icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints:...

14
Real Geometry of Matrix Completion Rainer Sinn (Freie Universität Berlin)

Transcript of icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints:...

Page 1: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

Real Geometry of Matrix Completion

Rainer Sinn (Freie Universität Berlin)

Positive Semidefinite Matrix Completion

Let G = ([n], E) be a simple graph on n vertices.

Definition. A G-partial matrix is the image of a symmetric matrix under the pro-jection

πG∶� Symn(R)→ Rn ⊕R�E

A = (ai j)� ((a��, . . . , ann), (ai j∶{i , j} ∈ E))Example.

������

a�� a�� a�� ∗a�� a�� a�� ∗a�� a�� a�� a��∗ ∗ a�� a��

������∗ ∗ ∗ ?∗ ∗ ∗ ?∗ ∗ ∗ ∗? ? ∗ ∗

��������

��������Example.

������

a�� a�� ∗ a��a�� a�� a�� ∗∗ a�� a�� a��a�� ∗ a�� a��

�������

� � �

∗ ∗ ? ∗∗ ∗ ∗ ?? ∗ ∗ ∗∗ ? ∗ ∗

��������

���������

� �

Question. Given a G-partial matrix, does there exist a positive semidefinite sym-metric completion?

Positive Semidefinite Matrix Completion

Let G = ([n], E) be a simple graph on n vertices.

Definition. A G-partial matrix is the image of a symmetric matrix under the pro-jection

πG∶� Symn(R)→ Rn ⊕R�E

A = (ai j)� ((a��, . . . , ann), (ai j∶{i , j} ∈ E))Example.

������

a�� a�� a�� ∗a�� a�� a�� ∗a�� a�� a�� a��∗ ∗ a�� a��

������∗ ∗ ∗ ?∗ ∗ ∗ ?∗ ∗ ∗ ∗? ? ∗ ∗

��������

��������Example.

������

a�� a�� ∗ a��a�� a�� a�� ∗∗ a�� a�� a��a�� ∗ a�� a��

�������

� � �

∗ ∗ ? ∗∗ ∗ ∗ ?? ∗ ∗ ∗∗ ? ∗ ∗

��������

���������

� �

Question. Given a G-partial matrix, does there exist a positive semidefinite sym-metric completion?

Page 2: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

Symmetric matrix completion

����� ? −�? � ?−� ? −�

����Constraints:(�) Rank (smallest possible, fixed, …)(�) Positive semidefinite

����∗ ? ∗? ∗ ?∗ ? ∗

����Spectral Theorem

Definition. A real symmetric matrix is positive semidefinite if all of its real eigen-values are nonnegative.

Theorem. A real symmetricmatrixM is positive semidefinite if andonly if there existsa real matrix B such thatM = BTB.

Real symmetric matricesM are in one-to-one correspondence with real quadraticforms QM = (x�, x�, . . . , xn)M(x�, x�, . . . , xn)T .Example.

M =������

� � −� �� � � −��−� � � ��� −�� �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� − �x�x� + �x�x� + ��x�x� − ��x�x� + ��x�x�

Symmetric matrix completion

����� ? −�? � ?−� ? −�

����Constraints:(�) Rank (smallest possible, fixed, …)(�) Positive semidefinite

����∗ ? ∗? ∗ ?∗ ? ∗

����Spectral Theorem

Definition. A real symmetric matrix is positive semidefinite if all of its real eigen-values are nonnegative.

Theorem. A real symmetricmatrixM is positive semidefinite if andonly if there existsa real matrix B such thatM = BTB.

Real symmetric matricesM are in one-to-one correspondence with real quadraticforms QM = (x�, x�, . . . , xn)M(x�, x�, . . . , xn)T .Example.

M =������

� � −� �� � � −��−� � � ��� −�� �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� − �x�x� + �x�x� + ��x�x� − ��x�x� + ��x�x�

Symmetric matrix completion

����� ? −�? � ?−� ? −�

����Constraints:(�) Rank (smallest possible, fixed, …)(�) Positive semidefinite

����∗ ? ∗? ∗ ?∗ ? ∗

����Spectral Theorem

Definition. A real symmetric matrix is positive semidefinite if all of its real eigen-values are nonnegative.

Theorem. A real symmetricmatrixM is positive semidefinite if andonly if there existsa real matrix B such thatM = BTB.

Real symmetric matricesM are in one-to-one correspondence with real quadraticforms QM = (x�, x�, . . . , xn)M(x�, x�, . . . , xn)T .Example.

M =������

� � −� �� � � −��−� � � ��� −�� �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� − �x�x� + �x�x� + ��x�x� − ��x�x� + ��x�x�

Page 3: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

Symmetric matrix completion

����� ? −�? � ?−� ? −�

����Constraints:(�) Rank (smallest possible, fixed, …)(�) Positive semidefinite

����∗ ? ∗? ∗ ?∗ ? ∗

����Spectral Theorem

Definition. A real symmetric matrix is positive semidefinite if all of its real eigen-values are nonnegative.

Theorem. A real symmetricmatrixM is positive semidefinite if andonly if there existsa real matrix B such thatM = BTB.

Real symmetric matricesM are in one-to-one correspondence with real quadraticforms QM = (x�, x�, . . . , xn)M(x�, x�, . . . , xn)T .Example.

M =������

� � −� �� � � −��−� � � ��� −�� �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� − �x�x� + �x�x� + ��x�x� − ��x�x� + ��x�x�

Symmetric matrix completion

����� ? −�? � ?−� ? −�

����Constraints:(�) Rank (smallest possible, fixed, …)(�) Positive semidefinite

����∗ ? ∗? ∗ ?∗ ? ∗

����Spectral Theorem

Definition. A real symmetric matrix is positive semidefinite if all of its real eigen-values are nonnegative.

Theorem. A real symmetricmatrixM is positive semidefinite if andonly if there existsa real matrix B such thatM = BTB.

Real symmetric matricesM are in one-to-one correspondence with real quadraticforms QM = (x�, x�, . . . , xn)M(x�, x�, . . . , xn)T .Example.

M =������

� � −� �� � � −��−� � � ��� −�� �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� − �x�x� + �x�x� + ��x�x� − ��x�x� + ��x�x�

Symmetric matrix completion

����� ? −�? � ?−� ? −�

����Constraints:(�) Rank (smallest possible, fixed, …)(�) Positive semidefinite

����∗ ? ∗? ∗ ?∗ ? ∗

����Spectral Theorem

Definition. A real symmetric matrix is positive semidefinite if all of its real eigen-values are nonnegative.

Theorem. A real symmetricmatrixM is positive semidefinite if andonly if there existsa real matrix B such thatM = BTB.

Real symmetric matricesM are in one-to-one correspondence with real quadraticforms QM = (x�, x�, . . . , xn)M(x�, x�, . . . , xn)T .Example.

M =������

� � −� �� � � −��−� � � ��� −�� �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� − �x�x� + �x�x� + ��x�x� − ��x�x� + ��x�x�

Symmetric matrix completion

����� ? −�? � ?−� ? −�

����Constraints:(�) Rank (smallest possible, fixed, …)(�) Positive semidefinite

����∗ ? ∗? ∗ ?∗ ? ∗

����Spectral Theorem

Definition. A real symmetric matrix is positive semidefinite if all of its real eigen-values are nonnegative.

Theorem. A real symmetricmatrixM is positive semidefinite if andonly if there existsa real matrix B such thatM = BTB.

Real symmetric matricesM are in one-to-one correspondence with real quadraticforms QM = (x�, x�, . . . , xn)M(x�, x�, . . . , xn)T .Example.

M =������

� � −� �� � � −��−� � � ��� −�� �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� − �x�x� + �x�x� + ��x�x� − ��x�x� + ��x�x�

Page 4: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

Nonnegative Quadratic Forms and Sums of Squares

Theorem. A quadratic form QM = xTMx is positive semidefinite if and only if it is asum of squaresQM = (Bx)T(Bx).Example.

��x�� +���x��+���x��+���x�x�+���x�x�+���x�x� = (�x�−x�+�x�)�+(�x�+��x�−��x�)�

������ �� ����� ��� ������ ��� ���

���� =����� �−� ��� −��

������ −� �� �� −���

Smallest number of squares in QM = ℓ�� + . . . + ℓ�r = rank ofM

Obvious Necessary Condition

Theorem (Sylvester’s criterion). A symmetric matrix is positive semidefinite if andonly if all of its principal minors are nonnegative.

Example.

M = ����� � ?� � ?? ? �

����QM = x�� + x�� + x�� + �x�x�+?x�x�+?x�x�

QM(�,−�, �) = � + � − � = −�Example.

������

� � ? −�� � � ?? � � �−� ? � �

������

Aside: diagonal entries

����� ? ?? � �? � ?

����

Nonnegative Quadratic Forms and Sums of Squares

Theorem. A quadratic form QM = xTMx is positive semidefinite if and only if it is asum of squaresQM = (Bx)T(Bx).Example.

��x�� +���x��+���x��+���x�x�+���x�x�+���x�x� = (�x�−x�+�x�)�+(�x�+��x�−��x�)�

������ �� ����� ��� ������ ��� ���

���� =����� �−� ��� −��

������ −� �� �� −���

Smallest number of squares in QM = ℓ�� + . . . + ℓ�r = rank ofM

Obvious Necessary Condition

Theorem (Sylvester’s criterion). A symmetric matrix is positive semidefinite if andonly if all of its principal minors are nonnegative.

Example.

M = ����� � ?� � ?? ? �

����QM = x�� + x�� + x�� + �x�x�+?x�x�+?x�x�

QM(�,−�, �) = � + � − � = −�Example.

������

� � ? −�� � � ?? � � �−� ? � �

������

Aside: diagonal entries

����� ? ?? � �? � ?

����

Nonnegative Quadratic Forms and Sums of Squares

Theorem. A quadratic form QM = xTMx is positive semidefinite if and only if it is asum of squaresQM = (Bx)T(Bx).Example.

��x�� +���x��+���x��+���x�x�+���x�x�+���x�x� = (�x�−x�+�x�)�+(�x�+��x�−��x�)�

������ �� ����� ��� ������ ��� ���

���� =����� �−� ��� −��

������ −� �� �� −���

Smallest number of squares in QM = ℓ�� + . . . + ℓ�r = rank ofM

Obvious Necessary Condition

Theorem (Sylvester’s criterion). A symmetric matrix is positive semidefinite if andonly if all of its principal minors are nonnegative.

Example.

M = ����� � ?� � ?? ? �

����QM = x�� + x�� + x�� + �x�x�+?x�x�+?x�x�

QM(�,−�, �) = � + � − � = −�Example.

������

� � ? −�� � � ?? � � �−� ? � �

������

Aside: diagonal entries

����� ? ?? � �? � ?

����

Nonnegative Quadratic Forms and Sums of Squares

Theorem. A quadratic form QM = xTMx is positive semidefinite if and only if it is asum of squaresQM = (Bx)T(Bx).Example.

��x�� +���x��+���x��+���x�x�+���x�x�+���x�x� = (�x�−x�+�x�)�+(�x�+��x�−��x�)�

������ �� ����� ��� ������ ��� ���

���� =����� �−� ��� −��

������ −� �� �� −���

Smallest number of squares in QM = ℓ�� + . . . + ℓ�r = rank ofM

Obvious Necessary Condition

Theorem (Sylvester’s criterion). A symmetric matrix is positive semidefinite if andonly if all of its principal minors are nonnegative.

Example.

M = ����� � ?� � ?? ? �

����QM = x�� + x�� + x�� + �x�x�+?x�x�+?x�x�

QM(�,−�, �) = � + � − � = −�Example.

������

� � ? −�� � � ?? � � �−� ? � �

������

Aside: diagonal entries

����� ? ?? � �? � ?

����

Page 5: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

Nonnegative Quadratic Forms and Sums of Squares

Theorem. A quadratic form QM = xTMx is positive semidefinite if and only if it is asum of squaresQM = (Bx)T(Bx).Example.

��x�� +���x��+���x��+���x�x�+���x�x�+���x�x� = (�x�−x�+�x�)�+(�x�+��x�−��x�)�

������ �� ����� ��� ������ ��� ���

���� =����� �−� ��� −��

������ −� �� �� −���

Smallest number of squares in QM = ℓ�� + . . . + ℓ�r = rank ofM

Obvious Necessary Condition

Theorem (Sylvester’s criterion). A symmetric matrix is positive semidefinite if andonly if all of its principal minors are nonnegative.

Example.

M = ����� � ?� � ?? ? �

����QM = x�� + x�� + x�� + �x�x�+?x�x�+?x�x�

QM(�,−�, �) = � + � − � = −�Example.

������

� � ? −�� � � ?? � � �−� ? � �

������

Aside: diagonal entries

����� ? ?? � �? � ?

����

Nonnegative Quadratic Forms and Sums of Squares

Theorem. A quadratic form QM = xTMx is positive semidefinite if and only if it is asum of squaresQM = (Bx)T(Bx).Example.

��x�� +���x��+���x��+���x�x�+���x�x�+���x�x� = (�x�−x�+�x�)�+(�x�+��x�−��x�)�

������ �� ����� ��� ������ ��� ���

���� =����� �−� ��� −��

������ −� �� �� −���

Smallest number of squares in QM = ℓ�� + . . . + ℓ�r = rank ofM

Obvious Necessary Condition

Theorem (Sylvester’s criterion). A symmetric matrix is positive semidefinite if andonly if all of its principal minors are nonnegative.

Example.

M = ����� � ?� � ?? ? �

����QM = x�� + x�� + x�� + �x�x�+?x�x�+?x�x�

QM(�,−�, �) = � + � − � = −�Example.

������

� � ? −�� � � ?? � � �−� ? � �

������

Aside: diagonal entries

����� ? ?? � �? � ?

����

Nonnegative Quadratic Forms and Sums of Squares

Theorem. A quadratic form QM = xTMx is positive semidefinite if and only if it is asum of squaresQM = (Bx)T(Bx).Example.

��x�� +���x��+���x��+���x�x�+���x�x�+���x�x� = (�x�−x�+�x�)�+(�x�+��x�−��x�)�

������ �� ����� ��� ������ ��� ���

���� =����� �−� ��� −��

������ −� �� �� −���

Smallest number of squares in QM = ℓ�� + . . . + ℓ�r = rank ofM

Obvious Necessary Condition

Theorem (Sylvester’s criterion). A symmetric matrix is positive semidefinite if andonly if all of its principal minors are nonnegative.

Example.

M = ����� � ?� � ?? ? �

����QM = x�� + x�� + x�� + �x�x�+?x�x�+?x�x�

QM(�,−�, �) = � + � − � = −�Example.

������

� � ? −�� � � ?? � � �−� ? � �

������

Aside: diagonal entries

����� ? ?? � �? � ?

����

Nonnegative Quadratic Forms and Sums of Squares

Theorem. A quadratic form QM = xTMx is positive semidefinite if and only if it is asum of squaresQM = (Bx)T(Bx).Example.

��x�� +���x��+���x��+���x�x�+���x�x�+���x�x� = (�x�−x�+�x�)�+(�x�+��x�−��x�)�

������ �� ����� ��� ������ ��� ���

���� =����� �−� ��� −��

������ −� �� �� −���

Smallest number of squares in QM = ℓ�� + . . . + ℓ�r = rank ofM

Obvious Necessary Condition

Theorem (Sylvester’s criterion). A symmetric matrix is positive semidefinite if andonly if all of its principal minors are nonnegative.

Example.

M = ����� � ?� � ?? ? �

����QM = x�� + x�� + x�� + �x�x�+?x�x�+?x�x�

QM(�,−�, �) = � + � − � = −�Example.

������

� � ? −�� � � ?? � � �−� ? � �

������

Aside: diagonal entries

����� ? ?? � �? � ?

����

Nonnegative Quadratic Forms and Sums of Squares

Theorem. A quadratic form QM = xTMx is positive semidefinite if and only if it is asum of squaresQM = (Bx)T(Bx).Example.

��x�� +���x��+���x��+���x�x�+���x�x�+���x�x� = (�x�−x�+�x�)�+(�x�+��x�−��x�)�

������ �� ����� ��� ������ ��� ���

���� =����� �−� ��� −��

������ −� �� �� −���

Smallest number of squares in QM = ℓ�� + . . . + ℓ�r = rank ofM

Obvious Necessary Condition

Theorem (Sylvester’s criterion). A symmetric matrix is positive semidefinite if andonly if all of its principal minors are nonnegative.

Example.

M = ����� � ?� � ?? ? �

����QM = x�� + x�� + x�� + �x�x�+?x�x�+?x�x�

QM(�,−�, �) = � + � − � = −�Example.

������

� � ? −�� � � ?? � � �−� ? � �

������

Aside: diagonal entries

����� ? ?? � �? � ?

����

Nonnegative Quadratic Forms and Sums of Squares

Theorem. A quadratic form QM = xTMx is positive semidefinite if and only if it is asum of squaresQM = (Bx)T(Bx).Example.

��x�� +���x��+���x��+���x�x�+���x�x�+���x�x� = (�x�−x�+�x�)�+(�x�+��x�−��x�)�

������ �� ����� ��� ������ ��� ���

���� =����� �−� ��� −��

������ −� �� �� −���

Smallest number of squares in QM = ℓ�� + . . . + ℓ�r = rank ofM

Obvious Necessary Condition

Theorem (Sylvester’s criterion). A symmetric matrix is positive semidefinite if andonly if all of its principal minors are nonnegative.

Example.

M = ����� � ?� � ?? ? �

����QM = x�� + x�� + x�� + �x�x�+?x�x�+?x�x�

QM(�,−�, �) = � + � − � = −�Example.

������

� � ? −�� � � ?? � � �−� ? � �

������

Aside: diagonal entries

����� ? ?? � �? � ?

����

Page 6: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

Nonnegative Quadratic Forms and Sums of Squares

Theorem. A quadratic form QM = xTMx is positive semidefinite if and only if it is asum of squaresQM = (Bx)T(Bx).Example.

��x�� +���x��+���x��+���x�x�+���x�x�+���x�x� = (�x�−x�+�x�)�+(�x�+��x�−��x�)�

������ �� ����� ��� ������ ��� ���

���� =����� �−� ��� −��

������ −� �� �� −���

Smallest number of squares in QM = ℓ�� + . . . + ℓ�r = rank ofM

Obvious Necessary Condition

Theorem (Sylvester’s criterion). A symmetric matrix is positive semidefinite if andonly if all of its principal minors are nonnegative.

Example.

M = ����� � ?� � ?? ? �

����QM = x�� + x�� + x�� + �x�x�+?x�x�+?x�x�

QM(�,−�, �) = � + � − � = −�Example.

������

� � ? −�� � � ?? � � �−� ? � �

������

Aside: diagonal entries

����� ? ?? � �? � ?

����

����� ? ?? � �? � ?

����Patterns and graphs

Complete principal minors: Cliques

References

Thank you

����� ? ?? � �? � ?

����Patterns and graphs

Complete principal minors: Cliques

References

Thank you

Page 7: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

Positive Semidefinite Matrix Completion

Let G = ([n], E) be a simple graph on n vertices.

Definition. A G-partial matrix is the image of a symmetric matrix under the pro-jection

πG∶� Symn(R)→ Rn ⊕R�E

A = (ai j)� ((a��, . . . , ann), (ai j∶{i , j} ∈ E))Example.

������

a�� a�� a�� ∗a�� a�� a�� ∗a�� a�� a�� a��∗ ∗ a�� a��

������∗ ∗ ∗ ?∗ ∗ ∗ ?∗ ∗ ∗ ∗? ? ∗ ∗

��������

��������Example.

������

a�� a�� ∗ a��a�� a�� a�� ∗∗ a�� a�� a��a�� ∗ a�� a��

�������

� � �

∗ ∗ ? ∗∗ ∗ ∗ ?? ∗ ∗ ∗∗ ? ∗ ∗

��������

���������

� �

Question. Given a G-partial matrix, does there exist a positive semidefinite sym-metric completion?

Positive Semidefinite Matrix Completion

Let G = ([n], E) be a simple graph on n vertices.

Definition. A G-partial matrix is the image of a symmetric matrix under the pro-jection

πG∶� Symn(R)→ Rn ⊕R�E

A = (ai j)� ((a��, . . . , ann), (ai j∶{i , j} ∈ E))Example.

������

a�� a�� a�� ∗a�� a�� a�� ∗a�� a�� a�� a��∗ ∗ a�� a��

������∗ ∗ ∗ ?∗ ∗ ∗ ?∗ ∗ ∗ ∗? ? ∗ ∗

��������

��������Example.

������

a�� a�� ∗ a��a�� a�� a�� ∗∗ a�� a�� a��a�� ∗ a�� a��

�������

� � �

∗ ∗ ? ∗∗ ∗ ∗ ?? ∗ ∗ ∗∗ ? ∗ ∗

��������

���������

� �

Question. Given a G-partial matrix, does there exist a positive semidefinite sym-metric completion?

Positive Semidefinite Matrix Completion

Let G = ([n], E) be a simple graph on n vertices.

Definition. A G-partial matrix is the image of a symmetric matrix under the pro-jection

πG∶� Symn(R)→ Rn ⊕R�E

A = (ai j)� ((a��, . . . , ann), (ai j∶{i , j} ∈ E))Example.

������

a�� a�� a�� ∗a�� a�� a�� ∗a�� a�� a�� a��∗ ∗ a�� a��

������∗ ∗ ∗ ?∗ ∗ ∗ ?∗ ∗ ∗ ∗? ? ∗ ∗

��������

��������Example.

������

a�� a�� ∗ a��a�� a�� a�� ∗∗ a�� a�� a��a�� ∗ a�� a��

�������

� � �

∗ ∗ ? ∗∗ ∗ ∗ ?? ∗ ∗ ∗∗ ? ∗ ∗

��������

���������

� �

Question. Given a G-partial matrix, does there exist a positive semidefinite sym-metric completion?

Positive Semidefinite Matrix Completion

Let G = ([n], E) be a simple graph on n vertices.

Definition. A G-partial matrix is the image of a symmetric matrix under the pro-jection

πG∶� Symn(R)→ Rn ⊕R�E

A = (ai j)� ((a��, . . . , ann), (ai j∶{i , j} ∈ E))Example.

������

a�� a�� a�� ∗a�� a�� a�� ∗a�� a�� a�� a��∗ ∗ a�� a��

������∗ ∗ ∗ ?∗ ∗ ∗ ?∗ ∗ ∗ ∗? ? ∗ ∗

��������

��������Example.

������

a�� a�� ∗ a��a�� a�� a�� ∗∗ a�� a�� a��a�� ∗ a�� a��

�������

� � �

∗ ∗ ? ∗∗ ∗ ∗ ?? ∗ ∗ ∗∗ ? ∗ ∗

��������

���������

� �

Question. Given a G-partial matrix, does there exist a positive semidefinite sym-metric completion?

����� ? ?? � �? � ?

����Patterns and graphs

Complete principal minors: Cliques

References

Thank you

Page 8: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

����� ? ?? � �? � ?

����Patterns and graphs

Complete principal minors: Cliques

Stanley-Reisner Ideals

Definition. The Stanley-Reisner ideal associated to G = ([n], E) is the square-free monomial ideal

IG = �xix j∶ {i , j} ∉ E� ⊂ R[x�, x�, . . . , xn].Example.

IG = �x�x�, x�x��

M =������

� � ? �� � � ?? � � ��� ? �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� + �x�x� + ��x�x� + ��x�x�+?x�x�+?x�x�

Proposition. (�)G-partialmatricesM are inone-to-onecorrespondencewith residueclasses of quadratic formsQM inR[x�, x�, . . . , xn]�IG = R[XG].(�) AG-partial matrixM has a positive semidefinite completion if and only if the cor-responding quadratic formQM is a sum of squares of linear formsmodulo IG.

References

Thank you

Positive Semidefinite Matrix Completion

Let G = ([n], E) be a simple graph on n vertices.

Definition. A G-partial matrix is the image of a symmetric matrix under the pro-jection

πG∶� Symn(R)→ Rn ⊕R�E

A = (ai j)� ((a��, . . . , ann), (ai j∶{i , j} ∈ E))Example.

������

a�� a�� a�� ∗a�� a�� a�� ∗a�� a�� a�� a��∗ ∗ a�� a��

������∗ ∗ ∗ ?∗ ∗ ∗ ?∗ ∗ ∗ ∗? ? ∗ ∗

��������

��������Example.

������

a�� a�� ∗ a��a�� a�� a�� ∗∗ a�� a�� a��a�� ∗ a�� a��

�������

� � �

∗ ∗ ? ∗∗ ∗ ∗ ?? ∗ ∗ ∗∗ ? ∗ ∗

��������

���������

� �

Question. Given a G-partial matrix, does there exist a positive semidefinite sym-metric completion?

����� ? ?? � �? � ?

����Patterns and graphs

Complete principal minors: Cliques

Stanley-Reisner Ideals

Definition. The Stanley-Reisner ideal associated to G = ([n], E) is the square-free monomial ideal

IG = �xix j∶ {i , j} ∉ E� ⊂ R[x�, x�, . . . , xn].Example.

IG = �x�x�, x�x��

M =������

� � ? �� � � ?? � � ��� ? �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� + �x�x� + ��x�x� + ��x�x�+?x�x�+?x�x�

Proposition. (�)G-partialmatricesM are inone-to-onecorrespondencewith residueclasses of quadratic formsQM inR[x�, x�, . . . , xn]�IG = R[XG].(�) AG-partial matrixM has a positive semidefinite completion if and only if the cor-responding quadratic formQM is a sum of squares of linear formsmodulo IG.

References

Thank you

����� ? ?? � �? � ?

����Patterns and graphs

Complete principal minors: Cliques

Stanley-Reisner Ideals

Definition. The Stanley-Reisner ideal associated to G = ([n], E) is the square-free monomial ideal

IG = �xix j∶ {i , j} ∉ E� ⊂ R[x�, x�, . . . , xn].Example.

IG = �x�x�, x�x��

M =������

� � ? �� � � ?? � � ��� ? �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� + �x�x� + ��x�x� + ��x�x�+?x�x�+?x�x�

Proposition. (�)G-partialmatricesM are inone-to-onecorrespondencewith residueclasses of quadratic formsQM inR[x�, x�, . . . , xn]�IG = R[XG].(�) AG-partial matrixM has a positive semidefinite completion if and only if the cor-responding quadratic formQM is a sum of squares of linear formsmodulo IG.

References

Thank you

����� ? ?? � �? � ?

����Patterns and graphs

Complete principal minors: Cliques

Stanley-Reisner Ideals

Definition. The Stanley-Reisner ideal associated to G = ([n], E) is the square-free monomial ideal

IG = �xix j∶ {i , j} ∉ E� ⊂ R[x�, x�, . . . , xn].Example.

IG = �x�x�, x�x��

M =������

� � ? �� � � ?? � � ��� ? �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� + �x�x� + ��x�x� + ��x�x�+?x�x�+?x�x�

Proposition. (�)G-partialmatricesM are inone-to-onecorrespondencewith residueclasses of quadratic formsQM inR[x�, x�, . . . , xn]�IG = R[XG].(�) AG-partial matrixM has a positive semidefinite completion if and only if the cor-responding quadratic formQM is a sum of squares of linear formsmodulo IG.

References

Thank you

����� ? ?? � �? � ?

����Patterns and graphs

Complete principal minors: Cliques

Stanley-Reisner Ideals

Definition. The Stanley-Reisner ideal associated to G = ([n], E) is the square-free monomial ideal

IG = �xix j∶ {i , j} ∉ E� ⊂ R[x�, x�, . . . , xn].Example.

IG = �x�x�, x�x��

M =������

� � ? �� � � ?? � � ��� ? �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� + �x�x� + ��x�x� + ��x�x�+?x�x�+?x�x�

Proposition. (�)G-partialmatricesM are inone-to-onecorrespondencewith residueclasses of quadratic formsQM inR[x�, x�, . . . , xn]�IG = R[XG].(�) AG-partial matrixM has a positive semidefinite completion if and only if the cor-responding quadratic formQM is a sum of squares of linear formsmodulo IG.

References

Thank you

����� ? ?? � �? � ?

����Patterns and graphs

Complete principal minors: Cliques

Stanley-Reisner Ideals

Definition. The Stanley-Reisner ideal associated to G = ([n], E) is the square-free monomial ideal

IG = �xix j∶ {i , j} ∉ E� ⊂ R[x�, x�, . . . , xn].Example.

IG = �x�x�, x�x��

M =������

� � ? �� � � ?? � � ��� ? �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� + �x�x� + ��x�x� + ��x�x�+?x�x�+?x�x�

Proposition. (�)G-partialmatricesM are inone-to-onecorrespondencewith residueclasses of quadratic formsQM inR[x�, x�, . . . , xn]�IG = R[XG].(�) AG-partial matrixM has a positive semidefinite completion if and only if the cor-responding quadratic formQM is a sum of squares of linear formsmodulo IG.

References

Thank you

Page 9: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

����� ? ?? � �? � ?

����Patterns and graphs

Complete principal minors: Cliques

Stanley-Reisner Ideals

Definition. The Stanley-Reisner ideal associated to G = ([n], E) is the square-free monomial ideal

IG = �xix j∶ {i , j} ∉ E� ⊂ R[x�, x�, . . . , xn].Example.

IG = �x�x�, x�x��

M =������

� � ? �� � � ?? � � ��� ? �� �

������QM = x�� + x�� + x�� + x�� + ��x�x� + �x�x� + ��x�x� + ��x�x�+?x�x�+?x�x�

Proposition. (�)G-partialmatricesM are inone-to-onecorrespondencewith residueclasses of quadratic formsQM inR[x�, x�, . . . , xn]�IG = R[XG].(�) AG-partial matrixM has a positive semidefinite completion if and only if the cor-responding quadratic formQM is a sum of squares of linear formsmodulo IG.

Minimal Free Resolutions

Let S = R[x�, x�, . . . , xn].Definition. A minimal free resolution of a graded S-module M is an exact se-quence

� δt+��→ Ftδt�→ Ft−� δt−��→ � δ��→ F� → M → �,

where the Fi are free S-modules and the image of the map δi is contained in thesubmodule (x�, x�, . . . , xn)Fi−� of Fi−� for all i ≥ �.Example. ConsiderM� = S��x�x�, x�x�� andM� = S��x�x�, x�x��.

�→ S(x�,−x�)T����→ S�

(x�x�,x�x�)�����→ S → M� → �

�→ S(x�x�,−x�x�)T������→ S�

(x�x�,x�x�)�����→ S → M� → �

areminimal free resolutions ofM� andM�.

Definition. A graded S-module M is �-regular (Castelnuovo-Mumford) if everythe degree of every entry of the matrices representing δi is �.

Theorem (Fröberg). The ideal IG is �-regular if and only ifG is chordal.

Extreme Rays of the Dual Convex Cone

Proposition. LetG = ([m], E) be them-cycle. Thematrix

����������������

m−�m−� −� � � . . . � � �

m−�−� � −� � . . . � � �� −� � −� . . . ⋮ ⋮ ⋮⋮ � −� � � ⋮ ⋮ ⋮⋮ ⋮ � � � ⋮� � � � . . . � −� �� � � � . . . −� � −��

m−� � � � . . . � −� m−�m−�

����������������is an extreme ray of Σ∨G. It has rankm − �. It certifies that

�����������

� � ? ? . . . ? −�� � � ? . . . ? ?? � � � . . . ? ?⋮ � � � ⋮? ? ? ? . . . � �−� ? ? ? . . . � �

�����������does not have a positive semidefinite completion even though every complete princi-pal minor is positive semidefinite.

References

Thank you

Example. ConsiderM� = �x�x�, x�x�� andM� = �x�x�, x�x��.�→ S

(x�,−x�)T����→ S�(x�x�,x�x�)�����→ M� → �

�→ S(x�x�,−x�x�)T������→ S�

(x�x�,x�x�)�����→ M� → �

areminimal free resolutions ofM� andM�.

Definition. A graded S-moduleM is �-regular (Castelnuovo-Mumford) if the en-tries of δ� have degree � and the entries of δi have degree � for all i > �.Theorem (Fröberg). The ideal IG is �-regular if and only ifG is chordal.

Extreme Rays of the Dual Convex Cone

Proposition. LetG = ([m], E) be them-cycle. Thematrix

����������������

m−�m−� −� � � . . . � � �

m−�−� � −� � . . . � � �� −� � −� . . . ⋮ ⋮ ⋮⋮ � −� � � ⋮ ⋮ ⋮⋮ ⋮ � � � ⋮� � � � . . . � −� �� � � � . . . −� � −��

m−� � � � . . . � −� m−�m−�

����������������is an extreme ray of Σ∨G. It has rankm − �. It certifies that

�����������

� � ? ? . . . ? −�� � � ? . . . ? ?? � � � . . . ? ?⋮ � � � ⋮? ? ? ? . . . � �−� ? ? ? . . . � �

�����������does not have a positive semidefinite completion even though every complete princi-pal minor is positive semidefinite.

The Hankel Spectrahedron

Definition. The convex dual cone of ΣG is the Hankel spectrahedron of G.

Proposition.

Σ∨G = S+ ∩ (ker(πG))⊥Proof.

Σ∨G = (πG(S+))∨ = S+ ∩ (ker(πG))⊥

Example. ConsiderM� = �x�x�, x�x�� andM� = �x�x�, x�x��.�→ S

(x�,−x�)T����→ S�(x�x�,x�x�)�����→ M� → �

�→ S(x�x�,−x�x�)T������→ S�

(x�x�,x�x�)�����→ M� → �

areminimal free resolutions ofM� andM�.

Definition. A homogeneous ideal I ⊂ S is �-regular (Castelnuovo-Mumford) if itis generated by quadrics and the entries of δi have degree � for all i ≥ �.Theorem (Fröberg). The ideal IG is �-regular if and only ifG is chordal.

Extreme Rays of the Dual Convex Cone

Proposition. LetG = ([m], E) be them-cycle. Thematrix

����������������

m−�m−� −� � � . . . � � �

m−�−� � −� � . . . � � �� −� � −� . . . ⋮ ⋮ ⋮⋮ � −� � � ⋮ ⋮ ⋮⋮ ⋮ � � � ⋮� � � � . . . � −� �� � � � . . . −� � −��

m−� � � � . . . � −� m−�m−�

����������������is an extreme ray of Σ∨G. It has rankm − �. It certifies that

�����������

� � ? ? . . . ? −�� � � ? . . . ? ?? � � � . . . ? ?⋮ � � � ⋮? ? ? ? . . . � �−� ? ? ? . . . � �

�����������does not have a positive semidefinite completion even though every complete princi-pal minor is positive semidefinite.

The Hankel Spectrahedron

Definition. The convex dual cone of ΣG is the Hankel spectrahedron of G.

Proposition.

Σ∨G = S+ ∩ (ker(πG))⊥Proof.

Σ∨G = (πG(S+))∨ = S+ ∩ (ker(πG))⊥

Page 10: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

Example. ConsiderM� = S��x�x�, x�x�� andM� = S��x�x�, x�x��.�→ S

(x�,−x�)T����→ S�(x�x�,x�x�)�����→ S → M� → �

�→ S(x�x�,−x�x�)T������→ S�

(x�x�,x�x�)�����→ S → M� → �

areminimal free resolutions ofM� andM�.

Definition. A graded S-module M is �-regular (Castelnuovo-Mumford) if everythe degree of every entry of the matrices representing δi is �.

Theorem (Fröberg). The ideal IG is �-regular if and only ifG is chordal.

Extreme Rays of the Dual Convex Cone

Proposition. LetG = ([m], E) be them-cycle. Thematrix

����������������

m−�m−� −� � � . . . � � �

m−�−� � −� � . . . � � �� −� � −� . . . ⋮ ⋮ ⋮⋮ � −� � � ⋮ ⋮ ⋮⋮ ⋮ � � � ⋮� � � � . . . � −� �� � � � . . . −� � −��

m−� � � � . . . � −� m−�m−�

����������������is an extreme ray of Σ∨G. It has rankm − �. It certifies that

�����������

� � ? ? . . . ? −�� � � ? . . . ? ?? � � � . . . ? ?⋮ � � � ⋮? ? ? ? . . . � �−� ? ? ? . . . � �

�����������does not have a positive semidefinite completion even though every complete princi-pal minor is positive semidefinite.

The Hankel Spectrahedron

Definition. The convex dual cone of ΣG is the Hankel spectrahedron of G.

Proposition.

Σ∨G = S+ ∩ (ker(πG))⊥Proof.

Σ∨G = (πG(S+))∨ = S+ ∩ (ker(πG))⊥

Example. ConsiderM� = S��x�x�, x�x�� andM� = S��x�x�, x�x��.�→ S

(x�,−x�)T����→ S�(x�x�,x�x�)�����→ S → M� → �

�→ S(x�x�,−x�x�)T������→ S�

(x�x�,x�x�)�����→ S → M� → �

areminimal free resolutions ofM� andM�.

Definition. A graded S-module M is �-regular (Castelnuovo-Mumford) if everythe degree of every entry of the matrices representing δi is �.

Theorem (Fröberg). The ideal IG is �-regular if and only ifG is chordal.

Extreme Rays of the Dual Convex Cone

Proposition. LetG = ([m], E) be them-cycle. Thematrix

����������������

m−�m−� −� � � . . . � � �

m−�−� � −� � . . . � � �� −� � −� . . . ⋮ ⋮ ⋮⋮ � −� � � ⋮ ⋮ ⋮⋮ ⋮ � � � ⋮� � � � . . . � −� �� � � � . . . −� � −��

m−� � � � . . . � −� m−�m−�

����������������is an extreme ray of Σ∨G. It has rankm − �. It certifies that

�����������

� � ? ? . . . ? −�� � � ? . . . ? ?? � � � . . . ? ?⋮ � � � ⋮? ? ? ? . . . � �−� ? ? ? . . . � �

�����������does not have a positive semidefinite completion even though every complete princi-pal minor is positive semidefinite.

The Hankel Spectrahedron

Definition. The convex dual cone of ΣG is the Hankel spectrahedron of G.

Proposition.

Σ∨G = S+ ∩ (ker(πG))⊥Proof.

Σ∨G = (πG(S+))∨ = S+ ∩ (ker(πG))⊥

Example. ConsiderM� = S��x�x�, x�x�� andM� = S��x�x�, x�x��.�→ S

(x�,−x�)T����→ S�(x�x�,x�x�)�����→ S → M� → �

�→ S(x�x�,−x�x�)T������→ S�

(x�x�,x�x�)�����→ S → M� → �

areminimal free resolutions ofM� andM�.

Definition. A graded S-module M is �-regular (Castelnuovo-Mumford) if everythe degree of every entry of the matrices representing δi is �.

Theorem (Fröberg). The ideal IG is �-regular if and only ifG is chordal.

Extreme Rays of the Dual Convex Cone

Proposition. LetG = ([m], E) be them-cycle. Thematrix

����������������

m−�m−� −� � � . . . � � �

m−�−� � −� � . . . � � �� −� � −� . . . ⋮ ⋮ ⋮⋮ � −� � � ⋮ ⋮ ⋮⋮ ⋮ � � � ⋮� � � � . . . � −� �� � � � . . . −� � −��

m−� � � � . . . � −� m−�m−�

����������������is an extreme ray of Σ∨G. It has rankm − �. It certifies that

�����������

� � ? ? . . . ? −�� � � ? . . . ? ?? � � � . . . ? ?⋮ � � � ⋮? ? ? ? . . . � �−� ? ? ? . . . � �

�����������does not have a positive semidefinite completion even though every complete princi-pal minor is positive semidefinite.

The Hankel Spectrahedron

Definition. The convex dual cone of ΣG is the Hankel spectrahedron of G.

Proposition.

Σ∨G = S+ ∩ (ker(πG))⊥Proof.

Σ∨G = (πG(S+))∨ = S+ ∩ (ker(πG))⊥Question:What are the extreme rays of Σ∨G?Proposition. Extreme rays of Σ∨G that have rank � are in one-to-one correspondencewith points x ∈ V(IG) = XG.

Proof. R = xxT ∈ Σ∨G,M ∈ ker(πG)) ≅ IG:� = �M , R� = tr(MR) = tr(MxxT) = xTMx = QM(x)

Definition. The Hankel index of G is the smallest r > � such that Σ∨G has an ex-treme ray of rank r.

Hankel index and bpf linear series

Let R ∈ Σ∨G = S+ ∩ (ker(πG))⊥ be an extreme ray.(�) The kernel of R is a linear series on XG: ker(R) ⊂ R[x�, x�, . . . , xn]�.(�) If rk(R) > �, then ker(R) is a base-point free linear series on XG. (Blekherman)(�) �ker(R)�� is contained in R⊥.

Theorem(Blekherman, Sinn,Velasco). TheHankel indexofG isat least the (Green-Lazarsfeld index of IG) + �.

Theorem (Eisenbud, Green, Hulek, Popescu). TheGreen-Lazarsfeld index of IG isthe (smallest length of a chordless cycle inG) - �.

References

Thank you

Question:What are the extreme rays of Σ∨G?Proposition. Extreme rays of Σ∨G that have rank � are in one-to-one correspondencewith points x ∈ V(IG) = XG.

Proof. R = xxT ∈ Σ∨G,M ∈ ker(πG)) ≅ IG:� = �M , R� = tr(MR) = tr(MxxT) = xTMx = QM(x)

Definition. The Hankel index of G is the smallest r > � such that Σ∨G has an ex-treme ray of rank r.

Hankel index and bpf linear series

Let R ∈ Σ∨G = S+ ∩ (ker(πG))⊥ be an extreme ray.(�) The kernel of R is a linear series on XG: ker(R) ⊂ R[x�, x�, . . . , xn]�.(�) If rk(R) > �, then ker(R) is a base-point free linear series on XG. (Blekherman)(�) �ker(R)�� is contained in R⊥.

Theorem(Blekherman, Sinn,Velasco). TheHankel indexofG isat least the (Green-Lazarsfeld index of IG) + �.

Theorem (Eisenbud, Green, Hulek, Popescu). TheGreen-Lazarsfeld index of IG isthe (smallest length of a chordless cycle inG) - �.

References

Thank you

Question:What are the extreme rays of Σ∨G?Proposition. Extreme rays of Σ∨G that have rank � are in one-to-one correspondencewith points x ∈ V(IG) = XG.

Proof. R = xxT ∈ Σ∨G,M ∈ ker(πG)) ≅ IG:� = �M , R� = tr(MR) = tr(MxxT) = xTMx = QM(x)

Definition. The Hankel index of G is the smallest r > � such that Σ∨G has an ex-treme ray of rank r.

Hankel index and bpf linear series

Let R ∈ Σ∨G = S+ ∩ (ker(πG))⊥ be an extreme ray.(�) The kernel of R is a linear series on XG: ker(R) ⊂ R[x�, x�, . . . , xn]�.(�) If rk(R) > �, then ker(R) is a base-point free linear series on XG. (Blekherman)(�) �ker(R)�� is contained in R⊥.

Theorem(Blekherman, Sinn,Velasco). TheHankel indexofG isat least the (Green-Lazarsfeld index of IG) + �.

Theorem (Eisenbud, Green, Hulek, Popescu). TheGreen-Lazarsfeld index of IG isthe (smallest length of a chordless cycle inG) - �.

References

Thank you

Page 11: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

Extreme Rays of the Dual Convex Cone

Proposition. LetG = ([m], E) be them-cycle. Thematrix

����������������

m−�m−� −� � � . . . � � �

m−�−� � −� � . . . � � �� −� � −� . . . ⋮ ⋮ ⋮⋮ � −� � � ⋮ ⋮ ⋮⋮ ⋮ � � � ⋮� � � � . . . � −� �� � � � . . . −� � −��

m−� � � � . . . � −� m−�m−�

����������������is an extreme ray of Σ∨G. It has rankm − �. It certifies that

�����������

� � ? ? . . . ? −�� � � ? . . . ? ?? � � � . . . ? ?⋮ � � � ⋮? ? ? ? . . . � �−� ? ? ? . . . � �

�����������does not have a positive semidefinite completion even though every complete princi-pal minor is positive semidefinite.

References

Thank you

Theorem (Blekherman, Sinn).

gcr(Km,m) = mmlt(Km,m) =min�k∶ �k + �

�� ≥ �m� ∈ O(�√m).

GramDimension (SOS length)

Definition (Laurent,Varvitsiotis). TheGramdimensionof a labeled simplegraphG = ([n], E) is the smallest whole number k such that for any positive semidef-inite matrix M ∈ S+, there exists another M′ ∈ S+ of rank at most k such thatπG(M) = πG(M′).Theorem (Laurent, Varvitsiotis). (�) k ≤ �: The Gramdimension ofG is atmost k ifand only ifG has no Kk+�minor.(�) The Gram dimension ofG is at most � if and only ifG has no K� and K�,�,�minors.

Example. The Gram dimension of anym-cylce is �.

m = �References

Thank you

Page 12: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

Question:What are the extreme rays of Σ∨G?Proposition. Extreme rays of Σ∨G that have rank � are in one-to-one correspondencewith points x ∈ V(IG) = XG.

Proof. R = xxT ∈ Σ∨G,M ∈ ker(πG)) ≅ IG:� = �M , R� = tr(MR) = tr(MxxT) = xTMx = QM(x)

Hankel index and bpf linear series

Definition. The Hankel index of G is the smallest r > � such that Σ∨G has an ex-treme ray of rank r.

Let R ∈ Σ∨G = S+ ∩ (ker(πG))⊥ be an extreme ray.(�) The kernel of R is a linear series on XG: ker(R) ⊂ R[x�, x�, . . . , xn]�.(�) If rk(R) > �, then ker(R) is a base-point free linear series on XG. (Blekherman)(�) �ker(R)�� is contained in R⊥.

Theorem(Blekherman, Sinn,Velasco). TheHankel indexofG isat least the (Green-Lazarsfeld index of IG) + �.

Theorem (Eisenbud, Green, Hulek, Popescu). TheGreen-Lazarsfeld index of IG isthe (smallest length of a chordless cycle inG) - �.

References

Thank you

Question:What are the extreme rays of Σ∨G?Proposition. Extreme rays of Σ∨G that have rank � are in one-to-one correspondencewith points x ∈ V(IG) = XG.

Proof. R = xxT ∈ Σ∨G,M ∈ ker(πG)) ≅ IG:� = �M , R� = tr(MR) = tr(MxxT) = xTMx = QM(x)

Hankel index and bpf linear series

Definition. The Hankel index of G is the smallest r > � such that Σ∨G has an ex-treme ray of rank r.

Let R ∈ Σ∨G = S+ ∩ (ker(πG))⊥ be an extreme ray.(�) The kernel of R is a linear series on XG: ker(R) ⊂ R[x�, x�, . . . , xn]�.(�) If rk(R) > �, then ker(R) is a base-point free linear series on XG. (Blekherman)(�) �ker(R)�� is contained in R⊥.

Theorem(Blekherman, Sinn,Velasco). TheHankel indexofG isat least the (Green-Lazarsfeld index of IG) + �.

Theorem (Eisenbud, Green, Hulek, Popescu). TheGreen-Lazarsfeld index of IG isthe (smallest length of a chordless cycle inG) - �.

References

Thank you

Question:What are the extreme rays of Σ∨G?Proposition. Extreme rays of Σ∨G that have rank � are in one-to-one correspondencewith points x ∈ V(IG) = XG.

Proof. R = xxT ∈ Σ∨G,M ∈ ker(πG)) ≅ IG:� = �M , R� = tr(MR) = tr(MxxT) = xTMx = QM(x)

Hankel index and bpf linear series

Definition. The Hankel index of G is the smallest r > � such that Σ∨G has an ex-treme ray of rank r.

Let R ∈ Σ∨G = S+ ∩ (ker(πG))⊥ be an extreme ray.(�) The kernel of R is a linear series on XG: ker(R) ⊂ R[x�, x�, . . . , xn]�.(�) If rk(R) > �, then ker(R) is a base-point free linear series on XG. (Blekherman)(�) �ker(R)�� is contained in R⊥.

Theorem(Blekherman, Sinn,Velasco). TheHankel indexofG isat least the (Green-Lazarsfeld index of IG) + �.

Theorem (Eisenbud, Green, Hulek, Popescu). TheGreen-Lazarsfeld index of IG isthe (smallest length of a chordless cycle inG) - �.

References

Thank you

Question:What are the extreme rays of Σ∨G?Proposition. Extreme rays of Σ∨G that have rank � are in one-to-one correspondencewith points x ∈ V(IG) = XG.

Proof. R = xxT ∈ Σ∨G,M ∈ ker(πG)) ≅ IG:� = �M , R� = tr(MR) = tr(MxxT) = xTMx = QM(x)

Hankel index and bpf linear series

Definition. The Hankel index of G is the smallest r > � such that Σ∨G has an ex-treme ray of rank r.

Let R ∈ Σ∨G = S+ ∩ (ker(πG))⊥ be an extreme ray.(�) The kernel of R is a linear series on XG: ker(R) ⊂ R[x�, x�, . . . , xn]�.(�) If rk(R) > �, then ker(R) is a base-point free linear series on XG. (Blekherman)(�) �ker(R)�� is contained in R⊥.

Theorem(Blekherman, Sinn,Velasco). TheHankel indexofG isat least the (Green-Lazarsfeld index of IG) + �.

Theorem (Eisenbud, Green, Hulek, Popescu). TheGreen-Lazarsfeld index of IG isthe (smallest length of a chordless cycle inG) - �.

References

Thank you

Question:What are the extreme rays of Σ∨G?Proposition. Extreme rays of Σ∨G that have rank � are in one-to-one correspondencewith points x ∈ V(IG) = XG.

Proof. R = xxT ∈ Σ∨G,M ∈ ker(πG)) ≅ IG:� = �M , R� = tr(MR) = tr(MxxT) = xTMx = QM(x)

Hankel index and bpf linear series

Definition. The Hankel index of G is the smallest r > � such that Σ∨G has an ex-treme ray of rank r.

Let R ∈ Σ∨G = S+ ∩ (ker(πG))⊥ be an extreme ray.(�) The kernel of R is a linear series on XG: ker(R) ⊂ R[x�, x�, . . . , xn]�.(�) If rk(R) > �, then ker(R) is a base-point free linear series on XG. (Blekherman)(�) �ker(R)�� is contained in R⊥.

Theorem(Blekherman, Sinn,Velasco). TheHankel indexofG isat least the (Green-Lazarsfeld index of IG) + �.

Theorem (Eisenbud, Green, Hulek, Popescu). TheGreen-Lazarsfeld index of IG isthe (smallest length of a chordless cycle inG) - �.

References

Thank you

Question:What are the extreme rays of Σ∨G?Proposition. Extreme rays of Σ∨G that have rank � are in one-to-one correspondencewith points x ∈ V(IG) = XG.

Proof. R = xxT ∈ Σ∨G,M ∈ ker(πG)) ≅ IG:� = �M , R� = tr(MR) = tr(MxxT) = xTMx = QM(x)

Hankel index and bpf linear series

Definition. The Hankel index of G is the smallest r > � such that Σ∨G has an ex-treme ray of rank r.

Let R ∈ Σ∨G = S+ ∩ (ker(πG))⊥ be an extreme ray.(�) The kernel of R is a linear series on XG: ker(R) ⊂ R[x�, x�, . . . , xn]�.(�) If rk(R) > �, then ker(R) is a base-point free linear series on XG. (Blekherman)(�) �ker(R)�� is contained in R⊥.

Theorem(Blekherman, Sinn,Velasco). TheHankel indexofG isat least the (Green-Lazarsfeld index of IG) + �.

Theorem (Eisenbud, Green, Hulek, Popescu). TheGreen-Lazarsfeld index of IG isthe (smallest length of a chordless cycle inG) - �.

References

Thank you

Question:What are the extreme rays of Σ∨G?Proposition. Extreme rays of Σ∨G that have rank � are in one-to-one correspondencewith points x ∈ V(IG) = XG.

Proof. R = xxT ∈ Σ∨G,M ∈ ker(πG)) ≅ IG:� = �M , R� = tr(MR) = tr(MxxT) = xTMx = QM(x)

Hankel index and bpf linear series

Definition. The Hankel index of G is the smallest r > � such that Σ∨G has an ex-treme ray of rank r.

Let R ∈ Σ∨G = S+ ∩ (ker(πG))⊥ be an extreme ray.(�) The kernel of R is a linear series on XG: ker(R) ⊂ R[x�, x�, . . . , xn]�.(�) If rk(R) > �, then ker(R) is a base-point free linear series on XG. (Blekherman)(�) �ker(R)�� is contained in R⊥.

Theorem(Blekherman, Sinn,Velasco). TheHankel indexofG isat least the (Green-Lazarsfeld index of IG) + �.

Theorem (Eisenbud, Green, Hulek, Popescu). TheGreen-Lazarsfeld index of IG isthe (smallest length of a chordless cycle inG) - �.

Positive Definite Completions

Theorem. Letm the smallest length of a chordless cycle in G. A G-partial matrixMhas a positive definite completion if and only if(�) every complete principal minor ofM is positive, and(�)M has a positive semidefinite completion of rank at least n −m + �.

MLT vs. GCR

Let G = ([n], E) be a labeled simple graph.

Definition. (�) The maximum likelihood threshold of G is the smallest wholenumber k such that there exists a positive definite matrix D with πG(M) = πG(D)for almost all positive semidefinite matricesM of rank at most k.(�) The generic completion rank of G is the smallest whole number k such thatdim(πG(Vk)) = n + �E, where Vk is the variety of matrices of rank at most k.

Proposition (Sard’sTheoremand Implicit FunctionTheorem). gcr(G) ≥mlt(G)

Page 13: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

Question:What are the extreme rays of Σ∨G?Proposition. Extreme rays of Σ∨G that have rank � are in one-to-one correspondencewith points x ∈ V(IG) = XG.

Proof. R = xxT ∈ Σ∨G,M ∈ ker(πG)) ≅ IG:� = �M , R� = tr(MR) = tr(MxxT) = xTMx = QM(x)

Hankel index and bpf linear series

Definition. The Hankel index of G is the smallest r > � such that Σ∨G has an ex-treme ray of rank r.

Let R ∈ Σ∨G = S+ ∩ (ker(πG))⊥ be an extreme ray.(�) The kernel of R is a linear series on XG: ker(R) ⊂ R[x�, x�, . . . , xn]�.(�) If rk(R) > �, then ker(R) is a base-point free linear series on XG. (Blekherman)(�) �ker(R)�� is contained in R⊥.

Theorem(Blekherman, Sinn,Velasco). TheHankel indexofG isat least the (Green-Lazarsfeld index of IG) + �.

Theorem (Eisenbud, Green, Hulek, Popescu). TheGreen-Lazarsfeld index of IG isthe (smallest length of a chordless cycle inG) - �.

Positive Definite Completions

Theorem. Letm the smallest length of a chordless cycle in G. A G-partial matrixMhas a positive definite completion if and only if(�) every complete principal minor ofM is positive, and(�)M has a positive semidefinite completion of rank at least n −m + �.

MLT vs. GCR

Let G = ([n], E) be a labeled simple graph.

Definition. (�) The maximum likelihood threshold of G is the smallest wholenumber k such that there exists a positive definite matrix D with πG(M) = πG(D)for almost all positive semidefinite matricesM of rank at most k.(�) The generic completion rank of G is the smallest whole number k such thatdim(πG(Vk)) = n + �E, where Vk is the variety of matrices of rank at most k.

Proposition (Sard’sTheoremand Implicit FunctionTheorem). gcr(G) ≥mlt(G)

Question:What are the extreme rays of Σ∨G?Proposition. Extreme rays of Σ∨G that have rank � are in one-to-one correspondencewith points x ∈ V(IG) = XG.

Proof. R = xxT ∈ Σ∨G,M ∈ ker(πG)) ≅ IG:� = �M , R� = tr(MR) = tr(MxxT) = xTMx = QM(x)

Hankel index and bpf linear series

Definition. The Hankel index of G is the smallest r > � such that Σ∨G has an ex-treme ray of rank r.

Let R ∈ Σ∨G = S+ ∩ (ker(πG))⊥ be an extreme ray.(�) The kernel of R is a linear series on XG: ker(R) ⊂ R[x�, x�, . . . , xn]�.(�) If rk(R) > �, then ker(R) is a base-point free linear series on XG. (Blekherman)(�) �ker(R)�� is contained in R⊥.

Theorem(Blekherman, Sinn,Velasco). TheHankel indexofG isat least the (Green-Lazarsfeld index of IG) + �.

Theorem (Eisenbud, Green, Hulek, Popescu). TheGreen-Lazarsfeld index of IG isthe (smallest length of a chordless cycle inG) - �.

Positive Definite Completions

Theorem. Letm the smallest length of a chordless cycle in G. A G-partial matrixMhas a positive definite completion if and only if(�) every complete principal minor ofM is positive, and(�)M has a positive semidefinite completion of rank at least n −m + �.

MLT vs. GCR

Let G = ([n], E) be a labeled simple graph.

Definition. (�) The maximum likelihood threshold of G is the smallest wholenumber k such that there exists a positive definite matrix D with πG(M) = πG(D)for almost all positive semidefinite matricesM of rank at most k.(�) The generic completion rank of G is the smallest whole number k such thatdim(πG(Vk)) = n + �E, where Vk is the variety of matrices of rank at most k.

Proposition (Sard’sTheoremand Implicit FunctionTheorem). gcr(G) ≥mlt(G)Theorem (Blekherman, Sinn).

gcr(Km,m) = mmlt(Km,m) =min�k∶ �k + �

�� ≥ �m� ∈ O(�√m).

GramDimension (SOS length)

Definition (Laurent,Varvitsiotis). TheGramdimensionof a labeled simplegraphG = ([n], E) is the smallest whole number k such that for any positive semidef-inite matrix M ∈ S+, there exists another M′ ∈ S+ of rank at most k such thatπG(M) = πG(M′).Theorem (Laurent, Varvitsiotis). (�) k ≤ �: The Gramdimension ofG is atmost k ifand only ifG has no Kk+�minor.(�) The Gram dimension ofG is at most � if and only ifG has no K� and K�,�,�minors.

Example. The Gram dimension of anym-cylce is �.

References

Thank you

Page 14: icerm.brown.edu · Symmetricmatrixcompletion ? −? ? − ? − Constraints: ()Rank(smallestpossible,fixed,…) ()Positivesemidefinite ∗ ? ∗? ∗ ? ∗ ? ∗ SpectralTheorem

Theorem (Blekherman, Sinn).

gcr(Km,m) = mmlt(Km,m) =min�k∶ �k + �

�� ≥ �m� ∈ O(�√m).

GramDimension (SOS length)

Definition (Laurent,Varvitsiotis). TheGramdimensionof a labeled simplegraphG = ([n], E) is the smallest whole number k such that for any positive semidef-inite matrix M ∈ S+, there exists another M′ ∈ S+ of rank at most k such thatπG(M) = πG(M′).Theorem (Laurent, Varvitsiotis). (�) k ≤ �: The Gramdimension ofG is atmost k ifand only ifG has no Kk+�minor.(�) The Gram dimension ofG is at most � if and only ifG has no K� and K�,�,�minors.

Example. The Gram dimension of anym-cylce is �.

References

Thank you

Theorem (Blekherman, Sinn).

gcr(Km,m) = mmlt(Km,m) =min�k∶ �k + �

�� ≥ �m� ∈ O(�√m).

GramDimension (SOS length)

Definition (Laurent,Varvitsiotis). TheGramdimensionof a labeled simplegraphG = ([n], E) is the smallest whole number k such that for any positive semidef-inite matrix M ∈ S+, there exists another M′ ∈ S+ of rank at most k such thatπG(M) = πG(M′).Theorem (Laurent, Varvitsiotis). (�) k ≤ �: The Gramdimension ofG is atmost k ifand only ifG has no Kk+�minor.(�) The Gram dimension ofG is at most � if and only ifG has no K� and K�,�,�minors.

Example. The Gram dimension of anym-cylce is �.

References

Thank you

Theorem (Blekherman, Sinn).

gcr(Km,m) = mmlt(Km,m) =min�k∶ �k + �

�� ≥ �m� ∈ O(�√m).

GramDimension (SOS length)

Definition (Laurent,Varvitsiotis). TheGramdimensionof a labeled simplegraphG = ([n], E) is the smallest whole number k such that for any positive semidef-inite matrix M ∈ S+, there exists another M′ ∈ S+ of rank at most k such thatπG(M) = πG(M′).Theorem (Laurent, Varvitsiotis). (�) k ≤ �: The Gramdimension ofG is atmost k ifand only ifG has no Kk+�minor.(�) The Gram dimension ofG is at most � if and only ifG has no K� and K�,�,�minors.

Example. The Gram dimension of anym-cylce is �.

References

Thank you