Matrix Completion, G = ([n], E) Free Resolutions, and S → R ⊕R π ∶� (a ) � (a , . . . , a ) ⊕ (a ∶ {i, j} ∈ E) Sums of Squares Matrix Completio Throughout, we fix a simple graph This encodes the coordinate projection n G ij n �� nn . �E ij n : real vector space of symmetric n × n matrices. S Rainer Sinn (Georgia Institute of Technology) NB: Think of πG (a i j) as a partial matrix. � � � � Joint withLet G be the four cycle. Example. Grigoriy Blekherman (Georgia� aInstitute of Technology) a a a �� �� �� �� � � a�� a�� ∗ � a�� a�� de � a�� a�� a�� Mauricio Velasco (Universidad Andes) a�� los a�� � � � � a�� � ∗ � � πG � =� � � a a a a ∗ a a a � �� �� �� �� � � �� �� �� � �a�� a�� a�� a��� �a�� ∗ a�� a��� Throughout, we fix a simple graph G = ([n],Completion E). Matrix This encodes the coordinate projection Throughout,Swe G ([n],Completion E). n fix Matrix → a simple graph Rn ⊕ R=�E πG ∶ � the coordinate projection This encodes (a i j) � (a�� , . . . , a nn ) ⊕Matrix (a i j∶ {i,Completion j} ∈ E) Throughout,Swe a simple graph G R=�E ([n], E). n fix n⊕ → R n S : real space symmetric nG ×=n([n], matrices. Throughout, fix a of simple graph E). πG vector ∶ � we This encodes the coordinate projection (a . . , projection a nnmatrix. ) ⊕ (a i j∶ {i, j} ∈ E) ��a, .partial jG)(a� NB: of(aπithe ) as ThisThink encodes coordinate ij Throughout, we fix a simple graph n n �E S space → of symmetric R ⊕ R n matrices. � S�n : real vector n × n n �E This encodes the coordinate projec πG ∶ � S → R ⊕ R . . , a nnmatrix. ) ⊕ (a i j∶ {i, j} ∈ E) ��a, .partial πG ∶ �of(aπiGj)(a�i j)(a NB: Think as (a i j) � (a�� , . . . , a nn ) ⊕ (a i j∶ {i, j} ∈ E) n → n S R � πG ∶ � S�nn : real vector space of symmetric n × n matrices. (a i j) � (a�� , . . . , a nn ) ⊕ � � S : real vector space of symmetric n × n matrices. NB: Think of πG (a i j) as a partial matrix. Matrix Completion NB: Think�of πG (a i j) as a partial matrix. Sn : real vector space of symmetric � Example. � Let G be the four cycle. � NB: Think of πG (a i j) as a partial m Throughout, we fix a simple graph G = ([n], E). � � a�� coordinate a�� a�� � projection � a�� the � a�� a�� ∗ a�� � This encodes Example. G beatheafour�cycle. ��aLet � an�� a���Ea�� ∗ � � a �� �� �� �� n � � �R ⊕ R � = � πG �� S → � �∗ a a a � πG ∶��aa�� a a a �� �� �� �� a a a a)��⊕a(a aj}�� �� i j )��� �� ��. .�. , a� �� i j ∶∗�� �� ∈� E) (a (a , {i, nn �� � Example. G��beaathe four cycle. � �aaLet � � � a a a ∗ a a �� �� �� �� �� �� a a a a a ∗ �� G��be the �� four �� �cycle. �� �� �� Example. Let � � � n π = S : realG vector of symmetric n × n matrices. � a aspace � � � a a ∗ a a a �� �� �� �� �� �� �� a ∗ � � � � �� ��a G��-partial �� matrix �� to��a positive �� semidefinite symmetric Goal: Complete matrix (and control Example. Let G be the four cycle. NB: Think of π (a ) as a partial matrix. a a a a a a ∗ a � � � � G i j �� a�� �� a�� �� a�� �� � � a�� �� a∗�� a�� a∗ �� �a�� � �� �� �� �� � � �� �� �� ��� the rank of the completion). � � � =� �a�� a�� a�� a�� � � πG � a a a ∗ � � � �� �� �� � � � � � a�� a�� a�� a�� � � a�� a a a a ∗ a a a π = � � � � �� �� �� �� �� �� �� G � � � � Goal: Complete a aG��-partial matrix to��a apositive semidefinite symmetric matrix (and control a a a ∗ a a � � � � �� �� �� �� �� � � a�� Geometry of positive semidefinite matrix completion a �a�� a�� a�� a��� �a�� ∗ a�� a��� �� a�� a�� a�� � � � � π = the rank�ofa�� theacompletion). � � � a a a ∗ a a G �a a a a � � ∗ �� �� �� �� �� �� � �� �� �� �� � � � � n Goal: Describe Completethe a Gimage -partialofmatrix to aSpositive semidefinite symmetric matrix (and control the cone of positive semidefinite quadratic forms under the � a a a a ≥� �� �� �� ��� � a�� Geometry of positive semidefinite matrix completion Goal: Complete a G -partial matrix to a positive semidefinite symmetric matrix (and control the rank of πthe projection . completion). G Example. be the four cycle. the rank of Let theGcompletion). n a �G,under S cone of image quadratic forms x i positive x j such that p j ≥Complete � for all (p .-partial . . , p nthe ) matrix ∈ ∑i,Sjna i jof ∑i, j a iGoal: j p iquadratic Goal: Describe the of the cone semidefinite forms ≥�: convex Goal: Complete a G -partial matrix to a positive semidefinite symmetric matrix (and control Goal: Complete a G -partial matrix to a positive semidefinite symmetric matrix (and control the rank of the completion). the rank of the completion). Geometry of positive semidefinite matrix completion Geometry of positive semidefinite matrix completion n of positive semidefinite quadratic forms under the Goal: Describe the image of the cone S≥� n of positive semidefinite quadratic forms under the Goal: Describe the image of the cone S≥� projection πG . projection πG . n S≥� cone of quadratic forms ∑ i, j a i j x i x j such that ∑ i, j a i j p i p j ≥ � for all (p� , . . . , p n ) ∈ n :: convex S≥� convex cone of quadratic forms ∑ i, j a i j x i x j such that ∑ i, j a i j p i p j ≥ � for all (p� , . . . , p n ) ∈ n Rn . R . Theorem (Diagonalization of Quadratic Forms). A quadratic form q ∈ R[x� , . . . , x n ] is positive Theorem (Diagonalization of Quadratic Forms). A quadratic form q ∈ R[x� , . . . , x n ] is positive semidefinite if and only if it is a sum of squares of linear forms after a change of basis. semidefinite if and only if it is a sum of squares of linear forms after a change of basis. n ) a cone of sums of squares? Question: Is πG (S≥� n ) a cone of sums of squares? Question: Is πG (S≥� Stanley-Reisner Ideals Stanley-Reisner Ideals Recall G = ([n], E) is fixed. Recall G = ([n], E) is fixed. Let IG = �x i x j∶ {i, j} ∈� E� ⊂ R[x� , . . . , x n ] Let IG = �x i x j∶ {i, j} ∈� E� ⊂ R[x� , . . . , x n ] Note: The degree � part of IG is the kernel of πG . Note: The degree � part of IG is the kernel of πG . Then IG is the Stanley-Reisner ideal of the clique complex of the graph: Then IG is the Stanley-Reisner ideal of the clique complex of the graph: simplicial complex ∆ on [n], where H ∈ ∆ if and only if the induced subgraph of G on the simplicial complex ∆ on [n], where H ∈ ∆ if and only if the induced subgraph of G on the vertices in H is complete. vertices in H is complete. So a monomial x H = ∏ i∈H x i is in IG if and only if G�H is not a complete graph, i.e. there are So a monomial x H = ∏ i∈H x i is in IG if and only if G�H is not a complete graph, i.e. there are i, j ∈ H such that {i, j} ∈� E . But then x i x j ∈ IG and x i x j�x H . i, j ∈ H such that {i, j} ∈� E . But then x i x j ∈ IG and x i x j�x H . Subspace Arrangement Subspace Arrangement Question: Is πG (S≥� ) a cone of sums of squares? n Question: (S aa cone of sums of squares? n ) ≥� Question: Is Is π πG (S ) cone of sums of squares? G ≥� Stanley-Reisner Ideals Stanley-Reisner Ideals Stanley-Reisner Ideals Recall G = ([n], E) is fixed. Recall G = ([n], E) is fixed. Let IG = �x i x j∶ {i, j} ∈� E� ⊂ R[x� , . . . , x n ] Let I G = �x ∈�fixed. E� ⊂ R[x� , . . . , x n ] i x j ∶ {i, Recall =degree ([n], E)�j}is Note:GThe part of IG is the kernel of πG . Note: The degree �j}part of⊂IR[x is �the kernel of πG . G Let I = �x x ∶ {i, ∈ � E� , . . . , x ] n i G j Then IG is the Stanley-Reisner ideal of the clique complex of the graph: Then IIG is the ideal clique complex of graph: Then the Stanley-Reisner Stanley-Reisner ideal of ofHthe the complex of the the graph: subgraph of G on the G is complex simplicial ∆ on [n], where ∈ clique ∆ if and only if the induced simplicial complex ∆ on [n] ,, where H ∈∈ ∆ if and only if the induced subgraph of G on the simplicial complex ∆ on [n] where H ∆ if and only if the induced subgraph of G on the vertices in H is complete. vertices in vertices in H H is is complete. complete. So a monomial x H = ∏i∈H x i is in IG if and only if G�H is not a complete graph, i.e. there are So aa monomial xx H = xx i is in IIG if and only if G� is not aa complete graph, i.e. there are ∏ H i∈H So monomial = is in if and only if G� is not complete graph, i.e. there are ∏ i thenGx i x j ∈ IG and x i xH i, j ∈ H such that H{i, j} ∈�i∈H E . But j �x H . i, and x x �x .. i, jj ∈∈ H H such such that that {i, {i, j} j} ∈∈�� E E .. But But then then xx ii xx jj ∈∈ IIG G and x ii x jj �x H H Subspace Arrangement Note: The degree � part of IG is the kernel of π G. Subspace Arrangement The quadratic form �x i x j corresponds to the symmetric matrix E i j + E ji . Set XG = V(IG ) ⊂ Pn−� , the subscheme associated with the Stanley-Reisner ideal IG . n−� Set XG = V(IG ) ⊂ P , the subscheme associated with the Stanley-Reisner ideal IG . Subspace Arrangement XG = � span{x i ∶ i ∈ K}, XG = K⊂G � span{x i ∶ i ∈ K}, n−� � Set XG = V(I , the subscheme associated with the Stanley-Reisner ideal IG . K⊂G G) ⊂ P � Subspace Arrangements Subspace Arrangements Throughout, we fix a simple graph Set XG = V(IG ) ⊂ Pn−�, the subscheme associated with the Stanley-Reisner ideal IGcoordinate . This encodes the proje Set XG = V(IG ) ⊂ Pn−�, the subscheme associated with the Stanley-Reisner ideal IG . Subspace Arrangements n n S → R Subspace Arrangements πG ∶ � XG = � span{e ∶ i ∈ K}, (a i j) � (a�� , . . . , a nn ) i n−� Set XX ⊂ P ,∶ the subscheme associated with the Stanley-Reisner ideal IG . G ==V(I G )span{e K⊂G i ∈ K}, � n−� Set XGG= V(IG ) ⊂ P i , the subscheme associated with the Stanley-Reisner idealspace IG . of symmetric n : real vector S K⊂Gover all complete subgraphs of G . where K runs NB: Think of πG (a i j) as a partial m whereXG K =runs over all complete subgraphs of G . � span{e i ∶ i ∈ K}, Example. Let be thei four cycle. � � XG = K⊂G ∶ i ∈ K}, � Gspan{e Example. Let fourXcycle. K⊂G Then IG = �x xG , xbe� xthe � and is the union of four lines in P�. � � � G where K runs over all complete subgraphs of G . �. Then I = �x x , x x � and X is the union of four lines in P � � � � G G where K runs over all complete subgraphs of G . Observation: The projection πG is the degree � part of the map of coordinate rings R[x� , . . . , x n ] → � � Example. Let G be the four cycle. Observation: projection πG is the degree � part of the map of coordinate rings R[x� , . . . , x n ] → R[X ] . Let The Example. G be the four cycle. � G Then IG = �x� x� , x� x�� and XG is the union of four lines in P . �. R[X ThenGI]G. = �x� x� , x� x�� and XG is the union of four lines in P Example. Let G beisthe four cycle. n ) of the cone of positive semidefinite Proposition. The image π (S quadratic forms the cone G π≥�is the degree � part of the map of coordinate rings R[x , . . . , x ] → Observation: The projection n G ) of the cone of positive semidefinite quadratic forms is�the cone n Proposition. The image π (S Σ XG of].all quadratic forms inG R[X are sums of of linear forms.� arings Observation: The projection π≥� is degree � partofofsquares the map coordinate R[x a a � that � � a� n] → G ]the � , . .a.��, x G �� �� �� R[X G Σ XG Gof].all quadratic forms in R[XG ]� that are sums of squares of linear forms.� a a a a � � a R[X �� �� �� �� � � � � π =� n G � a forms � Proposition. TheSums imageof πGSquares (S≥� ) of the cone of positive semidefinite quadratic is the cone a�� a�� a�� � � ∗ and Nonnegative Polynomials � ��forms n ) of the Proposition. The image π (S cone of positive semidefinite quadratic is the cone G ≥� Σ XG of all quadratic forms in R[X ] that are sums of squares of linear forms. Sums of Squares G � and Nonnegative Polynomials �a�� a�� a�� a��� �a� Σ XG of all quadratic forms in R[XG ]� that are sums of squares of linear forms. Observation: Every sum of squares of linear forms is a nonnegative quadric. Goal: Complete a G -partial matri Observation:r Every sumof of linear forms is a nonnegative quadric. Sums Squares and Nonnegative Polynomials rof squares � the rank of the completion). of and Nonnegative Polynomials r ℓ�Sums r (ℓSquares q(p) = � (p) = (p)) � i �i (p) = � q(p) = ℓ (ℓ (p)) � � i=� i=� i Observation: Every sum of squares of linear forms is a nonnegative quadric. i i=�Every sumi=� Observation: of squares of linear forms is a nonnegative quadric. Geometry of po r of nonnegative quadrics. We write PXG rfor the cone rfor r (ℓ We write of inonnegative quadrics. q(p)P=XG� ℓ�ithe (p)cone =� (p))� Σ XG of all quadratic forms in R[X ]of� that are sums of squares of linear forms. n G Matrix Completion Proposition. The image π (S ) the cone of positive semidefinite quadratic forms is the cone R[X ] . G ≥� G Σ XG of all quadratic forms in R[X G ]� that are sums of squares of linear forms. n Throughout, we fix acone simple graph G = ([n], E).quadratic forms is the cone Proposition. The image πG (S≥�)Summary of the of semidefinite ofpositive the setup This encodes coordinate Σ XG of all quadratic forms in R[XG the ]� that are sumsprojection of squares of linear forms. Summary of�xthe setup Let G = ([n], E) be a simple graph and let I = x ∶ {i, ∈� E� be the Stanley-Reisner ideal of i R G j n ⊕ j} Sn → R�E the clique complex of G . πG ∶ � (aSummary the setup ) n�let (aI��nG, = .of . �x . ,�E ai xnn ⊕ j} (a∈�i j∶E� {i,bej}the ∈ E) Let G = ([n], E) be a simple graphi jand Stanley-Reisner ideal of j ∶ ){i, Then the coordinate projection πG ∶ S → R ⊕ R is the same as the degree � part of the map the clique complex of nG . Matrix Completion S : real vector space of symmetric n × n matrices. R[x , .=. .([n], , x n ] E) → be R[X ] = R[x , x n ]�I coordinate rings. Let G aG simple graph letGIGofn=homogeneous �xRi x�Ej∶ {i, j} ∈� E� be the Stanley-Reisner ideal of � , .π. . and Then� the coordinate projection ∶ S → R ⊕ is equal to the degree � part of the map G (a ) as a partial matrix. n NB: Think of π Therefore, the imageofπGGThroughout, ofgraph squares Σ X([n], ⊂ R[X ]� on XG . G cone i j fixofa sums the clique complex .(S≥�) is the Grings. G we simple G = E) . R[x� , . . . , x n ] → R[X ] = R[x , . . . , x ]�I of homogeneous coordinate n Gn � G n �E is the same as the degree � part of the map � � Then the coordinate projection π ∶ S → R ⊕ R n G the coordinate Therefore, the image πGThis (S≥�encodes ) is the cone of sums of projection squares Σ XG ⊂ R[XG ]� on XG . R[x� , . . . , x n ] → R[X homogeneous coordinate rings. Sums ofR[x Squares and Nonnegative Polynomials � , . . . , xnn ]�I G] = G of n ⊕ R�E n S → R Therefore, the image πG (S≥�π) is∶ � the cone of sums of squares Subspace Σ XG ⊂ R[XGArrangements ]� on XG . G Sums Squares and Nonnegative � (aofi j)linear � (a ) ⊕ (aPolynomials j} ∈ E) � of �� , . . .is, aann i j ∶ {i, quadric. Observation: Every sum of squares forms nonnegative n−�, the subscheme rn : real Sums of Squares and Nonnegative S vector space of symmetric n ×Polynomials n matrices. Set X = V(I ) ⊂ P associated with the Stanley-Re G G Observation: Every sum of squares of linear forms is a nonnegative quadric. � � r Let G be the four cycle. q(p) = � ℓ i (p)Example. = NB: �(ℓThink i (p))of πG (a i j ) as a partial matrix. r r i=� i=� Observation: Every of linear forms isaa��nonnegative quadric. �a�� a a a a ∗ a � � � � � �of(ℓsquares �� �� �� �� �� q(p) = � ℓ�i (p)sum =� (p)) i XG = �quadrics. span{e ∶ i ∈ K}, i We write PXGi=� of nonnegative � � � rfor the cone r a�� � � a�� a�� a�� ∗ � i=� � a�� a�� a�� � K⊂G � =� G i�(p)) � � q(p) = � ℓ�i (p) = �π(ℓ a a a a ∗ a a a � � � � . of �� �� �� �� �� �� �� We write PXGi=� for the cone of nonnegative quadrics. Proposition (Obvious necessary condition). Assume π (A) ∈ Σ Then every completely where K runs over all complete subgraphs G . G X i=� G � �a��� a�� a�� a��� �a�� ∗ a�� a��� specified of A is positivequadrics. semidefinite. We writesymmetric PXG (Obvious for thesubmatrix cone of nonnegative Proposition necessary condition). G (A) ∈ Σ XG . Then every completely Example. Let G beAssume the fourπcycle. Goal: Complete aGGbe -partial matrix to a positive semidefinite symmetric ma Example. Let the four cycle. � specified symmetric submatrix of A is positive semidefinite. Proof. Σ ⊂ P . � Then IG = �x� x� , x�Assume x � and π XGG(A) is the∈union four every lines incompletely P. XG XG Proposition (Obvious Σ XG .ofThen thenecessary rank of thecondition). completion). � ais��positive a�� a��semidefinite. a�� � � a�� a�� ∗ a�� � specified submatrix of � AThen Proof. Σ Xsymmetric ⊂ P . Example. Let G be the four cycle. X Observation: The projection πG is the degree � part of the map of�coord G G � aGeometry �positive � a�� a��semidefinite a�� a�� aof a�� ∗ � matrix completio �� �� � � � � R[X ] . π = G G Proof. Σ XG ⊂ .a��the � � �∗ a a a � a��four∗cycle. a��� Example. LetPX G� Then � Gbe a a a a � �� �� �� �� � � �� �� �� � � a�� Goal: n∗n) a a�� a�� ∗�� � � � of Describe the image of the cone S of positive semidefinite quadraticquf a a a a a a�� Proposition. The image π (S of cone positive semidefinite � � �� �� �� �� �� ��the G ≥� Example. G be the four cycle. Then πG (a iLet ) = . ≥� a a ∗ a � � j ∗�� a �� a a �� G � XG Sums Sums of of Squares Squares and and Nonnegative Nonnegative Polynomials Polynomials Sums of Squares and Nonnegative Polynomials Observation: Observation: Every Every sum sum of of squares squares of of linear linear forms forms is is aa nonnegative nonnegative quadric. quadric. Observation:r Every sumrof squares of linear forms is a nonnegative quadric. r � r �� � r r q(p) = ℓ (p) = (ℓ (p)) � � i q(p) = � ℓ�ii (p) = �(ℓ i (p))� q(p) = � i=� i=� i=� ℓ i (p) = � i=� (ℓ i (p)) i=� i=� i=� i=� Matrix Co We for the cone of nonnegative quadrics. We write write P PX for the cone of nonnegative quadrics. Throughout, we fix a simple graph G = ([n], E) G X X G G We write PXG for the cone of nonnegative quadrics. This encodes the coordinate projection Proposition (Obvious necessary condition). Assume π (A) ∈ Σ . Then every completely X Proposition (Obvious necessary condition). Assume πG (A) ∈ Σ . Then every completely G (A) ∈ G X G G X Proposition (Obvious necessary condition). Assume πG (A) ∈n Σ XGG . Then every completely n ⊕ R�E specified symmetric submatrix of A is positive semidefinite. S → R specified symmetric submatrix of A is positive semidefinite. specified specified symmetric symmetric submatrix submatrix of of A A is is positive positive semidefinite. semidefinite. πG ∶ � (a i j) � (a�� , . . . , a nn ) ⊕ (a i j∶ {i, j} Proof. Σ ⊂ P .. � X X Proof. Σ ⊂ P � G G X X Proof. Σ ⊂ P . � G G X Proof. Σ X ⊂ P . � n G XG X G G S : real vector space of symmetric n × n matrice Example. Let G be the four cycle. Then Example. Let G be the four cycle. NB: Think of πG (a i j) as a partial matrix. Example. Let G be the four cycle. Then Then � � aa�� aa�� ∗ aa�� � � ∗ � � �� �� �� a�� a�� ∗ a�� � � � � aa�� a a ∗ � � �� �� a a ∗ � � �� �� �� a a a ∗ π (a ) = .. � � �� �� �� G i j � � π (a ) = G i j � � ∗ aa�� aa�� aa�� πG (a i j) = � . � ∗ � � �� �� �� ∗ a a a � �� �� �� � � � � � a ∗ a a �� �� �� � � ∗ a�� aa�� �aa�� �� ∗ a�� ��� The four fully specified submatrices correspond to the four maximal cliques of G ,, namely the edges. Example. Let G be the four cycle. The four fully specified submatrices correspond to the four maximal cliques of G namely the The four fully specified submatrices correspond to the four maximal cliques of G , namely the edges. edges. Question: When is this obvious necessary condition sufficient? � a�� a�� a�� a�� � � a�� a�� ∗ a�� Question: When is this obvious necessary condition Question: When is this obvious necessary condition sufficient? sufficient? � a�� a�� a�� a�� � � a�� a�� a�� ∗ � � πG � �a a a a � = � ∗ a a a � �� �� �� �� � � �� �� �� Two long lost friends Two long lost friends Two long lost friends�a�� a�� a�� a��� �a�� ∗ a�� a�� Recall: G = ([n], E) is aa simple graph. Recall: G = ([n], E) is simple Goal: Complete a G -partial matrix to a positiv Recall: G = ([n], E) is a simple graph. graph. Proof. Σ XG ⊂ PXG . Question: is this necessary condition sufficient? Recall: G = When ([n], E) is a obvious simple graph. Example. Letchordal G be theif four cycle. Then Two lost A graph G is every cycle in G long of length at friends least � has a chord. � spac Sn : real vector NB: Think of πG (a i � � Two long lost friends a a ∗ a � � �� �� �� Theorem (Agler-Helton-McCullough-Rodman) and Theorem and (Paulsen-Power-Smith). (Paulsen-Power-Smith). The The obvious obvious necnecRecall: G =(Agler-Helton-McCullough-Rodman) ([n], E) is a simple graph. � � a��positive a�� a�� ∗ � essary condition for semidefinite matrix completion is sufficient (equivalently, Σ =P )) � X X essary condition for positive semidefinite matrix completion is sufficient (equivalently, Σ = P A graph G=iisj([n], cycle in G of length at least � has a chord. πGG(a )chordal = �E) isifaevery . G G X X Recall: simple graph. � G � G ∗ a�� a�� a�� � � � if and only if G is chordal. if only is chordal. A and graph G ifisGchordal if∗every cycle in G of length at least � has a chord. � � a a a Theorem (Agler-Helton-McCullough-Rodman) and (Paulsen-Power-Smith). The obvious nec�� �� �� Theorem (Blekherman-Sinn-Velasco). Every nonnegative is sum on X (in Example. Let G be G Theorem (Blekherman-Sinn-Velasco). Every completion nonnegative quadric is a a(equivalently, sum of of squares squares on X (in essary condition for positive semidefinite matrix isquadric sufficient Σobvious = P ) G Theorem (Agler-Helton-McCullough-Rodman) and (Paulsen-Power-Smith). The necX X G edges. G The four Σ fully specified submatrices correspond to the four maximal cliques of G , namely the symbols = P ) if and only if I is � -regular. X symbols ΣX = Pchordal. ) if and only if IG if and only XifG XG G is �-regular. essary condition matrix completion is sufficient (equivalently, Σ XG =�PaX��G ) a�� a GG is for G positive semidefinite Question: is thisThe obvious necessary condition sufficient? if and only(Fröberg). ifWhen G is chordal. � a�� Theorem monomial ideal I is � -regular if and only ifisGa issum chordal. a�� a G �Theorem -regular:(Blekherman-Sinn-Velasco). The minimal free resolution is Every linear:nonnegative quadric of squares on X (in � G πG � a�� a � a�� symbols = P ) if and only if I is � -regular. φΣt+� φ φ φ Theorem (Blekherman-Sinn-Velasco). Every nonnegative quadric is a sum of squares on X (in t t−� � X X G G Two long � ��→G Ft �→G Ft−� ��→ � �→ F� → I → �lost friends � a�� a�� a symbols Σ XG = PXG ) if and only if IG is �-regular. Theorem (Fröberg). The monomial ideal IG is �-regular if and only if G is chordal. Theorem monomial I is �-regular if and only if G is chordal. Goal: Complete a Recall: G =(Fröberg). ([n], E) is The a simple graph.ideal Theorem (Fröberg). The monomial ideal IG G is �-regular if and only if G is chordal. the rank of the com A graph G is chordal if every cycle in G of length at least � has a chord. Theorem (Eisenbud-Green-Hulek-Popescu). The ideal I is �-regular if and only if V(I) is a linearly joined arrangement of subspaces if and only and if V(I) is small. Theorem (Agler-Helton-McCullough-Rodman) (Paulsen-Power-Smith). The obvious necGeom essary condition for positive semidefinite matrix completion is sufficient (equivalently, Σ XG = PXG ) References if and only if G is chordal. Goal: Describe the projection πG . Blekherman, Sinn, Velasco. Small Schemes and Sums of Squares, to appear on arXiv soon. Theorem (Blekherman-Sinn-Velasco). Every nonnegative quadric is a sum of squares on XG (in n S≥�: convex cone o Agler, Positive semidefinite matrices with a given sparsity patsymbolsHelton, Σ XG = PMcCullough, if IG is �-regular. n XG ) if and onlyRodman. R . tern, Proceedings of the Victoria Conference on Combinatorial Matrix Analysis (Victoria, BC, ����) Theorem (Fröberg). The monomial ideal IG is �-regular if and only if G is chordal. (����) Theorem (Diagon Fröberg.On Stanley-Reisner rings, Topics in algebra, Part � (Warsaw, ����) (����) semidefinite if and o Paulsen, Power, Smith. Schur products and matrix completions, Journal of Functional Analysis Question: Is π (S Theorem Theorem (Eisenbud-Green-Hulek-Popescu). (Eisenbud-Green-Hulek-Popescu). The The ideal ideal II is is ��-regular -regular if if and and only only if if V(I) V(I) is is a a linlinTheorem (Eisenbud-Green-Hulek-Popescu). The ideal I is � -regular if and only if V(I) is a linTheorem (Eisenbud-Green-Hulek-Popescu). The ideal I is � -regular if and only if V(I) is a linearly joined arrangement of subspaces if and only if V(I) is small. early joined arrangement of subspaces if and only if V(I) is small. early early joined joined arrangement arrangement of of subspaces subspaces if if and and only only if if V(I) V(I) is is small. small. Minimal Free Resolutions and Matrix Completion Minimal Minimal Free Free Resolutions Resolutions and and Matrix Matrix Completion Completion The is The dual dual convex convex cone cone to to Σ ΣX is G X G The dual convex cone to Σ is X The dual convex cone to Σ G XG is ∨ ∗ ∨ ∗ Σ = {ℓ ∈∈ R[X ]]�∗∶∶ ℓ( ff )) ≥ �� for }. G X Σ∨X = {ℓ R[X ℓ( ≥ for all all ff ∈∈ Σ Σ }. G G X G � X G G Σ = {ℓ ∈∈ R[X ]]∗� ∶∶ ℓ( ff )) ≥ �� for all ff ∈∈ Σ }. G X G }. Σ∨X = {ℓ R[X ℓ( ≥ for all Σ G G X � XG ∨G � Σ∨∨ . So there is an extreme ray ℓ ∈ Σ ∨ Convex Biduality: If Σ � P , then dually P Convex Biduality: If Σ X Σ∨XXG. So there is an extreme ray ℓ ∈ Σ XXGG, XG XGG, then dually PX G � PX XGG � ∨ G . So there is an extreme ray ℓ ∈ Σ X , Convex Biduality: If Σ � P , then dually P � ∨ � Σ ∨ X X G G X X Convex Biduality: If Σ � P , then dually P Σ , which .. XG G G . So there is an extreme ray ℓ ∈ Σ XG XG which is is not not in in PX G X X G X G G G. which is not in P ∗ with a space of symmetric matrices: X ⊥ R[xI ⊥ ∗]∗= (R[x G. G ∗Gwith which not in PR[X We canisidentify identify = (R[x , . . . , x ] �I ) ⊂. , x n ] XR[X n � � We can ] , . . . , x ] �I ) a space of symmetric matrices: I , . . G G ∗ n � � � � G ⊥ G ∗ �∗ = (R[x� , . . . , x n ]��IG )∗ with a space of symmetric matrices: IG We can identify R[X ] ⊥ R[x ∗ n ]] � ,, .. .. .. ,, x G � G We can identify R[X ] = (R[x , . . . , x ] �I ) with a space of symmetric matrices: I R[x x R[x , . . . , x ] . n n � � ℓG ∈ Σ X � PX ? � Grank � is the nsmallest What ray � of an extreme G � G G What � P XG ?? G ΣX � P What is is the the smallest smallest rank rank of of an an extreme extreme ray ray ℓℓ ∈∈ Σ XG XG Theorem (Blekherman-Sinn-Velasco). Suppose m is the Green-Lazarsfeld index of IG , i.e. the Theorem (Blekherman-Sinn-Velasco). Suppose m is the Green-Lazarsfeld index of IIG ,, largest i.e. the Theorem (Blekherman-Sinn-Velasco). Suppose m is the Green-Lazarsfeld index of i.e. the (Blekherman-Sinn-Velasco). Let m be the Green-Lazarsfeld index of I , i.e. the largest integer p such that IG has property N�,p (the minimal free resolution is linear G forGp steps). largest integer pp such that IIG has property N minimal free resolution is linear for pp steps). �,p (the largest integer such that has property N (the minimal free resolution is linear for steps). integer p such that I has property N (the minimal free resolution is linear for p steps). The smallThe smallest rank ofGan extreme ∈ Σ XG ��,p PXG is m + �. In terms of the graph, it is the smallest G ray ℓ�,p The smallest rank of anray extreme ray �ℓ P ∈∈ Σ � P m + ��.. of In terms of theit graph, it is the smallest Xis X�G. is G est rank of an extreme ℓ ∈ Σ m + In terms the graph, is the smallest number The smallest rank of an extreme ray ℓ Σ � P is m + In terms of the graph, it is the smallest number m ≥ � such that G contains cycle XG an induced XGXG XG of length m + �. number m ≥ � such that G contains an induced cycle of length m + �. m ≥ � such G contains an induced cycle of length + �. m + �. number m ≥that � such that G contains an induced cycle ofmlength Theorem (Eisenbud-Green-Hulek-Popescu). The Green-Lazarsfeld index of IG is the largest p Theorem (Eisenbud-Green-Hulek-Popescu). The Green-Lazarsfeld index of IG is the largest p Theorem index of IG is the largest p such that G(Eisenbud-Green-Hulek-Popescu). has no induced cycle of length at mostThe p +Green-Lazarsfeld �. such that G has no induced cycle of length at most p + �. such that G has no induced cycle of length at most p + �. References References References Blekherman, Sinn, Velasco. Small Schemes and Sums of Squares, to appear on arXiv soon. Blekherman, Sinn, Velasco. Small Schemes and Sums of Squares, to appear on arXiv soon. Blekherman, Velasco. Small Schemes and Sums of Squares, to appear arXiv soon. Agler, Helton,Sinn, McCullough, Rodman. Positive semidefinite matrices with aon given sparsity patAgler, Helton, McCullough, Rodman. Positive semidefinite matrices with a given sparsity patAgler, Helton, McCullough, Rodman. Positive semidefinite matrices with a given sparsity pattern, Proceedings of the Victoria Conference on Combinatorial Matrix Analysis (Victoria, BC, ����) tern, Proceedings of the Victoria Conference on Combinatorial Matrix Analysis (Victoria, BC, ����) tern, (����)Proceedings of the Victoria Conference on Combinatorial Matrix Analysis (Victoria, BC, ����) (����) The smallest rank of an extreme ray ℓ ∈ Σ XG � PXG is m + �. In terms of the graph, it is the smallest number m ≥ � such that G contains an induced cycle of length m + �. Theorem (Eisenbud-Green-Hulek-Popescu). The Green-Lazarsfeld index of IG is the largest p such that G has no induced cycle of length at most p + �. References Blekherman, Sinn, Velasco. Small Schemes and Sums of Squares, to appear on arXiv soon. Agler, Helton, McCullough, Rodman. Positive semidefinite matrices with a given sparsity pattern, Proceedings of the Victoria Conference on Combinatorial Matrix Analysis (Victoria, BC, ����) (����) Eisenbud, Green, Hulek, Popescu. Restricting linear syzygies: algebra and geometry, Compositio Mathematica (����) Fröberg.On Stanley-Reisner rings, Topics in algebra, Part � (Warsaw, ����) (����) Paulsen, Power, Smith. Schur products and matrix completions, Journal of Functional Analysis (����) Thank you