ON THE STABILITY OF FOURIER PHASE RETRIEVAL

STEFAN STEINERBERGER
arXiv:2004.06671v1 [math.FA] 14 Apr 2020

Abstract. Phase retrieval is concerned with recovering a function f from the

absolute value of its Fourier transform |fb|. We study the stability properties
of this problem in Lp (Rn ) ∩ L2 (Rn ) for 1 ≤ p < 2. The simplest result is as
follows: if f ∈ L1 (Rn ) ∩ L2 (Rn ) has a real-valued Fourier transform supported
on a set of measure L < ∞, then for all all g ∈ L1 (Rn ) ∩ L2 (Rn )
√
kf − gkL2 ≤ 2 · k|fb| − |b
g |kL2 + 30 L · kf − gkL1 + 2k Im g bkL2 .
This is a form of stability of the phase retrieval problem for band-limited
functions (up to the translation symmetry captured by the last term). The
inequality follows from a general result for f, g ∈ Lp (Rn ) ∩ L2 (Rn ).

1. Introduction and Results

1.1. Introduction. Phase retrieval refers to a broad class of problems where one
is given incomplete information about an object (often the size of the coefficients
with respect to some basis expansion but not their phase) and then tries to re-
construct the object. In the case of the Fourier transform, the challenge is to re-
cover a function f from knowing only the modulus of its Fourier transform |fb|.
The problem itself is classical and first arose, implicitly, a century ago in the
setting of x-ray crystallography. It has since appeared in a variety of different
fields [18, 28, 30, 35, 36, 39, 40]. There is a vast literature, one possible start-
ing point is [1, 2, 12, 13, 14, 16, 17, 20, 22, 23, 24, 25, 29, 31, 32, 33, 37, 38].
The question has recently been studied in more abstract settings (say, recovering
Hilbert space elements for which one knows the size of certain inner products), see
[3, 4, 5, 8, 9, 10, 11, 15, 19, 26, 27, 34] and references therein. We cannot possibly
hope to summarize the existing literature but we emphasize two very recent and
excellent surveys, one by Grohs, Koppensteiner & Rathmair [25] about theoretical
aspects and one by Fannjiang & Strohmer [20] about the numerical side of things.

We study the stability problem and begin by recalling the translation symmetry:
the functions f (x) and its shift f (x+ε) cannot be distinguished from looking at the
modulus of their Fourier transform and this has to play a role in all the stability
results. We also recall that if f, g ∈ L2 (R) are both compactly supported, then there
exists a convenient characterization of all pairs (f, g) ∈ L2 (R) × L2 (R) for which
|fb| = |b
g | in terms of complex analysis (since both fb and gb are entire) [1, 2, 29, 40].
As a consequence of this characterization, we have a basic uniqueness result, see
for example [25, Theorem 4.9] or [33, Proposition 3.3]

2010 Mathematics Subject Classification. 42A63, 45Q05.

Key words and phrases. Phase Retrieval, Stability, Fourier Transform.
S.S. is supported by the NSF (DMS-1763179) and the Alfred P. Sloan Foundation.
1
2

Proposition (see e.g. [1, 2, 25, 29, 33, 40]). If f ∈ L2 (R) is compactly supported
and satisfies fb(−ξ) = fb(ξ), then it is uniquely determined by |fb|.
We are interested in the stability question: let us fix f, g ∈ L2 (R). If |fb| ∼ |b
g|
are close in L2 , does this necessarily imply that f and g themselves are close in
L2 ? Without any further assumptions, this is certainly wrong: for any function
h : R → {−1, 1}, we can define
gb(ξ) = h(ξ)fb(ξ)
which results in them having the same modulus but there is absolutely no reason
for them to be close to one another in L2 . This shows that at least one more
assumption is needed. We will show that such stability results become possible if
we assume that f and g are close in Lp for some 1 ≤ p < 2.

2. The Result
2.1. The simplest case. We present the principle first in its simplest form. This is
not the most general formulation but maybe the one that is most easily visualized.
Fact. For any two functions f, g ∈ L1 (Rn ) ∩ L2 (Rn ), if
(1) supp(fb) has finite measure,
(2) g is close to f in L1 and
g | in L2 ,
(3) |fb| is close to |b
then f is close to g in L2 (up to the translation symmetry).

Figure 1. An even function f and g(x) = f (x − ε). They are

close in L1 and |fb| = |b
g | but they are not close in L2 . Any stability
estimate needs to compensate for this translation symmetry.

We quickly describe how this could be interpreted. Let us first specify what it
means for two function f, g to be close in L1 but not in L2 : it indicates that f − g
has concentrations of L1 −mass or, equivalently, that |f − g| assumes large values
over a small interval scaled in such a way that the contribution in L1 is not sub-
stantial but becomes substantial in L2 . One way of phrasing the general principle
is that in such a case we are either dealing with a translation symmetry, g is close
to a shift of f , or this behavior becomes visible in k|fb| − |b
g |kL2 : strong localized
mass translates into slow decay of the Fourier transform. Assuming a smoothness
condition on fb we obtain for large frequencies ξ that ||fb(ξ)| − |bg (ξ)|| ∼ |b
g (ξ)| and
we can recover the slow decay this way. Naturally, there are other interpretations.
 3

We introduced this principle in the case where f satisfies a very strong smoothness
condition and has a real-valued Fourier transform that is compactly supported.
The general result does not require any such conditions.
Corollary 1. Let f ∈ L1 (Rn ) ∩ L2 (Rn ) have a real-valued Fourier transform sup-
ported on a set of measure L = {ξ ∈ Rn : fb(ξ) 6= 0}|. For all g ∈ L1 (Rn ) ∩ L2 (Rn )
√
kf − gkL2 (Rn ) ≤ 2 · k|fb| − |b
g |kL2 (Rn ) + 30 L · kf − gkL1 (Rn ) + 2k Im gbkL2 (Rn ) .
This result shows stability of the phase retrieval problem of a function f with
real-valued and compactly supported Fourier transform in L1 (Rn ) ∩ L2 (Rn ) up to
the translation symmetry. If the Fourier transform was not real-valued and merely
compactly supported, we obtain the same result with k Im gbkL2 , the term accounting
for the translation symmetry, replaced by a slightly more general expression which
we discuss below. It is not difficult to see that Corollary 1 has the sharp scaling
and the optimal dependence on L: let φ : R → R be an even, nonnegative and
compactly supported C ∞ −function in [−1, 1] and set
 
1 ξ
fb(ξ) = φ and gb(ξ) = −f (ξ).
L L

We have |fb| = |b
g |, Im gb ≡ 0, f (x) = φ(Lx)
b and thus, as L becomes large,
1 1
kf − gkL2 ∼ √ and kf − gkL1 ∼ .
L L
Therefore Corollary 1 is optimal up to constants. Once f is less smooth, the
L1 −distance stops acting linearly, the bound moves from Lipschitz to Hölder.
Corollary 2. Let f ∈ L1 (Rn ) ∩ L2 (Rn ) have a real-valued Fourier transform and
f ∈ C k (Rn ) where k > (n + 2)/2. Then, for some constant cf > 0 depending only
on f and all g ∈ L1 (Rn ) ∩ L2 (Rn ) ,
n
1− 2k
kf − gkL2 (Rn ) ≤ 2 · k|fb| − |b
g |kL2 (Rn ) + cf kf − gkL1 (R bkL2 (Rn ) .
n ) + 2k Im g

Figure 2. (1 − cos x)x−2 has super-linear stability in L1 .

4

The result is also applicable to specific functions (see Fig. 2). We discuss f (x) =
(1 − cos x)x−2 satisfying fb(ξ) = max {0, 1 − |x|}. An application of the Theorem
shows that for some universal constant c > 0 and any even function g : R → R,
n o
3/2
kf − gkL2 (R) ≤ 2 · k|fb| − |b
g |kL2 (R) + c min kf − gkL1 (R) , kf − gkL1 (R) .
The second estimate follows from Corollary 1 since f is band-limited. The first
bound, decaying faster than linearly for f close to g in L1 , follows from using the
explicit form of the Theorem. This inequality could also easily be directly proven
since all the terms are real. We see that the Fourier Phase Retrieval problem for
f is quite stable in the space of symmetric perturbations as soon as f and g are
quite close in L1 (and g is even; otherwise the translation symmetry makes any
approximation faster than linear impossible).
2.2. The General Result. We will now describe the general stability result for
functions in Lp (Rn ) ∩ L2 (Rn ) where 1 ≤ p < 2. The result has the same form as the
results as above: the L2 −distance of two functions f and g is bounded from above
by the sum of three terms: (a) the L2 −distance of |fb| and |b g |, (b) the distance of
f and g in Lp where 1 ≤ p < 2, in a way depending on the smoothness of f , and
(c) a term accounting for the invariance under translations.
Theorem. Let 1 ≤ p < 2 and f ∈ Lp (Rn ) ∩ L2 (Rn ). We define hf : R≥0 → R≥0
Z !1/2 (
2 x if p > 1
hf (x) = 8 |fb(ξ)| dξ +
|f (ξ)|≤10x
b 0 if p = 1.
Then, for all g ∈ L1 (Rn ) ∩ L2 (Rn ),
  2
g |kL2 (Rn ) + hf kf − gkLp (Rn ) + 2 Im fb|fb|−1 gb 2
kf − gkL2 (Rn ) ≤ 2 · k|fb| − |b .

L (Rn )

The result is very much in the same flavor as the results above: we have quantitative
dependence on kf − gkLp whose rate depends on the smoothness of f . If fb is
supported on a set of measure L, then hf grows at most linearly since
!1/2 !1/2
Z Z √
8 |fb(ξ)|2 dξ ≤ 8 (10x)2 dξ ≤ 30 Lx.
|fb(ξ)|≤10x |fb(ξ)|≤10x

In the special case p = 1, it might even have smaller growth (as seen in the example
above). If fb is real-valued, then the last term simplifies to
Im fb|fb|−1 gb = Im gb
which recovers the previous results. We quickly illustrate the meaning of this term
for smooth functions f, g ∈ L2 by considering the case g(x) = f (x − ε). Their
Fourier transforms have the same modulus. As ε → 0 and for fixed ξ ∈ R the term
1 − cos (εξ) is quadratic in ε while i sin (εξ) is linear in ε and thus
kf − gkL2 = kfb(ξ) − e−iεξ fb(ξ)kL2 = kfb(ξ) (1 − cos (εξ) + i sin (εξ)) k|L2
∼ kfb(ξ) (i sin (εξ)) k|L2 = kfb(ξ) sin (εξ)kL2
We see that this term is a genuine L2 −quantity that is unlikely to be controlled by
Lp for p < 2. Indeed, it is controlled by the third quantity, since
| Im fb(ξ)|fb(ξ)|−1 gb(ξ)| = | Im fb(ξ)|fb(ξ)|−1 fb(ξ)e−iεξ | = |fb(ξ) sin (εξ)|.
 5

2.3. Main Idea and Extensions. We quickly illustrate the main idea and how
it would allow for even more general results. We focus on the case p = 1 and
f, g ∈ L1 (R) ∩ L2 (R). Let us suppose ε = kf − gkL1 (R) is small. Then
|fb(ξ) − gb(ξ)| ≤ kfb − gbkL∞ ≤ kf − gkL1 = ε
is uniformly small in ξ. Let us now consider a value ξ where |fb(ξ)| ≥ 10ε.

fb(ξ)

gb(ξ)

Figure 3. A sketch of the main idea.

g (ξ) − fb(ξ)) and then consider the quantity (b

We can express gb(ξ) = fb(ξ) + (b g (ξ) −
2 ∼
f (ξ)) as a vector in R = C. It is relatively small compared to the size of f (ξ). If
b b
it points roughly in the direction of fb(ξ), then |fb(ξ) − gb(ξ)| ∼ ||fb(ξ)| − |b
g (ξ)|| and
the difference in L2 shows up in the modulus. If that is not the case, then gb(ξ) −
fb(ξ) points roughly in the direction ifb(ξ) which means that there is a nontrivial
contribution to the subspace corresponding to translations of f . What is interesting
about this idea is that very few properties of the absolute value | · | : C → R are
being used (though the fact that the ‘critical’ subspace that cannot be recovered
corresponds to translations of the function f is very much connected to using the
absolute value; the more general problem will have other ‘critical’ subspaces that
do not have such an easy interpretation). Under some regularity assumptions, the
same type of arguments could be used to deal with more general problems of this
type. This would allow one, for example, to define h : C → R via
h(z) = ((Re z)4 + (Im z)4 )1/4
and then try to study the phase retrieval problem for h(fb) ∼ h(b
g ) for which similar
stability estimates could be obtained. The classical Fourier phase retrieval is well
motivated and this is maybe not (or not yet) the case for this generalized problem;
however, it does seem interesting that the methods extend.

3. Proofs
3.1. A Lemma.
Lemma 1. For all 0 ≤ w ∈ R and all z ∈ C satisfying |z − w| ≤ |w|/2
 
2 2
z − w
|w − Re z| ≤ |w − |z|| + 2 
  · | Im z|2 .
w 
6

Proof. Both sides of the inequality are invariant under multiplication with scalars,
so we can assume w.l.o.g. that w = 1. It then remains to show that
1
|1 − Re z|2 ≤ |1 − |z||2 + 2 |z − 1| · | Im z|2 for all |z − 1| ≤ .
2
We make the ansatz z = (1 + x) + iy which reduces the desired inequality to
p p
2 + 2x + y 2 + 2y 2 x2 + y 2 ≥ 2 (1 + x)2 + y 2
for all x2 + y 2 ≤ r2 for some r to be determined. The left-hand side is certainly
positive in the regime that we consider, so we can square both sides and try to
verify that identity instead. The desired inequality factors into
 p p p 
X = y 2 y 2 + 8 x2 + y 2 + 4x + 4x2 y 2 + 8x x2 + y 2 + 4y 2 x2 + y 2 ≥ 0.
We define r via x2 + y 2 = r2 and note that for any r ≤ 1/2
X p p
2
≥ 8 x2 + y 2 + 4x + 8x x2 + y 2 ≥ 4r − 8r2 ≥ 0.
y

3.2. Proof of the Theorem.
Proof. Let 1 ≤ p < 2. We fix the parameter
ε = kf − gkLp (Rn )
and write Z
kf − gk2L2 (Rn ) = |fb(ξ) − gb(ξ)|2 dξ.
Rn
We split this integral into two integrals over disjoint regions
Z Z
2
|f (ξ) − gb(ξ)| dξ =
b |fb(ξ) − gb(ξ)|2 dξ
Rn |fb(ξ)|≥10ε
Z
+ |fb(ξ) − gb(ξ)|2 dξ.
|fb(ξ)|≤10ε

The remainder of the argument is comprised of estimating these two integrals.

First Integral. We split the first integral once more

Z Z
2
|f (ξ) − gb(ξ)| dξ =
b |fb(ξ) − gb(ξ)|2 dξ
b |f (ξ)|≥10ε
|fb(ξ)|≥10ε
|fb(ξ)−b g (ξ)|≤ε
Z
+ |fb(ξ) − gb(ξ)|2 dξ.
|fb(ξ)|≥10ε
|fb(ξ)−b g (ξ)|≥ε

First term. We start by analyzing the first term. Fix a ξ ∈ R such that |fb(ξ)| ≥ 10ε
and |fb(ξ) − gb(ξ)| ≤ ε. We interpret fb(ξ) (which is not 0) and ifb(ξ) as the directions
of two orthogonal vectors in R2 and use them to express
fb(ξ) ifb(ξ)
gb(ξ) = fb(ξ) + a +b
|fb(ξ)| |fb(ξ)|
for some unique a, b ∈ R. We see that a and b cannot be very large since
p |fb(ξ)|
a2 + b2 = |fb(ξ) − gb(ξ)| ≤ ε ≤ .
10
 7

Multiplying on both sides with fb(ξ)/|fb(ξ)| results in

fb(ξ)
gb(ξ) = |fb(ξ)| + a + bi.
|fb(ξ)|
However, to this equation we can apply Lemma 1 with

fb(ξ)
w = |fb(ξ)| and z= gb(ξ)
|fb(ξ)|
and we obtain
1
|a|2 ≤ ||fb(ξ)| − |b
g (ξ)||2 + |b|2 .
5
This, in turn, implies that
6
|fb(ξ) − gb(ξ)|2 = a2 + b2 ≤ ||fb(ξ)| − |bg (ξ)||2 + |b|2
5
and thus, recalling the definition of b, we get, in this regime, the pointwise estimate
 2
2 2 2 2 6  fb(ξ) 
|f (ξ) − gb(ξ)| = a + b ≤ ||f (ξ)| − |b g (ξ)|| + Im gb(ξ) .
b b 
5  |fb(ξ)| 
This concludes our analysis of the first term and we arrive at
Z
6 2
| b(ξ) − gb(ξ)|2 dξ ≤ k|fb| − |b
f g |k 2
2 n +
b|fb|−1 gb
f 2 n .
L (R )
5
|fb(ξ)|≥10ε
Im
L (R )
|f (ξ)−b
b g (ξ)|≤ε

Second term. We now estimate the second term in the first integral. We recall that
if p = 1, then
kfb − gbkL∞ ≤ kf − gkL1 = ε
and the domain of integration
n o
X = ξ ∈ Rn : |fb(ξ)| ≥ 10ε ∧ |fb(ξ) − gb(ξ)| ≥ ε
is the empty set. This leads to the slight improvement if p = 1. Let us thus assume
that 1 < p < 2. We use Hölder’s inequality to argue that
Z Z  2p−2
p
p 2−p
2
|f (ξ) − gb(ξ)| dξ ≤
b |f (ξ) − gb(ξ)|
b p−1 dξ |X| p .
X X
The set X cannot be too big, note that
p
Z
p
p−1
kf − gbk p =
b |fb(ξ) − gb(ξ)| p−1 dξ
L p−1 n
ZR
p p
≥ |fb(ξ) − gb(ξ)| p−1 dξ ≥ ε p−1 |X|.
X
The Hausdorff-Young inequality, valid for any h ∈ Lp (Rn ) for 1 ≤ p ≤ 2,
kb
hk p ≤ khkLp (Rn )
L p−1 (Rn )

then implies, recalling the definition of ε,

p
p
kfb − gbk p−1p kf − gkLp−1
L p−1 p
|X| ≤ p ≤ p = 1.
ε p−1 ε p−1
8

We remark that we could get a slightly better constant from using Beckner’s in-
equality [6, 7] but do not pursue sharp constants in this paper. Therefore
Z Z  2p−2
p
p
|fb(ξ) − gb(ξ)|2 dξ ≤ |fb(ξ) − gb(ξ)| p−1 dξ .
X X

Employing the Hausdorff-Young inequality once more, we obtain

Z Z  2p−2
p
p
|fb(ξ) − gb(ξ)|2 dξ ≤ |fb(ξ) − gb(ξ)| p−1 dξ
X X
Z 2 p−1
p
p
≤ |fb(ξ) − gb(ξ)| p−1 dξ ≤ kf − gk2Lp (Rn ) .
Rn

Second Integral. This estimate is simple, we use the elementary inequality

|a − b|2 ≤ ((|a| − |b|) + 2|b|)2 ≤ 2(|a| − |b|)2 + 8|b|2 .

to argue that
Z Z
|fb(ξ) − gb(ξ)|2 dξ ≤ 8 |fb(ξ)|2 dξ
|fb(ξ)|≤10ε |fb(ξ)|≤10ε
Z
+2 g (ξ)||2 dξ.
||fb(ξ)| − |b
|fb(ξ)|≤10ε

Conclusion. Collecting all these estimates, we obtain two different bounds depend-
ing on the value of p. If 1 < p < 2, we obtain
6 2
g |k2L2 (Rn ) + Im fb|fb|−1 gb
kf − gk2L2 (Rn ) ≤ 2 · k|fb| − |b

5 L2 (Rn )
Z
+ kf − gk2Lp (Rn ) + 8 |fb(ξ)|2 dξ.
|fb(ξ)|≤10ε

If p = 1, then the kf − gk2L1 (Rn ) term can be omitted because we can estimate the
second term in the first integral by 0. In either case, recalling the definition of ε,
Z Z
|fb(ξ)|2 dξ = |fb(ξ)|2 dξ
|fb(ξ)|≤10ε |fb(ξ)|≤10kf −gkLp (Rn )

which results in the desired statement. 

3.3. Proof of Corollary 2.

Lemma 2. Let f ∈ L1 (Rn ) ∩ L2 (Rn ) be k−times continuously differentiable where

k > (n + 2)/2. Then, as ε → 0, we have for some constant c > 0 depending on f ,
Z
|fb(ξ)|2 dξ ≤ c · ε2−n/k .
|fb(ξ)|≤10ε

Proof. We know that, for some implicit constant depending only on f ,

1
|fb(ξ)| . .
1 + |ξ|k
 9

We observe that this is . ε as soon as |ξ| ≥ ε−1/k . This allows us to estimate

Z Z Z
2 2
|f (ξ)| dξ ≤
b |f (ξ)| 1|fb(ξ)|≤10ε dξ +
b |fb(ξ)|2 dξ
|fb(ξ)|≤10ε |ξ|≤ε−1/k |ξ|≥ε−1/k
Z
2 −n/k
≤ε ε + |fb(ξ)|2 dξ
|ξ|≥ε−1/k
Z ∞
−2k n−1
. ε2−n/k + |r| r dξ . ε2−n/k .
ε−1/k


References
[1] E. J. Akutowicz. On the determination of the phase of a Fourier integral, I. Transactions of
the American Mathematical Society, pages 179–192, 1956.
[2] E. J. Akutowicz. On the determination of the phase of a Fourier integral. II. Proc. Amer.
Math. Soc., 8:234–238, 1957.
[3] R. Alaifari and P. Grohs, Gabor phase retrieval is severely ill-posed, App. Comp. Harm. Anal,
to appear
[4] R. Alaifari, I. Daubechies, P. Grohs and G. Thakur, Recunstructing real-valued functions from
unsigned coefficients with respect to wavelet and other frames, Journal of Fourier Analysis
and Applications 23 (2017), p. 1480–1494
[5] B. Alexeev, A. S. Bandeira, M. Fickus and D. G. Mixon, Phase retrieval with polarization
SIAM Journal on Imaging Sciences 7 (2014), p. 35–66
[6] I. Babenko, An inequality in the theory of Fourier integrals, Izvestiya Akademii Nauk SSSR.
Seriya Matematicheskaya, 25 (1961): 531–542
[7] W. Beckner, William, Inequalities in Fourier analysis, Annals of Mathematics, Second Series,
102 (1975): p. 159–182
[8] R. Balan, B. G. Bodmann, P. G. Casazza, and D. Edidin. Painless reconstruction from
magnitudes of frame coefficients. J. Fourier Anal. Appl., 15 (2009): pp. 488–501.
[9] R. Balan, P. Casazza, and D. Edidin. On signal reconstruction without phase. Appl. Comput.
Harmon. Anal., 20(3):345–356, 2006.
[10] R. Balan and Y. Wang. Invertibility and robustness of phaseless reconstruction. Appl. Com-
put. Harmon. Anal., 38(3):469–488, 2015.
[11] A. S. Bandeira, J. Cahill, D. G. Mixon, and A. A. Nelson. Saving phase: Injectivity and
stability for phase retrieval. Appl. Comput. Harmon. Anal., 37(1):106–125, 2014.
[12] R. E. Burge, M. A. Fiddy, A. H. Greenaway, and G. Ross. The application of dispersion
relations (Hilbert transforms) to phase retrieval. Journal of Physics D: Applied Physics,
7(6):65, 1974.
[13] R. E. Burge, M. A. Fiddy, A. H. Greenaway, G. Ross, and W. C. Price. The phase prob-
lem. Proceedings of the Royal Society of London. A. Mathematical and Physical Sciences,
350(1661):191–212, 1976.
[14] Y. Bruck and L. Sodin, On the Ambiguity of the Image Reconstruction Problem, Optics
Communication 30 (1979), p. 304–308.
[15] J. Cahill, P. G. Casazza and I. Daubechies, Phase retrieval in infinite-dimensional Hilbert
spaces. Trans. Amer. Math. Soc. Ser. B 3 (2016), p. 63–76.
[16] T. R. Crimmins and J. R. Fienup. Ambiguity of phase retrieval for functions with disconnected
support. J. Opt. Soc. Am., 71 (1981): p. 1026–1028.
[17] ] T. R. Crimmins and J. R. Fienup. Uniqueness of phase retrieval for functions with sufficiently
disconnected support. J. Opt. Soc. Am., 73(1983): p. 218–221.
[18] J. Dainty and J. Fienup. Phase retrieval and image reconstruction for astronomy. Image
Recovery: Theory Appl., 13:231–275, 1987.
[19] Y. Eldar, P. Sidorenko, D. G. Mixon, S. Barel and O. Cohen, Sparse Phase Retrieval from
Short-Time Fourier Measurements, IEEE Signal Processing Letters 22.5 (2014): 638–642.
[20] A. Fannjiang, Absolute uniqueness of phase retrieval with random illumination, Inverse Prob-
lems 28 (2012), 075008
[21] A. Fannjiang and T. Strohmer, The Numerics of Phase Retrieval, Acta Numerica, to appear
10

[22] J. R. Fienup. Reconstruction of an object from the modulus of its Fourier transform. Opt.
Lett., 3(1978): p. 27–29.
[23] D. Gabor. A New Microscopic Principle. Nature, 161 (1948): p. 777–778.
[24] A. H. Greenaway. Proposal for phase recovery from a single intensity distribution. Opt. Lett.,
1 (1977): p.10–12.
[25] P. Grohs, S. Koppensteiner and M. Rathmair, Phase Retrieval: Uniqueness and Stability,
SIAM Review, to appear.
[26] P. Grohs and M. Rathmair, Stable Gabor Phase Retrieval and Spectral Clustering, Comm.
Pure Appl. Math 72 (2019):p. 981–1043.
[27] P. Grohs and M. Rathmair, Stable Gabor phase retrieval for multivariate functions,
arXiv:1903.01104
[28] R. Harrison, Phase problem in crystallography, J. Opt. Soc. Am. A 10 (1993), p. 1045–1055.
[29] E. Hofstetter. Construction of time-limited functions with specified autocorrelation functions.
IEEE Transactions on Information Theory, 10 (1964): p. 119–126.
[30] P. Jaming. Phase retrieval techniques for radar ambiguity problems. Journal of Fourier Anal-
ysis and Applications 5 (1999): p. 309–329.
[31] P. Jaming Uniqueness results in an extension of Pauli’s phase retrieval. Applied and Compu-
tational Harmonic Analysis 37 (2014) 413–441.
[32] P. Jaming, K. Kellay and R. Perez, Phase Retrieval for Wide Band Signals, 13th International
conference on Sampling Theory and Applications (SampTA), Bordeaux, France, 2019,pp 1–4.
[33] M. V. Klibanov, P. E. Sacks, and A. V. Tikhonravov. The phase retrieval problem. Inverse
Problems, 11(1):1, 1995.
[34] S. Mallat and I. Waldspurger. Phase retrieval for the Cauchy wavelet transform. J. Fourier
Anal. Appl., 21: p. 1251–1309, 2015.
[35] J. Miao, D. Sayre, and H. N. Chapman, Phase retrieval from the magnitude of the Fourier
transforms of nonperiodic objects, J. Opt. Soc. Am. A 15 (1998), p. 1662–1669
[36] R. Millane), Phase retrieval in crystallography and optics, J. Opt. Soc. Am. A. 7 (1990),
394–411.
[37] J. Rosenblatt. Phase retrieval. Comm. Math. Phys., 95: p.317–343, 1984.
[38] J. Sanz, Mathematical Considerations for the Problem of Fourier Transform Phase Retrieval
from Magnitude, SIAM J. Appl. Math., 45 (1985), 651–664.
[39] Y. Shechtman, Y. C. Eldar, O. Cohen, H. N. Chapman, J. Miao and M. Segev, Phase re-
trieval with application to optical imaging: a contemporary overview, IEEE Signal Processing
Magazine 32 (2015), p. 87–109.
[40] A. Walther. The question of phase retrieval in optics. Journal of Modern Optics, 10(1):41–49,
1963.

Department of Mathematics, Yale University, New Haven, CT 06511, USA

E-mail address: [email protected]