetry is an additional condition. The second transformed row is h R 1 † hR_1^\dagger h R 1 † , since ( h R 1 † ) † = R 1 h † (hR_1^\dagger)^\dagger=R_1h^\dagger ( h R 1 † ) † = R 1 h † ; the order follows from reversing products under the adjoint. Thus the four-factor equation becomes
⟨ f L 1 , h R 1 † ⟩ μ ^ = ⟨ f L 0 , h R 0 † ⟩ μ . \langle fL_1,hR_1^\dagger\rangle_{\hat{\mu}}=\langle fL_0,hR_0^\dagger\rangle_\mu. ⟨ f L 1 , h R 1 † ⟩ μ ^ = ⟨ f L 0 , h R 0 † ⟩ μ .
For regular factors, the particular solution and mass decomposition therefore take the form ⟨ f , h ⟩ μ ^ = β ~ μ ( f S L , S R h † ) + ν ( f , h † ) \langle f,h\rangle_{\hat{\mu}}=\widetilde{\beta}_\mu(fS_L,S_Rh^\dagger)+\nu(f,h^\dagger) ⟨ f , h ⟩ μ ^ = β μ ( f S L , S R h † ) + ν ( f , h † ) . For arbitrary row spaces X , Y X,Y X , Y , including different component bounds, the Gram matrix is the bilinear one for X , Y † X,Y^\dagger X , Y † , so its determinant criterion, bordered formulas and correction ranks apply. Left rows are unchanged; right rows and their normalization are adjoints of the bilinear columns and normalization. For R j = L j † R_j=L_j^\dagger R j = L j † , j ∈ { 0 , 1 } j\in\{0,1\} j ∈ { 0 , 1 } , one has S R = L 0 † ( L 1 † ) − 1 = ( L 1 − 1 L 0 ) † = S L † S_R=L_0^\dagger(L_1^\dagger)^{-1}=(L_1^{-1}L_0)^\dagger=S_L^\dagger S R = L 0 † ( L 1 † ) − 1 = ( L 1 − 1 L 0 ) † = S L † . The rational part is then ⟨ f S L , h S L ⟩ μ \langle fS_L,hS_L\rangle_\mu ⟨ f S L , h S L ⟩ μ : the same change acts on both row arguments.
For a balanced scalar functional u : A ⟶ C u:A\longrightarrow\mathbb{C} u : A ⟶ C with all filtered Gram matrices invertible, choose a basis column χ [ n ] \chi_{[n]} χ [ n ] of a ⋆ \star ⋆ -stable complement of F n − 1 A F_{n-1}A F n − 1 A in F n A F_nA F n A , with F − 1 A : = 0 F_{-1}A:=0 F − 1 A := 0 . Write χ [ n ] ⋆ = C [ n ] χ [ n ] \chi_{[n]}^\star=C_{[n]}\chi_{[n]} χ [ n ] ⋆ = C [ n ] χ [ n ] , so C ‾ [ n ] C [ n ] = I \overline{C}_{[n]}C_{[n]}=I C [ n ] C [ n ] = I . The monic block P [ n ] P_{[n]} P [ n ] equals χ [ n ] \chi_{[n]} χ [ n ] up to lower-degree terms and is orthogonal to F n − 1 A F_{n-1}A F n − 1 A ; its norm is H [ n ] : = u ( P [ n ] P [ n ] T ) H_{[n]}:=u(P_{[n]}P_{[n]}^\mathsf{T}) H [ n ] := u ( P [ n ] P [ n ] T ) , with u u u acting entrywise. The corresponding sesquilinear blocks are obtained without changing the prescribed leading classes: Φ [ n ] : = P [ n ] \Phi_{[n]}:=P_{[n]} Φ [ n ] := P [ n ] and Ψ [ n ] : = C [ n ] − 1 P [ n ] ⋆ \Psi_{[n]}:=C_{[n]}^{-1}P_{[n]}^\star Ψ [ n ] := C [ n ] − 1 P [ n ] ⋆ . The identity C ‾ [ n ] C [ n ] = I \overline{C}_{[n]}C_{[n]}=I C [ n ] C [ n ] = I gives C ‾ [ n ] − 1 = C [ n ] \overline{C}_{[n]}^{-1}=C_{[n]} C [ n ] − 1 = C [ n ] , so Ψ [ n ] ⋆ = C [ n ] P [ n ] \Psi^\star_{[n]}=C_{[n]}P_{[n]} Ψ [ n ] ⋆ = C [ n ] P [ n ] . Consequently
u ( Φ [ n ] ( Ψ [ j ] ⋆ ) T ) = u ( P [ n ] P [ j ] T ) C [ j ] T = δ n , j H [ n ] C [ n ] T . u\left(\Phi_{[n]}(\Psi^\star_{[j]})^{\mathsf T}\right)=u\left(P_{[n]}P_{[j]}^{\mathsf T}\right)C_{[j]}^{\mathsf T}=\delta_{n,j}H_{[n]}C_{[n]}^{\mathsf T}. u ( Φ [ n ] ( Ψ [ j ] ⋆ ) T ) = u ( P [ n ] P [ j ] T ) C [ j ] T = δ n , j H [ n ] C [ n ] T .
The construction preserves the leading column in both families, namely χ [ n ] \chi_{[n]} χ [ n ] , and the displayed pairing determines their sesquilinear norm. The same change of convention acts by a fixed basis matrix on each finite Gram: in a finite basis column χ \chi χ with χ ⋆ = C χ \chi^\star=C\chi χ ⋆ = C χ , put G b i l : = u ( χ χ T ) G^{\mathrm{bil}}:=u(\chi\chi^{\mathsf T}) G bil := u ( χ χ T ) and G s e s : = u ( χ ( χ ⋆ ) T ) = G b i l C T G^{\mathrm{ses}}:=u(\chi(\chi^\star)^{\mathsf T})=G^{\mathrm{bil}}C^{\mathsf T} G ses := u ( χ ( χ ⋆ ) T ) = G bil C T . The fixed right factor cancels in the ratio of the perturbed and original Gram determinants when both use the same finite space and basis. It also cancels between the column basis and inverse Gram in the kernel:
( χ ⋆ ( P ) ) T ( G s e s ) − 1 χ ( Q ) = χ ( P ) T C T ( C T ) − 1 ( G b i l ) − 1 χ ( Q ) . (\chi^\star(P))^{\mathsf T}(G^{\mathrm{ses}})^{-1}\chi(Q)=\chi(P)^{\mathsf T}C^{\mathsf T}(C^{\mathsf T})^{-1}(G^{\mathrm{bil}})^{-1}\chi(Q). ( χ ⋆ ( P ) ) T ( G ses ) − 1 χ ( Q ) = χ ( P ) T C T ( C T ) − 1 ( G bil ) − 1 χ ( Q ) .
The resulting kernel is therefore the one in the bilinear convention. For Hermitian data the two polynomial families also coincide: conjugating the orthogonality conditions and using u ⋆ = u u^\star=u u ⋆ = u shows that C [ n ] − 1 P [ n ] ⋆ C_{[n]}^{-1}P^\star_{[n]} C [ n ] − 1 P [ n ] ⋆ has the same leading class and lower orthogonality as P [ n ] P_{[n]} P [ n ] . Uniqueness gives equality of the two families. Conjugation acts compatibly on the functional equation as well: conjugating b v = a u bv=au b v = a u gives b ⋆ v ⋆ = a ⋆ u ⋆ b^\star v^\star=a^\star u^\star b ⋆ v ⋆ = a ⋆ u ⋆ , where v ⋆ ( f ) : = v ( f ⋆ ) v^\star(f):=v(f^\star) v ⋆ ( f ) := v ( f ⋆ ) ; the conjugate of an evaluation mass at c c c is at c ⋆ : = ( 1 / c j ‾ ) j c^\star:=(1/\overline{c_j})_j c ⋆ := ( 1/ c j ) j .
The effect of the degree filtration can be calculated on the subtorus ( s , t ) ↦ ( s , t , s t ) (s,t)\mapsto(s,t,st) ( s , t ) ↦ ( s , t , s t ) of T 3 \mathbb{T}^3 T 3 . Its algebra is A = C [ s ± 1 , t ± 1 ] A=\mathbb{C}[s^{\pm1},t^{\pm1}] A = C [ s ± 1 , t ± 1 ] , while the inherited degree counts the three coordinates s , t , s t s,t,st s , t , s t ; thus s t st s t has degree one. Writing s a t b = s a − c ( s t ) c s^at^b=s^{a-c}(st)^c s a t b = s a − c ( s t ) c shows that its inherited degree is the minimum of ∣ a − c ∣ + ∣ b − c ∣ + ∣ c ∣ |a-c|+|b-c|+|c| ∣ a − c ∣ + ∣ b − c ∣ + ∣ c ∣ over integers c c c . A median of 0 , a , b 0,a,b 0 , a , b minimizes this sum, whose value is the largest minus the smallest of these three numbers. Equivalently, ℓ ( a , b ) = max ( ∣ a ∣ , ∣ b ∣ , ∣ a − b ∣ ) \ell(a,b)=\max(|a|,|b|,|a-b|) ℓ ( a , b ) = max ( ∣ a ∣ , ∣ b ∣ , ∣ a − b ∣ ) . Write H N : = { ( a , b ) ∈ Z 2 : ℓ ( a , b ) ≤ N } \mathcal{H}_N:=\{(a,b)\in\mathbb{Z}^2:\ell(a,b)\le N\} H N := {( a , b ) ∈ Z 2 : ℓ ( a , b ) ≤ N } . For fixed b ∈ { − N , … , N } b\in\{-N,\ldots,N\} b ∈ { − N , … , N } the admissible a a a form the integer interval
max { − N , b − N } ≤ a ≤ min { N , b + N } , \max\{-N,b-N\}\le a\le\min\{N,b+N\}, max { − N , b − N } ≤ a ≤ min { N , b + N } ,
of size 2 N + 1 − ∣ b ∣ 2N+1-|b| 2 N + 1 − ∣ b ∣ . Summing these sizes gives ( 2 N + 1 ) 2 − N ( N + 1 ) = 3 N ( N + 1 ) + 1 (2N+1)^2-N(N+1)=3N(N+1)+1 ( 2 N + 1 ) 2 − N ( N + 1 ) = 3 N ( N + 1 ) + 1 . Subtracting the corresponding count at N − 1 N-1 N − 1 gives 6 N 6N 6 N boundary monomials for N ≥ 1 N\ge1 N ≥ 1 . Let w w w be uniform averaging of both angular variables over [ 0 , 2 π ] [0,2\pi] [ 0 , 2 π ] and λ \lambda λ uniform averaging on the circle s = 1 s=1 s = 1 ; these are the normalized Haar functionals. Put d : = s − 1 d:=s-1 d := s − 1 , q G : = d d ⋆ = 2 − s − s − 1 q_G:=dd^\star=2-s-s^{-1} q G := d d ⋆ = 2 − s − s − 1 , u = q G w u=q_Gw u = q G w and v M : = w + M λ v_M:=w+M\lambda v M := w + M λ , M ∈ C M\in\mathbb{C} M ∈ C . Then d v M d ⋆ = u dv_Md^\star=u d v M d ⋆ = u : both denominator factors are nonconstant. For M ≥ 0 M\ge0 M ≥ 0 both forms are positive definite, since q G > 0 q_G>0 q G > 0 off the Haar-null circle and v M v_M v M contains w w w . For b ∈ { − N , … , N } b\in\{-N,\ldots,N\} b ∈ { − N , … , N } set I N , b : = { a ∈ Z : ( a , b ) ∈ H N } I_{N,b}:=\{a\in\mathbb{Z}:(a,b)\in\mathcal{H}_N\} I N , b := { a ∈ Z : ( a , b ) ∈ H N } , n N , b : = 2 N + 1 − ∣ b ∣ n_{N,b}:=2N+1-|b| n N , b := 2 N + 1 − ∣ b ∣ and A N , b ( s , t ) : = ∑ a ∈ I N , b s a t b A_{N,b}(s,t):=\sum_{a\in I_{N,b}}s^at^b A N , b ( s , t ) := ∑ a ∈ I N , b s a t b . The two denominator factors thus express the scalar division as a bilateral transformation. The associated kernels follow the convention established above, with conjugation in the first variable.
This indexing separates the monomials into groups according to their exponent b b b of t t t . Full-torus averaging distinguishes every pair of exponents, whereas restriction to s = 1 s=1 s = 1 forgets the exponent of s s s . Consequently the mass couples all monomials within one group and none between different groups. Each group therefore contributes a rank-one mass matrix. The determinant and inverse can then be computed group by group, as in the following proposition.
Proposition 6.6 (A distributed mass on a restricted torus). Use the Laurent monomial basis [ s a t b ] ( a , b ) ∈ H N [s^at^b]_{(a,b)\in\mathcal{H}_N} [ s a t b ] ( a , b ) ∈ H N of F N A F_NA F N A . Let G N ( v ) G_N(v) G N ( v ) and K N v K_N^v K N v be its sesquilinear Gram matrix and kernel as defined above. Then
det G N ( v M ) = ∏ b = − N N ( 1 + M n N , b ) , det G N ( v M ) det G N ( u ) = ∏ b = − N N 1 + M n N , b n N , b + 1 . (107) \det G_N(v_M)=\prod_{b=-N}^{N}(1+Mn_{N,b}),\qquad\frac{\det G_N(v_M)}{\det G_N(u)}=\prod_{b=-N}^{N}\frac{1+Mn_{N,b}}{n_{N,b}+1}. \tag*{(107)} det G N ( v M ) = b = − N ∏ N ( 1 + M n N , b ) , det G N ( u ) det G N ( v M ) = b = − N ∏ N n N , b + 1 1 + M n N , b . (107) Normality holds precisely when the displayed factors are nonzero. For all N ≥ 0 N\ge0 N ≥ 0 simultaneously, this means M ∉ { − 1 / j : j ∈ Z ≥ 0 } M\notin\{-1/j:j\in\mathbb{Z}_{\ge0}\} M ∈ / { − 1/ j : j ∈ Z ≥ 0 } . At every normal degree,
K N v M ( P , Q ) = K N w ( P , Q ) − ∑ b = − N N M A N , b ⋆ ( P ) A N , b ( Q ) 1 + M n N , b . (108) K_N^{v_M}(P,Q)=K_N^w(P,Q)-\sum_{b=-N}^{N}\frac{M A_{N,b}^\star(P)A_{N,b}(Q)}{1+Mn_{N,b}}. \tag*{(108)} K N v M ( P , Q ) = K N w ( P , Q ) − b = − N ∑ N 1 + M n N , b M A N , b ⋆ ( P ) A N , b ( Q ) . (108)
For M ≠ 0 M \ne0 M = 0 this correction has rank 2 N + 1 2N+1 2 N + 1 , whereas the kernel multiplied by both rational factors satisfies
rank ( K N v M ( P , Q ) d ⋆ ( P ) d ( Q ) − K N u ( P , Q ) ) = 4 N + 2. (109) \operatorname{rank}\left(\frac{K_N^{v_M}(P,Q)}{d^{\star}(P)d(Q)}-K_N^{u}(P,Q)\right)=4N+2. \tag*{(109)} rank ( d ⋆ ( P ) d ( Q ) K N v M ( P , Q ) − K N u ( P , Q ) ) = 4 N + 2. (109)
For N ≥ 1 N \ge1 N ≥ 1 , assume also that G N − 1 ( v M ) G_{N-1}(v_M) G N − 1 ( v M ) is invertible. Then the monic blocks with prescribed leading classes are unique. For ( a , b ) ∈ H N ∖ H N − 1 (a,b) \in\mathcal{H}_N \setminus\mathcal{H}_{N-1} ( a , b ) ∈ H N ∖ H N − 1 , the left monic polynomial is Q N ; a , b = s a t b − M A N − 1 , b / ( 1 + M n N − 1 , b ) Q_{N;a,b}=s^at^b-MA_{N-1,b}/(1+Mn_{N-1,b}) Q N ; a , b = s a t b − M A N − 1 , b / ( 1 + M n N − 1 , b ) if ∣ b ∣ < N |b|<N ∣ b ∣ < N , and s a t b s^at^b s a t b otherwise; the right one replaces M M M by M ‾ \overline{M} M .
Proof. Haar orthogonality makes the Gram matrix of w w w the identity. On the circle s = 1 s=1 s = 1 , every monomial in the group with exponent b b b restricts to t b t^b t b . Different groups remain orthogonal under λ \lambda λ . Hence the block of G N ( v M ) G_N(v_M) G N ( v M ) for this group is I + M 11 T I+M\mathbf{1}\mathbf{1}^{\mathsf{T}} I + M 1 1 T , where 1 \mathbf{1} 1 has length n N , b n_{N,b} n N , b . This matrix is the identity on the subspace with coordinate sum zero, and multiplies 1 \mathbf{1} 1 by 1 + M n N , b 1+Mn_{N,b} 1 + M n N , b . Its determinant is that factor, and, when nonzero, its inverse is
I − M 1 + M n N , b 11 T . I-\frac{M}{1+Mn_{N,b}}\mathbf{1}\mathbf{1}^{\mathsf{T}}. I − 1 + M n N , b M 1 1 T .
Since the groups are mutually orthogonal, their determinants multiply and their inverse matrices contribute separately to the kernel expansion. This proves the numerator in (6.10) and (6.11).
For u = ( 2 − s − s − 1 ) w u=(2-s-s^{-1})w u = ( 2 − s − s − 1 ) w , the block is tridiagonal with diagonal 2 2 2 and adjacent diagonals − 1 -1 − 1 . Expanding its order-j j j determinant gives the recurrence d j = 2 d j − 1 − d j − 2 d_j=2d_{j-1}-d_{j-2} d j = 2 d j − 1 − d j − 2 , with d 0 = 1 d_0=1 d 0 = 1 , d 1 = 2 d_1=2 d 1 = 2 . Thus d j = j + 1 d_j=j+1 d j = j + 1 , proving the denominator in (6.10). For each N N N the group sizes are the integers from N + 1 N+1 N + 1 to 2 N + 1 2N+1 2 N + 1 . Their union over N ≥ 0 N\ge0 N ≥ 0 is all positive integers: 1 1 1 occurs at N = 0 N=0 N = 0 , and any j ≥ 2 j\ge2 j ≥ 2 occurs at N = j − 1 N=j-1 N = j − 1 in an extreme group. Thus the excluded parameters are exactly M = − 1 / j M=-1/j M = − 1/ j , j ∈ Z > 0 j\in\mathbb{Z}_{>0} j ∈ Z > 0 , proving the simultaneous normality condition.
For M ≠ 0 M\ne0 M = 0 , each kernel correction block has rank one, and the different groups have disjoint monomial supports. The group decomposition therefore makes their ranks add to 2 N + 1 2N+1 2 N + 1 . The same inverse matrices determine the monic polynomials: for a leading monomial with ∣ b ∣ < N |b|<N ∣ b ∣ < N , its Haar pairing with the lower group vanishes and its mass pairing is M 1 T M\mathbf{1}^{\mathsf{T}} M 1 T . If n = n N − 1 , b n=n_{N-1,b} n = n N − 1 , b , multiplication by the inverse lower Gram gives the coefficient row
M 1 T [ I − M 1 + M n 11 T ] = M 1 + M n 1 T . M\mathbf{1}^{\mathsf{T}}\left[I-\frac{M}{1+Mn}\mathbf{1}\mathbf{1}^{\mathsf{T}}\right]=\frac{M}{1+Mn}\mathbf{1}^{\mathsf{T}}. M 1 T [ I − 1 + M n M 1 1 T ] = 1 + M n M 1 T .
Thus the projection to subtract is exactly M A N − 1 , b / ( 1 + M n N − 1 , b ) MA_{N-1,b}/(1+Mn_{N-1,b}) M A N − 1 , b / ( 1 + M n N − 1 , b ) . If ∣ b ∣ = N |b|=N ∣ b ∣ = N , there is no lower group to subtract. The right orthogonality equations conjugate M M M , giving the two monic formulas.
The bilateral comparison involves the rationally transformed spaces, whose intersection with the original space controls the additional rank. In this case take E = X = F N A E=X=F_NA E = X = F N A and conjugate right spaces. A function in this intersection has f = d g f=dg f = d g with both g g g and d g dg d g in F N A F_NA F N A . Within a fixed group, multiplication by d = s − 1 d=s-1 d = s − 1 raises the largest s s s exponent by one unless its coefficient vanishes. Precisely one condition is imposed, leaving n N , b − 1 n_{N,b}-1 n N , b − 1 dimensions. Thus each side has excess 2 N + 1 2N+1 2 N + 1 . Every such f f f vanishes at s = 1 s=1 s = 1 , so the restricted mass is zero. This vanishing determines the rank of the bilateral kernel difference through a finite linear-algebra calculation. Write D : = dim F N A D:=\dim F_NA D := dim F N A and s : = 2 N + 1 s:=2N+1 s := 2 N + 1 . In minimal ambient bases the two kernels have difference matrix B R C B L T B_RCB_L^{\mathsf{T}} B R C B L T , where B L , B R B_L,B_R B L , B R concatenate the original and bases multiplied by the rational factors, and C = diag ( − G u − 1 , G v M − 1 ) C=\operatorname{diag}(-G_u^{-1},G_{v_M}^{-1}) C = diag ( − G u − 1 , G v M − 1 ) is invertible. Choose independent bases of the two sums of original and rationally transformed functions. Each sum has dimension D + s D+s D + s , by the overlap count, so B L , B R B_L,B_R B L , B R have sizes ( D + s ) × 2 D (D+s)\times2D ( D + s ) × 2 D , rank D + s D+s D + s , and kernels of dimension D − s D-s D − s . Each kernel describes the relations between the two basis lists; a vector in it therefore represents a function common to the original spaces and their rational images. The pairing u T C − 1 v u^{\mathsf{T}}C^{-1}v u T C − 1 v on these two kernels subtracts their original pairing from their target pairing. The rational terms cancel; the mass is zero because both candidates vanish at s = 1 s=1 s = 1 .
Consequently, for every u ∈ ker B L u\in\ker B_L u ∈ ker B L and v ∈ ker B R v\in\ker B_R v ∈ ker B R the cancellation just proved reads u T C − 1 v = 0 u^{\mathsf{T}}C^{-1}v=0 u T C − 1 v = 0 . Hence C − 1 v C^{-1}v C − 1 v lies in the annihilator of ker B L \ker B_L ker B L , which is im B L T \operatorname{im}B_L^{\mathsf{T}} im B L T . The latter image is the annihilator: it is contained there and both have dimension D + s D+s D + s . Thus ker B R ⊂ C im B L T \ker B_R\subset C\operatorname{im}B_L^{\mathsf{T}} ker B R ⊂ C im B L T . The first map C B L T CB_L^{\mathsf{T}} C B L T has rank D + s D+s D + s , and applying B R B_R B R kills exactly its D − s D-s D − s dimensional kernel. The difference matrix has rank ( D + s ) − ( D − s ) = 2 s = 4 N + 2 (D+s)-(D-s)=2s=4N+2 ( D + s ) − ( D − s ) = 2 s = 4 N + 2 . Independent ambient bases make this its separation rank as well. The cancellation used in this rank calculation is justified by the rational extension
β ~ u ( f / d , g / d ⋆ ) = w ( f g ) , \widetilde{\beta}_{u}(f/d,g/d^\star)=w(fg), β u ( f / d , g / d ⋆ ) = w ( f g ) ,
since the factor d d ⋆ dd^\star d d ⋆ in u u u cancels the two denominators. The same rank is visible in the coefficient matrix of each group of size n n n . After removal of the common Laurent monomial factor, the remaining powers are ordered from 0 0 0 to n − 1 n-1 n − 1 . Multiplying the kernel difference by the two denominators gives the coefficient matrix
B n : = E n ( I n + M 1 n 1 n T ) − 1 E n T − D n ( D n T D n ) − 1 D n T . B_n:=E_n(I_n+M1_n1_n^{\mathsf T})^{-1}E_n^{\mathsf T}-D_n(D_n^{\mathsf T}D_n)^{-1}D_n^{\mathsf T}. B n := E n ( I n + M 1 n 1 n T ) − 1 E n T − D n ( D n T D n ) − 1 D n T .
where E n : = [ I n ; 0 ] E_n:=[I_n;0] E n := [ I n ; 0 ] and the j j j th column of D n D_n D n is e j + 1 − e j e_{j+1}-e_j e j + 1 − e j , j ∈ { 1 , … , n } j\in\{1,\ldots,n\} j ∈ { 1 , … , n } . The matrix D n T D n D_n^{\mathsf T}D_n D n T D n is precisely the tridiagonal original Gram. Its column image is the hyperplane of coordinate sum zero, so
D n ( D n T D n ) − 1 D n T = I n + 1 − 1 n + 1 1 n + 1 T n + 1 . D_n(D_n^{\mathsf T}D_n)^{-1}D_n^{\mathsf T}=I_{n+1}-\frac{1_{n+1}1_{n+1}^{\mathsf T}}{n+1}. D n ( D n T D n ) − 1 D n T = I n + 1 − n + 1 1 n + 1 1 n + 1 T .
Put e : = [ 1 n ; 0 ] e:=[1_n;0] e := [ 1 n ; 0 ] and k : = e n + 1 k:=e_{n+1} k := e n + 1 . Substituting the already computed target inverse gives
B n = [ e k ] [ 1 n + 1 − M 1 + M n 1 n + 1 1 n + 1 − n n + 1 ] [ e T k T ] . B_n=
\begin{bmatrix}e&k\end{bmatrix}
\begin{bmatrix}
\frac{1}{n+1}-\frac{M}{1+Mn} & \frac{1}{n+1}\\
\frac{1}{n+1} & -\frac{n}{n+1}
\end{bmatrix}
\begin{bmatrix}e^{\mathsf T}\\ k^{\mathsf T}\end{bmatrix}. B n = [ e k ] [ n + 1 1 − 1 + M n M n + 1 1 n + 1 1 − n + 1 n ] [ e T k T ] .
In this factorization, the middle determinant is − 1 / ( ( n + 1 ) ( 1 + M n ) ) -1/((n+1)(1+Mn)) − 1/ (( n + 1 ) ( 1 + M n )) , nonzero at every normal parameter, and e , k e,k e , k are independent. Each group therefore has rank two. Adding the 2 N + 1 2N+1 2 N + 1 independent groups again gives (109) . □
The distinction between the unrestricted quotient and the prescribed degree bounds can also be described in this example. Since q G = − ( s − 1 ) 2 / s q_G=-(s-1)^2/s q G = − ( s − 1 ) 2 / s , its quotient is A / ( q G ) = C [ t ± 1 , ϵ ] / ( ϵ 2 ) A/(q_G)=\mathbb{C}[t^{\pm1},\epsilon]/(\epsilon^2) A / ( q G ) = C [ t ± 1 , ϵ ] / ( ϵ 2 ) , where ϵ : = [ s − 1 ] \epsilon:=[s-1] ϵ := [ s − 1 ] and s a t b = ( 1 + a ϵ ) t b s^at^b=(1+a\epsilon)t^b s a t b = ( 1 + a ϵ ) t b . The formula for negative a a a follows as well from ( 1 + ϵ ) − 1 = 1 − ϵ (1+\epsilon)^{-1}=1-\epsilon ( 1 + ϵ ) − 1 = 1 − ϵ modulo ϵ 2 \epsilon^2 ϵ 2 . For N ≥ 1 N\ge1 N ≥ 1 , each group contains at least two consecutive a a a values. Their classes ( 1 + a ϵ ) t b (1+a\epsilon)t^b ( 1 + a ϵ ) t b and ( 1 + ( a + 1 ) ϵ ) t b (1+(a+1)\epsilon)t^b ( 1 + ( a + 1 ) ϵ ) t b span t b , ϵ t b t^b,\epsilon t^b t b , ϵ t b . These are the value and first transverse coefficient at s = 1 s=1 s = 1 ; the $2N+1$ different b b b values consequently give 4 N + 2 4N+2 4 N + 2 independent classes. Restriction to that circle removes the ϵ \epsilon ϵ terms; the remaining 2 N + 1 2N+1 2 N + 1 modes have nondegenerate Haar pairing. By contrast, dim ( F N A / q G F N − 1 A ) = 6 N \dim(F_NA/q_GF_{N-1}A)=6N dim ( F N A / q G F N − 1 A ) = 6 N removes only multiples with bounded multipliers, whereas the image above removes F N A ∩ q G A F_NA\cap q_GA F N A ∩ q G A . These can differ because multiplication by q G q_G q G need not raise inherited degree, as q G s t 2 = 2 s t 2 − s 2 − t 2 q_Gst^2=2st^2-s^2-t^2 q G s t 2 = 2 s t 2 − s 2 − t 2 shows. For instance s t 2 st^2 s t 2 has degree two, while each term of q G s t 2 q_Gst^2 q G s t 2 still has degree at most two. Thus this multiple is in F 2 A ∩ q G A F_2A\cap q_GA F 2 A ∩ q G A although its unique multiplier is outside $F_1A. This example identifies the difference between the image in the unrestricted quotient and the quotient by bounded multiples: the former also removes multiples whose multipliers exceed the prescribed bound. The quotient conditions must therefore be accompanied by the degree constraints.
The filtration also affects Christoffel multiplication. Replacing the division factors by the multiplier q C : = c − 1 2 ( s + s − 1 + t + t − 1 ) q_C:=c-\frac{1}{2}(s+s^{-1}+t+t^{-1}) q C := c − 2 1 ( s + s − 1 + t + t − 1 ) , c > 2 c>2 c > 2 , permits an exact degree calculation. Each linear weight ± a , ± b , ± ( a − b ) \pm a,\pm b,\pm(a-b) ± a , ± b , ± ( a − b ) has maximum one on the exponents of q C q_C q C . The maximum-weight parts multiply to a nonzero Laurent polynomial, so their weights add. Hence deg F ( q C f ) = 1 + deg F f \deg_F(q_Cf)=1+\deg_Ff deg F ( q C f ) = 1 + deg F f . Also q C ≥ c 2 − 2 > 0 q_C\ge c^2-2>0 q C ≥ c 2 − 2 > 0 . Proposition 6.3 , with ambient F N A F_NA F N A and target F N − 1 A F_{N-1}A F N − 1 A , therefore gives dim F N A − dim F N − 1 A = 6 N \dim F_NA-\dim F_{N-1}A=6N dim F N A − dim F N − 1 A = 6 N independent conditions on each image and kernel correction rank 6 N 6N 6 N . Under the usual Laurent-length filtration on T 2 \mathbb{T}^2 T 2 , the filtered space has dimension 2 N ( N + 1 ) + 1 2N(N+1)+1 2 N ( N + 1 ) + 1 and the corresponding count is 4 N 4N 4 N for the same q C q_C q C . Thus the filtration determines the number of image conditions, which is reflected in the perturbation formula.
Symmetric rational congruences
Symmetry and positivity impose additional structure on the finite systems obtained above. A congruence applies the same rational change f ↦ f S f\mapsto fS f ↦ f S to both arguments of a symmetric pairing, so symmetry first restricts the admissible homogeneous masses. When the rational pairing and the mass are positive, the part common to the original and transformed spaces admits an orthogonal decomposition. Eliminating this common part reduces the matrices needed to construct the normalized polynomials and their kernels. The reduction depends on an intersection of polynomial spaces, which can itself be determined from their coefficient matrices.
Throughout this section the determinants are assumed regular, that is, non-zero-divisors in the coordinate algebra; the corresponding matrix changes are therefore invertible over its total quotient ring. The congruence specialization is given by q = p = m q=p=m q = p = m and R i = L i T R_i=L_i^{\mathsf{T}} R i = L i T . In the bivariate case symmetry means β μ ( f , g T ) = β μ ( g , f T ) \beta_\mu(f,g^{\mathsf{T}})=\beta_\mu(g,f^{\mathsf{T}}) β μ ( f , g T ) = β μ ( g , f T ) , including interchange of the two arguments; in the balanced case it is μ T = μ \mu^{\mathsf{T}}=\mu μ T = μ . Then L 1 μ ^ L 1 T = L 0 μ L 0 T L_1\widehat{\mu}L_1^{\mathsf{T}}=L_0\mu L_0^{\mathsf{T}} L 1 μ L 1 T = L 0 μ L 0 T , μ ^ = S μ S T + ν \widehat{\mu}=S\mu S^{\mathsf{T}}+\nu μ = S μ S T + ν , with S : = L 1 − 1 L 0 S:=L_1^{-1}L_0 S := L 1 − 1 L 0 . With a symmetric rational extension, the rational term is symmetric, so μ ^ \widehat{\mu} μ is symmetric exactly when ν \nu ν is symmetric. The two polynomial spaces may consequently be taken as transposes, while their common component bounds remain arbitrary.
These statements use bilinear transposition; the assertions concerning positivity below are made for real data and real vectors. Massless adjugate congruences on the line appear in [1 Section 3.6.2] ; matrix Geronimus–Sobolev forms in [25 ] . The additional restriction imposed by positivity follows directly from the homogeneous equation. For a positive semidefinite homogeneous mass, ν ( f L 1 , ( f L 1 ) T ) = 0 \nu(fL_1,(fL_1)^{\mathsf{T}})=0 ν ( f L 1 , ( f L 1 ) T ) = 0 for every row f f f . Its pairing with an arbitrary row g g g therefore satisfies the Cauchy–Schwarz bound
∣ ν ( f L 1 , g T ) ∣ 2 ≤ ν ( f L 1 , ( f L 1 ) T ) ν ( g , g T ) = 0. \lvert\nu(fL_1,g^{\mathsf{T}})\rvert^2\leq\nu(fL_1,(fL_1)^{\mathsf{T}})\nu(g,g^{\mathsf{T}})=0. ∣ ν ( f L 1 , g T ) ∣ 2 ≤ ν ( f L 1 , ( f L 1 ) T ) ν ( g , g T ) = 0.
Consequently L 1 ν = 0 = ν L 1 T L_1\nu=0=\nu L_1^{\mathsf{T}} L 1 ν = 0 = ν L 1 T : vanishing on the product of the two denominator images extends to vanishing whenever either argument lies in its denominator image. The mass is therefore unilateral on both sides. For balanced masses this means that its class in Theorem 2.4 vanishes, in every dimension.
This annihilation property separates symmetric masses from positive ones. Symmetric masses may contain indefinite derivative terms, whereas a zero-norm vector for a positive semidefinite form pairs to zero with every vector. On a curve, applying this property to the denominator image permits passage to a finite-dimensional positive quotient. For balanced pairings, multiplication by the coordinates is self-adjoint on that quotient, excluding nontrivial nilpotent multiplication and hence the corresponding derivative directions. The resulting masses are described by point-evaluation matrices in the following proposition.
Proposition 7.1 (Symmetric and positive homogeneous masses). Write Sym m ( A ) : = { H ∈ Mat m ( A ) : H T = H } \operatorname{Sym}_m(A):=\{H\in\operatorname{Mat}_m(A):H^{\mathsf{T}}=H\} Sym m ( A ) := { H ∈ Mat m ( A ) : H T = H } and d L = dim C ( A 1 × m / A 1 × m L 1 ) d_L=\dim_{\mathbb{C}}(A^{1\times m}/A^{1\times m}L_1) d L = dim C ( A 1 × m / A 1 × m L 1 ) . For balanced pairings on a curve the symmetric homogeneous masses are the dual of
Q s y m : = Sym m ( A ) / { L 1 T H L 1 : H ∈ Sym m ( A ) } . (110) Q_{\mathrm{sym}}:=\operatorname{Sym}_m(A)/\{L_1^{\mathsf{T}}HL_1:H\in\operatorname{Sym}_m(A)\}. \tag*{(110)} Q sym := Sym m ( A ) / { L 1 T H L 1 : H ∈ Sym m ( A )} . (110) On a smooth affine curve their dimension is ( m + 1 ) d L (m+1)d_L ( m + 1 ) d L . Locally, if the denominator has Smith orders α 1 , … , α m \alpha_1,\ldots,\alpha_m α 1 , … , α m , the independent symmetric entry ( i , j ) (i,j) ( i , j ) has jets of orders less than α i + α j \alpha_i+\alpha_j α i + α j , for i ≤ j i\leq j i ≤ j .
For balanced pairings over a real curve coordinate algebra, the positive semidefinite homogeneous masses are exactly the finite sums
ν ( f , g ) = ∑ P ∈ C ∘ ( R ) f ( P ) C P g ( P ) , C P = C P T ≥ 0 , L 1 ( P ) C P = 0. (111) \nu(f,g)=\sum_{P\in C^{\circ}(\mathbb{R})}f(P)C_Pg(P),\qquad C_P=C_P^{\mathsf{T}}\geq0,\qquad L_1(P)C_P=0. \tag*{(111)} ν ( f , g ) = P ∈ C ∘ ( R ) ∑ f ( P ) C P g ( P ) , C P = C P T ≥ 0 , L 1 ( P ) C P = 0. (111) Only zeros of det L 1 \det L_1 det L 1 contribute. Such a mass has rank ∑ P rank C P ≤ d L \operatorname{rank}\sum_P\operatorname{rank}C_P\leq d_L rank ∑ P rank C P ≤ d L .
Proof. The decomposition of a matrix test H H H into its symmetric and skew-symmetric parts is ( H + H T ) / 2 + ( H − H T ) / 2 (H+H^{\mathsf{T}})/2+(H-H^{\mathsf{T}})/2 ( H + H T ) /2 + ( H − H T ) /2 . Since a symmetric functional annihilates the second summand, only the symmetric part contributes to its homogeneous freedom. This decomposition is preserved by congruence, as follows from ( L 1 T H L 1 ) T = L 1 T H T L 1 (L_1^{\mathsf{T}}HL_1)^{\mathsf{T}}=L_1^{\mathsf{T}}H^{\mathsf{T}}L_1 ( L 1 T H L 1 ) T = L 1 T H T L 1 . Moreover, if L 1 T H L 1 L_1^{\mathsf{T}}HL_1 L 1 T H L 1 is symmetric, multiplication by the rational inverses on both sides gives H = H T H=H^{\mathsf{T}} H = H T . Hence the symmetric part of the full homogeneous quotient is exactly the symmetric tests modulo the congruence images of symmetric tests. A functional on this quotient gives a unique symmetric homogeneous mass, and every such mass factors through it. This proves (7.1).
The local dimension is determined by a Smith factorization L 1 = U D V L_1=UDV L 1 = U D V at a smooth point, with D = diag ( t α i ) D=\operatorname{diag}(t^{\alpha_i}) D = diag ( t α i ) . The change H ↦ U T H U H\mapsto U^{\mathsf{T}}HU H ↦ U T H U is invertible on symmetric tests, and the outer factors V T , V V^{\mathsf{T}},V V T , V likewise induce an invertible change on the target. The quotient therefore has the same dimension as that of the diagonal action H ′ ↦ D H ′ D H'\mapsto DH'D H ′ ↦ D H ′ D . Since its independent entry ( i , j ) (i,j) ( i , j ) is multiplied by t α i + α j t^{\alpha_i+\alpha_j} t α i + α j , exactly that many Taylor coefficients survive in the quotient. The diagonal test ( i , i ) (i,i) ( i , i ) contributes 2 α i 2\alpha_i 2 α i jets. Each of the m − 1 m-1 m − 1 off-diagonal pairs containing i i i contributes one further α i \alpha_i α i . Thus each exponent occurs m + 1 m+1 m + 1 times, and
∑ i ≤ j ( α i + α j ) = ( m + 1 ) ∑ i α i . (112) \sum_{i\leq j}(\alpha_i+\alpha_j)=(m+1)\sum_i\alpha_i. \tag*{(112)} i ≤ j ∑ ( α i + α j ) = ( m + 1 ) i ∑ α i . (112)
Summing over all denominator zeros proves the dimension formula.
For a positive mass, the quotient of the real row space by the radical carries a positive inner product. The Cauchy–Schwarz argument above places A 1 × m L 1 A^{1\times m}L_1 A 1 × m L 1 in that radical, so the quotient has dimension at most d L d_L d L . The classification can therefore be obtained from operators on a finite-dimensional inner-product space. To define coordinate multiplication on this space, the radical must be invariant under scalar multiplication. Balance gives this invariance: if f f f pairs to zero with every g g g , then a f af a f does as well, since ν ( a f , g T ) = ν ( f , ( a g ) T ) \nu(af,g^{\mathsf T})=\nu(f,(ag)^{\mathsf T}) ν ( a f , g T ) = ν ( f , ( a g ) T ) . Thus each real coordinate defines an operator on the quotient, and the same identity makes that operator self-adjoint.
Commutativity of the coordinate algebra makes these self-adjoint operators commute. After the first operator is diagonalized, all the others preserve its eigenspaces; the second can then be diagonalized on each of them. Repetition through the finite list of coordinate generators gives a common orthonormal eigenbasis. The defining equations of the curve hold for the operators and therefore for their joint eigenvalues, which are real points P P P of the affine curve. On the joint eigenspace associated with P P P , a polynomial consequently acts by its value at P P P . If Π P \Pi_P Π P denotes the orthogonal projection onto that eigenspace at P P P , the expansion of a row as ∑ i f i e i T \sum_i f_i e_i^{\mathsf T} ∑ i f i e i T gives, by this evaluation property at P P P ,
Π P [ f ] = ∑ i f i ( P ) Π P [ e i T ] . \Pi_P[f]=\sum_i f_i(P)\Pi_P[e_i^{\mathsf T}]. Π P [ f ] = i ∑ f i ( P ) Π P [ e i T ] .
With ( C P ) i , j (C_P)_{i,j} ( C P ) i , j defined as the inner product of Π P [ e i T ] \Pi_P[e_i^{\mathsf T}] Π P [ e i T ] and Π P [ e j T ] \Pi_P[e_j^{\mathsf T}] Π P [ e j T ] , the contribution of each joint eigenspace is represented by a positive semidefinite Gram matrix. Taking inner products in the displayed expansion and summing over the mutually orthogonal eigenspaces then gives (7.2).
The homogeneity condition can now be imposed separately at each joint eigenvalue. Their distinct maximal ideals are comaximal, so the Chinese remainder theorem supplies a polynomial equal to one at any chosen point and zero at the others. Its insertion in the homogeneity identity isolates that point; varying the constant row factors yields L 1 ( P ) C P L 1 ( P ) T = 0 L_1(P)C_PL_1(P)^{\mathsf T}=0 L 1 ( P ) C P L 1 ( P ) T = 0 . The positive factorization C P = C P 1 / 2 C P 1 / 2 C_P=C_P^{1/2}C_P^{1/2} C P = C P 1/2 C P 1/2 then forces L 1 ( P ) C P 1 / 2 = 0 L_1(P)C_P^{1/2}=0 L 1 ( P ) C P 1/2 = 0 , hence L 1 ( P ) C P = 0 L_1(P)C_P=0 L 1 ( P ) C P = 0 . Only determinant zeros can therefore contribute. Conversely, positive matrices with this property give a positive homogeneous mass by direct substitution in (7.2). Polynomial evaluation separates the finitely many distinct support points, so their contributions have additive ranks. Their sum is bounded by the dimension of the quotient through which the mass factors: Cauchy–Schwarz places A 1 × m L 1 A^{1\times m}L_1 A 1 × m L 1 in its radical, and the quotient by this subspace has dimension d L d_L d L . This proves the stated rank bound. □ \square □
For L 1 = diag ( z , 1 ) L_1=\operatorname{diag}(z,1) L 1 = diag ( z , 1 ) , define c 1 ( f ) : = [ z ] f c_1(f):=[z]f c 1 ( f ) := [ z ] f . The symmetric mass ν : = E 1 , 1 c 1 \nu:=E_{1,1}c_1 ν := E 1 , 1 c 1 is homogeneous because its only transformed entry applies c 1 c_1 c 1 to a multiple of z 2 z^2 z 2 . On e 1 T , z e 1 T e_1^{\mathsf T},ze_1^{\mathsf T} e 1 T , z e 1 T its Gram is
( 0 1 1 0 ) \begin{pmatrix}
0 & 1\\
1 & 0
\end{pmatrix} ( 0 1 1 0 )
whose eigenvalues 1 1 1 and − 1 -1 − 1 show that the mass is indefinite. The sign of the homogeneous remainder in a full pairing depends on the particular solution chosen for its decomposition. For the positive form Θ ( f , g ) : = ∫ 1 2 f ( t ) g ( t ) T d t \Theta(f,g):=\int_1^2 f(t)g(t)^{\mathsf T}\,dt Θ ( f , g ) := ∫ 1 2 f ( t ) g ( t ) T d t , the choices L 0 = I 2 L_0=I_2 L 0 = I 2 , μ = L 1 Θ L 1 T \mu=L_1\Theta L_1^{\mathsf T} μ = L 1 Θ L 1 T give two solutions: both Θ \Theta Θ and Θ − ν \Theta-\nu Θ − ν satisfy L 1 μ L 1 T = μ L_1\mu L_1^{\mathsf T}=\mu L 1 μ L 1 T = μ . Choosing Θ − ν \Theta-\nu Θ − ν as the particular solution leaves the indefinite homogeneous term ν \nu ν in the positive target Θ = ( Θ − ν ) + ν \Theta=(\Theta-\nu)+\nu Θ = ( Θ − ν ) + ν . Thus positivity of the full pairing alone does not determine the sign of a chosen homogeneous remainder; the positive reduction below assumes positivity of its two terms separately.
For complex rows in real coordinates, the corresponding Hermitian form is ⟨ f , h ⟩ μ : = β ~ μ ( f , h ‾ T ) \langle f,h\rangle_\mu:=\widetilde{\beta}_\mu(f,\overline{h}^{\mathsf T}) ⟨ f , h ⟩ μ := β μ ( f , h T ) , where the bar conjugates coefficients. When S S S is real rational and the rational extension is Hermitian, separate positivity is likewise imposed on the rational space used in the congruence: if ⟨ f S , f S ⟩ μ > 0 \langle fS,fS\rangle_\mu>0 ⟨ f S , f S ⟩ μ > 0 for f ≠ 0 f\ne0 f = 0 and ν ( f , f ‾ T ) ≥ 0 \nu(f,\overline{f}^{\mathsf T})\ge0 ν ( f , f T ) ≥ 0 , then ⟨ f , f ⟩ μ ^ = ⟨ f S , f S ⟩ μ + ν ( f , f ‾ T ) > 0 \langle f,f\rangle_{\widehat{\mu}}=\langle fS,fS\rangle_\mu+\nu(f,\overline{f}^{\mathsf T})>0 ⟨ f , f ⟩ μ = ⟨ f S , f S ⟩ μ + ν ( f , f T ) > 0 .
Positive congruences: smaller matrices and the orthogonal polynomials
Separate positivity of the rational pairing and the homogeneous mass allows the congruence to be reduced by orthogonal projection. Since the same rational change acts on both arguments, the common part of the original and transformed spaces can be removed together from the two arguments. One positive matrix determines this projection; a second is the Gram matrix on the remaining polynomial rows. The latter matrix determines both the new orthogonal polynomials and the reproducing kernel. Its entries depend only on the original polynomials, the rational moments and the specified mass.
The construction is stated for real rows and symmetric bilinear forms; replacing transposes by adjoints gives the Hermitian version. Let X X X be the space of lower polynomial rows against which orthogonality is required, and let X + ⊃ X X^{+} \supset X X + ⊃ X contain the prescribed leading rows. The comparison space E E E is an original polynomial space of the same dimension N : = dim X = dim E N := \dim X = \dim E N := dim X = dim E ; the component bounds in E E E and X X X remain arbitrary. The transformed form is written as
⟨ f , g ⟩ μ ^ : = ⟨ f S , g S ⟩ μ ~ + ν ( f , g ) , S = L − 1 L 0 . \langle f,g\rangle_{\widehat{\mu}} := \langle fS,gS\rangle_{\widetilde{\mu}} + \nu(f,g), \qquad S = L^{-1}L_{0}. ⟨ f , g ⟩ μ := ⟨ f S , g S ⟩ μ + ν ( f , g ) , S = L − 1 L 0 .
Assume that the rational pairing is positive definite on E + X + S E+X^{+}S E + X + S , that S S S is invertible over the total quotient ring, and that ν \nu ν is positive semidefinite on X + X^{+} X + . The homogeneous condition is imposed whenever this is a solution of the bilateral functional equation. The projection argument itself applies to any positive additive form.
The common space and the contribution of the mass
The intersection of the two spaces is measured before and after imposing vanishing of the mass. These two requirements define
Z : = { f ∈ X : f S ∈ E } , s : = N − dim Z , Z 0 : = { f ∈ Z : ν ( f , f ) = 0 } , η : = rank ( ν ∣ Z × Z ) , F : = Z 0 S ⊂ E , t : = N − dim Z 0 . (113) \begin{aligned}
Z &:= \{f \in X : fS \in E\}, \qquad s := N - \dim Z,\\
Z_{0} &:= \{f \in Z : \nu(f,f) = 0\}, \qquad\eta:= \operatorname{rank}(\nu|_{Z \times Z}),\\
F &:= Z_{0}S \subset E, \qquad t := N - \dim Z_{0}. \tag*{(113)}
\end{aligned} Z Z 0 F := { f ∈ X : f S ∈ E } , s := N − dim Z , := { f ∈ Z : ν ( f , f ) = 0 } , η := rank ( ν ∣ Z × Z ) , := Z 0 S ⊂ E , t := N − dim Z 0 . (113)
Here Z Z Z consists of target rows whose rational images already belong to the original space. The number s s s counts the dimensions lost from X X X when that requirement is imposed. Within Z Z Z , the nullspace of the mass is Z 0 Z_{0} Z 0 , and η \eta η counts the additional directions removed by requiring the mass to vanish. Only the mass on this overlap enters η \eta η ; it can have smaller rank than the mass on all of X X X . Positivity implies that ν ( f , f ) = 0 \nu(f,f) = 0 ν ( f , f ) = 0 makes f f f pair to zero with every element of X + X^{+} X + , by Cauchy–Schwarz. Thus Z 0 Z_{0} Z 0 is a vector space, and the mass vanishes whenever one argument belongs to Z 0 Z_{0} Z 0 .
An orthonormal basis p 1 , … , p N p_{1},\ldots,p_{N} p 1 , … , p N of E E E for the original pairing gives the N × m N \times m N × m polynomial matrix p : = [ p i ] i ∈ { 1 , … , N } \boldsymbol{p} := [p_{i}]_{i \in\{1,\ldots,N\}} p := [ p i ] i ∈ { 1 , … , N } by stacking its rows. Such a basis is obtained from the original orthogonal polynomials and their norms, and a row in E E E then has the coordinate representation c T p c^{\mathsf{T}}\boldsymbol{p} c T p . In these coordinates the common subspace is the kernel of a full-row-rank t × N t \times N t × N matrix Λ \Lambda Λ , characterized by c T p ∈ F \boldsymbol{c}^{\mathsf{T}}\boldsymbol{p} \in F c T p ∈ F if and only if Λ c = 0 \Lambda\boldsymbol{c} = 0 Λ c = 0 .
The orthogonality of the chosen coordinates expresses the projection onto this kernel through the first reduced matrix:
A F : = Λ Λ T , H F : = I N − Λ T A F − 1 Λ . A_{F} := \Lambda\Lambda^{\mathsf{T}}, \qquad H_{F} := I_{N} - \Lambda^{\mathsf{T}}A_{F}^{-1}\Lambda. A F := Λ Λ T , H F := I N − Λ T A F − 1 Λ.
For any rational row u u u in the spaces under consideration, set
c ( u ) : = [ ⟨ u , p i ⟩ μ ~ ] i ∈ { 1 , … , N } , Π F u : = c ( u ) T H F p . c(u) := [\langle u,p_{i}\rangle_{\widetilde{\mu}}]_{i \in\{1,\ldots,N\}}, \qquad\Pi_{F}u := c(u)^{\mathsf{T}}H_{F}\boldsymbol{p}. c ( u ) := [⟨ u , p i ⟩ μ ] i ∈ { 1 , … , N } , Π F u := c ( u ) T H F p .
The vector c ( u ) c(u) c ( u ) is an N × 1 N \times1 N × 1 column. It gives the coordinates of the original projection onto E E E , since ⟨ u − c ( u ) T p , p j ⟩ = 0 \langle u-c(u)^{\mathsf{T}}\boldsymbol{p},p_{j}\rangle= 0 ⟨ u − c ( u ) T p , p j ⟩ = 0 for every j j j . The symmetric N × N N \times N N × N matrix H F H_{F} H F then projects those coordinates onto ker Λ \ker\Lambda ker Λ , which represents F F F . Thus Π F u \Pi_{F}u Π F u is the original orthogonal projection onto F F F .
For a complement f 1 , … , f t ∈ X f_{1},\ldots,f_{t} \in X f 1 , … , f t ∈ X to a basis of Z 0 Z_{0} Z 0 in X X X , subtraction of this projection gives the rows
ρ j : = f j S − Π F ( f j S ) , w j : = ρ j S − 1 = f j − Π F ( f j S ) S − 1 , j ∈ { 1 , … , t } . (114) \rho_{j} := f_{j}S - \Pi_{F}(f_{j}S), \qquad w_{j} := \rho_{j}S^{-1} = f_{j} - \Pi_{F}(f_{j}S)S^{-1}, \qquad j \in\{1,\ldots,t\}. \tag*{(114)} ρ j := f j S − Π F ( f j S ) , w j := ρ j S − 1 = f j − Π F ( f j S ) S − 1 , j ∈ { 1 , … , t } . (114)
The row w j w_{j} w j remains polynomial and belongs to X X X , because Π F ( f j S ) ∈ F = Z 0 S \Pi_{F}(f_{j}S) \in F = Z_{0}S Π F ( f j S ) ∈ F = Z 0 S and its product with S − 1 S^{-1} S − 1 therefore lies in Z 0 Z_{0} Z 0 . With the notation w : = [ w j ] j ∈ { 1 , … , t } \boldsymbol{w} := [w_{j}]_{j \in\{1,\ldots,t\}} w := [ w j ] j ∈ { 1 , … , t } and ρ : = [ ρ j ] j ∈ { 1 , … , t } \boldsymbol{\rho} := [\rho_{j}]_{j \in\{1,\ldots,t\}} ρ := [ ρ j ] j ∈ { 1 , … , t } , both w w w and ρ \rho ρ have t t t rows and m m m columns. They represent the same complementary directions before and after the rational change S S S , with w S = ρ wS = \rho w S = ρ . Their Gram matrix for the transformed pairing is the second reduced matrix,
T ν : = [ ⟨ ρ i , ρ j ⟩ μ ^ + ν ( f i , f j ) ] i , j ∈ { 1 , … , t } . T_{\nu} := [\langle\rho_{i},\rho_{j}\rangle_{\widehat{\mu}} + \nu(f_{i},f_{j})]_{i,j \in\{1,\ldots,t\}}. T ν := [⟨ ρ i , ρ j ⟩ μ + ν ( f i , f j ) ] i , j ∈ { 1 , … , t } .
Both reduced matrices are thus determined before the perturbed orthogonal polynomials are constructed. The rational pairings appearing in their entries are precisely the second-kind data of Subsubsection 3.2.3; for measure pairings, these are Cauchy data in one variable and Fantappiè data in several variables.
Theorem 7.2 (Positive reduction and normalized polynomials). Under the preceding assumptions, both A F A_F A F and T ν T_\nu T ν are positive definite, and their common order is
For a prescribed row v ∈ X + v\in X^+ v ∈ X + , put
v 0 : = v − Π F ( v S ) S − 1 , b ( v ) : = [ ⟨ v S , ρ j ⟩ μ ~ + ν ( v , f j ) ] j ∈ { 1 , … , t } . v_0:=v-\Pi_F(vS)S^{-1},\qquad b(v):=\left[\langle vS,\rho_j\rangle_{\widetilde{\mu}}+\nu(v,f_j)\right]_{j\in\{1,\ldots,t\}}. v 0 := v − Π F ( v S ) S − 1 , b ( v ) := [ ⟨ v S , ρ j ⟩ μ + ν ( v , f j ) ] j ∈ { 1 , … , t } . Then the unique row with the class of v v v modulo X X X and orthogonal to X X X for μ ~ \widetilde{\mu} μ is
p ^ = v 0 − b ( v ) T ν − 1 w . (115) \widehat{p}=v_0-b(v)T_\nu^{-1}w. \tag*{(115)} p = v 0 − b ( v ) T ν − 1 w . (115) In particular, every scalar component has the bordered formula
p ^ a = det [ T ν w : , a b ( v ) ( v 0 ) a ] det T ν , a ∈ { 1 , … , m } . (116) \widehat{p}_a=\frac{\det\begin{bmatrix}T_\nu& w_{:,a}\\ b(v) & (v_0)_a\end{bmatrix}}{\det T_\nu},\qquad a\in\{1,\ldots,m\}. \tag*{(116)} p a = det T ν det [ T ν b ( v ) w : , a ( v 0 ) a ] , a ∈ { 1 , … , m } . (116) Thus the denominator has order t t t and is positive, and each bordered numerator has order t + 1 t+1 t + 1 . The preliminary projection uses the additional order-t t t matrix A F A_F A F .
Let K F ( P , Q ) : = p ( P ) T H F p ( Q ) K_F(P,Q):=\mathbf{p}(P)^{\mathsf T}H_F\mathbf{p}(Q) K F ( P , Q ) := p ( P ) T H F p ( Q ) . The kernel on X X X for the transformed pairing is
K ^ X ( P , Q ) = S ( P ) − T [ K F ( P , Q ) + ρ ( P ) T T ν − 1 ρ ( Q ) ] S ( Q ) − 1 . (117) \widehat{K}_X(P,Q)=S(P)^{-\mathsf T}\left[K_F(P,Q)+\rho(P)^{\mathsf T}T_\nu^{-1}\rho(Q)\right]S(Q)^{-1}. \tag*{(117)} K X ( P , Q ) = S ( P ) − T [ K F ( P , Q ) + ρ ( P ) T T ν − 1 ρ ( Q ) ] S ( Q ) − 1 . (117) The apparent rational singularities cancel: this is a polynomial kernel, interpreted by the polynomial rows in (7.4) wherever an evaluated inverse is undefined.
Proof. Positivity identifies the kernel of the restricted mass on Z Z Z with Z 0 Z_0 Z 0 , so its rank is η = dim Z − dim Z 0 \eta=\dim Z-\dim Z_0 η = dim Z − dim Z 0 . The dimension definitions in (7.3) then give t = s + η t=s+\eta t = s + η . Since the change f ↦ f S f\mapsto fS f ↦ f S is injective, the common image has dimension dim F = N − t \dim F=N-t dim F = N − t .
The constraint matrix has full row rank, which determines the positivity of the first reduced matrix. For every nonzero column a ∈ R t a\in\mathbb{R}^t a ∈ R t , the equality a T A F a = ∥ Λ T a ∥ 2 > 0 a^{\mathsf T}A_Fa=\lVert\Lambda^{\mathsf T}a\rVert^2>0 a T A F a = ∥ Λ T a ∥ 2 > 0 follows from injectivity of Λ T \Lambda^{\mathsf T} Λ T . Hence A F A_F A F is positive definite, and the definition of H F H_F H F gives the identities
H F T = H F , Λ H F = 0 , H F c = c ( c ∈ ker Λ ) H_F^{\mathsf T}=H_F,\qquad\Lambda H_F=0,\qquad H_Fc=c\quad(c\in\ker\Lambda) H F T = H F , Λ H F = 0 , H F c = c ( c ∈ ker Λ )
which follow by substituting A F = Λ Λ T A_F=\Lambda\Lambda^{\mathsf T} A F = Λ Λ T . These identities show that H F H_F H F is the Euclidean orthogonal projection onto ker Λ \ker\Lambda ker Λ and in particular H F 2 = H F H_F^2=H_F H F 2 = H F . For v = d T p ∈ F v=d^{\mathsf T}\mathbf{p}\in F v = d T p ∈ F one then has ⟨ u − Π F u , v ⟩ = c ( u ) T ( I − H F ) d = 0 \langle u-\Pi_Fu,v\rangle=c(u)^{\mathsf T}(I-H_F)d=0 ⟨ u − Π F u , v ⟩ = c ( u ) T ( I − H F ) d = 0 . This verifies the projection formula for every rational row u u u in the stated finite space.
The form induced by μ ~ \widetilde{\mu} μ on X + X^+ X + is positive definite: the first term is positive on every nonzero row, and the second term is nonnegative. On Z 0 Z_0 Z 0 , the map z ↦ z S z\mapsto zS z ↦ z S preserves the inner product, since the mass vanishes there. For z ∈ Z 0 z\in Z_0 z ∈ Z 0 ,
⟨ w j , z ⟩ μ ^ = ⟨ ρ j , z S ⟩ μ ~ + ν ( w j , z ) = 0. \langle w_j,z\rangle_{\widehat{\mu}}=\langle\rho_j,zS\rangle_{\widetilde{\mu}}+\nu(w_j,z)=0. ⟨ w j , z ⟩ μ = ⟨ ρ j , z S ⟩ μ + ν ( w j , z ) = 0.
The two terms vanish for separate reasons: the definition of Π F \Pi_F Π F gives orthogonality of the rational term, and the mass radical property gives vanishing of the mass term. Moreover, w j w_j w j and f j f_j f j have the same classes modulo Z 0 Z_0 Z 0 , so they still span the complementary directions. It follows that X X X is the orthogonal direct sum of Z 0 Z_0 Z 0 and the span of w 1 , … , w t w_1,\ldots,w_t w 1 , … , w t . The latter rows are independent: a relation ∑ j a j w j = 0 \sum_j a_jw_j=0 ∑ j a j w j = 0 would give ∑ j a j f j ∈ Z 0 \sum_j a_jf_j\in Z_0 ∑ j a j f j ∈ Z 0 , and the chosen complement forces a j = 0 a_j=0 a j = 0 for every j j j . Their Gram matrix is determined without any further projection: subtracting elements of Z 0 Z_0 Z 0 changes no mass pairing, while w j S = ρ j w_jS=\rho_j w j S = ρ j , giving
⟨ w i , w j ⟩ μ ^ = ⟨ ρ i , ρ j ⟩ μ ~ + ν ( f i , f j ) = ( T ν ) i , j . \langle w_i,w_j\rangle_{\widehat{\mu}}=\langle\rho_i,\rho_j\rangle_{\widetilde{\mu}}+\nu(f_i,f_j)=(T_\nu)_{i,j}. ⟨ w i , w j ⟩ μ = ⟨ ρ i , ρ j ⟩ μ + ν ( f i , f j ) = ( T ν ) i , j .
Independence and positivity now imply positivity of the second reduced matrix. For a nonzero coefficient vector a a a , the row ∑ j a j w j \sum_j a_jw_j ∑ j a j w j is nonzero, so its squared norm for the transformed pairing is a T T ν a > 0 a^{\mathsf T}T_\nu a>0 a T T ν a > 0 . This proves that T ν T_\nu T ν is positive definite.
The orthogonal decomposition of the lower space also determines the correction of a prescribed leading row. The subtraction defining v 0 v_0 v 0 belongs to Z 0 Z_0 Z 0 and removes the pairings with that summand. Its image Π F ( v S ) ∈ F \Pi_F(v_S) \in F Π F ( v S ) ∈ F is orthogonal to ρ j \rho_j ρ j , while its pullback belongs to Z 0 Z_0 Z 0 and has zero mass pairing. The pairings that remain after this first subtraction are therefore
⟨ v 0 , w j ⟩ μ = ⟨ v S , ρ j ⟩ μ + ν ( v , f j ) = b ( v ) j . \langle v_0,w_j\rangle_{\mu}=\langle v_S,\rho_j\rangle_{\mu}+\nu(v,f_j)=b(v)_j. ⟨ v 0 , w j ⟩ μ = ⟨ v S , ρ j ⟩ μ + ν ( v , f j ) = b ( v ) j .
For a coefficient row a T a^{\mathsf T} a T , the polynomial v 0 − a T w v_0-a^{\mathsf T}w v 0 − a T w is orthogonal to all w j w_j w j exactly when a T T ν = b ( v ) T a^{\mathsf T}T_\nu=b(v)^{\mathsf T} a T T ν = b ( v ) T . Its unique solution is a T = b ( v ) T T ν − 1 a^{\mathsf T}=b(v)^{\mathsf T}T_\nu^{-1} a T = b ( v ) T T ν − 1 , proving (7.5). Both corrections belong to X X X , so the prescribed leading class is preserved. Two rows with this class and orthogonal to X X X differ by an element of X X X with zero norm; positivity makes their difference zero. The component formula can then be written as a bordered determinant: left multiplication of the matrix in (7.6) by [ I t 0 − b ( v ) T T ν − 1 1 ] \left[\begin{smallmatrix}I_t&0\\-b(v)^{\mathsf T}T_\nu^{-1}&1\end{smallmatrix}\right] [ I t − b ( v ) T T ν − 1 0 1 ] preserves its determinant, since the multiplier has determinant one. Its lower-left block vanishes, and its lower-right entry is ( v 0 ) ν a − b ( v ) T T ν − 1 w : , a (v_0)_\nu a-b(v)^{\mathsf T}T_\nu^{-1}\mathsf{w}_{:,a} ( v 0 ) ν a − b ( v ) T T ν − 1 w : , a . Taking the determinant of this block triangular matrix gives the bordered formula, with the stated positive denominator and sizes.
The same orthogonal decomposition yields the kernel. An orthonormal basis of F F F , pulled back by S − 1 S^{-1} S − 1 to Z 0 Z_0 Z 0 , combines with the w j w_j w j to give a basis with Gram matrix diag ( I N − t , T ν ) \operatorname{diag}(I_{N-t},T_\nu) diag ( I N − t , T ν ) . The contribution of its first summand is S ( P ) − T K F ( P , Q ) S ( Q ) − 1 S(P)^{-\mathsf T}K_F(P,Q)S(Q)^{-1} S ( P ) − T K F ( P , Q ) S ( Q ) − 1 , while inversion of the second Gram block gives the contribution w ( P ) T T ν − 1 w ( Q ) w(P)^{\mathsf T}T_\nu^{-1}w(Q) w ( P ) T T ν − 1 w ( Q ) . Substituting w = ρ S − 1 w=\rho S^{-1} w = ρ S − 1 and factoring the two outer inverses gives (7.7). Every basis row before and after this orthogonal decomposition belongs to X X X , which proves the polynomial cancellation assertion. Empty matrices have determinant one; the formulas also include t = 0 t=0 t = 0 . □
Applying (7.5) to a basis of X + / X X^+/X X + / X gives the whole normalized block, using the same T ν T_\nu T ν for every leading row. A nonzero constant linear combination of these rows has a nonzero class in X + / X X^+/X X + / X , hence is a nonzero polynomial row. Its positive squared norm proves that the block norm matrix is positive definite. Iterating along nested component bounds constructs the complete orthogonal family.
The decomposition used in this proof depends on positivity of the restricted form on the common subspace. Nondegeneracy of the full symmetric form alone does not ensure it: for the form with matrix diag ( 1 , − 1 ) \operatorname{diag}(1,-1) diag ( 1 , − 1 ) , the restriction to the line spanned by [ 1 1 ] [1\ 1] [ 1 1 ] is zero, so that line cannot be a summand in an orthogonal direct sum. Symmetric indefinite data are still governed by the determinant of the general reduced system, but a reduction by the common space requires a separate nondegeneracy assumption on its restriction.
Computing the order from polynomial coefficients.
The order of the projection construction can be determined from the polynomial spaces and factors, without first finding the new orthogonal polynomials. With δ : = det L 1 \delta:=\det L_1 δ := det L 1 and B : = adj ( L 1 ) L 0 B:=\operatorname{adj}(L_1)L_0 B := adj ( L 1 ) L 0 , in the displayed order, clearing the common denominator gives
s = rank [ δ E X B ] − N . (118) s=\operatorname{rank}[\delta E\ \ XB]-N. \tag*{(118)} s = rank [ δ E X B ] − N . (118)
As in (3.23), the displayed lists are coefficient columns in a polynomial basis on the support. Their rank computes the dimension of the sum of spaces because multiplication by the regular scalar δ \delta δ is injective on rational rows and maps E + X S E+XS E + X S onto δ E + X B \delta E+XB δ E + X B . The coefficient matrix consequently has rank dim ( E + X S ) \dim(E+XS) dim ( E + X S ) . The change f ↦ f S f\mapsto fS f ↦ f S identifies Z Z Z with E ∩ X S E\cap XS E ∩ X S ; since both E E E and X S XS X S have dimension N N N , the dimension formula for a sum of subspaces gives
dim ( E + X S ) = 2 N − dim ( E ∩ X S ) = 2 N − dim Z \dim(E+XS)=2N-\dim(E\cap XS)=2N-\dim Z dim ( E + X S ) = 2 N − dim ( E ∩ X S ) = 2 N − dim Z
and hence (7.8). The same coefficient calculation recovers the common space itself: the solutions of f B = δ e fB=\delta e f B = δ e with f ∈ X f\in X f ∈ X , e ∈ E e\in E e ∈ E give a basis of Z Z Z . The matrix of the mass in that basis determines its rank η \eta η and nullspace Z 0 Z_0 Z 0 , and consequently determines both F F F and Λ \Lambda Λ . Since this calculation uses the actual coefficient spaces on the support, it also applies to unequal component bounds, singular supports and noncommuting matrix factors.
Without a mass, t = s t=s t = s . With a positive mass of restricted rank h h h on X X X , one has t = s + η ≤ s + h t=s+\eta\le s+h t = s + η ≤ s + h , and always t ≤ N t\le N t ≤ N . For S = I S=I S = I and E = X E=X E = X , this recovers the ordinary additive update: s = 0 s=0 s = 0 , Z = X Z=X Z = X and t = h t=h t = h . A cancellation between rational factors therefore does not discard a remaining mass. The two order-t t t matrices above describe the orthogonal projection construction. As in Subsection 3.3, their size is determined by the particular spaces eliminated in this representation.
Exact degree counts under a coprimality condition.
For equal filtered spaces E = X = ( F n − 1 A ) 1 × m E = X = (F_{n-1}A)^{1 \times m} E = X = ( F n − 1 A ) 1 × m , the coefficient count reduces to a difference of Hilbert functions under coprimality and degree assumptions. Here H A ( j ) = dim F j A H_A(j) = \dim F_j A H A ( j ) = dim F j A is as in Subsection 3.3. The factors L 0 , L 1 L_0, L_1 L 0 , L 1 are assumed to have filtered degrees a 0 , a 1 a_0, a_1 a 0 , a 1 and to satisfy
L 0 U 0 + L 1 U 1 = I m (119) L_0U_0 + L_1U_1 = I_m \tag*{(119)} L 0 U 0 + L 1 U 1 = I m (119)
for polynomial matrices U 0 , U 1 U_0, U_1 U 0 , U 1 . This Bézout identity is the coprimality condition used here and is compatible with noncommuting L 0 , L 1 L_0, L_1 L 0 , L 1 . The additional degree assumption is
deg ( h L i ) = deg h + a i for every nonzero polynomial row h , i ∈ { 0 , 1 } . (120) \deg(hL_i) = \deg h + a_i \qquad\text{for every nonzero polynomial row }h,\qquad i \in\{0,1\}. \tag*{(120)} deg ( h L i ) = deg h + a i for every nonzero polynomial row h , i ∈ { 0 , 1 } . (120)
On the line, invertible leading coefficient matrices ensure this property. On an algebraic support, multiplication can interact with the defining relations, so the property is an independent assumption on the inherited filtration rather than a consequence of matrix size.
Corollary 7.3 (Orders for positive coprime congruences). Assume (119) –(120) and let M : = max { a 0 , a 1 } M := \max\{a_0,a_1\} M := max { a 0 , a 1 } . For every positive homogeneous mass the order in Theorem 7.2 is exactly
t n = m [ H A ( n − 1 ) − H A ( n − M − 1 ) ] . (121) t_n = m\left[H_A(n-1)-H_A(n-M-1)\right]. \tag*{(121)} t n = m [ H A ( n − 1 ) − H A ( n − M − 1 ) ] . (121) In particular, the mass does not increase this order. The same formula, with m = 1 m=1 m = 1 , holds for coprime scalar polynomials P , Q P,Q P , Q in A = R [ x 1 , … , x D ] A = \mathbb{R}[x_1,\ldots,x_D] A = R [ x 1 , … , x D ] , with S = P / Q S=P/Q S = P / Q , even when a polynomial Bézout identity is unavailable.
Proof. An element f ∈ Z f \in Z f ∈ Z determines the rational row h = f L 1 − 1 h = fL_1^{-1} h = f L 1 − 1 , whose products h L 1 = f hL_1=f h L 1 = f and h L 0 = f S hL_0=fS h L 0 = f S are both polynomial. The Bézout identity then gives h = ( h L 0 ) U 0 + ( h L 1 ) U 1 h=(hL_0)U_0+(hL_1)U_1 h = ( h L 0 ) U 0 + ( h L 1 ) U 1 , proving that h h h is itself polynomial. The degree assumptions can therefore be applied to it: for nonzero h h h , the bounds on h L 1 hL_1 h L 1 and h L 0 hL_0 h L 0 are deg h + a 1 ≤ n − 1 \deg h+a_1 \le n-1 deg h + a 1 ≤ n − 1 and deg h + a 0 ≤ n − 1 \deg h+a_0 \le n-1 deg h + a 0 ≤ n − 1 . Their conjunction is deg h ≤ n − M − 1 \deg h \le n-M-1 deg h ≤ n − M − 1 . Including the zero row, the two requirements are consequently equivalent to h ∈ ( F n − M − 1 A ) 1 × m h \in(F_{n-M-1}A)^{1 \times m} h ∈ ( F n − M − 1 A ) 1 × m , where a negative filtered bound means the zero space. Conversely, every row in this space gives f = h L 1 ∈ Z f=hL_1 \in Z f = h L 1 ∈ Z . Injectivity of multiplication by L 1 L_1 L 1 therefore yields
Z = ( F n − M − 1 A ) 1 × m L 1 , dim Z = m H A ( n − M − 1 ) . Z = (F_{n-M-1}A)^{1 \times m}L_1,\qquad\dim Z = mH_A(n-M-1). Z = ( F n − M − 1 A ) 1 × m L 1 , dim Z = m H A ( n − M − 1 ) .
A positive homogeneous mass annihilates the entire denominator image in either argument, as proved at the beginning of this section. Hence ν ∣ Z × Z = 0 \nu|_{Z \times Z}=0 ν ∣ Z × Z = 0 , so η = 0 \eta=0 η = 0 and t = s t=s t = s . Subtracting dimensions proves (121) .
In the scalar case, the factorization of Q Q Q into irreducibles replaces the Bézout argument. Each irreducible factor occurring in Q Q Q has exponent zero in the coprime polynomial P P P . Consequently Q ∣ f P Q \mid fP Q ∣ f P forces each such power to divide f f f , proving Q ∣ f Q \mid f Q ∣ f . With f = Q h f=Qh f = Q h , the remaining argument again concerns the degrees of polynomial products. The nonzero highest homogeneous parts of two polynomials have nonzero product in R [ x 1 , … , x D ] \mathbb{R}[x_1,\ldots,x_D] R [ x 1 , … , x D ] ; hence total degrees add. The bounds on Q h Qh Q h and P h Ph P h give the same space of h h h as above. Finally positivity and homogeneity annihilate Q A QA Q A , so the mass still vanishes on Z Z Z . This proves the scalar claim without using a Bézout identity. ▫
On the line the order is m min { n , M } m\min\{n,M\} m min { n , M } , hence m M mM m M for n ≥ M n \ge M n ≥ M . For two independent variables it is
t n = m M ( 2 n − M + 1 ) 2 , n ≥ M . t_n = \frac{mM(2n-M+1)}{2},\qquad n \ge M. t n = 2 m M ( 2 n − M + 1 ) , n ≥ M .
In D D D independent variables the exact expression is
t n = m [ ( n + D − 1 D ) − ( n − M + D − 1 D ) ] , t_n = m\left[\binom{n+D-1}{D}-\binom{n-M+D-1}{D}\right], t n = m [ ( D n + D − 1 ) − ( D n − M + D − 1 ) ] ,
where the second term is zero when n − M − 1 < 0 n-M-1<0 n − M − 1 < 0 . On a curve of degree d d d , once both arguments of H A H_A H A are in its linear range, the order is exactly m d M mdM m d M . On a support of dimension r ≥ 2 r \ge2 r ≥ 2 and degree d d d ,
t n = m d M ( r − 1 ) ! n r − 1 + O ( n r − 2 ) ( n → ∞ ) . t_n = \frac{mdM}{(r-1)!}n^{r-1}+O(n^{r-2})\qquad(n\to\infty). t n = ( r − 1 )! m d M n r − 1 + O ( n r − 2 ) ( n → ∞ ) .
The first three counts follow by substituting H A ( j ) = j + 1 H_A(j)=j+1 H A ( j ) = j + 1 , H A ( j ) = ( j + 1 ) ( j + 2 ) / 2 H_A(j)=(j+1)(j+2)/2 H A ( j ) = ( j + 1 ) ( j + 2 ) /2 , and H A ( j ) = ( j + D D ) H_A(j)=\binom{j+D}{D} H A ( j ) = ( D j + D ) , respectively. For a curve in its linear Hilbert range, H A ( j ) = d j + c H_A(j)=dj+c H A ( j ) = d j + c , whose difference across M M M consecutive degrees is d M dM d M . In dimension r ≥ 2 r\ge2 r ≥ 2 , the leading Hilbert term is d j r / r ! dj^r/r! d j r / r ! ; subtracting its translate by M M M gives d M n r − 1 / ( r − 1 ) ! dMn^{r-1}/(r-1)! d M n r − 1 / ( r − 1 )! , and the remaining terms have order at most n r − 2 n^{r-2} n r − 2 . Thus all displayed counts retain the assumptions of the corollary. For arbitrary matrix factors or component bounds, (7.8) and t = s + η t=s+\eta t = s + η give the exact order instead.
A noncommutative example with a positive derivative mass at a node
The degree and coprimality assumptions above admit an explicit realization on a nodal curve, where a derivative mass can also be included. The coordinate algebra and parametrization are
A = R [ x , y ] / ( y 2 − x 2 ( x + 1 ) ) , P ( t ) = ( t 2 − 1 , t ( t 2 − 1 ) ) , O = ( 0 , 0 ) . A=\mathbb{R}[x,y]/(y^2-x^2(x+1)),\qquad P(t)=(t^2-1,t(t^2-1)),\qquad O=(0,0). A = R [ x , y ] / ( y 2 − x 2 ( x + 1 )) , P ( t ) = ( t 2 − 1 , t ( t 2 − 1 )) , O = ( 0 , 0 ) .
Let
C = ( 0 1 0 1 / 4 ) , D = ( 2 0 0 3 ) , L 1 = y I 2 − C , L 0 = y I 2 − D . C=\begin{pmatrix}0&1\\0&1/4\end{pmatrix},\qquad D=\begin{pmatrix}2&0\\0&3\end{pmatrix},\qquad L_1=yI_2-C,\qquad L_0=yI_2-D. C = ( 0 0 1 1/4 ) , D = ( 2 0 0 3 ) , L 1 = y I 2 − C , L 0 = y I 2 − D .
Direct multiplication gives [ C , D ] = E 1 , 2 [C,D]=E_{1,2} [ C , D ] = E 1 , 2 and [ L 0 , L 1 ] = [ D , C ] = − E 1 , 2 [L_0,L_1]=[D,C]=-E_{1,2} [ L 0 , L 1 ] = [ D , C ] = − E 1 , 2 . Thus the two polynomial factors do not commute. Put δ : = y ( y − 1 / 4 ) \delta:=y(y-1/4) δ := y ( y − 1/4 ) and define the original balanced pairing by
⟨ f , g ⟩ μ : = ∫ 2 3 δ ( P ( t ) ) 2 f ( P ( t ) ) g ( P ( t ) ) T d t . \langle f,g\rangle_\mu:=\int_2^3\delta(P(t))^2f(P(t))g(P(t))^{\mathsf{T}}\,dt. ⟨ f , g ⟩ μ := ∫ 2 3 δ ( P ( t ) ) 2 f ( P ( t )) g ( P ( t ) ) T d t .
Along the integration arc, the polynomial y y y ranges from 6 to 24, keeping the denominator zeros away from the integration support. The parametrization identifies the coordinate algebra with a subalgebra of R [ t ] \mathbb{R}[t] R [ t ] , so a nonzero polynomial row cannot vanish on the whole arc. The positive density therefore gives a positive definite pairing. The same reasoning, after clearing denominators, proves positivity of its integral extension on all the rational spaces used in this example.
For α , β ≥ 0 \alpha,\beta\ge0 α , β ≥ 0 , choose the homogeneous mass
ν α , β ( f , g ) : = α f 1 ( O ) g 1 ( O ) + β ∂ x f 1 ( O ) ∂ x g 1 ( O ) . \nu_{\alpha,\beta}(f,g):=\alpha f_1(O)g_1(O)+\beta\partial_xf_1(O)\partial_xg_1(O). ν α , β ( f , g ) := α f 1 ( O ) g 1 ( O ) + β ∂ x f 1 ( O ) ∂ x g 1 ( O ) .
Both tests descend to the coordinate algebra. The defining polynomial and its x x x derivative vanish at O O O , so changing a representative by a multiple of the relation changes neither its value nor the derivative used in the mass. Nonnegative coefficients then make this a positive form built from evaluation and a first derivative at the singular point.
Proposition 7.4 (Positive orthogonal families with component degree bounds on the nodal curve). For the preceding data,
⟨ f , g ⟩ μ ~ = ⟨ f S , g S ⟩ μ ~ + ν α , β ( f , g ) , S = L 1 − 1 L 0 , \langle f,g\rangle_{\widetilde{\mu}}=\langle fS,gS\rangle_{\widetilde{\mu}}+\nu_{\alpha,\beta}(f,g),\qquad S=L_1^{-1}L_0, ⟨ f , g ⟩ μ = ⟨ f S , g S ⟩ μ + ν α , β ( f , g ) , S = L 1 − 1 L 0 , is a positive definite solution of the bilateral congruence equation. It has uniquely normalized orthogonal blocks for every nested choice of component bounds.
For equal bounds E = X = ( F n − 1 A ) 1 × 2 E=X=(F_{n-1}A)^{1\times2} E = X = ( F n − 1 A ) 1 × 2 , the two matrices in Theorem 7.2 have orders
t 1 = 2 , t 2 = 4 , t n = 6 ( n ≥ 3 ) . (122) t_1=2,\qquad t_2=4,\qquad t_n=6\quad(n\ge3). \tag*{(122)} t 1 = 2 , t 2 = 4 , t n = 6 ( n ≥ 3 ) . (122)
The orders are independent of α , β ≥ 0 \alpha,\beta\ge0 α , β ≥ 0 . For arbitrary positive bounds u , v u,v u , v on the two components, put E = X = F u − 1 A ⊕ F v − 1 A E=X=F_{u-1}A\oplus F_{v-1}A E = X = F u − 1 A ⊕ F v − 1 A . The exact order is
t u , v = H A ( u − 1 ) + H A ( v − 1 ) − H A ( min { u − 2 , v − 1 } ) − H A ( v − 2 ) , (123) t_{u,v}=H_A(u-1)+H_A(v-1)-H_A(\min\{u-2,v-1\})-H_A(v-2), \tag*{(123)} t u , v = H A ( u − 1 ) + H A ( v − 1 ) − H A ( min { u − 2 , v − 1 }) − H A ( v − 2 ) , (123)
where H A ( 0 ) = 1 H_A(0)=1 H A ( 0 ) = 1 , H A ( j ) = 3 j H_A(j)=3j H A ( j ) = 3 j for j ≥ 1 j\ge1 j ≥ 1 , and H A ( j ) = 0 H_A(j)=0 H A ( j ) = 0 for j < 0 j<0 j < 0 . In particular, the component degree bounds ( u , v ) = ( n + 3 , n ) (u,v)=(n+3,n) ( u , v ) = ( n + 3 , n ) give order 12 for n ≥ 3 n\ge3 n ≥ 3 .
Proof. Homogeneity follows from the first component of L 1 L_1 L 1 , namely y f 1 yf_1 y f 1 . Its value at O O O is zero, and ∂ x ( y f 1 ) ( O ) = y ( O ) ∂ x f 1 ( O ) = 0 \partial_x(yf_1)(O)=y(O)\partial_xf_1(O)=0 ∂ x ( y f 1 ) ( O ) = y ( O ) ∂ x f 1 ( O ) = 0 , since the ambient coordinates x , y x,y x , y are differentiated independently. Both calculations descend to the coordinate algebra because the defining relation and its derivative vanish at O O O . The mass therefore annihilates the denominator image in either argument. Together with L 1 S = L 0 L_1S=L_0 L 1 S = L 0 , which gives the functional equation for the rational part, this proves the congruence equation for the sum. Its rational term is positive definite and its mass is nonnegative, so the full form is positive definite. Existence and uniqueness of the normalized blocks follow at every finite level.
The matrix C − D C-D C − D is invertible, and
L 0 ( C − D ) − 1 − L 1 ( C − D ) − 1 = I 2 . L_0(C-D)^{-1}-L_1(C-D)^{-1}=I_2. L 0 ( C − D ) − 1 − L 1 ( C − D ) − 1 = I 2 .
This verifies the Bézout hypothesis for the noncommuting factors. The remaining hypothesis concerns degree, which is determined by normal forms for the coordinate algebra. Writing the relation as x 3 = y 2 − x 2 x^3=y^2-x^2 x 3 = y 2 − x 2 , polynomial division by the monic cubic in x x x gives the unique representative a 0 ( y ) + x a 1 ( y ) + x 2 a 2 ( y ) a_0(y)+xa_1(y)+x^2a_2(y) a 0 ( y ) + x a 1 ( y ) + x 2 a 2 ( y ) . Each reduction replaces a degree-three factor by terms of degree two, so the filtered basis consists precisely of x i y j x^iy^j x i y j with i ∈ { 0 , 1 , 2 } i\in\{0,1,2\} i ∈ { 0 , 1 , 2 } , j ∈ N j\in\mathbb{N} j ∈ N and i + j ≤ k i+j\leq k i + j ≤ k . For k ≥ 2 k\geq2 k ≥ 2 their count is ( k + 1 ) + k + ( k − 1 ) = 3 k (k+1)+k+(k-1)=3k ( k + 1 ) + k + ( k − 1 ) = 3 k ; for k = 0 , 1 k=0,1 k = 0 , 1 it is 1 , 3 1,3 1 , 3 . Multiplication by y y y leaves these normal forms unchanged except for increasing every j j j by one, so it increases the filtered degree by exactly one; the leading part of each matrix L i L_i L i is y I 2 yI_2 y I 2 , so the matrix degree property also holds. Corollary 7.3 with m = 2 m=2 m = 2 and M = 1 M=1 M = 1 gives (7.12).
The Bézout argument also identifies the common space for unequal bounds: any row in that space has the form f = h L 1 f=hL_1 f = h L 1 . The diagonal numerator makes the condition h L 0 ∈ E hL_0\in E h L 0 ∈ E equivalent to h 1 ∈ F u − 2 A h_1\in F_{u-2}A h 1 ∈ F u − 2 A , h 2 ∈ F v − 2 A h_2\in F_{v-2}A h 2 ∈ F v − 2 A . The second component of h L 1 hL_1 h L 1 is − h 1 + ( y − 1 / 4 ) h 2 -h_1+(y-1/4)h_2 − h 1 + ( y − 1/4 ) h 2 . The second term already has degree at most v − 1 v-1 v − 1 , so requiring the same bound for their sum is equivalent to h 1 ∈ F v − 1 A h_1\in F_{v-1}A h 1 ∈ F v − 1 A . Therefore
h 1 ∈ F min { u − 2 , v − 1 } A , h 2 ∈ F v − 2 A . h_1\in F_{\min\{u-2,v-1\}}A,\qquad h_2\in F_{v-2}A. h 1 ∈ F m i n { u − 2 , v − 1 } A , h 2 ∈ F v − 2 A .
Counting these independent choices proves (7.13). Every such f f f is in the denominator image, so the mass vanishes on the overlap and does not change the count. □
For β > 0 \beta>0 β > 0 this derivative mass is not balanced. Indeed, with e 1 T = [ 1 0 ] e_1^{\mathsf T}=[1\ 0] e 1 T = [ 1 0 ] ,
ν α , β ( x 2 e 1 T , e 1 T ) = 0 , ν α , β ( x e 1 T , x e 1 T ) = β . \nu_{\alpha,\beta}(x^2e_1^{\mathsf T},e_1^{\mathsf T})=0,\qquad\nu_{\alpha,\beta}(xe_1^{\mathsf T},xe_1^{\mathsf T})=\beta. ν α , β ( x 2 e 1 T , e 1 T ) = 0 , ν α , β ( x e 1 T , x e 1 T ) = β .
The rational part of the transformed pairing is balanced, so the same calculation shows that the full form is not balanced when β > 0 \beta>0 β > 0 . Its positive Sobolev term therefore lies outside the point-evaluation classification in Proposition 7.1 , which assumes balance as well as positivity. Orthogonal projection uses only positivity and the specified finite spaces, so Theorem 7.2 applies for every α , β ≥ 0 \alpha,\beta\geq0 α , β ≥ 0 . The coordinate identities in Theorem 8.1 require both pairings to be balanced: within this nodal family the evaluation case β = 0 \beta=0 β = 0 satisfies that requirement, whereas the derivative case β > 0 \beta>0 β > 0 is governed by the positive polynomial and kernel formulas proved here.
The order calculation determines the size of the formulas for every normalized block and its kernel. From degree three onward, all leading rows at a fixed level with equal component bounds use the same 6 × 6 6\times6 6 × 6 matrix T ν T_\nu T ν and scalar bordered determinants of order seven. Its dependence on the mass parameters can be separated from the projection. Since the mass vanishes on Z Z Z , one has Z 0 = Z Z_0=Z Z 0 = Z , and the spaces and matrices F F F and A F A_F A F are fixed. For a fixed complementary list f 1 , … , f l f_1,\ldots,f_l f 1 , … , f l , let a j : = f j , l ( O ) a_j:=f_{j,l}(O) a j := f j , l ( O ) and b j : = ∂ x f j , 1 ( O ) b_j:=\partial_xf_{j,1}(O) b j := ∂ x f j , 1 ( O ) . Then
T ν = T 0 + α a a T + β b b T , T_\nu=T_0+\alpha aa^{\mathsf T}+\beta bb^{\mathsf T}, T ν = T 0 + α a a T + β b b T ,
where T 0 > 0 T_0>0 T 0 > 0 is the massless matrix. The two additions are positive semidefinite and each has rank at most one; at levels where the respective test is nonzero its rank is one.
A consequence for extremal norms
The positive dependence on the mass also controls the extremal norms associated with the kernel. With the rational part and a finite polynomial space X X X fixed, increasing the positive additive mass increases the Gram matrix in the positive semidefinite order. The effect on the kernel follows from the reverse order of inverses: for G > 0 G>0 G > 0 and M ≥ 0 M\geq0 M ≥ 0 , one has
( G + M ) − 1 = G − 1 / 2 ( I + G − 1 / 2 M G − 1 / 2 ) − 1 G − 1 / 2 ≤ G − 1 . (G+M)^{-1}=G^{-1/2}(I+G^{-1/2}MG^{-1/2})^{-1}G^{-1/2}\leq G^{-1}. ( G + M ) − 1 = G − 1/2 ( I + G − 1/2 M G − 1/2 ) − 1 G − 1/2 ≤ G − 1 .
Evaluation at a real point against the same polynomial basis therefore decreases the kernel on the diagonal. For scalar real polynomials, when evaluation at a real point P P P is nonzero on X X X , the extremal identity is
λ X ( P ) : = min { ⟨ f , f ⟩ : f ∈ X , f ( P ) = 1 } = 1 K X ( P , P ) . \lambda_X(P):=\min\{\langle f,f\rangle:f\in X,\ f(P)=1\}=\frac{1}{K_X(P,P)}. λ X ( P ) := min {⟨ f , f ⟩ : f ∈ X , f ( P ) = 1 } = K X ( P , P ) 1 .
The reproducing identity gives 1 = f ( P ) = ⟨ f , K X ( ⋅ , P ) ⟩ 1=f(P)=\langle f,K_X(\cdot,P)\rangle 1 = f ( P ) = ⟨ f , K X ( ⋅ , P )⟩ for an admissible f f f . Cauchy–Schwarz and ∥ K X ( ⋅ , P ) ∥ 2 = K X ( P , P ) \|K_X(\cdot,P)\|^2=K_X(P,P) ∥ K X ( ⋅ , P ) ∥ 2 = K X ( P , P ) give ∥ f ∥ 2 ≥ 1 / K X ( P , P ) \|f\|^2\geq1/K_X(P,P) ∥ f ∥ 2 ≥ 1/ K X ( P , P ) . The assumption on evaluation makes K X ( P , P ) > 0 K_X(P,P)>0 K X ( P , P ) > 0 , and the polynomial K X ( ⋅ , P ) / K X ( P , P ) K_X(\cdot,P)/K_X(P,P) K X ( ⋅ , P ) / K X ( P , P ) attains equality. Thus increasing a positive mass increases this minimum. The reduced kernel formula computes the resulting change in the extremal norm as well as the orthogonal polynomials.
Matrix orthogonal polynomials and the Durán–Grünbaum weights
When the component bounds coincide on the line, the normalized rows of the preceding construction assemble into ordinary matrix orthogonal polynomials. For p = q = m p=q=m p = q = m and a real symmetric positive matrix measure d Σ \mathrm{d}\Sigma d Σ on R \mathbb{R} R , the corresponding matrix inner product is
⟨ P , Q ⟩ Σ : = ∫ R P ( x ) d Σ ( x ) Q ( x ) T . \langle P,Q\rangle_{\Sigma}:=\int_{\mathbb{R}} P(x)\,\mathrm{d}\Sigma(x)\,Q(x)^{\mathsf{T}}. ⟨ P , Q ⟩ Σ := ∫ R P ( x ) d Σ ( x ) Q ( x ) T .
Its ( a , b ) (a,b) ( a , b ) entry pairs row a a a of P P P with row b b b of Q Q Q . Assume that all polynomial moments exist and that every nonzero polynomial row has positive squared norm. The monic matrix orthogonal polynomials P n P_n P n then have leading coefficient I m I_m I m ; write H n : = ⟨ P n , P n ⟩ Σ > 0 H_n:=\langle P_n,P_n\rangle_{\Sigma}>0 H n := ⟨ P n , P n ⟩ Σ > 0 .
Let L 0 , L 1 L_0,L_1 L 0 , L 1 be regular matrix polynomials and S = L 1 − 1 L 0 S=L_1^{-1}L_0 S = L 1 − 1 L 0 . Assume that S S S is defined Σ \Sigma Σ -almost everywhere, that the required rational moments exist, and that the rational inner product satisfies the positivity hypotheses of Theorem 7.2 at every level. With positive semidefinite matrices C ξ C_{\xi} C ξ at finitely many real points, take
d Σ ^ ( x ) = S ( x ) d Σ ( x ) S ( x ) T + ∑ ξ C ξ δ ξ , L 1 ( ξ ) C ξ = 0. (124) \mathrm{d}\widehat{\Sigma}(x)=S(x)\,\mathrm{d}\Sigma(x)\,S(x)^{\mathsf{T}}+\sum_{\xi}C_{\xi}\,\delta_{\xi},\qquad L_1(\xi)C_{\xi}=0. \tag*{(124)} d Σ ( x ) = S ( x ) d Σ ( x ) S ( x ) T + ξ ∑ C ξ δ ξ , L 1 ( ξ ) C ξ = 0. (124)
Multiplication on the two sides gives the measure identity L 1 d Σ ^ L 1 T = L 0 d Σ L 0 T L_1\mathrm{d}\widehat{\Sigma}L_1^{\mathsf{T}}=L_0\mathrm{d}\Sigma L_0^{\mathsf{T}} L 1 d Σ L 1 T = L 0 d Σ L 0 T . The measure remains positive definite on polynomial rows, since its rational part is positive definite and each added term is nonnegative.
Corollary 7.5 (Monic matrix polynomials from the reduced matrix). Under these assumptions, take E = X n = ( R [ x ] < n ) 1 × m E=X_n=(\mathbb{R}[x]_{<n})^{1\times m} E = X n = ( R [ x ] < n ) 1 × m and X n + = ( R [ x ] ≤ n ) 1 × m X_n^+=(\mathbb{R}[x]_{\leq n})^{1\times m} X n + = ( R [ x ] ≤ n ) 1 × m , for n ≥ 1 n\geq1 n ≥ 1 . Let the spaces and matrices be those of Theorem 7.2 , with common reduced order t n t_n t n . For a ∈ { 1 , … , m } a\in\{1,\ldots,m\} a ∈ { 1 , … , m } , the leading row is v a = x n e a T v_a=x^n e_a^{\mathsf{T}} v a = x n e a T . The corresponding projected rows ( v a ) 0 (v_a)_0 ( v a ) 0 form the m × m m\times m m × m polynomial matrix V 0 , n V_{0,n} V 0 , n , while their pairings with the complementary rows form the m × t n m\times t_n m × t n matrix
B n : = [ b ( ( v a ) j ) ] a ∈ { 1 , … , m } , j ∈ { 1 , … , t n } . B_n:=\left[b((v_a)_j)\right]_{a\in\{1,\ldots,m\},\,j\in\{1,\ldots,t_n\}}. B n := [ b (( v a ) j ) ] a ∈ { 1 , … , m } , j ∈ { 1 , … , t n } . Write W n : = w W_n:=w W n := w and T n : = T ν T_n:=T_{\nu} T n := T ν for the polynomial rows after projection and their Gram matrix. Then
P ^ n = V 0 , n − B n T n − 1 W n (125) \widehat{P}_n=V_{0,n}-B_nT_n^{-1}W_n \tag*{(125)} P n = V 0 , n − B n T n − 1 W n (125) is the monic matrix orthogonal polynomial for Σ ^ \widehat{\Sigma} Σ . Its norm and the kernel for degrees less than n n n are
H ^ n = ⟨ V 0 , n , V 0 , n ⟩ Σ ^ − B n T n − 1 B n T > 0 , (126) \widehat{H}_n=\langle V_{0,n},V_{0,n}\rangle_{\widehat{\Sigma}}-B_nT_n^{-1}B_n^{\mathsf{T}}>0, \tag*{(126)} H n = ⟨ V 0 , n , V 0 , n ⟩ Σ − B n T n − 1 B n T > 0 , (126) K ^ n ( x , y ) = ∑ j = 0 n − 1 P ^ j ( x ) T H ^ j − 1 P ^ j ( y ) . (127) \widehat{K}_n(x,y)=\sum_{j=0}^{n-1}\widehat{P}_j(x)^{\mathsf{T}}\widehat{H}_j^{-1}\widehat{P}_j(y). \tag*{(127)} K n ( x , y ) = j = 0 ∑ n − 1 P j ( x ) T H j − 1 P j ( y ) . (127) The latter is also given directly by (117) . Every entry of P ^ n \widehat{P}_n P n has the bordered determinant (116) , with denominator of order t n t_n t n and numerator of order t n + 1 t_n+1 t n + 1 . The preliminary projection uses the additional order- t n t_n t n matrix A F A_F A F . If the Bézout and degree hypotheses of Corollary 7.3 hold, then
t n = m min { n , M } , M : = max { deg L 0 , deg L 1 } . t_n=m\min\{n,M\},\qquad M:=\max\{\deg L_0,\deg L_1\}. t n = m min { n , M } , M := max { deg L 0 , deg L 1 } . In particular, these orders are independent of n n n for n ≥ M n\geq M n ≥ M .
Proof. The rows of H j − 1 / 2 P j H_j^{-1/2}P_j H j − 1/2 P j , with j ∈ { 0 , … , n − 1 } j\in\{0,\ldots,n-1\} j ∈ { 0 , … , n − 1 } , form the original orthonormal basis of E E E needed in the projection construction. All its entries are consequently determined by the original matrix polynomials and norms, together with the rational moments and masses. Applied to each leading row v a v_a v a , Theorem 7.2 subtracts only elements of X n X_n X n , so the assembled matrix retains the leading term x n I m x^nI_m x n I m . Its row orthogonality to X n X_n X n means that its matrix inner product with every lower-degree matrix polynomial vanishes. The result is therefore a monic matrix orthogonal polynomial, whose uniqueness follows from uniqueness of each normalized row.
The matrix of pairings of V 0 , n V_{0,n} V 0 , n with W n W_n W n is B n B_n B n , and the Gram matrix of W n W_n W n is T n T_n T n . Expanding the inner product of (125) with itself gives
H ^ n = ⟨ V 0 , n , V 0 , n ⟩ − B n T n − 1 B n T − B n T n − 1 B n T + B n T n − 1 T n T n − 1 B n T = ⟨ V 0 , n , V 0 , n ⟩ − B n T n − 1 B n T , (128) \widehat{H}_n = \langle V_{0,n}, V_{0,n}\rangle- B_nT_n^{-1}B_n^{\mathsf{T}} - B_nT_n^{-1}B_n^{\mathsf{T}} + B_nT_n^{-1}T_nT_n^{-1}B_n^{\mathsf{T}} = \langle V_{0,n}, V_{0,n}\rangle- B_nT_n^{-1}B_n^{\mathsf{T}}, \tag*{(128)} H n = ⟨ V 0 , n , V 0 , n ⟩ − B n T n − 1 B n T − B n T n − 1 B n T + B n T n − 1 T n T n − 1 B n T = ⟨ V 0 , n , V 0 , n ⟩ − B n T n − 1 B n T , (128)
which is (7.16). Its positivity follows because any nonzero constant combination of the rows of P ^ n \widehat{P}_n P n has a nonzero leading row. Orthogonality of the matrix blocks gives (127) , whereas their rows form an orthogonal basis of X n X_n X n , identifying it with (117) . The componentwise bordered formulas give the determinant statement. For the order count, Corollary 7.3 applies with H A ( j ) = j + 1 H_A(j)=j+1 H A ( j ) = j + 1 for j ≥ 0 j\geq0 j ≥ 0 , yielding the stated expression. The construction begins with P ^ 0 = I m \widehat{P}_0=I_m P 0 = I m .
The argument extends to the Hermitian inner product ∫ P d Σ Q ∗ \int P\,\mathrm{d}\Sigma\,Q^* ∫ P d Σ Q ∗ because complex orthogonal projection has the same block decomposition. Positive Hermitian Gram matrices therefore give the Schur complement (125) , with B n ∗ B_n^* B n ∗ as the last factor in (7.16) and P ^ j ( x ) ∗ \widehat{P}_j(x)^* P j ( x ) ∗ as the first factor in (127) . Here a star denotes conjugate transpose; the measure congruence (124) uses the same adjoints and masses satisfying C ξ = C ξ ∗ ≥ 0 C_\xi=C_\xi^*\geq0 C ξ = C ξ ∗ ≥ 0 .
An explicit connection with the Hermite matrix weights
The factorized matrix weights studied by Durán and Grünbaum provide examples of this congruence construction. Their families of matrix orthogonal polynomials satisfy second-order differential equations [26 , 28 ] . A weight w ( x ) T ( x ) T ( x ) ∗ w(x)T(x)T(x)^* w ( x ) T ( x ) T ( x ) ∗ has the form (124) whenever T T T is polynomial or rational and the stated moments exist, with d Σ = w I m d x \mathrm{d}\Sigma=wI_m\mathrm{d}x d Σ = w I m d x , S = T S=T S = T and zero mass. The broader differential constructions also include nonrational T T T .
In particular, the two Hermite weights of [27 ] have the explicit factorization
H a , r ( x ) = e − x 2 [ 1 + ∣ a ∣ 2 x 2 r a x r a ‾ x r 1 ] = T r ( x ) ( e − x 2 I 2 ) T r ( x ) ∗ , T r ( x ) : = [ 1 a x r 0 1 ] , r ∈ { 1 , 2 } . \mathcal{H}_{a,r}(x)=e^{-x^2}\begin{bmatrix}1+|a|^2x^{2r} & ax^r\\ \overline{a}x^r & 1\end{bmatrix}=T_r(x)(e^{-x^2}I_2)T_r(x)^*,\qquad T_r(x):=\begin{bmatrix}1 & ax^r\\ 0 & 1\end{bmatrix},\qquad r\in\{1,2\}. H a , r ( x ) = e − x 2 [ 1 + ∣ a ∣ 2 x 2 r a x r a x r 1 ] = T r ( x ) ( e − x 2 I 2 ) T r ( x ) ∗ , T r ( x ) := [ 1 0 a x r 1 ] , r ∈ { 1 , 2 } .
They correspond to L 1 = I 2 L_1=I_2 L 1 = I 2 , L 0 = T r L_0=T_r L 0 = T r and ν = 0 \nu=0 ν = 0 , so these established matrix families are specializations of the bilateral polynomial transformation. Their triangular factors also make it possible to compute the reduced order directly from the common space. For a ∈ R ∖ { 0 } a\in\mathbb{R}\setminus\{0\} a ∈ R ∖ { 0 } and E = X n E=X_n E = X n as above, the image of a row f = [ f 1 f 2 ] f=[f_1\ f_2] f = [ f 1 f 2 ] is
f T r = [ f 1 a x r f 1 + f 2 ] . fT_r=[f_1\ ax^rf_1+f_2]. f T r = [ f 1 a x r f 1 + f 2 ] .
Since both components of f f f have degree less than n n n , the second component of f T r fT_r f T r has degree less than n n n exactly when deg f 1 < n − r \deg f_1<n-r deg f 1 < n − r . A negative bound means the zero polynomial. Consequently
dim Z = n + max { n − r , 0 } , t n = 2 n − dim Z = min { n , r } . \dim Z=n+\max\{n-r,0\},\qquad t_n=2n-\dim Z=\min\{n,r\}. dim Z = n + max { n − r , 0 } , t n = 2 n − dim Z = min { n , r } .
The reduced matrices consequently have stable order one in the first family and two in the second. These orders are obtained from the common-space calculation because the highest-degree coefficient of T r T_r T r is singular. The count m M mM m M in Corollary 7.5 is subject to its additional degree hypotheses.
For the first family, scalar Hermite orthogonality also gives a closed expression for the monic matrix polynomial. Let h n h_n h n denote the monic scalar Hermite polynomials, with normalization
∫ R h n ( x ) h j ( x ) e − x 2 d x = γ n δ n , j , γ n : = π n ! 2 n , D n : = 1 + a 2 n 2 , \int_{\mathbb{R}} h_n(x)h_j(x)e^{-x^2}\,\mathrm{d}x=\gamma_n\delta_{n,j},\qquad\gamma_n:=\frac{\sqrt{\pi}\,n!}{2^n},\qquad D_n:=1+\frac{a^{2n}}{2}, ∫ R h n ( x ) h j ( x ) e − x 2 d x = γ n δ n , j , γ n := 2 n π n ! , D n := 1 + 2 a 2 n ,
and set h j = 0 h_j=0 h j = 0 for j < 0 j<0 j < 0 . The result is
P ^ n ( x ) = [ h n ( x ) − a n 2 h n − 1 ( x ) − a n 2 D n h n − 1 ( x ) h n ( x ) + a 2 n ( n − 1 ) 4 D n h n − 2 ( x ) ] , H ^ n = γ n diag ( D n + 1 , D n − 1 ) . (129) \widehat{P}_n(x)=\begin{bmatrix}h_n(x)&-\dfrac{an}{2}h_{n-1}(x)\\-\dfrac{an}{2D_n}h_{n-1}(x)&h_n(x)+\dfrac{a^2n(n-1)}{4D_n}h_{n-2}(x)\end{bmatrix},\qquad\widehat{H}_n=\gamma_n\operatorname{diag}(D_{n+1},D_n^{-1}). \tag*{(129)} P n ( x ) = h n ( x ) − 2 D n an h n − 1 ( x ) − 2 an h n − 1 ( x ) h n ( x ) + 4 D n a 2 n ( n − 1 ) h n − 2 ( x ) , H n = γ n diag ( D n + 1 , D n − 1 ) . (129)
Multiplication by the triangular factor reduces verification of the expression to scalar Hermite orthogonality. In fact, the recurrence x h n = h n + 1 + ( n / 2 ) h n − 1 xh_n=h_{n+1}+(n/2)h_{n-1} x h n = h n + 1 + ( n /2 ) h n − 1 gives
P ^ n T 1 = [ h n a h n + 1 − a n 2 D n h n − 1 D n − 1 h n ] . \widehat{P}_nT_1=\begin{bmatrix}h_n&ah_{n+1}\\-\dfrac{an}{2D_n}h_{n-1}&D_n^{-1}h_n\end{bmatrix}. P n T 1 = [ h n − 2 D n an h n − 1 a h n + 1 D n − 1 h n ] .
A lower test [ u v ] [u\ v] [ u v ] has image [ u a x u + v ] [u\ axu+v] [ u a xu + v ] , which reduces the matrix orthogonality conditions to scalar Hermite pairings. For the first row these vanish by the bounds deg u < n \deg u<n deg u < n and deg ( x u ) , deg v ≤ n \deg(xu),\deg v\le n deg ( xu ) , deg v ≤ n . In the second row, the only possible contribution comes from the coefficient of h n − 1 h_{n-1} h n − 1 in u u u ; its two scalar pairings cancel because − a n γ n − 1 / ( 2 D n ) + a γ n / D n = 0 -a n\gamma_{n-1}/(2D_n)+a\gamma_n/D_n=0 − an γ n − 1 / ( 2 D n ) + a γ n / D n = 0 . The same scalar inner product gives squared norm γ n + a 2 γ n + 1 = γ n D n + 1 \gamma_n+a^2\gamma_{n+1}=\gamma_nD_{n+1} γ n + a 2 γ n + 1 = γ n D n + 1 for the first image row. For the second row it gives
a 2 n 2 γ n − 1 4 D n 2 + γ n D n 2 = γ n ( 1 + a 2 n / 2 ) D n 2 = γ n D n , \frac{a^2n^2\gamma_{n-1}}{4D_n^2}+\frac{\gamma_n}{D_n^2}=\frac{\gamma_n(1+a^2n/2)}{D_n^2}=\frac{\gamma_n}{D_n}, 4 D n 2 a 2 n 2 γ n − 1 + D n 2 γ n = D n 2 γ n ( 1 + a 2 n /2 ) = D n γ n ,
using γ n = n γ n − 1 / 2 \gamma_n=n\gamma_{n-1}/2 γ n = n γ n − 1 /2 when n ≥ 1 n\ge1 n ≥ 1 . Their cross pairing is zero: it pairs h n h_n h n with h n − 1 h_{n-1} h n − 1 in the first component and h n + 1 h_{n+1} h n + 1 with h n h_n h n in the second. At n = 0 n=0 n = 0 , the displayed image matrix gives the same norm without using the recurrence for its previous index. The diagonal and cross pairings therefore establish the entire norm matrix. Since the leading coefficient in (7.18) is I 2 I_2 I 2 , uniqueness of the monic orthogonal polynomial identifies this expression with (7.15).
The additional differential structure
The Hermite matrix families also possess the differential property studied in [26 , 27 , 28 ] : there are fixed polynomial matrices F 2 F_2 F 2 , F 1 F_1 F 1 , F 0 F_0 F 0 satisfying
P ^ n ′ ′ F 2 + P ^ n ′ F 1 + P ^ n F 0 = Λ n P ^ n \widehat{P}_n''F_2+\widehat{P}_n'F_1+\widehat{P}_nF_0=\Lambda_n\widehat{P}_n P n ′′ F 2 + P n ′ F 1 + P n F 0 = Λ n P n
for every n n n , with deg F 2 ≤ 2 \deg F_2\le2 deg F 2 ≤ 2 , deg F 1 ≤ 1 \deg F_1\le1 deg F 1 ≤ 1 and F 0 F_0 F 0 constant. Together with the matrix three-term recurrence, this is the differential and discrete spectral structure of those families. Its preservation under a further rational congruence requires compatibility of the transformed weight with such an operator. The rational congruence determines the orthogonal polynomials, norms and kernel through the formulas above, while the differential equation requires this additional compatibility. Coordinate multiplication provides the corresponding discrete relations: Section 8 expresses the change of recurrence matrices through four-factor intertwining identities.
A positive bivariate example on three nodes
The congruence argument also applies to positive bivariate pairings, where different support points can interact. A finite example is obtained from the nodes x 1 = − 1 x_1=-1 x 1 = − 1 , x 2 = 0 x_2=0 x 2 = 0 , x 3 = 1 x_3=1 x 3 = 1 and the matrix blocks
U ( x ) : = [ 1 x 0 1 ] , K i , j : = δ i , j I 2 + U ( x i ) U ( x j ) T . U(x):=\begin{bmatrix}1&x\\0&1\end{bmatrix},\qquad K_{i,j}:=\delta_{i,j}I_2+U(x_i)U(x_j)^{\mathsf T}. U ( x ) := [ 1 0 x 1 ] , K i , j := δ i , j I 2 + U ( x i ) U ( x j ) T .
For complex polynomial rows f , h f,h f , h of length two, define
⟨ f , h ⟩ : = ∑ i , j = 1 3 f ( x i ) K i , j h ( x j ) T . \langle f,h\rangle:=\sum_{i,j=1}^{3}f(x_i)K_{i,j}h(x_j)^{\mathsf T}. ⟨ f , h ⟩ := i , j = 1 ∑ 3 f ( x i ) K i , j h ( x j ) T .
The 6 × 6 6\times6 6 × 6 matrix of this pairing in evaluation coordinates is
K : = [ K i , j ] i , j ∈ { 1 , … , 3 } = I 6 + U U T , U : = [ U ( x 1 ) U ( x 2 ) U ( x 3 ) ] . K:=[K_{i,j}]_{i,j\in\{1,\ldots,3\}}=I_6+UU^{\mathsf T},\qquad U:=\begin{bmatrix}U(x_1)\\U(x_2)\\U(x_3)\end{bmatrix}. K := [ K i , j ] i , j ∈ { 1 , … , 3 } = I 6 + U U T , U := U ( x 1 ) U ( x 2 ) U ( x 3 ) .
For a nonzero real or complex column v v v , v ∗ K v = ∥ v ∥ 2 + ∥ U T v ∥ 2 > 0 v^*Kv=\lVert v\rVert^2+\lVert U^{\mathsf T}v\rVert^2>0 v ∗ K v = ∥ v ∥ 2 + ∥ U T v ∥ 2 > 0 , so it is positive definite. Evaluation at the three distinct nodes is an isomorphism from rows of component degrees below three to C 6 \mathbb{C}^6 C 6 , by scalar polynomial interpolation in each component. This interpolation isomorphism makes the polynomial pairing positive definite on the stated space. Its determinant is also explicit. The identity U ( x i ) T U ( x i ) = [ 1 x i ; x i 1 + x i 2 ] U(x_i)^{\mathsf T}U(x_i)=[1\ x_i;\ x_i\ 1+x_i^2] U ( x i ) T U ( x i ) = [ 1 x i ; x i 1 + x i 2 ] , summed over x i ∈ { − 1 , 0 , 1 } x_i\in\{-1,0,1\} x i ∈ { − 1 , 0 , 1 } , yields U T U = diag ( 3 , 5 ) U^{\mathsf T}U=\operatorname{diag}(3,5) U T U = diag ( 3 , 5 ) ; the determinant identity det ( I + A B ) = det ( I + B A ) \det(I+AB)=\det(I+BA) det ( I + A B ) = det ( I + B A ) then gives det K = 4 ⋅ 6 = 24 \det K=4\cdot6=24 det K = 4 ⋅ 6 = 24 . This gives the determinant in evaluation coordinates. If V V V has as columns the evaluations of a real monomial basis, the change to that basis gives Gram matrix V T K V V^{\mathsf T}KV V T K V and hence determinant ( det V ) 2 det K (\det V)^2\det K ( det V ) 2 det K .
A rational congruence of this pairing is defined by
S ( x ) = [ ( x + 2 ) − 1 x 0 x + 3 ] . S(x)=\begin{bmatrix}(x+2)^{-1}&x\\0&x+3\end{bmatrix}. S ( x ) = [ ( x + 2 ) − 1 0 x x + 3 ] .
It is obtained from the polynomial factors
L 1 = [ x + 2 0 0 1 ] , L 0 = [ 1 x ( x + 2 ) 0 x + 3 ] , R i = L i T L_1=\begin{bmatrix}x+2&0\\0&1\end{bmatrix},\qquad L_0=\begin{bmatrix}1&x(x+2)\\0&x+3\end{bmatrix},\qquad R_i=L_i^{\mathsf{T}} L 1 = [ x + 2 0 0 1 ] , L 0 = [ 1 0 x ( x + 2 ) x + 3 ] , R i = L i T
since S = L 1 − 1 L 0 S=L_1^{-1}L_0 S = L 1 − 1 L 0 and S R = S T S_R=S^{\mathsf{T}} S R = S T . These factors make the example a massless solution of the four-factor equation. Regularity and invertibility at all three nodes ensure that the transformed blocks K ^ i , j = S ( x i ) K i , j S ( x j ) T \widehat{K}_{i,j}=S(x_i)K_{i,j}S(x_j)^{\mathsf{T}} K i , j = S ( x i ) K i , j S ( x j ) T again define a positive pairing. In evaluation coordinates, D : = diag ( S ( x 1 ) , S ( x 2 ) , S ( x 3 ) ) D:=\operatorname{diag}(S(x_1),S(x_2),S(x_3)) D := diag ( S ( x 1 ) , S ( x 2 ) , S ( x 3 )) gives K ^ = D K D T > 0 \widehat{K}=DKD^{\mathsf{T}}>0 K = D K D T > 0 . The determinant transforms by the square of the determinant of this change: det D = 2 ⋅ ( 3 / 2 ) ⋅ ( 4 / 3 ) = 4 \det D=2\cdot(3/2)\cdot(4/3)=4 det D = 2 ⋅ ( 3/2 ) ⋅ ( 4/3 ) = 4 therefore gives det K ^ = 4 2 ⋅ 24 = 384 \det\widehat{K}=4^2\cdot24=384 det K = 4 2 ⋅ 24 = 384 . The two nonzero Gram determinants establish normality on the specified polynomial space. At the same time, for i ≠ j i\ne j i = j the blocks K i , j K_{i,j} K i , j are nonzero and stay nonzero under the invertible congruence. Positivity and normality in this example therefore coexist with interactions between distinct nodes, as permitted by the bivariate pairing convention.
Positive systems on real curves
A positive density on a real arc gives congruences that are normal on every finite row space. Let the support be an integral real curve with a nontrivial real smooth arc, and let a compact nontrivial subarc avoid det L 0 det L 1 = 0 \det L_0\det L_1=0 det L 0 det L 1 = 0 . For a continuous positive definite symmetric density W W W on that subarc, both W W W and S W S T SWS^{\mathsf{T}} S W S T , with any additional mass (7.2), define positive definite pairings on every finite row space and its transpose. In particular, for f ≠ 0 f\ne0 f = 0 ,
∫ f S W S T f T d t + ∑ P f ( P ) C P f ( P ) T > 0 \int fSWS^{\mathsf{T}}f^{\mathsf{T}}\,dt+\sum_P f(P)C_Pf(P)^{\mathsf{T}}>0 ∫ f S W S T f T d t + P ∑ f ( P ) C P f ( P ) T > 0
Since S S S is invertible over the function field, f S ≠ 0 fS\ne0 f S = 0 . After clearing its denominators, at least one component is a nonzero regular function on the integral curve. Such a function has only finitely many zeros on this compact arc unless it is identically zero: its zero locus is a proper zero-dimensional closed subset of the curve. Hence f S fS f S is nonzero on a subinterval. The positive density makes the integral strictly positive there, and the atomic term is nonnegative. The same argument with S = I S=I S = I proves the assertion for W W W .
For the elliptic curve y 2 = x 3 − x + 1 y^2=x^3-x+1 y 2 = x 3 − x + 1 , an explicit instance is given by
L 1 = [ x 1 0 1 ] and L 0 = [ 1 y 1 x + 1 ] . L_1=\begin{bmatrix}x&1\\0&1\end{bmatrix}\qquad\text{and}\qquad L_0=\begin{bmatrix}1&y\\1&x+1\end{bmatrix}. L 1 = [ x 0 1 1 ] and L 0 = [ 1 1 y x + 1 ] .
These factors determine S = L 1 − 1 L 0 S=L_1^{-1}L_0 S = L 1 − 1 L 0 . With W = I 2 W=I_2 W = I 2 on the arc 3 ≤ x ≤ 4 3\le x\le4 3 ≤ x ≤ 4 , y = x 3 − x + 1 y=\sqrt{x^3-x+1} y = x 3 − x + 1 , the required invertibility follows from det L 1 = x \det L_1=x det L 1 = x and det L 0 = x + 1 − y \det L_0=x+1-y det L 0 = x + 1 − y : the identity y 2 − ( x + 1 ) 2 = x ( x 2 − x − 3 ) > 0 y^2-(x+1)^2=x(x^2-x-3)>0 y 2 − ( x + 1 ) 2 = x ( x 2 − x − 3 ) > 0 excludes zeros of the second determinant as well. The inverse product has the explicit form
S = [ 0 y − x − 1 x 1 x + 1 x ] . S=\begin{bmatrix}0&\dfrac{y-x-1}{x}\\1&\dfrac{x+1}{x}\end{bmatrix}. S = 0 1 x y − x − 1 x x + 1 .
The original measure has density I 2 d x I_2\,dx I 2 d x on the selected arc, and the transformed density S S T d x SS^{\mathsf{T}}\,dx S S T d x is therefore positive and gives normality on every finite row space paired with its transpose. The associated right factors are R i = L i T R_i=L_i^{\mathsf{T}} R i = L i T , and either zero mass or any mass satisfying (7.2) preserves this conclusion. The integral and the added point masses are balanced, since scalar multiplication can be moved between their two arguments. This supplies the positive matrix measure used in Subsection 8.1: the factors and arc give positivity and normality directly, and balance permits the two coordinate recurrence identities there.
Coordinate multiplication and bilateral Darboux identities
The polynomial formulas obtained above also determine recurrence relations: multiplication by a coordinate produces a polynomial whose expansion in the orthogonal basis defines the corresponding multiplication matrix. Since the coordinates on an algebraic support satisfy polynomial relations, their multiplication matrices satisfy those relations as well. The bilateral transformation changes these matrices through the four connections associated with L 0 , L 1 , R 0 , R 1 L_0,L_1,R_0,R_1 L 0 , L 1 , R 0 , R 1 . The connections are determined by the factors and the polynomial bases, so a single set of them relates multiplication by every coordinate.
For scalar Jacobi matrices, Darboux transformations express spectral modifications through triangular factorizations and reversal of the factors; see [38 ] . For scalar bilinear functionals, Bueno and Marcellán [42 ] introduce polynomial factors on the two arguments and relate the associated Hessenberg matrices through LU and QR factorizations in the one-sided and symmetric cases developed there. These results provide the scalar functional precedent for changing the two arguments. Recurrence formulations for matrix and mixed-type orthogonality are developed in [1 , 3 , 4 ] . Here the term bilateral Darboux identities denotes the four-factor intertwining relations and the equality between the two products of connection matrices proved below. Their formulation retains four independent matrix factors on the two modules over the coordinate algebra, including numerator and denominator factors. The same connections act simultaneously on all coordinate multiplication matrices and preserve the equations of the support. The theorem is stated in terms of finite polynomial expansions; triangular factorization and reversal impose additional conditions on the connection matrices.
Write β : = β μ \beta:= \beta_{\mu} β := β μ and β ^ : = β μ ^ \widehat{\beta} := \beta_{\widehat{\mu}} β := β μ . Throughout this section both pairings are bilinear and balanced:
β ( a f , g ) = β ( f , a g ) , β ^ ( a f , g ) = β ^ ( f , a g ) , a ∈ A . \beta(af,g)=\beta(f,ag), \qquad\widehat{\beta}(af,g)=\widehat{\beta}(f,ag), \qquad a \in A. β ( a f , g ) = β ( f , a g ) , β ( a f , g ) = β ( f , a g ) , a ∈ A .
Balance identifies the row and column matrices of scalar multiplication and supplies the lower bound on their nonzero blocks. It is also the condition that makes these matrices symmetric for positive forms in orthonormal bases. Take nested spaces E n ⊂ A 1 × q E_n \subset A^{1\times q} E n ⊂ A 1 × q and H n ⊂ A p × 1 H_n \subset A^{p\times1} H n ⊂ A p × 1 of equal dimensions, exhausting the two modules, with both pairings normal on every E n × H n E_n \times H_n E n × H n . Put E − 1 = H − 1 = 0 E_{-1}=H_{-1}=0 E − 1 = H − 1 = 0 and r n : = dim E n − dim E n − 1 > 0 r_n := \dim E_n-\dim E_{n-1}>0 r n := dim E n − dim E n − 1 > 0 . Choose dual polynomial blocks B n , A n B_n,A_n B n , A n spanning successive levels, with B n B_n B n of size r n × q r_n \times q r n × q , A n A_n A n of size p × r n p \times r_n p × r n and β ( B n , A n ) = I r n \beta(B_n,A_n)=I_{r_n} β ( B n , A n ) = I r n and β ( B n , A m ) = 0 \beta(B_n,A_m)=0 β ( B n , A m ) = 0 for n ≠ m n \ne m n = m . The analogous blocks for β ^ \widehat{\beta} β are denoted with hats. The vertical arrays B , B ^ \mathcal{B},\widehat{\mathcal{B}} B , B collect the row blocks, whereas the horizontal arrays A , A ^ \mathcal{A},\widehat{\mathcal{A}} A , A collect the column blocks. Thus the rows of B 0 , B 1 , … B_0,B_1,\ldots B 0 , B 1 , … occur consecutively in the vertical array, and the columns of the A n A_n A n occur consecutively in the horizontal array. These arrays record polynomial bases constructed, one block at a time, from finite Gram matrices. Exhaustion places every polynomial in a finite level, so each expansion has only finitely many basis elements. The arrays therefore act algebraically on the polynomial modules; no bounded-operator interpretation on a completed function space is required.
For each scalar polynomial class a ∈ A a \in A a ∈ A , assume finite nonnegative shifts satisfying
a E n ⊂ E n + d L ( a ) , a H n ⊂ H n + d R ( a ) ( n ≥ 0 ) . aE_n \subset E_{n+d_L(a)}, \qquad aH_n \subset H_{n+d_R(a)} \qquad(n \ge0). a E n ⊂ E n + d L ( a ) , a H n ⊂ H n + d R ( a ) ( n ≥ 0 ) .
Define
J ( a ) : = [ β ( a B n , A m ) ] n , m ≥ 0 , J ^ ( a ) : = [ β ^ ( a B ^ n , A ^ m ) ] n , m ≥ 0 , J(a):=[\beta(aB_n,A_m)]_{n,m\ge0}, \qquad\widehat{J}(a):=[\widehat{\beta}(a\widehat{B}_n,\widehat{A}_m)]_{n,m\ge0}, J ( a ) := [ β ( a B n , A m ) ] n , m ≥ 0 , J ( a ) := [ β ( a B n , A m ) ] n , m ≥ 0 ,
and define the four connection matrices by the polynomial expansions
B L 0 = U 0 B , B L 1 = U 1 B ^ , R 0 A = A V 0 , R 1 A = A ^ V 1 . (130) \mathcal{B}L_0=U_0\mathcal{B}, \qquad\mathcal{B}L_1=U_1\widehat{\mathcal{B}}, \qquad R_0\mathcal{A}=\mathcal{A}V_0, \qquad R_1\mathcal{A}=\widehat{\mathcal{A}}V_1. \tag*{(130)} B L 0 = U 0 B , B L 1 = U 1 B , R 0 A = A V 0 , R 1 A = A V 1 . (130)
Their blocks are computed directly from the dual polynomials:
U 0 [ n , m ] = β ( B n L 0 , A m ) , U 1 [ n , m ] = β ^ ( B n L 1 , A ^ m ) , U_0[n,m]=\beta(B_nL_0,A_m), \qquad U_1[n,m]=\widehat{\beta}(B_nL_1,\widehat{A}_m), U 0 [ n , m ] = β ( B n L 0 , A m ) , U 1 [ n , m ] = β ( B n L 1 , A m ) ,
V 0 [ n , m ] = β ( B n , R 0 A m ) , V 1 [ n , m ] = β ^ ( B ^ n , R 1 A m ) . V_0[n,m]=\beta(B_n,R_0A_m), \qquad V_1[n,m]=\widehat{\beta}(\widehat{B}_n,R_1A_m). V 0 [ n , m ] = β ( B n , R 0 A m ) , V 1 [ n , m ] = β ( B n , R 1 A m ) .
Indeed, the n n n th block row of U 1 U_1 U 1 is the coefficient row in B n L 1 = ∑ m U 1 [ n , m ] B ^ m B_nL_1=\sum_m U_1[n,m]\widehat{B}_m B n L 1 = ∑ m U 1 [ n , m ] B m , so pairing with A ^ m \widehat{A}_m A m recovers the indicated block by duality. Applying this coefficient extraction to each of the other three expansions gives the remaining formulas. Since each polynomial has a finite expansion, every block row of U i U_i U i and every block column of V i V_i V i has finitely many nonzero blocks. The choice of bases in these definitions expands the numerators in the original system and the denominators in the transformed system, as required by the two sides of the functional equation.
Theorem 8.1 (Multiplication matrices and the four-factor identities). The dual arrays satisfy
a B = J ( a ) B , a A = A J ( a ) , a\mathcal{B}=J(a)\mathcal{B}, \qquad a\mathcal{A}=\mathcal{A}J(a), a B = J ( a ) B , a A = A J ( a ) , and the hatted counterparts. Their blocks vanish unless n − d R ( a ) ≤ m ≤ n + d L ( a ) n-d_R(a)\le m\le n+d_L(a) n − d R ( a ) ≤ m ≤ n + d L ( a ) . The maps a ↦ J ( a ) a\mapsto J(a) a ↦ J ( a ) and a ↦ J ^ ( a ) a\mapsto\widehat{J}(a) a ↦ J ( a ) are faithful unital algebra representations. Thus, for A = C [ x 1 , … , x d ] / I A=\mathbb{C}[x_1,\ldots,x_d]/I A = C [ x 1 , … , x d ] / I ,
f ( J ( x 1 ) , … , J ( x d ) ) = 0 if and only if f ∈ I , f(J(x_1),\ldots,J(x_d))=0 \quad\text{if and only if} \quad f\in I, f ( J ( x 1 ) , … , J ( x d )) = 0 if and only if f ∈ I , and likewise for the transformed matrices. The functional equation (2.1) holds if and only if
U 1 V 1 = U 0 V 0 . (131) U_1V_1=U_0V_0. \tag*{(131)} U 1 V 1 = U 0 V 0 . (131)
For every a ∈ A a \in A a ∈ A the connections satisfy
J ( a ) U 1 = U 1 J ^ ( a ) , J ^ ( a ) V 1 = V 1 J ( a ) (132) J(a)U_1 = U_1\widehat{J}(a), \qquad\widehat{J}(a)V_1 = V_1J(a) \tag*{(132)} J ( a ) U 1 = U 1 J ( a ) , J ( a ) V 1 = V 1 J ( a ) (132)
J ( a ) U 0 = U 0 J ( a ) , J ( a ) V 0 = V 0 J ( a ) . (133) J(a)U_0 = U_0J(a), \qquad J(a)V_0 = V_0J(a). \tag*{(133)} J ( a ) U 0 = U 0 J ( a ) , J ( a ) V 0 = V 0 J ( a ) . (133)
Consequently, under the functional equation,
U 1 J ^ ( a ) V 1 = U 0 J ( a ) V 0 . U_1\widehat{J}(a)V_1 = U_0J(a)V_0. U 1 J ( a ) V 1 = U 0 J ( a ) V 0 .
For real factors with q = p q=p q = p , R i = L i T R_i=L_i^{\mathsf{T}} R i = L i T , H n = E n T H_n=E_n^{\mathsf{T}} H n = E n T and real symmetric positive definite forms on real spaces at every level, real orthonormal blocks make J ( a ) J(a) J ( a ) , J ^ ( a ) \widehat{J}(a) J ( a ) symmetric for real a a a and give V i = U i T V_i=U_i^{\mathsf{T}} V i = U i T .
Under the functional equation,
U 1 U 1 T = U 0 U 0 T , U 1 J ^ ( a ) U 1 T = U 0 J ( a ) U 0 T . (134) U_1U_1^{\mathsf{T}}=U_0U_0^{\mathsf{T}}, \qquad U_1\widehat{J}(a)U_1^{\mathsf{T}}=U_0J(a)U_0^{\mathsf{T}}. \tag*{(134)} U 1 U 1 T = U 0 U 0 T , U 1 J ( a ) U 1 T = U 0 J ( a ) U 0 T . (134)
All products used here are finite entry by entry.
Proof. The dual blocks are constructed inductively by eliminating their pairings with the preceding levels. Denote the preceding row blocks by B < n B_{<n} B < n and their dual columns by A < n A_{<n} A < n , so that β ( B < n , A < n ) = I \beta(B_{<n},A_{<n})=I β ( B < n , A < n ) = I . If rows b + b^{+} b + complete the row basis to E n E_n E n and columns a + a^{+} a + complete the column basis to H n H_n H n , the additional Gram blocks and their Schur complement are
C : = β ( b + , A < n ) , D : = β ( B < n , a + ) , G : = β ( b + , a + ) , S : = G − C D . C:=\beta(b^{+},A_{<n}), \qquad D:=\beta(B_{<n},a^{+}), \qquad G:=\beta(b^{+},a^{+}), \qquad S:=G-CD. C := β ( b + , A < n ) , D := β ( B < n , a + ) , G := β ( b + , a + ) , S := G − C D .
Subtracting the components detected by the preceding dual bases gives
B n : = b + − C B < n , A n 0 : = a + − A < n D . B_n:=b^{+}-CB_{<n}, \qquad A_n^{0}:=a^{+}-A_{<n}D. B n := b + − C B < n , A n 0 := a + − A < n D .
These subtractions ensure that β ( B n , A < n ) = C − C = 0 \beta(B_n,A_{<n})=C-C=0 β ( B n , A < n ) = C − C = 0 and β ( B < n , A n 0 ) = D − D = 0 \beta(B_{<n},A_n^{0})=D-D=0 β ( B < n , A n 0 ) = D − D = 0 , while expansion of the four terms in the pairing between the new blocks gives
β ( B n , A n 0 ) = G − C D − C D + C D = G − C D = S . \beta(B_n,A_n^{0})=G-CD-CD+CD=G-CD=S. β ( B n , A n 0 ) = G − C D − C D + C D = G − C D = S .
The same elimination applied to the full Gram matrix reads
[ I 0 − C I ] [ I D C G ] [ I − D 0 I ] = [ I 0 0 S ] . \begin{bmatrix} I & 0 \\ -C & I \end{bmatrix}
\begin{bmatrix} I & D \\ C & G \end{bmatrix}
\begin{bmatrix} I & -D \\ 0 & I \end{bmatrix}
=
\begin{bmatrix} I & 0 \\ 0 & S \end{bmatrix}. [ I − C 0 I ] [ I C D G ] [ I 0 − D I ] = [ I 0 0 S ] .
Since the full Gram matrix is invertible by normality, its Schur complement S S S is invertible, and the normalization A n : = A n 0 S − 1 A_n:=A_n^{0}S^{-1} A n := A n 0 S − 1 makes the new blocks dual. At the initial level n = 0 n=0 n = 0 , the lower arrays are empty and S = G S=G S = G , so the same argument starts the induction. Applying it to each pairing constructs the two dual polynomial systems.
The multiplication matrix can now be read from these dual bases. The expansion of a B n aB_n a B n has no block above n + d L ( a ) n+d_L(a) n + d L ( a ) , and pairing with A m A_m A m gives its coefficient J ( a ) [ n , m ] J(a)[n,m] J ( a ) [ n , m ] . The other side of the band follows from balance, since β ( a B n , A m ) = β ( B n , a A m ) = 0 \beta(aB_n,A_m)=\beta(B_n,aA_m)=0 β ( a B n , A m ) = β ( B n , a A m ) = 0 when m + d R ( a ) < n m+d_R(a)<n m + d R ( a ) < n . Moreover, in the column expansion a A m = ∑ n A n C n , m aA_m=\sum_n A_nC_{n,m} a A m = ∑ n A n C n , m , pairing with B n B_n B n gives C n , m = β ( B n , a A m ) = J ( a ) [ n , m ] C_{n,m}=\beta(B_n,aA_m)=J(a)[n,m] C n , m = β ( B n , a A m ) = J ( a ) [ n , m ] . Thus the same matrix, with the two band bounds just established, represents multiplication in both arrays. Its compatibility with products follows from the successive expansions
a b B = a ( J ( b ) B ) = J ( b ) J ( a ) B . ab\mathcal{B}=a\bigl(J(b)\mathcal{B}\bigr)=J(b)J(a)\mathcal{B}. ab B = a ( J ( b ) B ) = J ( b ) J ( a ) B .
Interchanging a , b a,b a , b and using uniqueness of polynomial expansions therefore gives J ( a b ) = J ( a ) J ( b ) = J ( b ) J ( a ) J(ab)=J(a)J(b)=J(b)J(a) J ( ab ) = J ( a ) J ( b ) = J ( b ) J ( a ) ; uniqueness also gives additivity and J ( 1 ) = I J(1)=I J ( 1 ) = I . To prove faithfulness, suppose that J ( a ) = 0 J(a)=0 J ( a ) = 0 . Exhaustion then implies that multiplication by a a a annihilates every polynomial row, including each constant coordinate row, and hence a = 0 a=0 a = 0 in A A A . This argument applies even when the algebra has zero divisors, and identifies the polynomial relations among the coordinate matrices with the defining relations of the support. The same reasoning applies to the transformed bases.
For the denominator expansions, duality gives the explicit block calculation
β ^ ( B n L 1 , R 1 A m ) = ∑ j , l U 1 [ n , j ] β ^ ( B ^ j , A ^ l ) V 1 [ l , m ] = ∑ j U 1 [ n , j ] V 1 [ j , m ] = ( U 1 V 1 ) [ n , m ] . \begin{aligned}
\widehat{\beta}(B_nL_1,R_1A_m)=\sum_{j,l}U_1[n,j]\widehat{\beta}(\widehat{B}_j,\widehat{A}_l)V_1[l,m] \\
&=\sum_j U_1[n,j]V_1[j,m]=(U_1V_1)[n,m].
\end{aligned} β ( B n L 1 , R 1 A m ) = j , l ∑ U 1 [ n , j ] β ( B j , A l ) V 1 [ l , m ] = j ∑ U 1 [ n , j ] V 1 [ j , m ] = ( U 1 V 1 ) [ n , m ] .
Both sums are finite because both polynomial arguments have finite expansions. With the original dual bases, the corresponding calculation gives ( U 0 V 0 ) [ n , m ] (U_0V_0)[n,m] ( U 0 V 0 ) [ n , m ] , so equality of the two products is precisely
β ^ ( B n L 1 , R 1 A m ) = β ( B n L 0 , R 0 A m ) ( n , m ≥ 0 ) . \widehat{\beta}(B_nL_1,R_1A_m)=\beta(B_nL_0,R_0A_m)\qquad(n,m\ge0). β ( B n L 1 , R 1 A m ) = β ( B n L 0 , R 0 A m ) ( n , m ≥ 0 ) .
Since the bases exhaust the modules, this equality is equivalent to (2.2). By balance, testing that identity on constant coordinate rows multiplied by arbitrary elements of A A A recovers (2.1) on every matrix test, proving the asserted equivalence.
Each intertwining identity comes from two expansions of the same polynomials. Keeping the order of all matrix factors, these expansions are
a ( B L 1 ) = J ( a ) U 1 B ^ = U 1 J ^ ( a ) B ^ , a ( R 1 A ) = A ^ J ^ ( a ) V 1 = A ^ V 1 J ( a ) , a ( B L 0 ) = J ( a ) U 0 B = U 0 J ( a ) B , a ( R 0 A ) = A J ( a ) V 0 = A V 0 J ( a ) . \begin{aligned}
a(\mathcal{B}L_1)&=J(a)U_1\widehat{\mathcal{B}}=U_1\widehat{J}(a)\widehat{\mathcal{B}},\\
a(R_1\mathcal{A})&=\widehat{\mathcal{A}}\widehat{J}(a)V_1=\widehat{\mathcal{A}}V_1J(a),\\
a(\mathcal{B}L_0)&=J(a)U_0\mathcal{B}=U_0J(a)\mathcal{B},\\
a(R_0\mathcal{A})&=\mathcal{A}J(a)V_0=\mathcal{A}V_0J(a).
\end{aligned} a ( B L 1 ) a ( R 1 A ) a ( B L 0 ) a ( R 0 A ) = J ( a ) U 1 B = U 1 J ( a ) B , = A J ( a ) V 1 = A V 1 J ( a ) , = J ( a ) U 0 B = U 0 J ( a ) B , = A J ( a ) V 0 = A V 0 J ( a ) .
These equalities use only the fact that a a a , being a scalar in A A A , commutes with the polynomial matrix entries. Uniqueness of each row or column expansion proves the corresponding intertwining or commutation identity. Together with (8.2), they give
U 1 J ^ ( a ) V 1 = J ( a ) U 1 V 1 = J ( a ) U 0 V 0 = U 0 J ( a ) V 0 . U_1\widehat{J}(a)V_1=J(a)U_1V_1=J(a)U_0V_0=U_0J(a)V_0. U 1 J ( a ) V 1 = J ( a ) U 1 V 1 = J ( a ) U 0 V 0 = U 0 J ( a ) V 0 .
The finite rows of U i U_i U i , finite columns of V i V_i V i and finite bands of J ( a ) J(a) J ( a ) , J ^ ( a ) \widehat{J}(a) J ( a ) justify these sums and regroupings entry by entry, without taking the inverse of a semi-infinite matrix.
Finally, in the positive real case take A n = B n T A_n=B_n^{\mathsf T} A n = B n T and A ^ n = B ^ n T \widehat{A}_n=\widehat{B}_n^{\mathsf T} A n = B n T using orthonormal blocks. Balance and symmetry give β ( a B n , B m T ) = β ( a B m , B n T ) T \beta(aB_n,B_m^{\mathsf T})=\beta(aB_m,B_n^{\mathsf T})^{\mathsf T} β ( a B n , B m T ) = β ( a B m , B n T ) T . This proves symmetry of the multiplication matrices. In addition, transposition turns the left expansions in (8.1) into the right expansions, whose uniqueness gives V i = U i T V_i=U_i^{\mathsf T} V i = U i T . Substitution in the identities already proved yields (8.5). □
On the polynomial line A = C [ z ] A=\mathbb{C}[z] A = C [ z ] , the band bounds have a direct interpretation in terms of the numbers of components. Order the row monomials as
e 1 T , … , e q T , z e 1 T , … , z e q T , z 2 e 1 T , … , e_1^{\mathsf T},\ldots,e_q^{\mathsf T},ze_1^{\mathsf T},\ldots,ze_q^{\mathsf T},z^2e_1^{\mathsf T},\ldots, e 1 T , … , e q T , z e 1 T , … , z e q T , z 2 e 1 T , … ,
where e i e_i e i are the coordinate columns of C q \mathbb{C}^q C q , and order the column monomials in the same way with the p p p coordinate columns of C p \mathbb{C}^p C p . Let E n , H n E_n,H_n E n , H n consist of the first n + 1 n+1 n + 1 monomials in their respective lists. Assume, as above, that both pairings are normal on every such pair of spaces. Multiplication by z z z shifts the row list by q q q positions and the column list by p p p positions. Hence both recurrence matrices vanish outside n − p ≤ m ≤ n + q n-p\le m\le n+q n − p ≤ m ≤ n + q .
The two coordinate identities on an elliptic curve
Two choices of spaces on the same curve exhibit the effects of different row and column component counts and positivity. The rectangular calculation assumes the hypotheses of Theorem 8.1 and determines the resulting bands and identities. The square calculation uses the explicit positive matrix measure of Subsection 7.4, whose positivity supplies normality at every level.
The coordinate algebra is A = C [ x , y ] / ( y 2 − x 3 + x − 1 ) A=\mathbb{C}[x,y]/(y^2-x^3+x-1) A = C [ x , y ] / ( y 2 − x 3 + x − 1 ) . Its smooth projective completion has one point O = [ 0 : 1 : 0 ] O=[0:1:0] O = [ 0 : 1 : 0 ] at infinity in projective coordinates [ X : Y : Z ] [X:Y:Z] [ X : Y : Z ] , where the equation is Y 2 Z = X 3 − X Z 2 + Z 3 Y^2Z=X^3-XZ^2+Z^3 Y 2 Z = X 3 − X Z 2 + Z 3 . In the chart Y = 1 Y=1 Y = 1 , put u : = X / Y u:=X/Y u := X / Y and v : = Z / Y v:=Z/Y v := Z / Y . Near O O O the equation is
v ( 1 + u v − v 2 ) = u 3 . v(1+uv-v^2)=u^3. v ( 1 + uv − v 2 ) = u 3 .
The factor in parentheses is invertible at O O O , where it equals one. Moreover, the derivative of the left side minus u 3 u^3 u 3 with respect to v v v is one at O O O , so u u u is a local parameter and the equation forces v v v to have order three. It follows that x = u / v x=u/v x = u / v has a pole of order two and y = 1 / v y=1/v y = 1/ v a pole of order three. These orders determine the spaces L ( m O ) L(mO) L ( m O ) of rational functions regular away from O O O and with a pole of order at most m m m there: a basis of L ( m O ) L(mO) L ( m O ) consists of the monomials x a x^a x a with 2 a ≤ m 2a\le m 2 a ≤ m and x a y x^ay x a y with 2 a + 3 ≤ m 2a+3\le m 2 a + 3 ≤ m , where a ≥ 0 a\ge0 a ≥ 0 . Indeed, every element of A A A has a unique expression p ( x ) + y q ( x ) p(x)+yq(x) p ( x ) + y q ( x ) , so the monomials are independent; their distinct pole orders then characterize the bound without cancellation. For m ≥ 3 m\ge3 m ≥ 3 , the two lists contain respectively ⌊ m / 2 ⌋ + 1 \lfloor m/2\rfloor+1 ⌊ m /2 ⌋ + 1 and ⌊ ( m − 3 ) / 2 ⌋ + 1 \lfloor(m-3)/2\rfloor+1 ⌊( m − 3 ) /2 ⌋ + 1 monomials. Their sum is m m m , as is seen by taking m m m even and odd separately. For m = 1 , 2 m=1,2 m = 1 , 2 the bases are 1 1 1 and 1 , x 1,x 1 , x . Thus dim L ( m O ) = m \dim L(mO)=m dim L ( m O ) = m for every m ≥ 1 m \ge1 m ≥ 1 .
For the rectangular case, take two row components and three column components with spaces
X ℓ = L ( 3 ℓ O ) 1 × 2 , Y ℓ = L ( 2 ℓ O ) 3 × 1 , ℓ ≥ 1. X_\ell=L(3\ell O)^{1\times2}, \qquad Y_\ell=L(2\ell O)^{3\times1}, \qquad\ell\ge1. X ℓ = L ( 3 ℓ O ) 1 × 2 , Y ℓ = L ( 2 ℓ O ) 3 × 1 , ℓ ≥ 1.
The dimension calculation above shows that both spaces have dimension 6 ℓ 6\ell 6 ℓ ; increasing the pole bound makes them nested and exhaustive in the polynomial row and column modules. Set E n = X n + 1 E_n=X_{n+1} E n = X n + 1 and H n = Y n + 1 H_n=Y_{n+1} H n = Y n + 1 , so the block index is n = ℓ − 1 n=\ell-1 n = ℓ − 1 and every block, including the initial one, has size six. Suppose the two balanced pairings and the four polynomial matrix factors satisfy the hypotheses and the functional equation of Theorem 8.1 on these spaces. In particular, both pairings are normal at every level.
The shifts in the recurrence follow by comparing the poles of the coordinates with the increments in the chosen filtrations. Multiplication by x x x adds at most two to a pole order, whereas multiplication by y y y adds at most three. One level on the left allows three additional orders, and hence accommodates either coordinate. On the right one level allows two additional orders, so x x x requires one level and y y y requires two. Consequently,
x X ℓ ⊂ X ℓ + 1 , x Y ℓ ⊂ Y ℓ + 1 , y X ℓ ⊂ X ℓ + 1 , y Y ℓ ⊂ Y ℓ + 2 . xX_\ell\subset X_{\ell+1}, \qquad xY_\ell\subset Y_{\ell+1}, \qquad yX_\ell\subset X_{\ell+1}, \qquad yY_\ell\subset Y_{\ell+2}. x X ℓ ⊂ X ℓ + 1 , x Y ℓ ⊂ Y ℓ + 1 , y X ℓ ⊂ X ℓ + 1 , y Y ℓ ⊂ Y ℓ + 2 .
The theorem now gives block ranges n − 1 ≤ m ≤ n + 1 n-1 \le m \le n+1 n − 1 ≤ m ≤ n + 1 for J ( x ) J(x) J ( x ) , J ^ ( x ) \widehat{J}(x) J ( x ) and n − 2 ≤ m ≤ n + 1 n-2 \le m \le n+1 n − 2 ≤ m ≤ n + 1 for J ( y ) J(y) J ( y ) , J ^ ( y ) \widehat{J}(y) J ( y ) . Since these matrices represent multiplication in A A A , the two coordinate matrices commute and satisfy
J ( y ) 2 = J ( x ) 3 − J ( x ) + I , J ^ ( y ) 2 = J ^ ( x ) 3 − J ^ ( x ) + I . J(y)^2=J(x)^3-J(x)+I, \qquad\widehat{J}(y)^2=\widehat{J}(x)^3-\widehat{J}(x)+I. J ( y ) 2 = J ( x ) 3 − J ( x ) + I , J ( y ) 2 = J ( x ) 3 − J ( x ) + I .
The four connections in (130) are defined by the matrix factors and the polynomial bases; their definition does not involve a choice of coordinate. Thus the same U 0 , U 1 , V 0 , V 1 U_0,U_1,V_0,V_1 U 0 , U 1 , V 0 , V 1 satisfy all four intertwining and commutation identities (8.3)–(8.4) for both x x x and y y y . In particular,
U 1 J ^ ( x ) V 1 = U 0 J ( x ) V 0 , U 1 J ^ ( y ) V 1 = U 0 J ( y ) V 0 . U_1\widehat{J}(x)V_1=U_0J(x)V_0, \qquad U_1\widehat{J}(y)V_1=U_0J(y)V_0. U 1 J ( x ) V 1 = U 0 J ( x ) V 0 , U 1 J ( y ) V 1 = U 0 J ( y ) V 0 .
The bilateral transformation therefore acts simultaneously on the two multiplication operators and respects their elliptic relation.
The positive realization is square: both arguments have two components and are paired by the matrix measures in Subsection 7.4. Use X ℓ = L ( 3 ℓ O ) 1 × 2 X_\ell=L(3\ell O)^{1\times2} X ℓ = L ( 3 ℓ O ) 1 × 2 and Y ℓ = X ℓ T Y_\ell=X_\ell^{\mathsf T} Y ℓ = X ℓ T . The densities and point masses specified there are balanced, and positivity guarantees normality of both pairings at every level. The common pole bound increases by three on each side. Since either coordinate increases the pole order by at most three, all four multiplication matrices have block range n − 1 ≤ m ≤ n + 1 n-1 \le m \le n+1 n − 1 ≤ m ≤ n + 1 , again with blocks of size six. Choosing real orthonormal blocks also makes both coordinate matrices symmetric in each system, so the same connections have V i = U i T V_i=U_i^{\mathsf T} V i = U i T and satisfy
U 1 U 1 T = U 0 U 0 T , U 1 J ^ ( a ) U 1 T = U 0 J ( a ) U 0 T ( a ∈ { x , y } ) . U_1U_1^{\mathsf T}=U_0U_0^{\mathsf T}, \qquad U_1\widehat{J}(a)U_1^{\mathsf T}=U_0J(a)U_0^{\mathsf T} \qquad(a\in\{x,y\}). U 1 U 1 T = U 0 U 0 T , U 1 J ( a ) U 1 T = U 0 J ( a ) U 0 T ( a ∈ { x , y }) .
These are the positive four-factor Darboux identities for the two coordinates of the explicit elliptic weight, with any of the admissible positive masses specified there.
Conclusions
The bilateral functional equation leads to a determinant construction for mixed-type orthogonality on algebraic supports. Starting from the original biorthogonal data, the ordered matrix factors, the rational moments required by division and a chosen homogeneous mass, the construction expresses the transformed orthogonality conditions as a finite linear system. Its image conditions preserve the relations of the support and the prescribed component bounds. Consequently, the determinant of the reduced system characterizes normality, while its bordered determinants give the normalized polynomials and their reproducing kernel.
The functional equation also determines the full freedom in the transformed pairing. For balanced masses with regular denominator determinants, its exact sequence distinguishes the obstruction to unilateral decomposition from the ambiguity of the summands. Common components of the denominator zero sets can support mass directions that require the joint bilateral equation.
The size of these determinants reflects both the transformation and the polynomial spaces on which it acts. In the degree construction, the Hilbert function counts the independent polynomial classes on the support, and the finite mass rank accounts for the contribution of the chosen mass. Polynomial rank computations then remove redundant directions and give the size of the smaller system. These counts explain why fixed perturbations, under compatible growth of the component degrees, have bounded determinant orders on curves and at most linear orders on surfaces. The formulas based on consecutive scalar degrees, matrix division and reduced linear systems eliminate different polynomial spaces; their respective counts specify the sizes of those representations under the stated hypotheses.
To form these systems explicitly, the quotient conditions must themselves be expressed in computable coordinates. On integral curves, dualizing differentials and the geometric Cauchy kernel supply dual coordinates for matrix remainders, including at singular points. On surfaces, an additional scalar equation selects a finite quotient, and ordered matrix divided differences reconstruct each quotient class as a finite sum of its residue pairings. Applied to the respective quotients, these coordinates describe the masses and the divisibility tests required by the polynomial formulas. The explicit curve family also shows how this description behaves when the mass support reaches a singular point. The residue quotients remain nondegenerately paired and an explicit minor proves that the matrix mass keeps its rank. Normality of the finite Gram matrices, proved for the stated component bounds, then gives regular limits of the normalized polynomials and kernels.
For positive symmetric congruences, the intersection of the original space with its rationally transformed counterpart permits a further reduction. The resulting two positive matrices determine both the normalized polynomials and the kernel; the coprimality count and the nodal matrix family identify situations in which positive masses, including bivariate derivative masses, leave the reduced order unchanged. This reduction applies to both balanced forms and the nonbalanced Sobolev forms exhibited at the node, because the orthogonal decomposition uses the positive forms on the specified finite spaces. With equal component bounds on the line, this construction assembles the polynomial rows into monic matrix orthogonal polynomials and yields their matrix norms and Christoffel–Darboux kernel. In particular, the triangular Hermite weights of Durán and Grünbaum illustrate how the common space determines the small orders of the corresponding formulas.
The one-variable and Laurent reductions connect the general formulas with their standard specializations, while the description of positive masses identifies the admissible additions in the symmetric setting. When both the original and transformed pairings are balanced, their polynomial bases also give four-factor intertwining identities for coordinate multiplication. These bilateral Darboux identities use the same connection matrices for every coordinate and preserve the defining relations of the support. The explicit positive elliptic matrix measure supplies a square realization in which both coordinate matrices are symmetric. Thus, under the balance and normality hypotheses of the recurrence construction, the polynomial and kernel formulas determine the transformed recurrence matrices with the original order of the four matrix factors.
Funding
This work was supported by the Spanish Ministry of Science, Innovation and Universities and the Spanish State Research Agency [grant number PID2024-155133NB-I00], co-funded by the European Union through the European Regional Development Fund.
Data availability
The article gives the matrix factors, masses, polynomial spaces, coefficient formulas and finite linear systems used in its constructions. The accompanying verification package supplies exact symbolic checks of the examples and of the matrix identities. The general arguments are proved in the text; tests at individual degrees supplement those proofs. No external datasets are used.
Declaration of AI use
During the preparation of this manuscript, the author used ChatGPT and Codex (OpenAI) and Claude (Anthropic) as aids in exposition, mathematical exploration, and the discussion and review of mathematical arguments. These tools also assisted with computational checks, the development of formal verification code, and the review of notation and references.
References