acceptodds
Under review as a conference paper at ICLR 2027

Complex Phase Retrieval via Conditional Quantiles: Finite-Sample Recovery without Noise Moments

Abstract

Phase retrieval becomes statistically fragile when response noise is heavy-tailed or asymmetric: mean-based objectives can be dominated by extreme observations or even lose their population interpretation, while median-based methods fix attention on a single response quantile. We instead formulate complex phase retrieval through a prescribed conditional quantile, and propose robust quantile Wirtinger flow (RQWF). RQWF combines a clipped spectral estimate of the signal direction, quantile-based radial calibration, and bounded-score Wirtinger refinement, so that initialization and optimization target the same response functional. Under complex Gaussian sensing, we prove end-to-end finite-sample recovery assuming only local regularity of the noise distribution around the target quantile, with no noise-moment requirement. The resulting recovery rate is , and a shifted-Cauchy submodel gives a matching dimension–sample lower bound up to logarithmic factors. The analysis also quantifies the bias introduced by smoothing and allows the same update sample to be reused throughout optimization. On a compute-matched Cauchy benchmark, RQWF-P lowers median recovery error relative to every tested complete SRPR/AdaIPL baseline in all 12 conditions and relative to Median-TWF in 10 of 12; prescribed non-median fits also improve selected asymmetric-noise cases. Student- gains depend on the comparator, while Gaussian and near-noiseless settings expose limits of the advantage. Together, the results show that conditional quantiles provide a tractable statistical target for complex phase retrieval precisely in regimes where moment-based formulations become unreliable.

Primary area:
learning theory
Keywords:
Phase retrieval, quantile regression, robust estimation, heavy-tailed noise, nonconvex optimization.

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_C7J9OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes robust quantile Writinger flow (RQWF) for complex phase retrieval without noise moments. RQWF consistts a clipped spectral initialization with quantile-based radial calibration and a bounded-score Writinger flow refinement. Under complex random Gaussian measurement vectors and local regularity assumptions on noise, an end-to-end finite-sample recovery guarantee is proved, with rates matching a Cauchy lower bound up to log factors. Experimental results show benefits that vary across different noise regimes.

Critical strengths
  1. The end-to-end finite-sample recovery guarantee without noise-moment assumptions. The theory is proved under complex random Gaussian measurements and it is assumed only the local noise-density regularity. The recovery rate matches a Cauchy lower bound up to log factors. It establishes statistically accurate, computationally attainable recovery in regime excluded by finite-moment theory. It is a substantive theoretial contribution.

  2. The technical contributions in the proofs, including a moment-free spectral gap (Lemma 13), 2nd-order radial calibration error (Lemma 14), uniform gradient control and sample reuse (Lemma 11), and sharp bias analysis (Proposition 5). These results also provide technical depth of this paper.

Critical weaknesses
  1. Insufficient application and justification for the quantile-noise model. The noise model is primarily a theoretical statistical model for recovery without noise-moment assumptions. The practical relevance of the prescribed non-median quantile remains insufficiently demonstrated. The optical experiment (Section 6) adds simulated noise to the recorded measurements.

  2. Gaps between the analyzed algorithm and the one with the strongest emprical advantage. For examples, (a) the analyzed RQWF-T uses separate samples for spectral estimation, radial calibration, and refinement, but the emprical RQWF-P reuses all data; (b) the clipping threshold is fixed beforehand in RQWF-T but estimated from data in RQWF-P.

Questions for authors
  1. Can the authors describe a concrete measurement setting where the true intensity satisfies Qτ(Y∣a)=∣a∗x⋆∣2Q_{\tau}(Y|a)=|a^*x_{\star}|^2, particularly for τ≠12\tau\neq \frac12? How would you prescribed τ\tau be determined or calibrated before reconstruction?

  2. How essential are sample splitting and fixed clipping to the recovery analysis? Can the authors clarify how the guarantees relate to RQWF-P's stronger empirical performance, given its full data reuse and data-dependent clippling?

Reviewer_Zr6VOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a robust phase retrieval method for recovering complex-valued signals from dense, noisy measurements without imposing any moment assumptions on the noise. The authors provide theoretical convergence guarantees with optimal sample complexity for this setting—specifically, linear convergence to the optimal statistical bias (up to logarithmic factors, matching a Cauchy lower bound)—and complement their theoretical findings with extensive numerical experiments.

Although I have several editorial and technical readability concerns detailed below, I believe these results are valuable to share with the phase retrieval and robust statistics communities. Therefore, I lean toward a weak accept.

Critical strengths
  • Establishes recovery and convergence guarantees for robust phase retrieval under two challenging conditions: (1) complex-valued signals, and (2) dense noise without any moment assumptions.

  • Achieves optimal sample complexity up to logarithmic factors.

  • Provides a detailed theoretical comparison with existing baseline methods.

  • Includes extensive numerical experiments with thorough descriptions of the experimental settings, comparing the proposed approach against several baselines on both synthetic and real datasets. Overall, the experimental results demonstrate the effectiveness of the proposed method over other baselines.

Critical weaknesses
  • While the high-level theoretical results appear sound, I was unable to verify many details of the proofs in the appendix—which is a notable concern given that the theoretical contributions are a central pillar of this paper. In particular, several proofs (especially in Appendices A.2 and A.3) are overly condensed and difficult to follow compared to standard proofs in the phase retrieval literature. For example:
  1. On Page 12, it is stated that "Symmetry of Kδ and the residual convention imply Eε[−ψτ,δ(ε − s)] = mδ(s). Consequently Gδ(z) = Ea[mδ(Δ)aa∗z], Gδ(x⋆) = βδx⋆, 0 ≤ fδ ≤ M, Lip(fδ) ≤ ℓ, fδ(0) ≥ κ − ℓδ. (21)" — The intermediate steps leading to these conclusions are omitted, making them hard to verify.
  2. On Page 13, "If f is C2 near zero, fix a small constant R > 0 such that f(u) − f(−u) = 2f′(0)u + O(u2) for 0 < u < R. The error integral is O(δ2) because ∫ R 0 u2Rδ(u) du = O(δ2). The tail above R is also O(δ2). Finally ∫ R 0 uRδ(u) du = δ2 log(1/δ)/4 + O(δ2)." -> What precisely is the "error integral" here? While a reader might be able to infer its definition from context, it should be explicitly defined for clarity and rigor.
  3. On Page 14, although the ϵ-net argument follows a relatively standard approach, it is far too compressed compared to similar derivations in the phase retrieval literature. It warrants much more expansion than a single dense paragraph: "Use an ϵ-net of the radius-R ball in R2n, ϵ = δ/(CmB2), of size at most (1 + 2R/ϵ)2n, and a 1/2-net of its unit sphere of size at most 52n. A union bound for the unconditioned scalar Bernstein inequalities on these deterministic nets, with a sufficiently large constant in (14), makes every net deviation at most C[ √(ΛU/m) + ΛU/m ] except with probability ζ/2. A sphere net controls the vector norm within factor two. Intersecting with the Gaussian norm event and extending the parameter net costs at most C/m, absorbed in tU. In particular, we have not conditioned a concentration inequality on a row-truncation event and then assumed the rows remain independent. This proves (24). The factor 2n from the real coordinates of Cn is absorbed in the universal constants."

Handling dense noise without moment assumptions is inherently delicate and requires greater care than other noise regimes (such as sparse adversarial outliers or heavy-tailed noise with finite moments) in the phase retrieval and robust statistics literature. Consequently, more detailed elaborations and expanded proofs are necessary to ensure readability and verifiability.

  • Another concern is the frequent inclusion of defensive scope qualifications or justifications that read as abrupt or out of context. For example:
  1. "The proof combines population phase-aligned curvature with a gradient deviation bound uniform on ∥z∥ ≤ 3/2; it does not require an empirical Hessian bound independent of δ" (Page 11).
  2. "Taking its positive square root and using |fδ(0) − f(0)| ≤ ℓδ proves the claim. No assertion about other stationary points is needed" (Page 13).
  3. "Thus a generic O(δ2) bias claim is false for this smoothing kernel." (Page 11).
  • Similarly, certain passages appear disconnected from the surrounding text. For example, on Page 22, the following paragraph appears without clear connection to the preceding or subsequent sections: "For a fixed, theory-aligned experimental interface, use proper CN(0, In) sensing (real and imaginary variances 1/2), unit-energy test signals, and a predeclared population-centered noise law: ε = σ[Z − Qτ (Z)], with fixed σ, not an empirical quantile or a realized-sample SNR normalization. Use fixed index splits, for example thirds, and the same deterministic c, δ, η across repetitions of a cell; v uses only D. Compute the radial solution on V , then exactly T ungated updates on U with response residual y − |a∗z|2. Choose admissible η using declared model regularity bounds (MR for the local-density result), not the true reconstruction error. Log numerical certificates and report any failures; do not discard them and reinterpret the theorem as a success probability conditional on passing. The conservative schedule (46) is a sufficient reference, not a recommended practical iteration count for every baseline. For equal-budget empirical comparisons, predeclare a different budget and retain the optimization term in (45)."

I strongly recommend that the authors thoroughly revise the appendix to include omitted intermediate proof steps and perform a careful editorial pass to address the above concerns.

Questions for authors
  1. On page 3, ρτ,δ(u) = (τ − 1/2)u + (1/2)√(u2 + δ2), δ > 0, (4) and ψτ,δ(u) = ρ′τ,δ(u) = τ − 1/2 + u / (2√(u2 + δ2)), |ψτ,δ(u)| ≤ max(τ, 1 − τ) (5): Could you please cite or state the origin of these smoothing functions in the quantile regression literature?

  2. What are the specific technical challenges of analyzing complex-valued phase retrieval compared to the real-valued setting in your framework?

  3. What additional challenges would arise in generalizing the results and techniques of this paper to more general measurement ensembles beyond Gaussian measurements? (Note: I am not pointing this out as a weakness, just asking a question.)

Reviewer_pH9HOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

To recover complex signals from intensity measurements with heavy-tailed noise, the authors propose RQWF, a method based on quantile regression. It uses spectral initialization with clipped measurements, fits the signal norm on a separate sample by solving a scalar convex optimization problem, and then refines the estimate using a smoothed quantile loss with a Wirtinger flow algorithm. For Gaussian measurements, the paper proves an error bound of order nlog⁡N/N\sqrt{n\log N/N} without requiring the noise to have finite moments. The analysis needs only local regularity of the noise density near the target quantile. A lower bound for Cauchy noise shows that the sample complexity is tight up to logarithms. Experiments use both simulated data and recorded optical measurements. Compared with the theoretical version, practical RQWF reuses all data and uses empirical clipping, sample filtering, and bandwidth annealing. In the Cauchy benchmark, it achieves lower median recovery error than baseline algorithms in most of the test cases.

Critical strengths
  1. Theorem 2 gives a recovery rate of order nlog⁡N/N\sqrt{n\log N/N} under complex Gaussian sensing and independent noise whose density is regular only near the prescribed quantile. This mild regularity assumption permits noise distributions for which even the mean does not exist. The analysis gives both a sample complexity bound and an upper bound on the recovery error. The Cauchy lower bound in Theorem 3 matches the dependence on dimension and sample size up to logarithms at fixed noise scale and quantile. The proofs are a bit long, but they appear rigorous to me.

  2. In the synthetic experiments with Cauchy noise, RQWF-P has lower median recovery error than every tested complex implementation of SRPR and AdaIPL in all settings, and than Median-TWF in most settings. These results give concrete evidence that RQWF-P can improve recovery over existing robust baselines in the heavy-tailed setting motivating the paper.

Critical weaknesses

The largest issue for me is the theory-practice gap between RQWF-T and RQWF-P. Theorem 2 analyzes RQWF-T with independent data splits, fixed tuning, and ungated updates. RQWF-P gives the strongest experimental results, but it reuses all the data, changes the radius estimate and clipping rule, adds filtering, and adjusts the bandwidth during optimization. The difference in performance is substantial. For example, at n=256n=256, N/n=8N/n=8, and Cauchy scale 0.05, the median recovery error is 0.392 for RQWF-T and 0.0516 for RQWF-P. Even the step size used in the RQWF-T experiments is not certified to satisfy the theorem’s conditions. The authors acknowledge these differences, but I still find it hard to see how much the theory explains the gains of the practical version.

Questions for authors
  1. Reusing all observations improves the median error in all settings in Table 2 and Figures 3&4 suggest that this change alone retains much of RQWF-P's performance under Cauchy noise. Does the existing analysis extend to this simpler version, while keeping the other choices consistent with Theorem 2? In particular, could the uniform gradient bound in Lemma 11 control the radial fit when the spectral direction is estimated from the same observations, or is there a specific step where independent samples remain essential?

  2. How much of the performance gap between RQWF-T and RQWF-P is already present after initialization, and how much develops during refinement? Appendix D.3 (below Figure 8) shows that RQWF-T may improve with more iterations, but this can due to starting from a poorer estimate or making slower progress during optimization. Could the authors compare the errors immediately after initialization and their subsequent decrease under the reported computational budget? This would help explain the gap and its connection to the local recovery guarantee.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.