On the Injectivity of STFT Phase Retrieval with super-exponential decaying window function

Shuang Guan and Kasso A. Okoudjou · 2025

Read On the Injectivity of STFT Phase Retrieval with super-exponential decaying window function by Shuang Guan and Kasso A. Okoudjou (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-15894.txt.

Introduction

Shuang Guan and Kasso A. Okoudjou study when a signal can be uniquely recovered from its short-time Fourier spectrogram without phase information. Phase retrieval matters because many sensors record only magnitudes, and uniqueness guarantees that no other signal could produce the same recording. Their setting uses window functions whose Fourier transforms decay faster than any exponential, a very strong smoothness condition.

For such windows, they derive sufficient conditions on discrete sampling sets that ensure unique phase retrieval for square-integrable signals. The result generalizes earlier uniqueness guarantees that were proven only for Gaussian windows, the classic bell-shaped case. Knowing which sampling patterns suffice helps design measurement schemes where reconstruction is provably unambiguous.

Catalog record

CREATORShuang Guan and Kasso A. Okoudjou
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE33 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-15894

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

On the Injectivity of STFT Phase Retrieval with super-exponential decaying window function Shuang Guan and Kasso A. Okoudjou Abstract We investigate the uniqueness of short-time Fourier transform phase retrieval problems in L 2 ​ ( ℝ ) L^{2}(\mathbb{R}) . In particular, for underlying window functions whose Fourier transform decay faster than any exponential function, we derive sufficient conditions on discrete sampling sets for unique phase retrieval from the spectrogram. This result generalizes previous uniqueness guarantees on sampling sets for Gaussian windows. 1 Introduction The phase retrieval problem refers to the task of recovering a signal from only intensity measurements under a given operator. This problem arises naturally in a wide range of applications in science and engineering, including X-ray crystallography millane1990phase , coherent diffraction imaging walther1963question and speech recognition picone1996fundamentals . In recent years, the phase retrieval problem has attracted significant interests, involving tools from harmonic analysis, complex analysis, functional analysis and numerical optimization candes2015phase ; alaifari2021phase ; grohs2019stable . Mathematically the phase retrieval problem is a nonlinear inverse problem that is typically ill-posed due to the natural ambiguity introduced by the loss of phase information. Over the years, addressing this subtle problem has led to a rich literature seeking to examine conditions for stability and feasibility of certain reconstruction algorithms, spanning both discrete and continuous settings. In this paper, we specifically study the phase retrieval problem in the setting of short-time Fourier transform (STFT). For a signal f\in L^{2}(\mathbb{R}) and a window function g\in L^{2}(\mathbb{R}) , the STFT is defined as a function in L^{2}(\mathbb{R}^{2}) given by: V_{g}f(x,\omega)=\int_{\mathbb{R}}f(t)\overline{g(t-x)}e^{2\pi i\omega\cdot t}dt,\quad(x,\omega)\in\mathbb{R}^{2}. In many real-world applications, detectors can only capture the intensity, |V_{g}f(x,\omega)| , known as the spectrogram. The loss of phase information leads to the STFT phase retrieval problem: determine the unique signal f up to a global sign from a sampling set of spectrogram |V_{g}f(\Lambda)| for some \Lambda\subset\mathbb{R}^{2} . While the STFT phase retrieval problem is ill-posed in general, it is known that uniqueness can still be guaranteed, up to a global phase factor, by imposing certain conditions on the window function g and sampling on a sufficiently dense set \Lambda . However, the difficulty of this problem is two-fold: First, it was recently proved that if \Lambda is an integer lattice, i.e, \Lambda=A\mathbb{Z}^{2} for some A\in GL(2,\mathbb{R}) , then the uniqueness is never attained grohs2023injectivity . The second difficulty is concerning the window function g , in the sense that if V_{g}g(x,\omega) vanishes on a set of positive Lebesgue measure, then there’s no established uniqueness result for general f . Recently L. Liehr and P. Grohs grohs2024phaseless showed that when g is Gaussian, uniqueness is guaranteed for general f\in L^{2}(\mathbb{R}) if the sampling set \Lambda is a square-root lattice, i.e., \Lambda=\{A(\pm\sqrt{n},\pm\sqrt{n}),n\in\mathbb{N}\} for some A\in GL(2,\mathbb{R}) . In what follows, we investigate how a weaker decay property of window function g affects the injectivity result and determine the corresponding sampling set. More specifically, our main result can be stated as follows. Theorem 1.1 Let g\in L^{2}(\mathbb{R}) with Fourier transform satisfies the super-exponential decay condition: |\hat{g}(\xi)|\leq Ce^{-a|\xi|^{m}} for some constants m>1 , C>0 , a\in(1,\infty) . Let \tau_{1},\tau_{2}\in\mathbb{R}^{+} satisfy \tau_{1}<\big(\frac{2}{(2\pi)^{m/(m-1)}(ma)^{-1/(m-1)}e}\big)^{\frac{m-1}{m}},\quad\quad\quad\tau_{2}<(\frac{2}{ame})^{\frac{1}{m}}. Define the set \Lambda such that: \Lambda=\{(\pm\tau_{1}n^{\frac{m-1}{m}},\pm\tau_{2}n^{\frac{1}{m}}),n\in\mathbb{N}\} Then the following statements are equivalent for every f,h\in L^{2}(\mathbb{R}): 1. |V_{g}(f)(\lambda)|=|V_{g}h(\lambda)| for every \lambda\in\Lambda . 2. f=e^{i\alpha}h for some \alpha\in[0,2\pi). Figure 1: An illustration of sampling points \Lambda in time-frequency plane for g with Fourier decay parameter m=\frac{3}{2} and a=1 The rest of the paper is organized as follows: In Section 2 we introduce some necessary preliminaries. Subsequently we prove Theorem 1.1 in Section 3. To this end, we first establish a sufficient condition for a discrete set to be a uniqueness set for the relevant class of entire functions (Proposition 1) and use Proposition 2 to show that our result is in a sense ”sharp”. We then use these technical results to prove Theorem 1.1 , which provides a concrete discrete sampling set \Lambda that guarantees the unique recovery of f up to a global phase. 2 Preliminary A central principle in Fourier analysis is the connection between the smoothness of a function and the decay rate of its Fourier transform at infinity. The Paley-Wiener theorem and its variants are the primary tools that precisely characterize this relationship, providing the foundation for extending functions into the complex domain based on the behavior of their Fourier transforms. 2.1 Paley-Wiener Type Theorem The Fourier transform of a Lebesgue integrable function f on \mathbb{R} is given by \mathcal{F}(f):=\hat{f}(\xi)=\int_{\mathbb{R}}f(t)e^{-2\pi it\cdot\xi}dt,\quad\xi\in\mathbb{R}. A standard density argument can be used to extend this definition so that \mathcal{F} is an isometry on L^{2}(\mathbb{R}) . The Fourier transform of a function f\in L^{2}(\mathbb{R}) is said to be of super-exponential decay if |\hat{f}(\xi)|\lesssim Ce^{-a|\xi|^{m}} (1) for some m>1 , C>0 and a\in(1,\infty) . In this subsection we introduce a slightly modified version of Paley-Wiener theorem. In particular, we show that if the Fourier transform of a function f decays super-exponential, then f can be extended to an entire function whose order and type depend on the parameters a and m . To establish this result, we make use of Morera’s Theorem, which we now recall. Theorem A ( stein2010complex , Theorem 5.2) If \{f_{n}\}_{n\in\mathbb{N}} is a sequence of holomorphic functions that converges uniformly to a function f in every compact subset of \Omega , then f is holomorphic in \Omega . The next result shows that if f\in L^{2}(\mathbb{R}) satisfies ( 1 ), then f is entire. Lemma 1 If a function f\in L^{2}(\mathbb{R}) has Fourier transform of super-exponential decay, then f can be extended analytically to an entire function. For completeness, we include the proof of this result which is based on arguments found in ( stein1971introduction , Chapter 3, Theorem 4.1) and in ( stein2010complex , Chapter 4, Problem 1) . Proof Let f\in L^{2}(\mathbb{R}) satisfy ( 1 ) for some m>1,C>0, and a\in(1,\infty) . The \hat{f}\in L^{2}(\mathbb{R}) from which we conclude that \hat{f}\in L^{1}[-n,n] for each n\in\mathbb{N} . For each n\in\mathbb{N} and z=x+iy\in\mathbb{C} , define F_{n}(z)=F_{n}(x+iy):=\int_{-n}^{n}\hat{f}(\xi)e^{2\pi i\xi z}d\xi. \{F_{n}(z)\}_{n\in\mathbb{N}} forms a sequence of well-defined entire functions. Next, for z=x+iy\in\mathbb{C} , observe that by ( 1 ) we have that |\hat{f}(\xi)e^{2\pi iz\xi}|=|\hat{f}(\xi)e^{2\pi i(x+iy)\xi}|\leq C^{\prime}e^{-a|\xi|^{m}}e^{-2\pi y|\xi|}. Let K\subset\mathbb{C} be a compact set. For each z=x+iy\in K , F(z)=F(x+iy)=\int_{\mathbb{R}}\hat{f}(\xi)e^{2\pi i(x+iy)\xi}d\xi (2) is well defined. Moreover, for all z\in K , we have |F_{n}(z)-F(z)|\leq C^{\prime}\int_{|\xi|>n}e^{-a|\xi|^{m}}e^{-2\pi y\xi}d\xi\leq C^{\prime}\int_{|\xi|>n}e^{-a|\xi|^{m}}e^{-2\pi R\xi}d\xi. But the last expression tends to 0 as n approaches \infty . Hence the sequence \{F_{n}(z)\} converges uniformly to F(z) for z\in K . Because of K is arbitrary, F(z) is entire by Morera’s Theorem. In addition, \lim_{y\to 0}F(x+iy)=\int_{\mathbb{R}}\hat{f}(\xi)e^{2\pi ixix}d\xi=f(x) in the L^{2} sense. Remark 1 For m=1 , one can show that f can be analytically extended to \mathbb{C} up to a strip \mathcal{S}_{a} consisting of complex numbers whose imaginary part is between a and -a , but not necessarily entire. The following lemma is a corollary of the Paley-Wiener Theorem from the relation between L^{2}- integrability on the real line and entire functions: Corollary 1 Assume f(z) is an entire function satisfying |f(z)|\leq Ce^{A|z|^{\rho}} for some A,C,\rho>0 and the restriction of f to the real line is square-integrable. Then |f(x+iy)|\leq C^{\prime}e^{-a|x|^{\rho}+b|y|^{\rho}} for some a,b,C^{\prime}>0 . Proof Notice that since f(z) is entire,

Read the complete text