External zonotopal algebra Nan Li (MIT) Joint work with Amos Ron (UW-Madison) Sep 24 2011, @ Wake Forest University Outline • From box splines to central zonotopal algebra • External zonotopal algebra: • focusing on a geometric interpretation of the duality. Starting from the box splines Given a list X of m vectors in Rd , consider the variable polytope Π1X (t) = {y ∈ [0, 1]m | X · y = t}. Define the box spline: Bx (t) = Vol Π1X (t). It is well known that Bx (t) is a piece-wise polynomial in t = (t1 , . . . , td ), supported on the zonotope Z (X ). Example Let X = { 10 , 01 , 11 }. Then the box spline Bx (t) = Vol{(y1 , y2 , y3 ) ∈ [0, 1]3 | y1 10 + y2 01 + y3 11 = t2 O −t2 + 2 t − t 1 2 −t + 2 +1 1 t2 − t1 t1 +1 t2 o (1, 0) / t1 t1 t2 } is Starting from the box splines Given a list X of m vectors in Rd , consider the variable polytope Π1X (t) = {y ∈ [0, 1]m | X · y = t}. Define the box spline: Bx (t) = Vol Π1X (t). It is well known that Bx (t) is a piece-wise polynomial in t = (t1 , . . . , td ), supported on the zonotope Z (X ). Example Let X = { 10 , 01 , 11 }. Then the box spline Bx (t) = Vol{(y1 , y2 , y3 ) ∈ [0, 1]3 | y1 10 + y2 01 + y3 11 = t2 O −t2 + 2 t − t 1 2 −t + 2 +1 1 t2 − t1 t1 +1 t2 o (1, 0) / t1 t1 t2 } is D-space and J -ideal Define the D-space to be the linear space in C[t1 , . . . , td ] generated by all the polynomials in Bx (t) and their partial derivatives. Example In the previous example, we have D(X ) = span {t1 , t2 , t1 −t2 +1, t2 −t1 +1, −t2 +2, −t1+2, 1} = span {1, t1 , t2 }. There is an equivalent definition using inverse systems. Given a list of vectors X , first define the ideal of cocircuits: Y J (X ) = Ideal{ ℓx | Z ⊂ X , Z ∩ B 6= ∅, ∀B ∈ B(X )}, then let x∈Z D(X ) = ker J (X ) := {g ∈ C[t1 , . . . , td ] | hf , g i = 0, ∀f ∈ J (X )}, where the inner product is defined as ∂ hf , g i = f g |t=0 . ∂t D-space and J -ideal Define the D-space to be the linear space in C[t1 , . . . , td ] generated by all the polynomials in Bx (t) and their partial derivatives. Example In the previous example, we have D(X ) = span {t1 , t2 , t1 −t2 +1, t2 −t1 +1, −t2 +2, −t1+2, 1} = span {1, t1 , t2 }. There is an equivalent definition using inverse systems. Given a list of vectors X , first define the ideal of cocircuits: Y J (X ) = Ideal{ ℓx | Z ⊂ X , Z ∩ B 6= ∅, ∀B ∈ B(X )}, then let x∈Z D(X ) = ker J (X ) := {g ∈ C[t1 , . . . , td ] | hf , g i = 0, ∀f ∈ J (X )}, where the inner product is defined as ∂ hf , g i = f g |t=0 . ∂t D-space and J -ideal Example As in the previous examples, take X = { 10 , 01 , 11 }. Then 1 0 1 1 0 1 B(X ) = , , , , , . 0 1 0 1 1 1 So all the possible Z ’s such that Z ⊂ X , Z ∩ B 6= ∅, ∀B ∈ B(X ) are 1 0 1 1 0 1 1 0 1 , , , , , and , , . 0 1 0 1 1 1 0 1 1 Therefore, J (X ) = Ideal{t1 t2 , t1 (t1 +t2 ), t2 (t1 +t2 ), t1 t2 (t1 +t2 )} = Ideal{t1 t2 , t12 , t22 }, and D(X ) = ker J (X ) = span {1, t1 , t2 }, the same as computed by the definition from box splines. Hyperplane arrangement and I-ideal Example For X = { 1 0 0 1 , 1 , 1 }, consider the hyperplane arrangement ?? O t2 ?? ?? ?? H2 ?? / t1 ?? H3 H 1 where each vector in X defines the normal vector of one hyperplane. Then define a power ideal I = Ideal{t12 , t22 , (t1 − t2 )2 }, where • each generator corresponds to a line in the hyperplane arrangement; • and its power is the number of hyperplanes in the arrangement that do not contain that line. In this case I(X ) = Ideal{t12 , t22 , t1 t2 }. I-ideal and P-space From the power ideal I, define P(X ) = ker I(X ). In the above example, I = Ideal{t12 , t22 , t1 t2 }, so we have P(X ) = span {1, t1, t2 }. Theorem (de Boor-Dyn-Ron) P(X ) = span { Y ℓx | Z ⊂ X , rank(X \Z ) = rank(X )}. x∈Z As in the example, for X = { 10 , 01 , 11 }, all the possible Z ’s are: 1 0 1 ∅, , , , 0 1 1 so P(X ) = span {1, t1, t2 , t1 + t2 } = span {1, t1, t2 }, the same as we computed from the power ideal I. I-ideal and P-space From the power ideal I, define P(X ) = ker I(X ). In the above example, I = Ideal{t12 , t22 , t1 t2 }, so we have P(X ) = span {1, t1, t2 }. Theorem (de Boor-Dyn-Ron) P(X ) = span { Y ℓx | Z ⊂ X , rank(X \Z ) = rank(X )}. x∈Z As in the example, for X = { 10 , 01 , 11 }, all the possible Z ’s are: 1 0 1 ∅, , , , 0 1 1 so P(X ) = span {1, t1, t2 , t1 + t2 } = span {1, t1, t2 }, the same as we computed from the power ideal I. Central zonotopal algebra Given a list a vectors X , construct • group one: cocircuit ideal J → D-space; • group two: power ideal I → P-space. • The two groups are dual to each other. In particular, the two spaces D and P have the same Hilbert function. Central zonotopal algebra • is not an algebra; • is the idea of connecting a matriod with the geometry (zonotopes, hyperplane arrangement) and the four algebraic objects. External zonotopal algebra Given a list of vectors X , append an extra list of vectors Y to X , and consider some collection of bases B′ ⊂ B(X ∪ Y ). Now define Y J B′ = Ideal{ ℓx | Z ⊂ X ∪ Y , Z ∩ B 6= ∅, ∀B ∈ B′ }. x∈Z and D B′ = ker J B′ . Remark • Introducing this extra Y will help us understand X better; • the definition of D B′ depends on Y , but the properties (e.g. dimension and Hilbert function) we are interested in will be independent of Y . Question • In general, we have dim DB′ ≥ #B′ (de Boor-Ron). When do we have the equality (B′ is coherent)? • How to define the corresponding P-space, which is dual to the D-space? External zonotopal algebra Given a list of vectors X , append an extra list of vectors Y to X , and consider some collection of bases B′ ⊂ B(X ∪ Y ). Now define Y J B′ = Ideal{ ℓx | Z ⊂ X ∪ Y , Z ∩ B 6= ∅, ∀B ∈ B′ }. x∈Z and D B′ = ker J B′ . Remark • Introducing this extra Y will help us understand X better; • the definition of D B′ depends on Y , but the properties (e.g. dimension and Hilbert function) we are interested in will be independent of Y . Question • In general, we have dim DB′ ≥ #B′ (de Boor-Ron). When do we have the equality (B′ is coherent)? • How to define the corresponding P-space, which is dual to the D-space? Coherency It is known that in the following cases, B′ is coherent, i.e., dim DB′ = #B′ . • central case, by Holtz-Ron; • external case, by Holtz-Ron; • semi-external, by Holtz-Ron-Xu. Our set up: Assume that Y = {y1 , y2 , . . .} contains sufficiently many vectors in general position in X ∪ Y . Consider a function κ : 2X → N. We consider a class of B′ = Bκ depending κ. Theorem (L.-Ron) dim DBκ = #Bκ if and only if κ is solid, i.e., span Z ⊂ span Z ′ implies κ(Z ) ≤ κ(Z ′ ), for any Z , Z ′ ⊂ X . Remark We can pick special κ’s to recover the above three cases. Approaching the second question: family of P-spaces It is known that the P-spaces defined as follows are dual to their corresponding D-spaces: • central case, by Holtz-Ron; • external case, by Holtz-Ron, Postnikov-Shapiro-Shapiro; • semi-external, by Holtz-Ron-Xu. There are some recent development of external P-spaces from another approach without introducing the D-spaces. They studied the Hilbert functions of the P-spaces and their connection to the power ideals. • (Ardila-Postnikov) For k ≥ 0, define X Y P k (X ) = ℓ x Πk , Z ⊂X x∈X \Z where Πk is the set of polynomials in d variables with degree at most k. • (Lenz) Generalized the results by Holtz-Ron-Xu and Ardila-Postnikov. Defining a unified P-space Here, we generalize the above P-spaces using a solid assignment κ : 2X → N, X Y P κ := ℓx Πκ(Z ) , Z ⊂X x∈X \Z where Πκ(Z ) is the set of polynomials in d variables with degree at most κ(Z ). Theorem (L.-Ron) P κ is dual to D κ if and only if κ is incremental, i.e., for every Z ⊂ X and x ∈ X , we have κ(Z ∪ x) ≤ κ(Z ) + 1. • Assume the assignment κ to be solid and incremental. We studied the homogeneous basis for P κ , and its Hilbert function. Defining a unified P-space Here, we generalize the above P-spaces using a solid assignment κ : 2X → N, X Y P κ := ℓx Πκ(Z ) , Z ⊂X x∈X \Z where Πκ(Z ) is the set of polynomials in d variables with degree at most κ(Z ). Theorem (L.-Ron) P κ is dual to D κ if and only if κ is incremental, i.e., for every Z ⊂ X and x ∈ X , we have κ(Z ∪ x) ≤ κ(Z ) + 1. • Assume the assignment κ to be solid and incremental. We studied the homogeneous basis for P κ , and its Hilbert function. Geometric interpretation of the duality Given a list of vectors X , for example, X = { 10 , 01 }. Append a list Y = {y1 , y2 , y3 } to X and consider the corresponding generic affine hyperplane arrangement. y2 x1 ? ? ? ? ? ? i y3 ? i i i ?i? i i ? i i i i ? x2 i i ? ? ? y 1 Geometric interpretation of the duality Given a list of vectors X , for example, X = { 10 , 01 }. Append a list Y = {y1 , y2 , y3 } to X and consider the corresponding generic affine hyperplane arrangement. y2 x1 ? ? ? ? i y3 ? i i i ?i? i i ? i i i i x2 i •? ? i ? ? y 1 ? ? Each basis B ∈ B(X ∪ Y ) defines a vertex v(B) ∈ Rn , viz, the common zero of the polynomials (qx )x∈B . For example, {x2 , y1 } defines the black vertex. Geometric interpretation of the duality Given a list of vectors X , for example, X = { 10 , 01 }. Append a list Y = {y1 , y2 , y3 } to X and consider the corresponding generic affine hyperplane arrangement. y2 x1 ? ? •? ? i y3 ? i i i•?•i? i i ? i i • i i•• • x2 i • •? ? i ? • ? y 1 ? ? Since the arrangement is generic, there is a bijection between B(X ∪ Y ) and the set of all the vertices (0-dim intersections) of the hyperplane arrangement. Geometric interpretation of the duality Given a list of vectors X , for example, X = { 10 , 01 }. Append a list Y = {y1 , y2 , y3 } to X and consider the corresponding generic affine hyperplane arrangement. y2 x1 ? ? •? ? i y3 ? i i i•?•i? i i ? i i • i i◦• • x2 i ◦ •? ? i ? • ? y 1 ? ? Now in our set up, we choose some Bκ ⊂ B(X ∪ Y ), which corresponds to a subset of vertices v(Bκ ). Geometric interpretation of the duality Given a list of vectors X , for example, X = { 10 , 01 }. Append a list Y = {y1 , y2 , y3 } to X and consider the corresponding generic affine hyperplane arrangement. y2 x1 ? ? •? ? i y3 ? i i i•?•i? i i ? i i • i i◦• • x2 i ◦ •? ? i ? • ? y 1 ? ? Recall that for a solid and incremental κ, we have #v(Bκ ) = #Bκ = dim Pκ = dim Dκ . We will build connections from the above points set v(Bκ ) to both D κ -space and P κ -space. Geometric interpretation of the duality For a solid and incremental κ, let v(Bκ ) be the point set of the vertices in the hyperplane arrangement corresponding to Bκ . Theorem (L.-Ron) • Π(v(Bκ )) = Dκ , where Π(v(Bκ )) is the least space of the set v(Bκ ), where the least space is a polynomial space assigned to a point set, whose dimension is the cardinality of the point set; • there exists a Lagrange basis of Pκ = span {qB | B ∈ Bκ } corresponding to the set v(Bκ ). Namely, for each v(B) ∈ v(Bκ ), qB (v(B)) 6= 0 and qB (v(B ′ )) = 0 for all B ′ 6= B. t2O T A K H N Y U O /t o (1,0) 1 Geometric interpretation of the duality For a solid and incremental κ, let v(Bκ ) be the point set of the vertices in the hyperplane arrangement corresponding to Bκ . Theorem (L.-Ron) • Π(v(Bκ )) = Dκ , where Π(v(Bκ )) is the least space of the set v(Bκ ), where the least space is a polynomial space assigned to a point set, whose dimension is the cardinality of the point set; • there exists a Lagrange basis of Pκ = span {qB | B ∈ Bκ } corresponding to the set v(Bκ ). Namely, for each v(B) ∈ v(Bκ ), qB (v(B)) 6= 0 and qB (v(B ′ )) = 0 for all B ′ 6= B. t2O T A K H N Y U O /t o (1,0) 1