{"id":"35710186-f653-403c-846a-16d0d259288d","arxiv_id":"2501.03549","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper frames multi-reference alignment and cryo-EM as generalized phase retrieval over compact groups and conjectures a bi-Lipschitz stability condition for recovery from the second moment.","lead":"This paper introduces a general framework for recovering signals from quadratic measurements, extending classical phase retrieval to any compact group symmetry and covering cryo-EM and multi-reference alignment. It surveys uniqueness results for semialgebraic priors, adapts projection algorithms, and proposes a conjecture on when recovery is stable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corollary III.6's effective-dimension estimate K≈L^2(R+2L/3) has a sign error: from Eq. (III.1) and the cryo-EM decomposition, K=dim V−k(H)≈L^2(R−2L/3), so the stated uniqueness threshold is too optimistic.","rationale":"The reader's weakest assumption was GL(V)-genericity. That is a legitimate concern, but it is an explicitly stated condition of Theorem III.1, so the paper is internally consistent on this point. The more decisive issue is the concrete sign error in Corollary III.6's effective-dimension formula. Since K is defined in the same paper as dim V−k(H), the printed K≈L^2(R+2L/3) is inconsistent with the authors' own decomposition and orbit-dimension computation: the L^3 term must be negative. This does not invalidate the general framework or the earlier theorems, but it does mean the cryo-EM uniqueness threshold as stated is not supported by the definitions. Because the error is localized and easily fixable, it does not change the overall conditional verdict; the paper should correct the formula and restate the resulting threshold before acceptance.","tokens_in":8068,"tokens_out":21879,"duration_ms":225686,"concrete_test":"Recompute Corollary III.6 for L=10, R=21 (so R=2L+1). Exact values: dim V=21·11^2=2541, k(H)=Σ_{ℓ=0}^{10} ℓ(2ℓ+1)=825, hence K=1716. The printed formula gives L^2(R+2L/3)=100·(21+20/3)=2766.7. If K is instead computed from Eq. (III.1), the plus sign must be a minus sign; this exact recomputation settles whether the printed threshold is a typo.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper defines K=dim V−k(H) in Eq. (III.1). In the cryo-EM model of §II-B, V=⊕_{ℓ=0}^L V_ℓ^{⊕R} with dim V_ℓ=2ℓ+1, so dim V=R(L+1)^2≈RL^2. When R≥2L+1, the generic H-orbit under H=∏_{ℓ=0}^L O(2ℓ+1) has dimension k(H)=Σ_{ℓ=0}^L ℓ(2ℓ+1)≈2L^3/3 (each O(2ℓ+1) acts freely on the full-rank X_ℓ). Thus K=dim V−k(H)≈L^2(R−2L/3), not L^2(R+2L/3) as printed in Corollary III.6. The sign error overstates the effective dimension by about 4L^3/3 in the regime R≥2L+1, meaning the corollary claims uniqueness for priors of dimension up to roughly twice the actual generic threshold. This is an internal inconsistency with the paper's own definitions, not a matter of scientific opinion, and it weakens the cryo-EM application unless the printed formula is corrected. The GL(V)-genericity issue raised by the reader is also real, but this numeric error is definite and checkable.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a generalized phase retrieval problem over compact groups, in which the goal is to recover a signal x in a finite-dimensional representation V of a compact group G from the second moment of the multi-reference alignment model. The second moment is shown, following earlier work of the authors, to determine the tuple of Gram matrices X_ℓ^* X_ℓ up to the action of an ambiguity group H = ∏ U(N_ℓ). The paper then states a transversality theorem for semialgebraic priors (Theorem III.1) and derives uniqueness-up-to-sign guarantees under a dimension inequality K > M, with specializations to classical phase retrieval and to a cryo-EM model. It also describes projection-based algorithms borrowed from [14], presents numerical experiments on a linear-prior example, and proposes a bi-Lipschitz stability conjecture (Conjecture V.1). The exposition is clear, but the main theoretical results are surveyed rather than proved in this manuscript, and one of the stated corollaries contains an internal inconsistency that affects its conclusion.","tokens_in":8346,"tokens_out":6357,"duration_ms":61748,"significance":"If the stated results are taken at face value, the paper offers a useful unifying algebraic framework for phase retrieval, MRA, and cryo-EM, and the effective-dimension criterion K = dim V − k(H) provides a clean heuristic for when semialgebraic priors remove the ambiguity group. The transversality-based approach is elegant and the specialization to cryo-EM is timely. However, the novelty of the present manuscript is mostly expository: Theorem II.1 is attributed to [12], Theorem III.1 to [10], and the algorithms to [14]. The paper's original contributions are the unified presentation, the numerical experiments, and the bi-Lipschitz conjecture. The experiments are reproducible in principle but lack error bars and code, and the conjecture is supported only by a narrow linear-prior example. The corrected effective dimension for cryo-EM changes the quantitative uniqueness threshold, so the paper needs revision before its claims can be relied upon.","major_comments":[{"comment":"The effective-dimension formula in Corollary III.6 contains a sign error that is internal to the paper's own definitions. Equation (III.1) defines K = dim V − k(H). For the cryo-EM representation V = ⊕_{ℓ=0}^L V_ℓ^{⊕R}, we have dim V = R(L+1)^2 ≈ R L^2. With R ≥ 2L+1, the generic H-orbit under H = ∏_{ℓ=0}^L O(2ℓ+1) has dimension k(H) = Σ_{ℓ=0}^L ℓ(2ℓ+1) ≈ 2L^3/3, because each O(2ℓ+1) acts freely on a full-rank (2ℓ+1)×R matrix. Hence K ≈ L^2(R − 2L/3), not L^2(R + 2L/3) as printed. The printed expression overstates K by about 4L^3/3, so the condition K > M is claimed for priors of dimension up to roughly twice the actual generic threshold. This is a load-bearing error in the cryo-EM uniqueness statement and must be corrected, together with any downstream discussion of the threshold.","section":"§III-B, Corollary III.6"},{"comment":"The uniqueness statements are conditional on a GL(V)-genericity hypothesis that is not carried through to the applications. Theorem III.1 states the transversality result for a GL(V)-generic translate A(M) of a semialgebraic set M, but Corollary III.4 is phrased as if any semialgebraic set of dimension M satisfying K > M is sufficient. For the motivating priors—exact sparsity with known support, or a fixed deep generative model—there is no proof that the set is GL(V)-generic, and a generic linear translate is not the same as the original prior. The paper should either prove the genericity condition for concrete instances of sparsity and generative-model priors, or explicitly state in Corollaries III.4–III.6 that the guarantee applies only after a generic linear translate of the prior. Without this qualification, the practical claim that a fixed prior ensures uniqueness is not established.","section":"Theorem III.1 and Corollary III.4"}],"minor_comments":[{"comment":"The proof of Theorem III.1 is only sketched, and the text refers to [10] for the detailed formulation. Since the paper's abstract describes itself as a survey, this is acceptable, but the introduction should state explicitly that the main theorems are surveyed from prior work, and the present manuscript's original contribution should be clearly delineated.","section":"Theorem III.1"},{"comment":"The numerical experiments report median errors over 10,000 runs but provide no error bars, no confidence intervals, and no code or detailed parameter settings beyond the matrix size and noise level. Adding these would substantially strengthen the empirical support for Conjecture V.1.","section":"Numerical experiments, §IV"},{"comment":"The notation is inconsistent between the complex and real settings: Corollary II.2 uses H = ∏ U(N_ℓ), while the cryo-EM discussion uses H = ∏ O(2ℓ+1). Please clarify that the unitary group is used for complex representations and the orthogonal group for real representations.","section":"Corollary II.2 and §II-B"},{"comment":"The target space in Conjecture V.1 is written as R^N = ∏_{i=1}^L R^{N_ℓ×R_ℓ}, which conflates the total dimension with the product of matrix spaces. The intended meaning is a direct sum of matrix spaces, so the notation should be cleaned up to avoid dimension confusion.","section":"Conjecture V.1"},{"comment":"The text says that a precise statement of Corollary III.6 is provided in [10]. After correcting the sign error, it would be helpful to include the precise statement in the main text as well, so the corollary is self-contained.","section":"Corollary III.6"}],"recommendation":"major_revision","confidential_remarks":"The paper is essentially a survey of the authors' own recent results, and the self-citation burden is heavy; the journal should confirm that this format is appropriate. The sign error in Corollary III.6 is definite and must be fixed; the genericity qualification in Theorem III.1 is also a substantive limitation that the current corollaries gloss over. If the authors address these two points and add the missing experimental details, the paper could be acceptable as an expository/position paper, but in its current form the quantitative claims overstate what is proven."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is best read as a well-organized survey of Bendory and Edidin's program on second-moment recovery in MRA and cryo-EM, recast as generalized phase retrieval over compact groups. The representation-theoretic framing is clear, and the paper does a good job explaining why the second moment reduces the ambiguity to a product of unitary/orthogonal groups. The genuinely new items are Conjecture V.1 on bi-Lipschitz stability and the numerical experiments that support it; those are worth taking seriously, even though the experiments are small and no code is provided.\n\nThe main theoretical results are restatements from the authors' prior papers (Theorem II.1 from [12], Theorem III.1 from [10], the algorithm from [14]). That is fine for a survey, but the paper should be more explicit about what is new versus expository.\n\nThe soft spots are real. First, there is a definite sign error in Corollary III.6. The paper claims K ≈ L^2(R + 2L/3). From the definitions, dim V = R(L+1)^2 and k(H) = Σ_{ℓ=0}^L ℓ(2ℓ+1) ≈ 2L^3/3, so K = dim V − k(H) ≈ L^2(R − 2L/3). The printed plus sign overstates the effective dimension by roughly 4L^3/3, which means the corollary claims uniqueness for priors of dimension up to about twice the actual generic threshold. This is an internal inconsistency with the paper's own definition (III.1), and it weakens the cryo-EM application until corrected.\n\nSecond, the GL(V)-genericity assumption in Corollary III.4 is a genuine caveat. The paper motivates sparsity and generative models as semialgebraic priors, but the theorem only applies to a generic linear translate of the prior, not to a fixed natural prior. Whether exact sparsity with known support or a specific trained generative model satisfies the genericity condition is unproven. This is not a fatal flaw, but it should be stated more prominently.\n\nThe numerical experiments are honest but minimal: median errors over 10,000 trials, no error bars, no code. They support the conjecture but do not prove it.\n\nOverall, this is a competent synthesis with one new conjecture and one clear numerical error. I would send it to peer review, but the sign error must be fixed and the authors should clarify the scope of the genericity assumption. The paper will be useful to researchers working on MRA or cryo-EM who want a compact introduction to the second-moment framework.","headline":"Useful survey of the authors' own second-moment program, but the genuinely new content is a conjecture plus numerics, and Corollary III.6 contains a definite sign error that overstates the cryo-EM uniqueness threshold.","tokens_in":8889,"tokens_out":2399,"would_cite":false,"duration_ms":22537,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["22C05","14P10","94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"Phase retrieval is a special case of recovering matrices from their Gram matrices; the second moment fixes signals up to a product of unitary groups, and low-dimensional semialgebraic priors make that recovery unique up to sign.","keywords":["phase retrieval","multi-reference alignment","cryo-electron microscopy","compact groups","second moment","semialgebraic priors","Gram matrices","bi-Lipschitz stability"],"falsifier":"Take a fixed semialgebraic prior $M$ of dimension $m<K$, for instance a specific low-dimensional subspace or a union of subspaces, and numerically compute the second-moment map $\\Psi(x)=(X_1^*X_1,\\ldots,X_L^*X_L)$ restricted to $M$; finding two points $x,y\\in M$ with $y\\neq \\pm x$ and $\\Psi(x)=\\Psi(y)$ would contradict the all-vectors claim of Corollary III.4. For the generic claim, the same search over random linear translates $A(M)$ would settle whether the $GL(V)$-generic condition delivers the promised uniqueness.","tokens_in":7831,"feed_emoji":"🔬","tokens_out":11086,"duration_ms":96119,"temperature":0.7,"pith_summary":"The paper establishes that phase retrieval—recovering a signal from its Fourier magnitudes—is a special case of a broader inverse problem: recovering a tuple of matrices from their Gram matrices, with the missing phases replaced by missing unitary or orthogonal factors coming from a compact group action. The central result is that the second moment of the multi-reference alignment model determines the signal exactly up to the ambiguity group $H=\\prod_{\\ell=1}^L U(N_\\ell)$, where the $N_\\ell$ are the dimensions of the irreducible representations in the signal space. For signals lying in a semialgebraic prior of dimension $M$, a transversality theorem shows that generic signals are recovered up to sign when the effective dimension $K=\\dim V-k(H)$ exceeds $M$, and every signal is recovered up to sign when $K>2M$. These results unify uniqueness guarantees for crystallographic phase retrieval, multi-reference alignment, and cryo-electron microscopy, and the paper adapts classical alternating-projection algorithms to the group setting via Procrustes steps. It closes with a conjecture, backed by numerical experiments, that the second-moment recovery map is bi-Lipschitz under linear priors.","feed_headline":"Second moment fixes signals up to sign under low-dimensional priors","feed_subtitle":"For signals in semialgebraic sets, the second moment yields unique recovery up to sign, unifying MRA, crystallography, and cryo-EM.","key_machinery":"The load-bearing object is the second moment viewed as a $G$-equivariant endomorphism of the signal space $V$. Schur's lemma makes this endomorphism block-diagonal over the isotypic decomposition $V=\\bigoplus_{\\ell=1}^L V_\\ell^{\\oplus R_\\ell}$, with each block a scalar multiple of the identity, and a direct trace computation shows those scalars are exactly the inner products defining the Gram matrices $X_\\ell^* X_\\ell$; this is what turns the measurement into a Gram-matrix tuple. The uniqueness results then rest on a transversality statement: in an orthogonal representation $V$ of a compact Lie group $H$, a $GL(V)$-generic semialgebraic set $M$ of dimension $M$ is transverse to the $H$-orbits, in the sense that the orbit of a generic point meets $M$ only at $\\pm x$ when $K=\\dim V-\\max_x\\dim Hx$ exceeds $M$, and only at $\\pm x$ for all points when $K>2M$. The algorithm side is carried by the Procrustes projection, which replaces the classical 'match the measured magnitudes' projection by an orthogonal matching of a current estimate to the Gram-matrix constraint.","core_discovery":"On the paper's own terms, the discovery is that the generalized phase retrieval problem over a compact group $G$ is the problem of lifting a tuple of Gram matrices $X_\\ell^* X_\\ell$ back to the matrices $X_\\ell$, where the missing data are unitary matrices rather than phases. Theorems II.1 and Corollary II.2 show that the second moment of the MRA observation model is a $G$-invariant element of $\\operatorname{Hom}(V,V)$; Schur's lemma forces such an element to act as a scalar multiple of the identity between copies of each irreducible representation, and a trace calculation identifies those scalars with the entries of the Gram matrices. Consequently the second moment determines $x$ only up to $H=\\prod_{\\ell=1}^L U(N_\\ell)$, and the remaining work is to pin down the unitaries. The paper's main new tool is a transversality theorem for semialgebraic sets: for a $GL(V)$-generic prior $M$ of dimension $M$, if $K>M$ then a generic point of $M$ has its $H$-orbit meet $M$ only at $\\pm x$, yielding uniqueness up to sign from the second moment; if $K>2M$, this holds for every point. The same machinery gives explicit thresholds for phase retrieval and cryo-EM, and numerical experiments on a linear prior support the conjecture that the recovery map is bi-Lipschitz.","pith_inferences":["The $GL(V)$-generic assumption is the fragile bridge to practice: for a fixed natural prior such as exact sparsity with known support, or a specific trained generative model, the paper provides no proof of genericity. A direct test would be to apply the second-moment map to random linear embeddings of that prior and check numerically whether any two distinct signals share a Gram tuple at $K>M$.","The dimension thresholds are probably not improvable without extra structure: at $K=M$ the intersection of an $M$-dimensional prior with generic orbits of dimension $k(H)$ should generically have isolated self-intersections, so uniqueness up to sign is the best one can expect—this is our inference, not a claim of the paper.","If the bi-Lipschitz conjecture is proven, it would convert the existing uniqueness statements into finite-sample guarantees by standard concentration of the empirical second moment: the number of samples needed to reach a target error would scale as $\\sigma^4$ times a condition-number factor, which is exactly the regime the paper motivates.","The same Procrustes-projection framework should extend to third-moment recovery, where the missing objects are not unitary matrices but elements of larger representation-theoretic tensor products; that would give algorithms for signals whose second moment is not injective, a direction the authors flag but do not develop."],"forward_implications":["In the high-noise regime, signals that satisfy the dimension conditions are recoverable from the second moment with $n=\\omega(\\sigma^4)$ samples, improving on the $\\omega(\\sigma^6)$ cost of third-moment methods.","Cryo-EM structure determination becomes a second-moment problem whenever the radial discretization satisfies $R\\ge 2L+1$ and the molecule lies in a sufficiently low-dimensional semialgebraic prior: generic uniqueness up to sign follows from the dimension count $K\\approx L^2(R+2L/3)$.","Classical phase-retrieval software can be lifted to any compact-group setting by swapping the magnitude-matching projection for a Procrustes projection, so sparsity, support, and generative-model priors plug in through their usual projection operators.","Phase retrieval's known dimension thresholds ($N\\ge 2M$ generic, $N\\ge 4M$ for all signals) reappear as the special case of one-dimensional irreducible representations, giving a uniform explanation across applications.","If the bi-Lipschitz conjecture holds, the recovery map from Gram matrices to $x$ has a noise-robustness constant, so small perturbations of the empirical second moment translate to linearly controlled recovery error under linear priors."],"supporting_citations":[{"why":"Proves Theorem II.1: the second moment of the MRA model determines the tuple of Gram matrices $X_\\ell^* X_\\ell$.","marker":"[12]"},{"why":"Supplies the detailed transversality theorem for semialgebraic sets and the precise cryo-EM uniqueness formulation that Theorem III.1 and Corollaries III.4–III.6 rely on.","marker":"[10]"},{"why":"Introduces the projection-based algorithm that generalizes classical phase retrieval to second-moment recovery via Procrustes projections, which Section IV adapts.","marker":"[14]"},{"why":"Establishes that the second moment is invariant to the tomographic projection, linking the cryo-EM model to the MRA second-moment analysis.","marker":"[26]"},{"why":"Proves the bi-Lipschitz property for standard phase retrieval, the base case that Conjecture V.1 extends to higher-dimensional representations.","marker":"[2]"},{"why":"Provides a bi-Lipschitz quotient embedding theorem for group actions that motivates the conjectured stability of the recovery map.","marker":"[15]"},{"why":"Shows sample complexity in high noise is governed by the highest moment order used, motivating recovery from the second moment alone for $\\omega(\\sigma^4)$ samples.","marker":"[1]"}],"fun_headline_variants":["Second moment solves phase retrieval over compact groups","Uniqueness up to sign from second moments in semialgebraic priors","Generalized phase retrieval: second moment uniquely recovers signals","Semialgebraic priors give unique recovery in generalized phase retrieval","Compact groups extend phase retrieval to Gram matrices"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the semialgebraic prior $M$ is $GL(V)$-generic, and for cryo-EM that $R\\ge 2L+1$; a fixed natural prior (exact sparsity with known support, or a trained generative model) is not proven to satisfy this genericity, so the dimension thresholds may not apply to it.","fun_headline_variants_meta":{"raw":{"variants":["Second moment solves phase retrieval over compact groups","Uniqueness up to sign from second moments in semialgebraic priors","Generalized phase retrieval: second moment uniquely recovers signals","Semialgebraic priors give unique recovery in generalized phase retrieval","Compact groups extend phase retrieval to Gram matrices"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000995,"raw_usage":{"total_tokens":4247,"prompt_tokens":1009,"completion_tokens":3238,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":3156}},"tokens_in":625,"tokens_out":3238,"duration_ms":21409,"temperature":1.0,"reasoning_tokens":3156,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:51:51.401015+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed semialgebraic prior $M$ of dimension $m<K$, for instance a specific low-dimensional subspace or a union of subspaces, and numerically compute the second-moment map $\\Psi(x)=(X_1^*X_1,\\ldots,X_L^*X_L)$ restricted to $M$; finding two points $x,y\\in M$ with $y\\neq \\pm x$ and $\\Psi(x)=\\Psi(y)$ would contradict the all-vectors claim of Corollary III.4. For the generic claim, the same search over random linear translates $A(M)$ would settle whether the $GL(V)$-generic condition delivers the promised uniqueness.","supporting_citations":[{"cited_title":"The sample complexity of sparse multireference alignment and single-particle cryo-electron microscopy","cited_arxiv_id":null,"evidence_quote":"Proves Theorem II.1: the second moment of the MRA model determines the tuple of Gram matrices $X_\\ell^* X_\\ell$."},{"cited_title":"Autocorrelation analysis for cryo-EM with sparsity constraints: improved sample complexity and projection-based algorithms","cited_arxiv_id":null,"evidence_quote":"Introduces the projection-based algorithm that generalizes classical phase retrieval to second-moment recovery via Procrustes projections, which Section IV adapts."},{"cited_title":"The reconstruction of structure from electron micrographs of randomly oriented particles","cited_arxiv_id":null,"evidence_quote":"Establishes that the second moment is invariant to the tomographic projection, linking the cryo-EM model to the MRA second-moment analysis."},{"cited_title":"On Lipschitz analysis and Lipschitz synthesis for the phase retrieval problem","cited_arxiv_id":null,"evidence_quote":"Proves the bi-Lipschitz property for standard phase retrieval, the base case that Conjecture V.1 extends to higher-dimensional representations."},{"cited_title":"Bi-Lipschitz Quotient embedding for Euclidean Group actions on Data","cited_arxiv_id":"2409.06829","evidence_quote":"Provides a bi-Lipschitz quotient embedding theorem for group actions that motivates the conjectured stability of the recovery map."},{"cited_title":"Estimation in the group action channel","cited_arxiv_id":null,"evidence_quote":"Shows sample complexity in high noise is governed by the highest moment order used, motivating recovery from the second moment alone for $\\omega(\\sigma^4)$ samples."}],"review_version":1}