Reshaped Wirtinger Flow (RWF)

• The paper shows that RWF replaces the quartic magnitude loss with an amplitude-based loss, enabling exact signal recovery with linear sample complexity.
• Reshaped Wirtinger Flow is defined by a gradient-like iteration combined with a truncated spectral initializer that achieves geometric convergence and stability under noise.
• RWF offers lower computational complexity than standard WF, supporting efficient stochastic, minibatch, and Kaczmarz-type variants in practical phase retrieval.

Reshaped Wirtinger Flow (RWF) is a gradient-like method for phase retrieval, the problem of recovering an unknown signal $x\in\mathbb{R}^n$ or $\mathbb{C}^n$ from phaseless magnitude measurements $y_i=|\langle a_i,x\rangle|$, $i=1,\dots,m$. It was introduced as a nonconvex, nonsmooth alternative to standard Wirtinger Flow (WF), replacing WF’s quartic magnitude-fitting objective by an amplitude-based loss of lower algebraic order. Under i.i.d. Gaussian measurements $a_i\sim N(0,I_n)$, the original analysis establishes exact recovery with $m=O(n)$ samples, geometric convergence from a spectral initializer, stability under bounded additive noise, and an incremental variant with linear convergence in expectation (Zhang et al., 2016).

1. Problem formulation and identifiability

RWF is formulated for quadratic systems arising in phase retrieval, where the observations are

$y_i=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$

Because only amplitudes are observed, the signal is identifiable only up to a global phase factor: if $x$ is feasible, then so is $xe^{j\phi}$. In the real case, which is the main setting for the original theory, this reduces to the sign ambiguity $\pm x$. The recovery error is therefore measured through

$\mathbb{C}^n$0

or, in the real case,

$\mathbb{C}^n$1

The measurement model used in the main theoretical analysis is Gaussian,

$\mathbb{C}^n$2

and the central claim is that exact recovery is possible with only $\mathbb{C}^n$3 measurements. This places RWF in the line of nonconvex phase retrieval methods that seek sample complexity linear in the ambient dimension while avoiding convex lifting formulations (Zhang et al., 2016).

2. Reshaped amplitude loss and gradient-like iteration

The defining feature of RWF is the replacement of the standard WF quartic loss

$\mathbb{C}^n$4

by the amplitude loss

$\mathbb{C}^n$5

This is the “reshaping” in Reshaped Wirtinger Flow: instead of matching $\mathbb{C}^n$6 to $\mathbb{C}^n$7, it matches $\mathbb{C}^n$8 directly to $\mathbb{C}^n$9. The resulting objective is nonconvex and nonsmooth because of the absolute value at $y_i=|\langle a_i,x\rangle|$0, but it depends only on the second power of the variable rather than the fourth. The paper’s main intuition is that, near the solution, the residual

$y_i=|\langle a_i,x\rangle|$1

behaves similarly to the linear residual one would use if the unknown signs were available, giving the loss favorable local geometry.

At points where $y_i=|\langle a_i,x\rangle|$2, the gradient is

$y_i=|\langle a_i,x\rangle|$3

with the convention

$y_i=|\langle a_i,x\rangle|$4

At nonsmooth points, the corresponding term is set to zero; the paper justifies this through the Fréchet superdifferential viewpoint.

The batch RWF iteration is

$y_i=|\langle a_i,x\rangle|$5

For the real Gaussian case, the recommended parameters are

$y_i=|\langle a_i,x\rangle|$6

and for complex Gaussian cases the paper uses $y_i=|\langle a_i,x\rangle|$7. A recurrent misconception is that RWF is simply WF without truncation; more precisely, RWF eliminates truncation in the gradient loop, but its initializer remains explicitly truncated from both sides (Zhang et al., 2016).

3. Spectral initialization and the role of two-sided truncation

RWF uses a spectral initializer built from amplitude measurements rather than squared magnitudes. It first estimates the norm through

$y_i=|\langle a_i,x\rangle|$8

and then forms

$y_i=|\langle a_i,x\rangle|$9

The initialization is

$i=1,\dots,m$0

where $i=1,\dots,m$1 is the leading eigenvector of $i=1,\dots,m$2.

The truncation is two-sided: the initializer discards both unusually small and unusually large amplitudes. The paper emphasizes that this is distinctive. Because the measurements are $i=1,\dots,m$3 rather than $i=1,\dots,m$4, small amplitudes can still distort the top eigenvector, so low-amplitude samples must also be removed. This is one of the main structural differences between RWF initialization and spectral initializers based only on squared observations.

The initialization guarantee is stated as follows: for any fixed $i=1,\dots,m$5, the initializer satisfies

$i=1,\dots,m$6

with probability at least

$i=1,\dots,m$7

provided $i=1,\dots,m$8, for suitable constants $i=1,\dots,m$9. Thus the initializer reaches the basin of attraction with linear sample complexity in the dimension (Zhang et al., 2016).

4. Local regularity, geometric convergence, and noise stability

The central theoretical result for batch RWF is geometric convergence under Gaussian measurements. The theorem states that there exist universal constants $a_i\sim N(0,I_n)$0 (practically $a_i\sim N(0,I_n)$1), $a_i\sim N(0,I_n)$2, and $a_i\sim N(0,I_n)$3 such that if

$a_i\sim N(0,I_n)$4

then, with probability at least

$a_i\sim N(0,I_n)$5

the iterates obey

$a_i\sim N(0,I_n)$6

This gives exact recovery from a proper initializer with $a_i\sim N(0,I_n)$7 samples and a linear rate in the logarithmic error scale.

The proof is based on a regularity condition of the form

$a_i\sim N(0,I_n)$8

shown to hold with high probability in a local neighborhood $a_i\sim N(0,I_n)$9. The analysis further derives the quantitative estimates

$m=O(n)$0

and

$m=O(n)$1

These bounds produce an admissible step-size range. The proof gives a conservative upper bound around $m=O(n)$2, while the reported practical choice is $m=O(n)$3.

The original paper also treats bounded additive noise in the amplitude domain: $m=O(n)$4 Under the same Gaussian sampling assumptions and $m=O(n)$5, RWF contracts until it reaches a noise floor proportional to $m=O(n)$6, with

$m=O(n)$7

up to constants. A common misunderstanding is that the nonsmooth amplitude loss precludes sharp convergence theory; the original analysis shows instead that the nonsmoothness is manageable in both proof and implementation because the iterates almost never land exactly at the singular set under Gaussian measurements (Zhang et al., 2016).

5. Incremental RWF, minibatching, and the Kaczmarz connection

The stochastic counterpart of RWF is Incremental Reshaped Wirtinger Flow (IRWF). Using the same initialization as batch RWF, IRWF updates one randomly chosen sample at each step: $m=O(n)$8 where $m=O(n)$9 is sampled uniformly from $y_i=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$0.

Its convergence theorem states that if

$y_i=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$1

then, with probability at least

$y_i=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$2

IRWF satisfies

$y_i=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$3

Thus IRWF has linear convergence in expectation with a fixed step size of order $y_i=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$4; the paper suggests $y_i=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$5 in practice.

A minibatch version is also given: $y_i=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$6 where $y_i=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$7 is a random minibatch of size $y_i=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$8. The corresponding guarantee is

$y_i=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$9

An important structural observation is that randomized Kaczmarz for phase retrieval is a sample-dependent-step-size instance of IRWF: $x$0 Since Gaussian vectors satisfy $x$1, this is close to IRWF with $x$2. Using the IRWF analysis, the paper derives the convergence guarantee

$x$3

for iterates in the basin of attraction, thereby supplying a linear convergence theorem for a Kaczmarz-type phase retrieval method that had previously been supported mainly empirically (Zhang et al., 2016).

6. Computational profile, empirical behavior, and subsequent developments

RWF was introduced partly to reduce the computational burden of WF. Because its gradient uses only the second power of the inner product rather than the fourth power implicit in the quartic WF loss, the paper states a total complexity of

$x$4

to reach $x$5-accuracy, compared with WF’s

$x$6

The method also avoids the truncation checks required by TWF during the gradient iterations. Empirically, the reported findings are that RWF beats WF and is often competitive with or better than TWF, especially for complex Gaussian and coded diffraction pattern settings; IRWF is described as the best overall in sample efficiency and often the fastest among stochastic methods; RWF is typically about $x$7 faster than TWF and $x$8–$x$9 faster than WF in time and iteration count; and RWF is robust to Poisson noise, outperforming TWF in noisy imaging experiments (Zhang et al., 2016).

Later work extended the theoretical picture beyond spectral initialization. A 2025 analysis studies RWF with random initialization

$xe^{j\phi}$0

under Gaussian measurements and proves that, when

$xe^{j\phi}$1

RWF with random initialization reaches $xe^{j\phi}$2-accuracy within

$xe^{j\phi}$3

iterations. The proof uses a resampling strategy during an initial phase, establishes a two-phase dynamics with $xe^{j\phi}$4, and shows that convergence remains stable even with larger step sizes; in the paper’s discussion, this removes the need for spectral initialization and yields total complexity $xe^{j\phi}$5 rather than an alternative involving principal-eigenvector computation (Li et al., 21 Jul 2025).

Related reshaping-style algorithms help delimit what is specific to RWF. “Tanh Wirtinger Flow” is conceptually similar in seeking a better-conditioned nonconvex landscape, but it is not RWF: it uses a likelihood-derived tanh/log-cosh weighting and a tanh-weighted spectral initializer rather than RWF’s amplitude loss $xe^{j\phi}$6. The comparison there is primarily conceptual and experimental rather than a formal reduction from one method to the other (Luo et al., 2019).

RWF therefore occupies a specific place in phase retrieval methodology: it retains the nonconvex first-order character of WF, matches the $xe^{j\phi}$7 Gaussian sample complexity associated with TWF, removes truncation from the gradient loop, admits stochastic and minibatch variants with linear convergence guarantees, and supports later analyses showing that even random initialization can suffice under stronger polylogarithmic sampling conditions.