Jul 15, 2010 - probabilities than subgaussian and gaussian random variables. Geometrically, perpendicular in l2 needs to
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Non-convex Optimization for Linear System with Pregaussian Matrices and Recovery from Multiple Measurements Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
University of Georgia July 15, 2010
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
`q -minimization
Introduction Null Space Property of `q -minimization for MMV Proof of the main theorem
problem
In compressed sensing, we want to recover a sparse or compressible signal via solving the minimization problem minimizex∈RN where
kxk0
kxk0
subject to
Ax = b
is the the number of non-zero entries of vector
namely the sparsity of
x,
and
A
is a matrix of size
m×N
x,
with
m N.
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
`q -minimization
Introduction Null Space Property of `q -minimization for MMV Proof of the main theorem
problem
In compressed sensing, we want to recover a sparse or compressible signal via solving the minimization problem minimizex∈RN where
kxk0
kxk0
subject to
Ax = b
is the the number of non-zero entries of vector
namely the sparsity of
x,
and
A
is a matrix of size
m×N
x,
with
m N. `q -approach with 0 < q ≤ 1, The
is to consider the
minimizex∈RN
kxkq
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
`q -minimization
subject to
problem
Ax = b.
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Null Space Property of `q -minimization for MMV Proof of the main theorem
Null space property for SMV The null space property has been used to quantify the error of approximations (Cohen, Dahmen, and DeVore, 2009), and it also guarantees the exact recovery. Proposition (Null space property)
Let S ⊆ {1, 2, · · · , N } be a xed index set. Then a vector x with support (x) ⊆ S can be uniquely recovered from Ax = b using `1 -minimization if for all non-zero v in the null space of A, kvS k1 < kvS c k1 ,
where S c is the complement of S in {1, 2, · · · , N }.
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Null Space Property of `q -minimization for MMV Proof of the main theorem
Proof of the proposition Proof. We know for any non-zero
z
in the null space of
A,
kxS k1 ≤ kzS k1 + k(z + x)S k1 . By the assumption
kzS k1 < kzS c k1 ,
we then have
kxS k1 < k(z + x)S k1 + kzS c k1 . But
x
vanishes on
Sc,
thus
kxk1 = kxS k1 < k(z + x)S k1 + k(z + x)S c k1 = kz + xk1 , and so
x ∈ RN
is the unique solution to the
`1 -minimization
problem. Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Null Space Property of `q -minimization for MMV Proof of the main theorem
Remarks to the proposition
Remark
`1
can be replaced by
`q
.
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Null Space Property of `q -minimization for MMV Proof of the main theorem
Remarks to the proposition
Remark
`1
can be replaced by
`q
.
Remark The converse is also true.
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Null Space Property of `q -minimization for MMV Proof of the main theorem
MMV problem
Denition Given a set of
r
measurements
Ax(k) = b(k) Find the vectors
x(k)
for
k = 1, · · · , r.
which are jointly sparse.
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Null Space Property of `q -minimization for MMV Proof of the main theorem
MMV problem
Denition Given a set of
r
measurements
Ax(k) = b(k) Find the vectors
x(k)
for
k = 1, · · · , r.
which are jointly sparse.
The MMV problem arises in biomedical engineering, especially in neuromagnetic imaging.
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Null Space Property of `q -minimization for MMV Proof of the main theorem
`1,2 -minimization
The approach of mixed
`1,2 -minimization
to recover jointly sparse
vectors from MMV is solving the optimization problem
minimize
N q X x21,j + · · · + x2r,j
subject to
Ax(k) = b(k)
j=1 for
k = 1, · · · , r.
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Null Space Property of `q -minimization for MMV Proof of the main theorem
Null space property for MMV Theorem (BergFriedlander, 2009)
Let A be a real matrix of m × N and S ⊂ {1, 2, · · · , N } be a xed index set. Denote by S c the complement set of S in {1, 2, · · · , N }. Let k · k be any norm. Then all x(k) with support x(k) in S for k = 1, · · · , r can be uniquely recovered using the following minimize
N X
k(x1,j , · · · , xr,j )k subject to Ax(k) = b(k)
j=1
for k = 1, · · · , r if and only if for all vectors
(u(1) , · · · , u(r) ) ∈ (N (A))r \{(0, 0, · · · , 0)} satisfy the following X X k(u1,j , · · · , ur,j )k < k(u1,j , · · · , ur,j )k, . j∈S Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
j∈S c Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Null Space Property of `q -minimization for MMV Proof of the main theorem
Real vs. complex null space properties Theorem (FoucartGribonval, 2009)
Let A be a matrix of size m × N and S ⊂ {1, · · · , N } be the support of the sparse vector y. The complex null space property: for any u ∈ N (A), w ∈ N (A) with (u, w) 6= 0, Xq Xq u2j + wj2 < u2j + wj2 , j∈S c
j∈S
where u = (u1 , u2 , · · · , uN )T and w = (w1 , w2 , · · · , wN )T is equivalent to the following standard null space property: for any u in the null space N (A) with u 6= 0, X
|uj |
0, there exists K > 0 such that (1) P s1 (A) ≤ Km ≤ ε
where K only depends on ε and the pregaussian variable ξ .
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Upper Tail Probability of the Largest q -singular Value Lower Tail Probability of the Largest q -singular Value
Lower tail probability of largest
1-singular
value
Proof. Since
aij
is pregaussian with variance
exists some
δ>0
1,
then for any
(1)
s1 (A) =
there
ε . 8
P (|aij | ≤ δ) ≤ Since
ε > 0,
such that
Pm
i=1 |aij0 | for some
j0 ,
by the previous lemma, we
have
! m X mδ mδ (1) P s1 (A) ≤ ≤ P ≤ 8P (|aij | ≤ δ) ≤ ε. |aij0 | ≤ 2 2 i=1
Finally let
K=
δ 2 , then the claim follows.
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Upper Tail Probability of the Largest q -singular Value Lower Tail Probability of the Largest q -singular Value
Lower tail probability of largest
For general
0 < q ≤ 1,
q -singular
value
we have
Theorem (Lower tail probability of the largest
q -singular
value )
Let ξ be a pregaussian variable normalized to have variance 1 and A be an m × N matrix with i.i.d. copies of ξ in its entries, then for every 0 < q ≤ 1 and any ε > 0, there exists K > 0 such that 1 (q) P s1 (A) ≤ Km q ≤ ε
where K only depends on q , ε and the pregaussian variable ξ .
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Smallest
q -singular
Introduction Lower Tail Probability of the Smallest q -singular Value Upper Tail Probability of the Smallest q -singular Value Additional Results: Probability Estimates on the Largest p-sing
value
Denition The smallest
q -singular
value of an
s(q) n (A) : =
n×n inf
matrix
x∈Rn , kxkq =1
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
kAxkq .
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Lower Tail Probability of the Smallest q -singular Value Upper Tail Probability of the Smallest q -singular Value Additional Results: Probability Estimates on the Largest p-sing
Smallest singular value in `2
Rudelson and Vershynin rst showed the following result Theorem (RudelsonVershynin, 2008)
If A is a matrix of size n × n whose entries are independent random variables with variance 1 and bounded fourth moment. Then for any δ > 0, there exists > 0 and integer n0 > 0 such that P sn (A) ≤ √ ≤ δ, n
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
∀n ≥ n0 .
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Lower Tail Probability of the Smallest q -singular Value Upper Tail Probability of the Smallest q -singular Value Additional Results: Probability Estimates on the Largest p-sing
Smallest singular value in `2
Later, they proved the following result Theorem (RudelsonVershynin, 2008)
Let A be an n × n matrix whose entries are i.i.d. centered random variables with unit variance and fourth moment bounded by B. Then, for every δ > 0 there exist K > 0 and n0 which depend (polynomially) only on δ and B, and such that K ≤ δ, P sn (A) > √ n
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
∀n ≥ n0
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Lower Tail Probability of the Smallest q -singular Value Upper Tail Probability of the Smallest q -singular Value Additional Results: Probability Estimates on the Largest p-sing
Lower tail probabilistic estimate on the smallest
q -singular
value For tall rectangular matrices, we have the lower tail probabilistic of exponential decay on the smallest
q -singular
value
Theorem
Given any 0 < q ≤ 1, and let ξ be the pregaussian random variable with variance 1 and A be an m × n matrix with i.i.d. copies of ξ in its entries. Then for any ε > 0, there exist some γ > 0 and c > 0 and r ∈ (0, 1) dependent on q and ε, such that 1 q P s(q) (A) < γm < e−cm n
if n ≤ rm. Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Lower Tail Probability of the Smallest q -singular Value Upper Tail Probability of the Smallest q -singular Value Additional Results: Probability Estimates on the Largest p-sing
Lower tail probabilistic estimate on the smallest
q -singular
value
Theorem
Given any 0 < q ≤ 1, and let ξ be the pregaussian random variable with variance 1 and A be an n × n matrix with i.i.d. copies of ξ in its entries. Then for any ε > 0 and 0 < q ≤ 1, there exist some K > 0 and c > 0 dependent on q and ε, such that 1 − 1q P s(q) < Cε + Cαn + P kAk > Kn− 2 . n (A) < εn
where α ∈ (0, 1) and C > 0 depend only on the pregaussian variable and K . Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Lower Tail Probability of the Smallest q -singular Value Upper Tail Probability of the Smallest q -singular Value Additional Results: Probability Estimates on the Largest p-sing
Lower tail probabilistic estimate on the smallest
q -singular
value
Theorem (Lower tail probabilistic estimate on the smallest
q -singular
value )
Given any 0 < q ≤ 1, and let ξ be the pregaussian random variable with variance 1 and A be an n × n matrix with i.i.d. copies of ξ in its entries. Then for any ε > 0, there exists some γ > 0 such that − 1q P s(q) (A) < γn < ε, n
where γ only depends on q , ε and the pregaussian variable ξ .
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Lower Tail Probability of the Smallest q -singular Value Upper Tail Probability of the Smallest q -singular Value Additional Results: Probability Estimates on the Largest p-sing
Upper tail probability of the smallest
q -singular
value
Theorem (Upper tail probabilistic estimate on the smallest
q -singular
value)
Given any 0 < q ≤ 1, and let ξ be the pregaussian random variable with variance 1 and A be an n × n matrix with i.i.d. copies of ξ in its entries. Then for any K > e, there exist some C > 0, 0 < c < 1, and α > 0 only dependent on pregaussian variable ξ , q , such that − 21 P s(q) (A) > Kn ≤ n
C (ln K)α + cn . Kα
In particular, for any ε > 0, there exist some K > 0 and n0 , such that for all n ≥ n0 , − 21 < ε. P s(q) (A) > Kn n Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Lower Tail Probability of the Smallest q -singular Value Upper Tail Probability of the Smallest q -singular Value Additional Results: Probability Estimates on the Largest p-sing
Lower tail probability of the largest
p-singular
value
Theorem (Lower tail probability of the largest -singular value,
p > 1)
Let ξ be a pregaussian random variable normalized to have variance 1 and A be an m × N matrix with i.i.d. copies of ξ in its entries, then for every p > 1 and any ε > 0, there exists γ > 0 such that 1 (p) ≤ ε P s1 (A) ≤ γm p
where γ only depends on p, ε and the pregaussian random variable
ξ.
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
Introduction Lower Tail Probability of the Smallest q -singular Value Upper Tail Probability of the Smallest q -singular Value Additional Results: Probability Estimates on the Largest p-sing
A Remark
Remark By the duality that for any
p≥1
and
(p)
(q)
s1 (A) = s1 where
1 p
+
1 q
= 1,
in particular, for
n×n AT
matrix
A,
(∞)
p = ∞, s1
(A)
is of order
n
with high probability.
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
2-singular For
value
p = 2,
we plot the largest
matrices of size
n × n,
Figure: Largest
where
2-singular
2-singular value of Gaussian n runs from 1 through 100.
random
value of Gaussian random matrices
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
1-singular
value
For p = 1, in the rst numerical experiment we plot the largest 1-singular value of Gaussian random of size n × n, where n runs from 1 through 100.
Figure: Largest
1-singular
value of Gaussian random matrices:
Experiment 1
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
1-singular
value
In the second numerical experiment for
1-singular value of Gaussian n runs from 1 through 200.
Figure: Largest
1-singular
p = 1,
we plot the largest
random matrices of size
n × n,
where
value of Gaussian random matrices:
Experiment 2
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
1-singular
value
In the third experiment for
p = 1,
we plot the largest
value of Gaussian random matrices of size from
1
through
Figure: Largest
n × n,
1-singular n runs
where
400.
1-singular
value of Gaussian random matrices:
Experiment 3
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
∞-singular For
value
p = ∞,
we plot the largest
random matrices of size
Figure: Largest
n × n,
∞-singular
∞-singular value of Gaussian where n runs from 1 through 500.
value of Gaussian random matrices
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
1 3 -singular value
For
p=
1 1 3 , we plot the largest 3 -singular value of Gaussian random
matrices of size
n × n,
Figure: Largest
where
n
runs from
1
through
500.
1 3 -singular value of Gaussian random matrices
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
1 4 -singular value
For
p=
1 1 4 , we plot the largest 4 -singular value of Gaussian random
matrices of size
n × n,
Figure: Largest
where
n
runs from
1
through
300.
1 4 -singular value of Gaussian random matrices
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values
Null Space Property for Sparse Recovery from MMV Probability Estimates on Largest q -singular Value Probability Estimates on Smallest q -singular Value Some Numerical Experiments
1 4 -singular value
For rectangular matrices, we also plot the largest of Gaussian random matrices of size from
1
through
Figure: Largest
m × n,
1 4 -singular value
where
m
and
n
run
100.
1 4 -singular value of rectangular Gaussian random matrices
Louis Yang Liu (Advisor: Prof. Ming-Jun Lai)
Sparse Recovery by `q -minimization and q -Singular Values