An introduction to finite frames

Matthew Fickus

Department of Mathematics and Statistics

Air Force Institute of Technology
Wright-Patterson Air Force Base, Ohio

July 28, 2015

The views expressed in this talk are those of the speaker and do not reflect the official policy
or position of the United States Air Force, Department of Defense, or the U.S. Government.
Outline

• Part I: Finite Frames

− Motivation
− Notation
− Terminology

• Part II: Unit Norm Tight Frames (UNTFs)

− Generalizations of orthonormal bases
− Often studied with differential and algebraic geometry
− Major open problems

• Part III: Equiangular Tight Frames (ETFs)

− Optimal packings of lines
− Closely related to combinatorial design
− Major open problems

1/27
 Part I:
Finite Frames
A Brief History of Frame Theory
1950s: Frames are introduced to study nonharmonic Fourier series.
Infinite-dimensional generalization of standard linear algebra.

1960s-1970s: “Frames” is an obscure term used by harmonic analysts.

Time-frequency analysis routinely used in real-world applications.

1980s-1990s: Wavelets (time-scale analysis) invented to address shortcomings

of time-frequency analysis.
Frame theory used to compare these two competing methods.
Frames popularized as “painless nonorthogonal expansions.”

2000s-2010s: Finite frame theory developed to study packing and covering

problems in Euclidean geometry.
It overlaps with compressed sensing, which is invented to
address shortcomings of wavelets.

Common theme: In what ways (and to what degree) can

nonorthonormal vectors behave like orthonormal vectors?

2/27
Matrix Notation
Definition: Let M, N be positive integers and let either F = R or F = C.
Given N vectors {ϕn }N M
n=1 in F , consider the

 
• M × N synthesis operator Φ = ϕ1 · · · ϕN ,

ϕ∗1


• N × M analysis operator Φ∗ =  ... ,

 

ϕ∗N

• M × M frame operator ΦΦ∗ = ϕ1 ϕ∗1 + · · · + ϕN ϕ∗N ,

ϕ∗1 ϕ1 · · · ϕ∗1 ϕN
 

• N × N Gram matrix Φ∗ Φ =  ... .. .. .


. . 
∗ ∗
ϕN ϕ1 · · · ϕN ϕN

3/27
Orthonormal Bases
Notes:
• Vectors {ϕn }N
n=1 in F
M
are orthonormal if and only if Φ∗ Φ = I.

• Vectors {ϕn }N
n=1 are an orthonormal basis for F
M
if and only if
they’re orthonormal and M = N.

• In that case, Φ is square and Φ∗ = Φ−1 implying that ∀x ∈ FM ,

N
X  N
X
x = ΦΦ∗ x = ϕn ϕ∗n x = (ϕ∗n x)ϕn .
n=1 n=1

• This implies the Pythagorean theorem: ∀x ∈ FM ,

N
X  N
X
kxk2 = x∗ x = x∗ ΦΦ∗ x = x∗ ϕn ϕ∗n x = |ϕ∗n x|2 .
n=1 n=1

4/27
Finite Frames
• Now suppose your real-world application prohibits you from having
{ϕn }N M
n=1 be an orthonormal basis for F .

• As long as {ϕn }N M
n=1 spans F , you can still “painlessly” expand any
x in terms of them: in this case, ΦΦ∗ is invertible and so
N
X  N
X
x = ΦΦ∗ (ΦΦ∗ )−1 x = ΦΨ∗ x = ϕn ψ ∗n x = (ψ ∗n x)ϕn .
n=1 n=1

• This expansion is numerically stable when Φ is well conditioned, i.e.

when {ϕn }N
n=1 satisfies a relaxed Pythagorean theorem:

N
X
αkxk2 ≤ x∗ ΦΦ∗ x = |ϕ∗n x|2 ≤ βkxk2 , ∀x ∈ FM ,
n=1

for “close” scalars 0 < α ≤ β < ∞. Here, we call {ϕn }N

n=1 a frame
for FM with lower and upper frame bounds α and β, respectively.

5/27
Tight Frames

• We say {ϕn }Nn=1 is a tight frame for F

M
if Φ is optimally well
conditioned, namely when there exists α > 0 such that
N
X
αkxk2 = x∗ Φ∗ Φx = |ϕ∗n x|2 , ∀x ∈ FM .
n=1

• This is equivalent to ΦΦ∗ = αI, i.e. to when the rows of Φ are

orthogonal and have constant norm.

• Naimark’s Theorem: Every tight frame is a scalar multiple of an

orthogonal projection of an orthonormal basis.

6/27
Example: 6 × 16 Tight Frame
 
1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 
0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
 
0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0
 
0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
 
0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
Φ =
0

 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
 
0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
 
0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0
 
0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0
 
0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0
 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1

7/27
Example: 6 × 16 Tight Frame
 
1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 
0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
 
0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0
 
0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
 
0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0
Φ =
0

 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0
 
0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0
 
0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0
 
0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0
 
0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0
 
0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1

7/27
Example: 6 × 16 Tight Frame
 
1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 
0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
Φ =
0

 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0

7/27
Example: 6 × 16 Tight Frame
 
1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 
0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0
Φ =
0

 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0

Question: This is one of many tight frames of 16 vectors in R6 ... can we

find others that are even more like orthonormal bases in some sense?

7/27
Example: Better 6 × 16 Tight Frame
 
+ + + + + + + + + + + + + + + +
+
 − + − + − + − + − + − + − + −
+
 + − − + + − − + + − − + + − −
+
 − − + + − − + + − − + + − − +
+
 + + + − − − − + + + + − − − −
+
 − + − − + − + + − + − − + − +
+
 + − − − − + + + + − − − − + +
1 + − − + − + + − + − − + − + + −
Φ= √  
+
6 + + + + + + + − − − − − − − −
+
 − + − + − + − − + − + − + − +
+
 + − − + + − − − − + + − − + +
+
 − − + + − − + − + + − − + + −
+
 + + + − − − − − − − − + + + +
+
 − + − − + − + − + − + + − + −
+ + − − − − + + − − + + + + − −
+ − − + − + + − − + + − + − − +

Notation: “ + ” = 1, “ − ” = −1
8/27
Example: Better 6 × 16 Tight Frame
 
+ + + + + + + + + + + + + + + +
+
 − + − + − + − + − + − + − + −
+
 + − − + + − − + + − − + + − −
+
 − − + + − − + + − − + + − − +
+
 + + + − − − − + + + + − − − −
+
 − + − − + − + + − + − − + − +
+
 + − − − − + + + + − − − − + +
1 + − − + − + + − + − − + − + + −
Φ= √  
+
6 + + + + + + + − − − − − − − −
+
 − + − + − + − − + − + − + − +
+
 + − − + + − − − − + + − − + +
+
 − − + + − − + − + + − − + + −
+
 + + + − − − − − − − − + + + +
+
 − + − − + − + − + − + + − + −
+ + − − − − + + − − + + + + − −
+ − − + − + + − − + + − + − − +

8/27
Example: Better 6 × 16 Tight Frame
 
+ + + + + + + + + + + + + + + +
+
 − + − + − + − + − + − + − + −
1 + + − − + + − − + + − − + + − −
Φ= √  
+
6 − − + + − − + + − − + + − − +
+ + + + − − − − + + + + − − − −
+ − + − − + − + + − + − − + − +

8/27
Example: Better 6 × 16 Tight Frame
 
+ + + + + + + + + + + + + + + +
+
 − + − + − + − + − + − + − + −
1 + + − − + + − − + + − − + + − −
Φ= √  
+
6 − − + + − − + + − − + + − − +
+ + + + − − − − + + + + − − − −
+ − + − − + − + + − + − − + − +

Note: All columns are unit norm.

8/27
Example: Better 6 × 16 Tight Frame
 
3 0 + 0 + 0 − 0 3 0 + 0 + 0 − 0
 0
 3 0 + 0 + 0 − 0 3 0 + 0 + 0 −
+
 0 3 0 − 0 + 0 + 0 3 0 − 0 + 0
 0
 + 0 3 0 − 0 + 0 + 0 3 0 − 0 +
+
 0 − 0 3 0 + 0 + 0 − 0 3 0 + 0
 0
 + 0 − 0 3 0 + 0 + 0 − 0 3 0 +
− 0 + 0 + 0 3 0 − 0 + 0 + 0 3 0
 
1 0 − 0 + 0 + 0 3 0 − 0 + 0 + 0 3
Φ∗ Φ =  
 3
3 0 + 0 + 0 − 0 3 0 + 0 + 0 − 0
 0
 3 0 + 0 + 0 − 0 3 0 + 0 + 0 −
+
 0 3 0 − 0 + 0 + 0 3 0 − 0 + 0
 0
 + 0 3 0 − 0 + 0 + 0 3 0 − 0 +
+
 0 − 0 3 0 + 0 + 0 − 0 3 0 + 0
 0
 + 0 − 0 3 0 + 0 + 0 − 0 3 0 +
− 0 + 0 + 0 3 0 − 0 + 0 + 0 3 0
0 − 0 + 0 + 0 3 0 − 0 + 0 + 0 3

8/27
Example: Even Better 6 × 16 Tight Frame
 
+ + + + + + + + + + + + + + + +
+
 − + − + − + − + − + − + − + −
+
 + − − + + − − + + − − + + − −
+
 − − + + − − + + − − + + − − +
+
 + + + − − − − + + + + − − − −
+
 − + − − + − + + − + − − + − +
+
 + − − − − + + + + − − − − + +
1 + − − + − + + − + − − + − + + −
Φ= √  
+
6 + + + + + + + − − − − − − − −
+
 − + − + − + − − + − + − + − +
+
 + − − + + − − − − + + − − + +
+
 − − + + − − + − + + − − + + −
+
 + + + − − − − − − − − + + + +
+
 − + − − + − + − + − + + − + −
+ + − − − − + + − − + + + + − −
+ − − + − + + − − + + − + − − +

9/27
Example: Even Better 6 × 16 Tight Frame
 
+ + + + + + + + + + + + + + + +
+
 − + − + − + − + − + − + − + −
+
 + − − + + − − + + − − + + − −
+
 − − + + − − + + − − + + − − +
+
 + + + − − − − + + + + − − − −
+
 − + − − + − + + − + − − + − +
+
 + − − − − + + + + − − − − + +
1 + − − + − + + − + − − + − + + −
Φ= √  
+
6 + + + + + + + − − − − − − − −
+
 − + − + − + − − + − + − + − +
+
 + − − + + − − − − + + − − + +
+
 − − + + − − + − + + − − + + −
+
 + + + − − − − − − − − + + + +
+
 − + − − + − + − + − + + − + −
+ + − − − − + + − − + + + + − −
+ − − + − + + − − + + − + − − +

9/27
Example: Even Better 6 × 16 Tight Frame
 
+ + + + + + + + + + + + + + + +
+
 − + − + − + − + − + − + − + −
1 + − − + + − − + + − − + + − − +
Φ= √  
+
6 + + + − − − − + + + + − − − −
+ − + − + − + − − + − + − + − +
+ − − + − + + − − + + − + − − +

9/27
Example: Even Better 6 × 16 Tight Frame
 
3 − + + + − + − + + + + + − − +
− 3 + + − + − + + + + + − + + −
 
+
 + 3 − + − + − + + + + − + + −
+
 + − 3 − + − + + + + + + − − +
+
 − + − 3 − + + + − − + + + + +
− + − + − 3 + + − + + − + + + +
 
+
 − + − + + 3 − − + + − + + + +
1 − + − + + + − 3 + − − + + + + +
Φ∗ Φ =  
3+ + + + + − − + 3 − + + + − + −
+
 + + + − + + − − 3 + + − + − +
+
 + + + − + + − + + 3 − + − + −
+
 + + + + − − + + + − 3 − + − +
+
 − − + + + + + + − + − 3 − + +
− + + − + + + + − + − + − 3 + +
 
− + + − + + + + + − + − + + 3 −
+ − − + + + + + − + − + + + − 3

9/27
 Part II:
Unit Norm Tight Frames
Unit Norm Tight Frames (UNTFs)

Definition: Vectors {ϕn }N n=1 are a unit norm tight frame (UNTF) for
FM if kϕn k = 1 for all n and there exists α > 0 such that
N
X
αI = ΦΦ∗ = ϕn ϕ∗n ,
n=1

i.e. if the orthogonal projection operators onto these lines sum to a scalar
multiple of the identity.

N
Note: Here α is necessarily the redundancy M since

N
X
Mα = Tr(αI) = Tr(ΦΦ∗ ) = Tr(Φ∗ Φ) = ϕ∗n ϕn = N.
n=1

Questions: For what M and N do UNTFs exist? How many of them are
there? What does the set of all M × N UNTFs look like?

10/27
Frame Potential
Theorem: For any unit vectors {ϕn }N M
n=1 in F ,

N X
X N
N(N−M)
M ≤ |ϕ∗n ϕn0 |2 = kΦ∗ Φ − Ik2Fro
n=1 n0 =1
0
n 6=n

with equality if and only if {ϕn }N M

n=1 is a UNTF for F .

Proof:
N2
0 ≤ Tr(ΦΦ∗ − N 2
M I) = Tr[(ΦΦ∗ )2 ] − 2 M
N
Tr(ΦΦ∗ ) + M2 Tr(I)
∗ 2 N ∗ N2
= Tr[(Φ Φ) ] − 2M Tr(Φ Φ) + M
N X
N
N2
X
= |ϕ∗n ϕn0 |2 − M.
n=1 n0 =1

Theorem: [Benedetto & F 03] Local minimizers of this potential are

UNTFs, and so they exist for any N ≥ M.

11/27
Example: UNTFs of 5 vectors in R3

A technique called spectral tetris gives the following 3 × 5 real UNTF:

 q q 
1 1
1 0 0
 q3 q3 q q 
Φ = 0 2 2 1 1 .
−

3 3 q6 q6 


5 5
0 0 0 6 − 6

But it’s not the only one. For example, we can rotate (multiply Φ by a
3 × 3 orthogonal matrix) to obtain others. But that’s not all...

12/27
Example: UNTFs of 5 vectors in R3
Every 3 × 5 real UNTF consists of 15 real unknowns:
 
Φ(1, 1) Φ(1, 2) Φ(1, 3) Φ(1, 4) Φ(1, 5)
Φ =  Φ(2, 1) Φ(2, 2) Φ(2, 3) Φ(2, 4) Φ(2, 5) 
Φ(3, 1) Φ(3, 2) Φ(3, 3) Φ(3, 4) Φ(3, 5)

which satisfy a system of 10 quadratic equations:

• 3 row orthogonality conditions,
• 3 row norm conditions,
• 5 column norm conditions (but one of these is redundant).
Modulo the 3-dimensional orthogonal group O(3), we thus expect a
15 − 10 − 3 = 2-dimensional set of UNTFs modulo rotations.

Theorem: [Dykema & Strawn 06] If N > M are relatively prime, the set
of all M × N UNTFs modulo O(M) is a manifold of dimension

(N − M − 1)(M − 1).

12/27
Paulsen Problem

• Open Problem: If {ϕn }N n=1 is “close” to unit norm, and “close” to

tight, how close is {ϕn }N
n=1 to a UNTF?

• Note: [Bodmann & Casazza 2010] and [Casazza, F & Mixon 2012]
give solutions to this problem when M and N are relatively prime.
As noted in [Dykema & Strawn 06], this prevents the frame from
being orthodecomposable (where the variety “crosses itself”).

• The fact that the Paulsen problem is open tells us we still do not
really know good ways of “moving around” frames in ways that
simultaneously control the norms of our vectors and the spectrum of
our frame operator.

13/27
Eigensteps Motivation (3 × 5 UNTFs Example Continued)
• Given any unit vectors {ϕn }5n=1 in R3 , consider their partial frame
operators (partial sums of their rank-one orthogonal projections):

Φ1 Φ∗1 = ϕ1 ϕ∗1 ,
Φ2 Φ∗2 = ϕ1 ϕ∗1 + ϕ2 ϕ∗2 ,
..
.
Φ5 Φ∗5 = ϕ1 ϕ∗1 + ϕ2 ϕ∗2 + ϕ3 ϕ∗3 + ϕ4 ϕ∗4 + ϕ5 ϕ∗5 .

• For every n, consider the Rayleigh quotient over the unit sphere:

x 7→ xϕn ϕ∗n x = |ϕ∗n x|2 ,

which has a max of 1 at ±ϕn and min of 0 at the “equator.”

We want 5 of these distributions that sum to 53 everywhere.
Pn
• The “hot spots” of xΦn Φ∗n x = i=1 |ϕ∗i x|2 are given in terms of
the eigenvalues/vectors of Φn Φ∗n ...

14/27
Eigensteps

Definition: The eigensteps of a UNTF {ϕn }N n=1 for F

M
is the array
M N M
{λm,n }m=1, n=0 where for any n, {λm,n }m=1 is the nondecreasing
Pn
spectrum of Φn Φ∗n = i=1 ϕi ϕ∗i .

Theorem: [Cahill, F, Mixon, Poteet & Strawn 13]

The eigensteps of any UNTF {ϕn }Nn=1 for F
M
satisfy
N
• λm,0 = 0 and λm,N = M for all m = 1, . . . , M;
PM
• m=1 λm,n = n for all n = 0, . . . , N (trace condition);
• λm+1,n ≤ λm,n−1 ≤ λm,n for all m = 1, . . . , M, n = 1, . . . , N − 1.

Conversely, for any {λm,n }M N

m=1, n=0 that satisfies these properties, there
N M M
exists a UNTF Pn{ϕn }n=1∗ for F with the property that {λm,n }m=1 is the
spectrum of i=1 ϕi ϕi for all n = 0, . . . , N. This construction is
explicit, and almost unique up to rotations.

15/27
Example: Eigensteps of 3 × 5 UNTFs

λ3,0 λ3,1 λ3,2 λ3,3 λ3,4 λ3,5

λ2,0 λ2,1 λ2,2 λ2,3 λ2,4 λ2,5

λ1,0 λ1,1 λ1,2 λ1,3 λ1,4 λ1,5

16/27
Example: Eigensteps of 3 × 5 UNTFs

5
0 λ3,1 λ3,2 λ3,3 λ3,4 3

5
0 λ2,1 λ2,2 λ2,3 λ2,4 3

5
0 λ1,1 λ1,2 λ1,3 λ1,4 3

16/27
Example: Eigensteps of 3 × 5 UNTFs

5
0 λ3,1 λ3,2 λ3,3 λ3,4 3

5
0 λ2,1 λ2,2 λ2,3 λ2,4 3

5
0 λ1,1 λ1,2 λ1,3 λ1,4 3

0 1 2 3 4 5

16/27
Example: Eigensteps of 3 × 5 UNTFs

5
0 ≤
≥ λ3,1 ≤ λ3,2 ≤ λ3,3 ≤ λ3,4 ≤ 3

≥
5
0 ≤ λ2,1 ≤ λ2,2 ≤ λ2,3 ≤ λ2,4 ≤ 3
≥

≥
5
0 ≤ λ1,1 ≤ λ1,2 ≤ λ1,3 ≤ λ1,4 ≤ 3

0 1 2 3 4 5

16/27
Example: Eigensteps of 3 × 5 UNTFs

5
0 ≤
≥ 0 ≤ λ3,2 ≤ λ3,3 ≤ λ3,4 ≤ 3

≥
5
0 ≤ 0 ≤ λ2,2 ≤ λ2,3 ≤ λ2,4 ≤ 3
≥

≥
5
0 ≤ λ1,1 ≤ λ1,2 ≤ λ1,3 ≤ λ1,4 ≤ 3

0 1 2 3 4 5

16/27
Example: Eigensteps of 3 × 5 UNTFs

5
0 ≤
≥ 0 ≤ 0 ≤ λ3,3 ≤ λ3,4 ≤ 3

≥
5
0 ≤ 0 ≤ λ2,2 ≤ λ2,3 ≤ λ2,4 ≤ 3
≥

≥
5
0 ≤ 1 ≤ λ1,2 ≤ λ1,3 ≤ λ1,4 ≤ 3

0 1 2 3 4 5

16/27
Example: Eigensteps of 3 × 5 UNTFs

5
0 ≤
≥ 0 ≤ 0 ≤ λ3,3 ≤ λ3,4 ≤ 3

≥
5 5
0 ≤ 0 ≤ λ2,2 ≤ λ2,3 ≤ 3 ≤ 3
≥

≥
5 5
0 ≤ 1 ≤ λ1,2 ≤ λ1,3 ≤ 3 ≤ 3

0 1 2 3 4 5

16/27
Example: Eigensteps of 3 × 5 UNTFs

2 5
0 ≤
≥ 0 ≤ 0 ≤ λ3,3 ≤ 3 ≤ 3

≥
5 5
0 ≤ 0 ≤ λ2,2 ≤ λ2,3 ≤ 3 ≤ 3
≥

≥
5 5 5
0 ≤ 1 ≤ λ1,2 ≤ 3 ≤ 3 ≤ 3

0 1 2 3 4 5

16/27
Example: Eigensteps of 3 × 5 UNTFs

2 5
0 ≤
≥ 0 ≤ 0 ≤ x ≤ 3 ≤ 3

≥
5 5
0 ≤ 0 ≤ y ≤ λ2,3 ≤ 3 ≤ 3
≥

≥
5 5 5
0 ≤ 1 ≤ λ1,2 ≤ 3 ≤ 3 ≤ 3

0 1 2 3 4 5

16/27
Example: Eigensteps of 3 × 5 UNTFs

2 5
0 ≤
≥ 0 ≤ 0 ≤ x ≤ 3 ≤ 3

≥
4 5 5
0 ≤ 0 ≤ y ≤ 3 −x ≤ 3 ≤ 3
≥

≥
5 5 5
0 ≤ 1 ≤ 2−y ≤ 3 ≤ 3 ≤ 3

0 1 2 3 4 5

16/27
Example: Eigensteps of 3 × 5 UNTFs
4 4
3 3

1 1

2 2
y

y
3 3

1 1
3 3

1 2 4 1 2 4
0 3 3 1 3 0 3 3 1 3
x x
(a) (b)

16/27
Eigensteps Polytope

Note: The set of all eigensteps arising from M × N UNTFs forms a

convex polytope. [F, Mixon, Strawn & Poteet 13] gives an algorithm for
constructing a particular type of “extreme” eigensteps which correspond
to one of the corner points of this polytope.

Open Problems:
• How many corner points does this polytope have in general?
• What “strategies” do each of these corner points correspond to?
• What special properties do “corner point” UNTFs have?
• Can we use eigensteps to solve the Paulsen problem?

Disclaimer: I have only briefly read the paper Tim Haga and Christoph
Pegel posted to arXiv on July 15. (and will present on Thursday?)

17/27
Weaver’s Conjecture Theorem

• In 2013, Marcus, Spielman & Srivastava proved the famous

Kadison-Singer conjecture by using the probabilistic method to
prove the stronger Weaver’s conjecture:

There exists universal constants α > 2 and β > 0 so that if {ϕn }N

n=1
is any α-tight frame for FM where kϕn k ≤ 1 for all n, then the
frame elements can be partitioned into two frames {ϕn }n∈N1 and
{ϕn }n∈N2 whose frame bounds lie between β and α − β.

• In particular, they proved Weaver’s conjecture holds for α = 18 and

β = 2, implying any UNTF of redundancy 18 can always be
decomposed into two frames whose condition number is at most 8.

• Open Problem: How good of a partition can we compute

deterministically (practically, numerically)?
Is there a “square root bottleneck” á la deterministic RIP?

18/27
 Part III:
Equiangular Tight Frames
Optimal Packings of Lines

Definition: The coherence of a set of unit vectors {ϕn }N

n=1 in F
M
is

max0 |ϕ∗n ϕn0 |.

n6=n

Frames of minimal coherence are called Grassmannian frames.

Note: Minimizing coherence is equivalent to packing lines: letting θn,n0

denote the interior angle between the lines spanned by ϕn and ϕn0 ,
   
arg min max0 |ϕ∗n ϕn0 | = arg max min0 θn,n0 .
{ϕn } n6=n {ϕn } n6=n

19/27
Welch Bound
Theorem: [Rankin 56, Welch 74, Strohmer & Heath 03]
The coherence of any unit vectors {ϕn }N
n=1 in F
M
satisfies
q
N−M ∗
M(N−1) ≤ max0 |ϕn ϕn0 |,
n6=n

with equality ⇔ {ϕn }N M

n=1 is an equiangular tight frame (ETF) for F :

• {ϕn }N
n=1 is a UNTF and
• the modulus of inner products of distinct ϕn ’s is constant, i.e.

1, n = n0 ,

∗ 0 ∗
|(Φ Φ)(n, n )| = |ϕn ϕn0 | =
β, n 6= n0 .

N X
X N
N(N−M)
Proof: M ≤ |ϕ∗n ϕn0 |2 ≤ N(N − 1) max0 |ϕ∗n ϕn0 |2 .
n6=n
n=1 n0 =1
n0 6=n

20/27
Example: Optimally Packing 16 Lines In R6

 
+ + + + + + + + + + + + + + + +
+
 − + − + − + − + − + − + − + −
1  + − − + + − − + + − − + + − − +
Φ= √  
6 +
 + + + − − − − + + + + − − − −
+ − + − + − + − − + − + − + − +
+ − − + − + + − − + + − + − − +

21/27
Example: Optimally Packing 16 Lines In R6

 
1 0 0 0 0 0
0 1 0 0 0 0
 
∗ 16 
0 0 1 0 0 0
ΦΦ = 
0
6  0 0 1 0 0
0 0 0 0 1 0
0 0 0 0 0 1

21/27
Example: Optimally Packing 16 Lines In R6
 
3 − + + + − + − + + + + + − − +
− 3 + + − + − + + + + + − + + −
 
+
 + 3 − + − + − + + + + − + + −
+
 + − 3 − + − + + + + + + − − +
+
 − + − 3 − + + + − − + + + + +
− + − + − 3 + + − + + − + + + +
 
+
 − + − + + 3 − − + + − + + + +
∗ 1 − + − + + + − 3 + − − + + + + +
Φ Φ=  
3+ + + + + − − + 3 − + + + − + −
+
 + + + − + + − − 3 + + − + − +
+
 + + + − + + − + + 3 − + − + −
+
 + + + + − − + + + − 3 − + − +
+
 − − + + + + + + − + − 3 − + +
− + + − + + + + − + − + − 3 + +
 
− + + − + + + + + − + − + + 3 −
+ − − + + + + + − + − + + + − 3

21/27
The Grassmannian Frame Problem

• For any M ≤ N, we want Grassmannian frames. If there exists an

M × N ETF, it is Grassmannian.
• However, for many most choices of M and N, we do not know
whether an M × N ETF exists. Moreover, for many choices of M
and N, we know that M × N ETF cannot exist.
• Almost all research in this area has used the following program:
− Find as many explicit constructions of ETFs as possible.
− Find the strongest possible necessary conditions on ETF existence.

Open Problem: Find Grassmannian ETFs for cases of M and N for

which no ETF exists. In particular, find ways of proving that {ϕn }N
n=1
has optimal coherence that do not involve equiangularity.

22/27
Some Known Constructions of ETFs
Fact: All known infinite families of ETFs involve some type of
combinatorial design, including:
• harmonic ETFs arising from difference sets in abelian groups, e.g.
{(0, 0, 0, 0), (0, 0, 1, 0), (1, 0, 0, 0), (1, 0, 0, 1), (1, 1, 0, 0), (1, 1, 1, 1)}
regarded as a subset of Z2 × Z2 × Z2 × Z2 :
− (0, 0, 0, 0) (0, 0, 1, 0) (1, 0, 0, 0) (1, 0, 0, 1) (1, 1, 0, 0) (1, 1, 1, 1)

(0, 0, 0, 0) (0, 0, 0, 0) (0, 0, 1, 0) (1, 0, 0, 0) (1, 0, 0, 1) (1, 1, 0, 0) (1, 1, 1, 1)

(0, 0, 1, 0) (0, 0, 1, 0) (0, 0, 0, 0) (1, 0, 1, 0) (1, 0, 1, 1) (1, 1, 1, 0) (1, 1, 0, 1)

(1, 0, 0, 0) (1, 0, 0, 0) (1, 0, 1, 0) (0, 0, 0, 0) (0, 0, 0, 1) (0, 1, 0, 0) (0, 1, 1, 1)

(1, 0, 0, 1) (1, 0, 0, 1) (1, 0, 1, 1) (0, 0, 0, 1) (0, 0, 0, 0) (0, 1, 0, 1) (0, 1, 1, 0)

(1, 1, 0, 0) (1, 1, 0, 0) (1, 1, 1, 0) (0, 1, 0, 0) (0, 1, 0, 1) (0, 0, 0, 0) (0, 0, 1, 1)

(1, 1, 1, 1) (1, 1, 1, 1) (1, 1, 0, 1) (0, 1, 1, 1) (0, 1, 1, 0) (0, 0, 1, 1) (0, 0, 0, 0)

Singer and McFarland difference sets give harmonic ETFs of size

 j+1   j+2   j+1   j+1 
q −1 q −1 q −1 q −1
× , qj ×q j+1 +1 ,
q−1 q−1 q−1 q−1

respectively, for any prime power q and any positive integer j.

23/27
Some Known Constructions of ETFs
Fact: All known infinite families of ETFs involve some type of
combinatorial design, including:
• Steiner ETFs from balanced incomplete block designs e.g.

Fano plane

23/27
Some Known Constructions of ETFs
Fact: All known infinite families of ETFs involve some type of
combinatorial design, including:
• Steiner ETFs from balanced incomplete block designs e.g.
 
++ 0 + 0 0 0
0 ++ 0 + 0 0
 
0 0 ++ 0 + 0
 
 0 0 0 + + 0 +
 
+ 0 0 0 + + 0 
 
 0 + 0 0 0 + +
+ 0 + 0 0 0 +

Incidence matrix of the corresponding Steiner triple system

23/27
Some Known Constructions of ETFs
Fact: All known infinite families of ETFs involve some type of
combinatorial design, including:
• Steiner ETFs from balanced incomplete block designs e.g.

 
+ + 0 + 0 0 0
 0 + + 0 + 0 0   
 0 0 + + 0 + 0  + − + −
 
 0 0 0 + + 0 + “ ⊗ ” + + − − 
 
+
 0 0 0 + + 0 
 + − − +
 0 + 0 0 0 + +
+ 0 + 0 0 0 +
− − − − − −
 
+ + + + 0 0 0 0 + + 0 0 0 0 0 0 0 0 0 0 0 0
 0 0 0 0 + − + − + + − − 0 0 0 0 + − − + 0 0 0 0 0 0 0 0 
 0 0 0 0 0 0 0 0 + − + − + + − − 0 0 0 0 + − − + 0 0 0 0 
 
= 0 0 0 0 0 0 0 0 0 0 0 0 + − + − + + − − 0 0 0 0 + − − +
 
+
 − − + 0 0 0 0 0 0 0 0 0 0 0 0 + − + − + + − − 0 0 0 0 

 0 0 0 0 + − − + 0 0 0 0 0 0 0 0 0 0 0 0 + − + − + + − −
+ + − − 0 0 0 0 + − − + 0 0 0 0 0 0 0 0 0 0 0 0 + − + −

“Inflating” the Steiner system by a regular simplex to form an ETF

23/27
Some Known Constructions of ETFs
Fact: All known infinite families of ETFs involve some type of
combinatorial design, including:
• Steiner ETFs from balanced incomplete block designs e.g.

ETFs of size:
q j+1 −1 q j+2 −1



− q−1
× q−1
from affine geometries,
j+1 j+2 j+2
(q −1)(q −1) q −1 q j+1 −1
− (q+1)(q−1)2
× q−1
(1 + q−1
) from projective geometries,
(2s +1)(2r +s +2r −2s )
− 2r
× (2s + 2)(2r +s + 2r − 2s ) from Denniston designs,
for any prime power q and any positive integers j, 2 ≤ r < s.

23/27
Some Known Constructions of ETFs
Fact: All known infinite families of ETFs involve some type of
combinatorial design, including:
• Steiner ETFs from balanced incomplete block designs e.g.

ETFs of size:
q j+1 −1 q j+2 −1



− q−1
× q−1
from affine geometries,
j+1 j+2 j+2
(q −1)(q −1) q −1 q j+1 −1
− (q+1)(q−1)2
× q−1
(1 + q−1
) from projective geometries,
(2s +1)(2r +s +2r −2s )
− 2r
× (2s + 2)(2r +s + 2r − 2s ) from Denniston designs,
for any prime power q and any positive integers j, 2 ≤ r < s.

Finding new explicit constructions of ETFs seems really hard and is

probably not well-suited to the “large collaborative group” setting of
workshops... maybe we should focus on necessary conditions instead?

23/27
Absolute Bounds
Theorem: [Gerzon in Lemmens & Seidel 73]
If unit vectors {ϕn }N
n=1 are equiangular and not collinear in F
M
then
 
M +1
N≤ if F = R, N ≤ M 2 if F = C.
2

If these bounds are achieved then {ϕn }N M

n=1 is necessarily an ETF for F .

Proof Sketch: {ϕn }Nn=1 being equiangular but not collinear implies their
projection operators {ϕn ϕ∗n }N
n=1 are linearly independent.

Note: When F = R, N = M+1

2 is known to not be achievable for many
M due to integrality conditions given on the next slide.
It is achievable for M = 3, 7, 23.

When F = C, N = M 2 is known to be achievable for many M, and is

conjectured to be always so (see Dustin’s talk tomorrow!)

24/27
Integrality Conditions

Theorem: [Sustik, Tropp, Dhillon & Heath 07]

If a real M × N ETF exists and 1 < M < N − 1 with N 6= 2M, then
 M(N−1)  21  (N−M)(N−1)  21
N−M , N−1

are necessarily odd integers.

Proof Sketch: Study the eigenvalues of the matrix obtained by

converting Φ∗ Φ into a {−1, 0, 1}-valued matrix.
This is closely related to a well known equivalence between real ETFs
and strongly regular graphs.

Note: These necessary conditions are not sufficient, e.g. 47 × 1128.

25/27
Complex Integrality Conditions?

• In the complex case, the only necessary conditions we have on ETFs

is the absolute bound on it and its Naimark complement

N ≤ M 2, N ≤ (N − M)2 .

• Well, that’s not quite true... in 2014, Ferenc Szöllősi used algebraic
geometry to prove that there does not exist a 3 × 8 complex ETF!

• Based on this “overwhelming” evidence (and a lot of explicit

constructions of complex ETFs, and my suspicion that it will be
extremely hard to prove it either true or false), I conjecture the
following:

Conjecture: If there exists a complex M × N ETF, then one of the

three integers M, N − 1 and N − M must divide the product of the
other two.

26/27
Summary

• A lot of finite frame theory is about generalizing orthonormal bases.

• Tight frames generalize the Pythagorean theorem, are commonplace

and (barring additional restrictions) are easy to construct.

• Unit norm tight frames (UNTFs) are much harder to construct.

− Nevertheless, UNTFs exist for every M ≤ N.
− The fact that the Paulsen problem is open tells us we still don’t
really understand the geometry of the set of all M × N UNTFs.
− Eigensteps are a way of parametrizing this set, and raise their own
questions.

• Equiangular tight frames are even more rare.

− In the complex case, we have almost no necessary conditions on their
existence.
− For M and N for which no ETF exists, we have almost no techniques
for proving given vectors are Grassmannian.

27/27