Clayton Shonkwiler PRO
Mathematician and artist
/kc24
this talk!
SIAM Minisymposium on Interactions Among Analysis, Optimization, and Network Science
October 5, 2024
Florida State University
National Science Foundation (DMS–2107700)
Definition.
A∈Cd×d is normal if AA∗=A∗A.
Equivalently,
0=AA∗−A∗A=[A,A∗].
Define the non-normal energy E:Cd×d→R by
E(A):=∥[A,A∗]∥2.
Obvious Fact.
The normal matrices are the global minima of E.
Theorem [with Needham]
The only critical points of E are the global minima; i.e., the normal matrices.
E is not quasiconvex!
Theorem [with Needham]
The only critical points of E are the global minima; i.e., the normal matrices.
Let F:Cd×d×R→Cd×d be negative gradient descent of E; i.e.,
F(A0,0)=A0dtdF(A0,t)=−∇E(F(A0,t)).
Theorem [with Needham]
For any A0∈Cd×d, the matrix A∞:=limt→∞F(A0,t) exists, is normal, has the same eigenvalues as A0, and is real if A0 is.
Cd×d is symplectic, with symplectic form ωA(X,Y)=−Im⟨X,Y⟩=−ImTr(Y∗X).
A symplectic manifold is a smooth manifold M together with a closed, non-degenerate 2-form ω∈Ω2(M).
Example: (R2,dx∧dy)=(C,2idz∧dzˉ)
Cd×d is symplectic, with symplectic form ωA(X,Y)=−Im⟨X,Y⟩=−ImTr(Y∗X).
Consider the conjugation action of SU(d) on Cd×d: U⋅A =UAU∗.
This action is Hamiltonian with associated momentum map μ:Cd×d→H0(d) given by
μ(A):=[A,A∗].
So E(A)=∥μ(A)∥2.
Frances Kirwan
Gert-Martin Greuel [CC BY-SA 2.0 DE], from Oberwolfach Photo Collection
Image by rawpixel.com on Freepik
This kind of function is really nice!
The GIT quotient consists of group orbits which can be distinguished by G-invariant (homogeneous) polynomials.
C∗↷CP2
t⋅[z0:z1:z2]=[z0:tz1:t1z2]
Roughly: identify orbits whose closures intersect, throw away orbits on which all G-invariant polynomials vanish.
CP2//C∗≅CP1
Let T≃U(1)d−1 be the diagonal subgroup of SU(d). The conjugation action of T on Cd×d is also Hamiltonian, with momentum map
A↦diag([A,A∗]).
[A,A∗]ii=∥Ai∥2−∥Ai∥2, where Ai is the ith row of A and Ai is the ith column.
If A=(aij)i,j∈Rd×d such that diag([A,A∗])=0, then A=(aij2)i,j is the adjacency matrix of a balanced multigraph.
Define the unbalanced energy B(A):=∥diag([A,A∗])∥2=∑(∥Ai∥2−∥Ai∥2)2.
Let F(A0,0)=A0,dtdF(A0,t)=−∇B(F(A0,t)) be negative gradient flow of B.
Theorem (with Needham)
For any A0∈Cd×d, the matrix A∞:=limt→∞F(A0,t) exists, is balanced, has the same eigenvalues and principal minors as A0, and has zero entries whenever A0 does.
If A0 is real, so is A∞, and if A0 has all non-negative entries, then so does A∞.
This is “local”: aij is updated by a multiple of (∥Aj∥2−∥Aj∥2)−(∥Ai∥2−∥Ai∥2).
Theorem (with Needham)
For any A0∈Cd×d, the matrix A∞:=limt→∞F(A0,t) exists, is balanced, has the same eigenvalues and principal minors as A0, and has zero entries whenever A0 does.
If A0 is real, so is A∞, and if A0 has all non-negative entries, then so does A∞.
Theorem (with Needham)
For any A0∈Cd×d, the matrix A∞:=limt→∞F(A0,t) exists, is balanced, has the same eigenvalues and principal minors as A0, and has zero entries whenever A0 does.
If A0 is real, so is A∞, and if A0 has all non-negative entries, then so does A∞.
Theorem (with Needham)
For any A0∈Cd×d, the matrix A∞:=limt→∞F(A0,t) exists, is balanced, has the same eigenvalues and principal minors as A0, and has zero entries whenever A0 does.
If A0 is real, so is A∞, and if A0 has all non-negative entries, then so does A∞.
∥A∥2=1
∥A∥2=0.569
Theorem (with Needham)
For any A0∈Cd×d, the matrix A∞:=limt→∞F(A0,t) exists, is balanced, has the same eigenvalues and principal minors as A0, and has zero entries whenever A0 does.
If A0 is real, so is A∞, and if A0 has all non-negative entries, then so does A∞.
Theorem (with Needham)
For any A0∈Cd×d, the matrix A∞:=limt→∞F(A0,t) exists, is balanced, has the same eigenvalues and principal minors as A0, and has zero entries whenever A0 does.
If A0 is real, so is A∞, and if A0 has all non-negative entries, then so does A∞.
By doing gradient flow F on the unit sphere, we can preserve weights:
Theorem (with Needham)
For any non-nilpotent A0∈Cd×d with ∥A∥2=1, the matrix A∞:=limt→∞F(A0,t) exists, is balanced, has Frobenius norm 1, and has zero entries whenever A0 does.
If A0 is real, so is A∞, and if A0 has all non-negative entries, then so does A∞.
By doing gradient flow F on the unit sphere, we can preserve weights:
Theorem (with Needham)
For any non-nilpotent A0∈Cd×d with ∥A∥2=1, the matrix A∞:=limt→∞F(A0,t) exists, is balanced, has Frobenius norm 1, and has zero entries whenever A0 does.
If A0 is real, so is A∞, and if A0 has all non-negative entries, then so does A∞.
By doing gradient flow F on the unit sphere, we can preserve weights:
Theorem (with Needham)
For any non-nilpotent A0∈Cd×d with ∥A∥2=1, the matrix A∞:=limt→∞F(A0,t) exists, is balanced, has Frobenius norm 1, and has zero entries whenever A0 does.
If A0 is real, so is A∞, and if A0 has all non-negative entries, then so does A∞.
By doing gradient flow F on the unit sphere, we can preserve weights:
Theorem (with Needham)
For any non-nilpotent A0∈Cd×d with ∥A∥2=1, the matrix A∞:=limt→∞F(A0,t) exists, is balanced, has Frobenius norm 1, and has zero entries whenever A0 does.
If A0 is real, so is A∞, and if A0 has all non-negative entries, then so does A∞.
shonkwiler.org/kc24
Fusion frame homotopy and tightening fusion frames by gradient descent
Tom Needham and Clayton Shonkwiler
Journal of Fourier Analysis and Applications 29 (2023), no. 4, 51
Three proofs of the Benedetto–Fickus theorem
Dustin Mixon, Tom Needham, Clayton Shonkwiler, and Soledad Villar
Sampling, Approximation, and Signal Analysis (Harmonic Analysis in the Spirit of J. Rowland Higgins), Stephen D. Casey, M. Maurice Dodson, Paulo J. S. G. Ferreira and Ahmed Zayed, eds., Birkhäuser, Cham, 2023, 371–391
Geometric approaches to matrix normalization and graph balancing
Tom Needham and Clayton Shonkwiler
See also
By Clayton Shonkwiler