Real Phase Retrieval Overview

• Real phase retrieval is the process of recovering real-valued signals from phaseless measurements, inherently accounting for a global sign ambiguity.
• Finite-dimensional uniqueness is secured by the complement property, ensuring injectivity at the sharp phase transition when the number of measurements meets or exceeds 2M-1.
• Almost injectivity reduces the minimal measurement requirement to M+1, while a range of algorithms—from convex relaxations to spectral methods—handle practical recovery in various domains.

Real phase retrieval is the problem of recovering a real-valued signal from magnitude-only or intensity-only measurements of linear transforms, with recovery understood modulo the unavoidable global sign ambiguity. In the finite-dimensional model emphasized in the modern uniqueness theory, one seeks $x\in\mathbb{R}^M$ from phaseless quadratic measurements

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$

with $\Phi=\{\varphi_n\}_{n=1}^N\subseteq\mathbb{R}^M$, so that $\mathcal A$ is naturally regarded as a map $\mathbb{R}^M/\{\pm1\}\to\mathbb{R}^N$ because $$|\langle x,\varphi_n\rangle|^2=|\langle -x,\varphi_n\rangle|^2$$ (Mixon, 2014). The same sign ambiguity appears in the magnitude model $y_i=|\langle a_i,x\rangle|$, in Fourier phase retrieval, in wavefront sensing, and in shift-invariant or bandlimited settings; what changes across these settings is the measurement operator, the relevant identifiability criterion, and the algorithmic strategy (Vaswani, 2020).

1. Measurement models, equivalence classes, and fundamental ambiguities

In the standard real finite-dimensional formulation, the sensing vectors are known and the unknown signal is real. The data may be written either as squared magnitudes,

$\mathcal{A}(x)(n)=|\langle x,\varphi_n\rangle|^2,$

or as magnitudes,

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

These formulations encode the same global ambiguity: $x$ and $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$0 generate identical measurements. Accordingly, the natural performance criterion is sign-invariant. One convenient distance is

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$1

which is the real specialization of the global-phase-invariant metric used in general phase retrieval (Vaswani, 2020).

In Fourier-based formulations, the forward map is typically the discrete or continuous Fourier transform, and the detector records intensity rather than complex field. A representative model is

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$2

where $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$3 is the complex field after interaction with a mask or object. In coherent diffraction imaging and related optical settings, the inverse problem is to infer the field from intensity-only diffraction data. In the real-signal setting, the same missing-phase mechanism persists, but the admissible ambiguity class collapses from global phase to sign when the object is known to be real-valued (White et al., 2021).

Fourier phase retrieval has additional trivial ambiguities beyond sign when no extra measurement diversity or structural prior is imposed. In one-dimensional Fourier-only formulations, the literature summarized in the overview paper identifies global phase, time-shift, and conjugate-flip as standard equivalences, and shows that oversampling alone does not make the problem well posed; by spectral factorization, the feasible set can contain up to $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$4 non-equivalent solutions in the 1D case (Jaganathan et al., 2015). For two-dimensional real signals, by contrast, generic uniqueness up to trivial ambiguities is far more common, and this 1D/2D gap is one of the defining structural facts of Fourier phase retrieval (Kogan et al., 2016).

A useful consequence is that “real phase retrieval” is not a single theorem but a family of inverse problems linked by one principle: the data constrain only magnitudes, and the correct recovery target is always an equivalence class. In finite-dimensional vector models that class is $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$5; in Fourier imaging it may also include shifts and reflections unless additional constraints, masks, support information, or other diversity are present. This suggests that identifiability questions must always be stated relative to the measurement architecture rather than in purely abstract terms.

2. Exact injectivity in finite dimensions

The central exact uniqueness criterion for real finite-dimensional phase retrieval is the complement property. For $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$6, one writes $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$7 for the statement that for every subset $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$8, either $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$9 or $\Phi=\{\varphi_n\}_{n=1}^N\subseteq\mathbb{R}^M$0 spans $\Phi=\{\varphi_n\}_{n=1}^N\subseteq\mathbb{R}^M$1. The key theorem attributed to Balan–Casazza–Edidin is

$\Phi=\{\varphi_n\}_{n=1}^N\subseteq\mathbb{R}^M$2

so injectivity modulo sign is exactly equivalent to a spanning/combinatorial condition (Mixon, 2014).

The constructive mechanism behind this equivalence is especially transparent in the real case. If the complement property fails, then there is a subset $\Phi=\{\varphi_n\}_{n=1}^N\subseteq\mathbb{R}^M$3 and nonzero vectors $\Phi=\{\varphi_n\}_{n=1}^N\subseteq\mathbb{R}^M$4 such that $\Phi=\{\varphi_n\}_{n=1}^N\subseteq\mathbb{R}^M$5 for $\Phi=\{\varphi_n\}_{n=1}^N\subseteq\mathbb{R}^M$6 and $\Phi=\{\varphi_n\}_{n=1}^N\subseteq\mathbb{R}^M$7 for $\Phi=\{\varphi_n\}_{n=1}^N\subseteq\mathbb{R}^M$8. It follows that

$\Phi=\{\varphi_n\}_{n=1}^N\subseteq\mathbb{R}^M$9

hence $\mathcal A$0 even though $\mathcal A$1. Conversely, if $\mathcal A$2 with $\mathcal A$3, one partitions the indices according to where the signs agree or disagree and deduces failure of spanning on one side or the other (Mixon, 2014). This is one of the cleanest exact characterizations in the phase-retrieval literature.

From this criterion one obtains the sharp real injectivity threshold. Since any spanning subset of $\mathcal A$4 must contain at least $\mathcal A$5 vectors, the complement property is impossible when $\mathcal A$6: choosing $\mathcal A$7 forces both $\mathcal A$8 and $\mathcal A$9 to have cardinality $\mathbb{R}^M/\{\pm1\}\to\mathbb{R}^N$0, so neither can span. Thus injective real phase retrieval cannot occur below $\mathbb{R}^M/\{\pm1\}\to\mathbb{R}^N$1 measurements. Conversely, if $\mathbb{R}^M/\{\pm1\}\to\mathbb{R}^N$2 and $\mathbb{R}^M/\{\pm1\}\to\mathbb{R}^N$3 is full spark, then $\mathbb{R}^M/\{\pm1\}\to\mathbb{R}^N$4 satisfies the complement property, where full spark means every $\mathbb{R}^M/\{\pm1\}\to\mathbb{R}^N$5 submatrix is invertible. This yields the exact phase transition statement

$\mathbb{R}^M/\{\pm1\}\to\mathbb{R}^N$6

At the extremal count $\mathbb{R}^M/\{\pm1\}\to\mathbb{R}^N$7, injectivity is equivalent to full sparkness (Mixon, 2014).

The set of full spark matrices is open and dense because the product of all $\mathbb{R}^M/\{\pm1\}\to\mathbb{R}^N$8 minors is a nonzero polynomial, and the Vandermonde matrix

$\mathbb{R}^M/\{\pm1\}\to\mathbb{R}^N$9

is a concrete full spark example (Mixon, 2014). This algebraic-geometric perspective explains generic injectivity above the threshold: the bad sets are determinantal varieties, and generic ensembles avoid them.

The contrast with the complex case is substantial. In complex phase retrieval the real complement property is not sufficient, real measurement vectors can fail to distinguish conjugate-related signals, and the chapter instead gives the characterization

$$|\langle x,\varphi_n\rangle|^2=|\langle -x,\varphi_n\rangle|^2$$0

or equivalently

$$|\langle x,\varphi_n\rangle|^2=|\langle -x,\varphi_n\rangle|^2$$1

The same source records the rank-1/rank-2 obstruction

$$|\langle x,\varphi_n\rangle|^2=|\langle -x,\varphi_n\rangle|^2$$2

and the conjectured complex transition at $$|\langle x,\varphi_n\rangle|^2=|\langle -x,\varphi_n\rangle|^2$$3. The real case therefore remains “much better understood” in the precise sense that exact injectivity is completely characterized by complement property and full sparkness, with a sharp threshold at $$|\langle x,\varphi_n\rangle|^2=|\langle -x,\varphi_n\rangle|^2$$4 (Mixon, 2014).

3. Almost injectivity and minimal redundancy

Real phase retrieval admits a second uniqueness notion, almost injectivity, which is weaker than exact injectivity and dramatically lowers the minimal measurement burden. In the real setting, almost injective means

$$|\langle x,\varphi_n\rangle|^2=|\langle -x,\varphi_n\rangle|^2$$5

for almost every $$|\langle x,\varphi_n\rangle|^2=|\langle -x,\varphi_n\rangle|^2$$6. This allows a lower-dimensional exceptional set of ambiguous signals while retaining generic uniqueness (Fickus et al., 2013).

The decisive characterization is a rank inequality. Assuming the measurement vectors are nonzero and span $$|\langle x,\varphi_n\rangle|^2=|\langle -x,\varphi_n\rangle|^2$$7,

$$|\langle x,\varphi_n\rangle|^2=|\langle -x,\varphi_n\rangle|^2$$8

for every nonempty proper subset $$|\langle x,\varphi_n\rangle|^2=|\langle -x,\varphi_n\rangle|^2$$9 (Mixon, 2014). The same paper derives this from a more geometric Minkowski-sum description involving

$y_i=|\langle a_i,x\rangle|$0

while the earlier “Phase retrieval from very few measurements” paper formulates the equivalent proper-subspace criterion and emphasizes that the real almost-injective theory is complete, unlike the complex almost-injective case, which is left open (Fickus et al., 2013).

This rank criterion yields an exact threshold: $y_i=|\langle a_i,x\rangle|$1 If $y_i=|\langle a_i,x\rangle|$2, almost injectivity is impossible. If $y_i=|\langle a_i,x\rangle|$3, almost injectivity is equivalent to full sparkness. More generally, every full spark real ensemble with $y_i=|\langle a_i,x\rangle|$4 is almost injective (Mixon, 2014). Thus the minimal redundancy for generic uniqueness is $y_i=|\langle a_i,x\rangle|$5, whereas exact injectivity requires $y_i=|\langle a_i,x\rangle|$6.

A particularly simple sufficient condition uses frame theory. If $y_i=|\langle a_i,x\rangle|$7 is a unit norm tight frame and $y_i=|\langle a_i,x\rangle|$8 and $y_i=|\langle a_i,x\rangle|$9 are relatively prime, then $\mathcal{A}(x)(n)=|\langle x,\varphi_n\rangle|^2,$0 is almost injective. The proof exploits

$\mathcal{A}(x)(n)=|\langle x,\varphi_n\rangle|^2,$1

and the conclusion is that the frame cannot be orthogonally partitionable; the chapter remarks that a UNTF is almost injective exactly when it is not orthogonally partitionable (Mixon, 2014). This provides a concrete sufficient condition beyond generic full spark constructions.

The reduced measurement count has a computational price. The NP-completeness theorem for full-spark $\mathcal{A}(x)(n)=|\langle x,\varphi_n\rangle|^2,$2-measurement families states that if $\mathcal{A}(x)(n)=|\langle x,\varphi_n\rangle|^2,$3 is a family of full spark ensembles with rational entries computable in polynomial time, then the decision problem

$\mathcal{A}(x)(n)=|\langle x,\varphi_n\rangle|^2,$4

is NP-complete (Fickus et al., 2013). The same source interprets this as evidence that polynomial-time phase retrieval at minimal almost-injective redundancy is unlikely unless $\mathcal{A}(x)(n)=|\langle x,\varphi_n\rangle|^2,$5, and therefore that algorithmic efficiency may require redundancy beyond the information-theoretic minimum.

A plausible implication is that the real theory separates into two notions of economy. The first is measurement economy, where $\mathcal{A}(x)(n)=|\langle x,\varphi_n\rangle|^2,$6 is optimal for almost injectivity. The second is computational economy, where more redundant ensembles—such as the explicit $\mathcal{A}(x)(n)=|\langle x,\varphi_n\rangle|^2,$7-measurement construction based on the standard basis and pairwise sums—can support straightforward recovery even though they are not minimally sparse in measurement count (Fickus et al., 2013).

4. Fourier phase retrieval, dimensional effects, and continuous-domain generalizations

Fourier phase retrieval is historically the canonical realization of the problem. In the discrete 1D formulation one seeks a length-$\mathcal{A}(x)(n)=|\langle x,\varphi_n\rangle|^2,$8 signal $\mathcal{A}(x)(n)=|\langle x,\varphi_n\rangle|^2,$9 from data

$y_i:=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$0

and, with oversampling, the measurements correspond to the Fourier transform of the autocorrelation

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

The review on recent developments emphasizes that in 1D, even with oversampling, the feasible set can contain up to $y_i:=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$2 non-equivalent solutions, while in dimensions two and higher, oversampled Fourier phase retrieval is generically unique up to trivial ambiguities except for a measure-zero set (Jaganathan et al., 2015).

The two-dimensional real case admits a particularly sharp reformulation. For a real $y_i:=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$3 matrix $y_i:=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$4, the measurements are

$y_i:=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$5

and, assuming sufficient oversampling $y_i:=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$6, these are equivalent to the 2D autocorrelation. The paper “On The 2D Phase Retrieval Problem” shows that after row-wise vectorization,

$y_i:=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$7

the 2D problem can be recast as a 1D phase retrieval problem plus $y_i:=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$8 extra constraints arising from the 2D geometry (Kogan et al., 2016).

The strongest theorem in that paper states that, for almost all real $y_i:=|\langle a_i,x\rangle|,\qquad i=1,\dots,m.$9 signals, one additional constraint is enough to eliminate the nontrivial 1D ambiguities. The specific constraint used is

$x$0

and the constrained problem

$x$1

has a unique solution for almost all $x$2 (Kogan et al., 2016). This identifies the mechanism by which higher-dimensional structure resolves the reciprocal-root ambiguities of 1D spectral factorization.

Real phase retrieval also extends beyond finite dimensions. In the shift-invariant setting

$x$3

the central structural notion is nonseparability. A real signal $x$4 is separable if $x$5 with $x$6 and $x$7; otherwise it is nonseparable. The main theorem is

$x$8

Thus phase retrievability on the whole line is exactly equivalent to nonseparability (Chen et al., 2016).

For spline signals in $x$9, the same paper proves recoverability from phaseless samples on a set $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$00 with sampling rate $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$01. More precisely, any nonseparable spline signal in $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$02 is determined up to sign from phaseless samples on $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$03, and the reconstruction remains stable under bounded noise via the MEPS algorithm, with error scaling like $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$04 in the main stability theorem (Chen et al., 2016). This places real phase retrieval in an infinite-dimensional functional-analytic setting while preserving the same global sign ambiguity seen in finite dimensions.

A related but distinct relaxation is conjugate phase retrieval. Real measurement vectors cannot perform ordinary complex phase retrieval on $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$05 because $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$06 for every real $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$07. The relaxed equivalence relation is

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$08

and generic real frames with at least $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$09 measurements are conjugate phase retrievable in $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$10 for $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$11 (Evans et al., 2017). In Paley–Wiener space, the analogous relaxation allows sampling-based conjugate phase retrieval at three times the Nyquist rate generically, whereas standard phase retrieval requires four times the Nyquist rate (Lai et al., 2019). These results do not change the real theory itself, but they clarify why real measurements are especially natural in the genuinely real setting: only sign ambiguity remains, rather than phase-or-conjugation ambiguity.

5. Algorithms, high-dimensional thresholds, and structured priors

Algorithmic work on real phase retrieval spans convex relaxation, nonconvex optimization, alternating projections, high-dimensional asymptotics, and structure-exploiting methods. A baseline convex strategy is PhaseLift, which lifts the unknown vector to the rank-one matrix

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$12

using

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$13

PhaseLift is provably correct with $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$14 Gaussian measurements, but it is computationally expensive because it lifts the problem to dimension $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$15 (Vaswani, 2020).

Nonconvex methods operate directly on the original objective. Spectral initialization uses

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$16

so its top eigenvector aligns with $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$17. This underlies AltMinPhase, Wirtinger Flow, and Truncated Wirtinger Flow. The latter achieves $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$18 independent Gaussian phaseless measurements and complexity $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$19 for $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$20-accurate recovery (Vaswani, 2020). The numerical review further emphasizes that success depends strongly on initialization and on favorable measurement geometry, often characterized locally through a spectral gap in the Hessian or Jacobian of the amplitude loss (Fannjiang et al., 2020).

Polytope-based formulations provide a particularly sharp real Gaussian theory. PhaseMax solves

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$21

where $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$22 is an anchor vector. Its recovery boundary is expressed through the input cosine similarity

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$23

and the paper derives the sharp asymptotic threshold

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$24

It then introduces PhaseLamp, an MM-based iterative scheme obtained by repeatedly solving PhaseMax subproblems, and proves a strictly weaker sufficient initialization condition $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$25 with $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$26 (Dhifallah et al., 2018). This yields an explicit algorithmic improvement within the real Gaussian model.

High-dimensional Bayesian analysis gives a different family of thresholds. For the real model

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$27

the weak-recovery threshold is

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$28

and the information-theoretic full-recovery threshold is

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$29

in the real case. For full-rank real sensing matrices this gives $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$30. The same paper shows that approximate message passing or G-VAMP can exhibit a statistical-to-algorithmic gap, so efficient algorithms need not achieve the information-theoretic limit (Maillard et al., 2020).

Structure can reduce sample complexity below the unstructured regime. In sparse phase retrieval, the unknown $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$31 is $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$32-sparse and the best nonconvex guarantees discussed in the survey require $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$33 Gaussian measurements. Methods such as AltMinSparse, SPARTA, Thresholded Wirtinger Flow, and CoPRAM all use variants of spectral initialization and support-aware refinement (Vaswani, 2020). For low-rank phase retrieval, where a set of signals $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$34 forms a low-rank matrix $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$35, AltMinLowRaP exploits the shared subspace and achieves per-signal measurement complexity

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$36

later improved to

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$37

under the paper’s incoherence assumptions (Vaswani, 2020).

Generative and implicit priors constitute a further shift from classical uniqueness theory toward underdetermined reconstruction. With a generator $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$38, the basic reconstruction problem becomes

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$39

“Optimizing Intermediate Representations of Generative Models for Phase Retrieval” proposes PRILO, a phase-retrieval adaptation of intermediate layer optimization that optimizes latent codes and intermediate activations under $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$40-ball constraints, together with magnitude-informed initialization and learned initialization. The paper reports that PRILO and especially PRILO-MII outperform ER and HIO on non-oversampled Fourier phase retrieval benchmarks, and that PRILO-MII improves DPR’s reconstruction performance for both real and complex Gaussian phase retrieval (Uelwer et al., 2022).

Training-free implicit priors have also been used. In the far-field setting, double deep image priors optimize

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$41

directly against Fourier magnitudes, with the paper arguing that FFPR is hard because of global phase, translation, and 2D conjugate-flipping symmetries (Zhuang et al., 2022). Alternating Phase Langevin Sampling combines alternating phase updates

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$42

with denoiser-driven Langevin sampling under the completed linear constraint $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$43, and reports competitive in-distribution performance together with improved out-of-distribution reconstruction on Fourier measurements (Agrawal et al., 2022). These methods are not specific to real signals, but their underlying principle—a strong structural prior to regularize a phaseless nonconvex inverse problem—applies directly to real phase retrieval as well.

6. Optical, wavefront, and physics-driven formulations

Real phase retrieval is also a physical inversion problem in optics, X-ray imaging, and wave propagation. In coherent diffractive wavefront sensing, a focused beam passes through a binary transmission mask and the detector records far-field intensity

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$44

A convolutional encoder–decoder can then learn the inverse map

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$45

predicting real and imaginary parts rather than amplitude and phase to avoid wrapping and discontinuities (White et al., 2021). In the reported XUV experiment at wavelength $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$46 nm with a $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$47 pinhole, the network achieved inference in about $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$48 ms on a Tesla V100 GPU, whereas iterative phase retrieval took several seconds, and the method remained usable on noisy diffraction patterns down to $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$49 peak counts where the iterative method failed badly (White et al., 2021). A plausible implication is that real-time wavefront sensing effectively repurposes phase retrieval from an offline inverse problem into a feedback component for adaptive optics.

A later XFEL-oriented deep method addresses imperfect diffraction patterns with beam-stop occlusion, detector gaps, partial coherence, and mixed Poisson–Gaussian noise. DPR learns the direct map from masked noisy diffraction to real-space object, using weighted partial convolution in the encoder and Fourier modulation/projection in the decoder. The reported speed is $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$50 ms per pattern on a single NVIDIA RTX 3090, or $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$51 ms per pattern with a short refinement stage, and the paper states that this is more than $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$52 faster than iterative phase-retrieval algorithms (Lee et al., 2024). The same work emphasizes that no support constraint is required at inference time, which is a major practical distinction from HIO- or GPS-based pipelines.

Transport-of-intensity and propagation-based formulations provide a different physical route. For X-ray phase-contrast imaging, the Transport of Intensity Equation is

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$53

For monomorphous objects with constant $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$54, one has

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$55

and therefore the monomorphous TIE

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$56

The monomorphous decomposition method extends this to arbitrary objects by decomposing the refractive index into a sum of two monomorphous components with bounded $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$57 ratios, using prior material knowledge to stabilize inversion at low spatial frequencies (Gureyev et al., 2015). This is a physically structured phase-retrieval method in which identifiability is strengthened not by random measurements but by constitutive constraints on the sample.

An even more explicit diversity principle appears in “Phase retrieval via media diversity.” There the incident field is $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$58, the data are multi-plane intensities after propagation through different media, and the propagation laws are linear, quadratic nonlinear, or Gross–Pitaevskii equations. In the linear case, if at least $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$59 media are chosen so that the matrix

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$60

is full rank pointwise, then equal intensity data imply

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$61

for some constant $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$62; thus the phase gradient is uniquely determined up to a constant phase shift (Cheng et al., 2024). In the quadratic nonlinear case, if $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$63 media are chosen so that the corresponding matrix

$$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$64

is full rank pointwise, then the result strengthens to $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$65, eliminating even the constant phase ambiguity (Cheng et al., 2024). The paper’s key claim is that media diversity supplies enough independent intensity information to avoid the prescribed boundary conditions required by standard TIE-based methods.

These physical formulations show that real phase retrieval is not restricted to abstract frame theory. It also includes real-time beam characterization, phase-contrast tomography, and PDE-constrained wave-field recovery. Across these settings the same conceptual dichotomy reappears: one either designs measurements to guarantee injectivity, or imposes structure—geometric, material, statistical, or learned—to select a unique and stable solution from a magnitude-only data set.

7. Open directions and conceptual summary

Several open directions are explicit or implicit across the literature. In the sharp finite-dimensional theory, the major unresolved practical issue is not uniqueness but certification: full sparkness is open and dense and yields exact injectivity at $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$66, yet testing full spark is NP-hard in general, so computationally efficient injectivity tests beyond complement property or brute-force minor checking remain desirable (Mixon, 2014). Closely related is the construction problem: minimal explicit injective ensembles are easy to state abstractly but harder to build with useful structure for applications.

Another persistent issue is robustness beyond exact injectivity. The exact threshold $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$67 and the almost-injective threshold $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$68 are uniqueness statements; they do not by themselves quantify stability under noise, model mismatch, or incomplete data. The optical and numerical literature addresses this gap by using redundancy, masks, overlap, support constraints, or learned priors, but a general real-valued robustness theory comparable in sharpness to the complement-property theorem is not isolated in the provided sources. This suggests that uniqueness theory and stable algorithm design remain partially decoupled subfields.

The Fourier setting continues to illustrate this distinction. In one dimension, uniqueness is poor without extra structure; in two dimensions, generic uniqueness is common; yet algorithmic success still depends on support accuracy, mask design, scan overlap, initialization, or prior modeling (Jaganathan et al., 2015, Fannjiang et al., 2020). In high dimensions, information-theoretic recovery can be possible at $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$69 for full-rank real random matrices while polynomial-time message-passing algorithms exhibit a statistical-to-algorithmic gap (Maillard et al., 2020). This suggests that “measurement threshold” and “algorithmic threshold” should be treated as distinct invariants of a real phase-retrieval model.

Taken together, the modern theory presents a remarkably stratified picture. At the level of exact finite-dimensional uniqueness, real phase retrieval is governed by the complement property and has a complete phase transition at $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$70 (Mixon, 2014). At the level of generic uniqueness, almost injectivity reduces the threshold to $$(\mathcal A(x))(n):=|\langle x,\varphi_n\rangle|^2,\qquad n=1,\dots,N,$$71 (Fickus et al., 2013). In Fourier and functional settings, nonseparability, dimensionality, masks, and additional geometric constraints determine identifiability (Chen et al., 2016, Kogan et al., 2016). In algorithmics, convex lifting, spectral initialization, polytope optimization, AMP-type methods, and structure-exploiting nonconvex solvers provide complementary regimes of tractability (Dhifallah et al., 2018, Vaswani, 2020, Maillard et al., 2020). In physical imaging, phase retrieval becomes a real-time or physics-constrained inference task in which diversity, propagation models, and learned priors compensate for missing phase (White et al., 2021, Lee et al., 2024, Cheng et al., 2024).

The most stable overarching interpretation is therefore exact. Real phase retrieval is the theory and practice of reconstructing real signals from magnitude-only data under sign ambiguity, with a uniquely clean finite-dimensional uniqueness theory, a markedly lower minimal threshold than in the complex case, and a broad algorithmic landscape in which structural priors and measurement design determine whether those theoretical guarantees can be realized in computation and experiment.