Gelfand Widths in Compressive Sensing Optimal Number of Measurements
Björn Bringmann Technische Universität München
15.01.2015
Björn Bringmann
Gelfand Widths in Compressive Sensing
1 / 26
Outline 1
Introduction to m-Widths Kolmogorov and Gelfand Widths Compressive m-widths
2
Gelfand Widths of l1 -balls Upper bound Lower bound
3
Applications Optimal Number of Measurements Kashin’s Decomposition Theorem
Björn Bringmann
Gelfand Widths in Compressive Sensing
2 / 26
Introduction to m-Widths
Outline 1
Introduction to m-Widths Kolmogorov and Gelfand Widths Compressive m-widths
2
Gelfand Widths of l1 -balls Upper bound Lower bound
3
Applications Optimal Number of Measurements Kashin’s Decomposition Theorem
Björn Bringmann
Gelfand Widths in Compressive Sensing
3 / 26
Introduction to m-Widths
Kolmogorov and Gelfand Widths
Motivation Let X = {f ∈ C (R) : f is 2π periodic} and define
n
Sn (f )(x ) :=
∑
bf (k )eikx .
k =−n
How fast does kf − Sn (f )k2 converge to 0? Approximate for f ∈ C ([a, b]) Z a
n
b
f (x )dx ≈
∑ w k f ( xk ) .
k =0
What is a reasonable error estimate? Conclusion: Approximation by linear subspaces is a good idea! Björn Bringmann
Gelfand Widths in Compressive Sensing
4 / 26
Introduction to m-Widths
Kolmogorov and Gelfand Widths
Motivation Let X = {f ∈ C (R) : f is 2π periodic} and define
n
Sn (f )(x ) :=
∑
bf (k )eikx .
k =−n
How fast does kf − Sn (f )k2 converge to 0? Approximate for f ∈ C ([a, b]) Z a
n
b
f (x )dx ≈
∑ w k f ( xk ) .
k =0
What is a reasonable error estimate? Conclusion: Approximation by linear subspaces is a good idea! Björn Bringmann
Gelfand Widths in Compressive Sensing
4 / 26
Introduction to m-Widths
Kolmogorov and Gelfand Widths
Motivation Let X = {f ∈ C (R) : f is 2π periodic} and define
n
Sn (f )(x ) :=
∑
bf (k )eikx .
k =−n
How fast does kf − Sn (f )k2 converge to 0? Approximate for f ∈ C ([a, b]) Z a
n
b
f (x )dx ≈
∑ w k f ( xk ) .
k =0
What is a reasonable error estimate? Conclusion: Approximation by linear subspaces is a good idea! Björn Bringmann
Gelfand Widths in Compressive Sensing
4 / 26
Introduction to m-Widths
Kolmogorov and Gelfand Widths
Motivation Let X = {f ∈ C (R) : f is 2π periodic} and define
n
Sn (f )(x ) :=
∑
bf (k )eikx .
k =−n
How fast does kf − Sn (f )k2 converge to 0? Approximate for f ∈ C ([a, b]) Z a
n
b
f (x )dx ≈
∑ w k f ( xk ) .
k =0
What is a reasonable error estimate? Conclusion: Approximation by linear subspaces is a good idea! Björn Bringmann
Gelfand Widths in Compressive Sensing
4 / 26
Introduction to m-Widths
Kolmogorov and Gelfand Widths
Kolmogorov m-Width Definition Let X be a real (or complex) Banach space. For any subset K ⊂ X and any m-dimensional linear subspace Xm ⊂ X define d (K , Xm ; X ) := sup inf kx − z kX . x ∈K z ∈Xm
Now define the Kolmogorov m-width as dm (K ; X ) := inf{d (K , Xm ; X ) : Xm ⊂ X m-dimensional linear subspace} . y
X1
K x
Björn Bringmann
Gelfand Widths in Compressive Sensing
5 / 26
Introduction to m-Widths
Kolmogorov and Gelfand Widths
Kolmogorov m-Width Definition Let X be a real (or complex) Banach space. For any subset K ⊂ X and any m-dimensional linear subspace Xm ⊂ X define d (K , Xm ; X ) := sup inf kx − z kX . x ∈K z ∈Xm
Now define the Kolmogorov m-width as dm (K ; X ) := inf{d (K , Xm ; X ) : Xm ⊂ X m-dimensional linear subspace} . y
X1
K x
Björn Bringmann
Gelfand Widths in Compressive Sensing
5 / 26
Introduction to m-Widths
Kolmogorov and Gelfand Widths
Properties of the Kolmogorov m-width
Lemma Let X be a Banach space and K ⊂ X ,m ∈ N. ¯ ; X ), where K¯ is the closure of K . 1 dm (K ; X ) = dm (K 2
For every scalar α there holds dm (α K ; X ) = |α|dm (K ; X ).
3
dm (co(K ); X ) = dm (K ; X ), where co(K ) is the convex hull of K .
4
dm (K ; X ) ≥ dm+1 (K ; X ).
Björn Bringmann
Gelfand Widths in Compressive Sensing
6 / 26
Introduction to m-Widths
Kolmogorov and Gelfand Widths
Gelfand m-Width Definition Let X be a real or complex Banach space and K a subset of X. The Gelfand m-width is defined as d m (K ; X ) = inf{ sup kx kX : Lm ⊂ X linear subspace with codim(Lm ) ≤ m} x ∈K ∩Lm
We say that a closed linear subspace Lm ⊂ X has codim(Lm ) = m if dim(X /Lm ) = m . Lemma A linear subspace Lm ⊂ X has codim(Lm ) ≤ m if and only if there exist f1 , . . . , fm ∈ X ∗ such that Lm = {x ∈ X : fi (x ) = 0 ∀i = 1, . . . , m} . Björn Bringmann
Gelfand Widths in Compressive Sensing
7 / 26
Introduction to m-Widths
Kolmogorov and Gelfand Widths
Duality of Kolmogorov and Gelfand m-widths Theorem For 1 ≤ p, q ≤ ∞ let p∗ , q ∗ be such that
1 p
+ p1∗ = 1 and
1 q
+ q1∗ = 1. Then
m N N dm (BpN ; `N q ) = d (Bq ∗ ; `p∗ ) .
Lemma Let Y be a finite-dimensional subspace of a Banach space X. Given x ∈ X \Y and y ∗ ∈ Y , the following properties are equivalent: ∗ 1 y is a best approximation to x from Y . ∗ 2 For some λ ∈ X with kλ kX ∗ ≤ 1 and λ |Y ≡ 0, there holds
kx − y ∗ k = λ (x ) .
Björn Bringmann
Gelfand Widths in Compressive Sensing
8 / 26
Introduction to m-Widths
Kolmogorov and Gelfand Widths
Connection with linear operators Note that
( m
d (K ; X ) = inf
) sup
m
kx k : A : X → K linear, continuous
.
x ∈K ∩Ker(A)
If Lm is a subspace with codim(Lm ) ≤ m, then choose f1 , . . . , fm ∈ X ∗ as in the previous Lemma and define
A : X → Km , x 7→ [f1 (x ), . . . , fm (x )]t .
If A : X → Km is given, define the corresponding linear subspace Lm := Ker(A). Björn Bringmann
Gelfand Widths in Compressive Sensing
9 / 26
Introduction to m-Widths
Kolmogorov and Gelfand Widths
Connection with linear operators Note that
( m
d (K ; X ) = inf
) sup
m
kx k : A : X → K linear, continuous
.
x ∈K ∩Ker(A)
If Lm is a subspace with codim(Lm ) ≤ m, then choose f1 , . . . , fm ∈ X ∗ as in the previous Lemma and define
A : X → Km , x 7→ [f1 (x ), . . . , fm (x )]t .
If A : X → Km is given, define the corresponding linear subspace Lm := Ker(A). Björn Bringmann
Gelfand Widths in Compressive Sensing
9 / 26
Introduction to m-Widths
Kolmogorov and Gelfand Widths
Connection with linear operators Note that
( m
d (K ; X ) = inf
) sup
m
kx k : A : X → K linear, continuous
.
x ∈K ∩Ker(A)
If Lm is a subspace with codim(Lm ) ≤ m, then choose f1 , . . . , fm ∈ X ∗ as in the previous Lemma and define
A : X → Km , x 7→ [f1 (x ), . . . , fm (x )]t .
If A : X → Km is given, define the corresponding linear subspace Lm := Ker(A). Björn Bringmann
Gelfand Widths in Compressive Sensing
9 / 26
Introduction to m-Widths
Compressive m-widths
Compressive m-width Definition The compressive m-width of a subset K of a (real) Banach space X is defined as
E (K ; X ) := inf sup kx − ∆(A x )kX : A ∈ L (X , R ), ∆ : R → X m
m
m
.
x ∈K
A is the measurement map and ∆ : Rm → X is the arbitrary reconstruction map.
∆
A Björn Bringmann
Gelfand Widths in Compressive Sensing
10 / 26
Introduction to m-Widths
Compressive m-widths
Adaptive compressive m-width Definition The adaptive map F : X → Rm is defined by
F (x ) :=
λ1 (x ) λ2;λ1 (x ) (x )
. .. . λm;λ1 (x ),...,λm−1 (x ) (x )
for λ1 (·), λ2;λ1 (x ) (·), . . . , λm;λ1 (x ),...,λm−1 (x ) (·) ∈ X ∗ . The adaptive compressive m-width of a subset K of a Banach space X is defined as m Eada (K ; X ) := inf
Björn Bringmann
m
m
sup kx − ∆(F (x ))k : F : X → R adaptive, ∆ : R → X x ∈K
Gelfand Widths in Compressive Sensing
11 / 26
Introduction to m-Widths
Compressive m-widths
Connection with the Gelfand m-width Theorem If K is a subset of a Banach space X, then m Eada (K ; X ) ≤ E m (K ; X ) .
If the subset K satisfies −K = K , then m d m (K ; X ) ≤ Eada (K , X ) .
If the set K further satisfies K + K ⊂ a K for some positive constant a, then E m (K ; X ) ≤ a d m (K ; X ) . Therefore under these assumptions 1 m m (K ; X ) ≤ E m (K ; X ) . E (K ; X ) ≤ d m (K ; X ) ≤ Eada a Björn Bringmann
Gelfand Widths in Compressive Sensing
12 / 26
Gelfand Widths of l1 -balls
Outline 1
Introduction to m-Widths Kolmogorov and Gelfand Widths Compressive m-widths
2
Gelfand Widths of l1 -balls Upper bound Lower bound
3
Applications Optimal Number of Measurements Kashin’s Decomposition Theorem
Björn Bringmann
Gelfand Widths in Compressive Sensing
13 / 26
Gelfand Widths of l1 -balls
Main result
Theorem For 1 < p ≤ 2 and m < N, there exist constants c1 , c2 > 0 depending only on p such that
c1 min 1,
ln(eN /m) m
Björn Bringmann
1− p1
1− p1 ln(eN /m) ≤ d m (B1N , `N ) ≤ c min 1 , . 2 p m
Gelfand Widths in Compressive Sensing
14 / 26
Gelfand Widths of l1 -balls
Upper bound
Upper bound
Theorem There is a constant C > 0 such that, for 1 < p ≤ 2 and m < N
d m (B1N , `N p ) ≤ C min 1,
Björn Bringmann
ln(eN /m) m
1− p1 .
Gelfand Widths in Compressive Sensing
15 / 26
Gelfand Widths of l1 -balls
Lower bound
Lower bound
Theorem There is a constant c > 0 such that, for 1 < p ≤ ∞ and m < N
d m (B1N , `N p ) ≥ c min 1,
Björn Bringmann
ln(eN /m) m
1− p1 .
Gelfand Widths in Compressive Sensing
16 / 26
Gelfand Widths of l1 -balls
Lower bound
Recovery of 2s-sparse vectors
Theorem Given a matrix A ∈ Rm×N , if every 2s-sparse vector x ∈ RN is a minimizer of kz k1 subject to A z = A x, then
m ≥ c1 s ln where c1 =
1 ln(9)
N c2 s
and c2 = 4.
Björn Bringmann
Gelfand Widths in Compressive Sensing
17 / 26
Gelfand Widths of l1 -balls
Lower bound
Preparation
Lemma Given integers s < N, there exist
n≥
N 4s
s/2
subsets S1 , . . . , Sn of [N ] such that each Sj has cardinality s and
card(Si ∩ Sj )
0 by the maximum number P of points xk ∈ T , k ∈ [P], which are t-separated, i.e. kxk − xl k > t for all k , l ∈ [P], k 6= l.
•
• •
• •
•
• •
•
Lemma Then for any norm k · k on Rm there holds
P (B1 (0), k · k, t ) ≤ Björn Bringmann
2 1+ t
•
m . Gelfand Widths in Compressive Sensing
19 / 26
Gelfand Widths of l1 -balls
Lower bound
Preparation Definition Let X be a Banach space. For a subset T ⊂ X define the packing number P (T , k · kX , t ) for t > 0 by the maximum number P of points xk ∈ T , k ∈ [P], which are t-separated, i.e. kxk − xl k > t for all k , l ∈ [P], k 6= l.
•
• •
• •
•
• •
•
Lemma Then for any norm k · k on Rm there holds
P (B1 (0), k · k, t ) ≤ Björn Bringmann
2 1+ t
•
m . Gelfand Widths in Compressive Sensing
19 / 26
Gelfand Widths of l1 -balls
Lower bound
Kolmogorov widths Revisited
Corollary For 2 ≤ p < ∞ and m < N, there exist constants c1 , c2 > 0 depending only on p such that the Kolmogorov widths satisfy
( c1 min 1,
eN m
ln
m
Björn Bringmann
)1−1/p
( ≤
dm (BpN , `N ∞)
≤ c2 min 1,
eN m
ln
)1−1/p
m
Gelfand Widths in Compressive Sensing
20 / 26
Applications
Outline 1
Introduction to m-Widths Kolmogorov and Gelfand Widths Compressive m-widths
2
Gelfand Widths of l1 -balls Upper bound Lower bound
3
Applications Optimal Number of Measurements Kashin’s Decomposition Theorem
Björn Bringmann
Gelfand Widths in Compressive Sensing
21 / 26
Applications
Optimal Number of Measurements
Estimates for the Compressive m-widths
Corollary For 1 < p ≤ 2 and m < N, the adaptive and nonadaptive compressive m-widths satisfy m Eada (B1N , `N p)
Björn Bringmann
E
m
(B1N , `N p)
ln(eN /m) min 1, m
1− p1
Gelfand Widths in Compressive Sensing
.
22 / 26
Applications
Optimal Number of Measurements
Optimal Number of Measurements Theorem Let 1 < p ≤ 2. Suppose that the matrix A ∈ Rm×N and the map ∆ : Rm → RN satisfy
kx − ∆(A x )kp ≤
C σs (x )1 s1−1/p
∀x ∈ RN .
Then for some constant c > 0 depending only on C there holds
m ≥ c s ln
eN s
.
(1)
In particular, if A ∈ Rm×N satisfies δ2s (A) < 0.6246, then necessarily (1) holds with c = c (δ2s ).
Björn Bringmann
Gelfand Widths in Compressive Sensing
23 / 26
Applications
Optimal Number of Measurements
Donoho-Tanner Phase Transition
Figure: Success of L1 −Minimization from Random Partial Fourier Measurements. x-Axis: δ = m/N undersampling fraction y-Axis: ρ = s/m sparsity fraction Björn Bringmann
Gelfand Widths in Compressive Sensing
24 / 26
Applications
Kashin’s Decomposition Theorem
Kashin’s Decomposition Theorem
Theorem There exist universal constants α, β > 0 such that, for any m ≥ 1 the space R2m can be split into an orthogonal sum of two m-dimensional subspaces E and E ⊥ such that
√ √ α mkx k2 ≤ kx k1 ≤ β mkx k2 for all x ∈ E and for all x ∈ E ⊥ .
Björn Bringmann
Gelfand Widths in Compressive Sensing
25 / 26
References
References Holger Rauhut and Simon Foucart A Mathematical Introduction to Compressive Sensing Birkhäuser, 2013 Allan Pinkus n-Widths in Approximation Theory Springer, 1985 David L. Donoho and Jared Tanner Observed Universality of Phase Transitions in High Dimensional Geometry, with Implications for Modern Data Analysis and Signal Processing Philosophical Transactions of the Royal Society, 2009 Björn Bringmann
Gelfand Widths in Compressive Sensing
26 / 26