{"id":"229b1c1a-4acb-4765-9a68-b330175e8bc4","arxiv_id":"2505.02221","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For entangled photons crossing a scattering medium, shaping one photon gives η≈(π/4)N, shaping both before the medium gives η≈(π/4)^2 N, while symmetric same-mode detection can restore correlations with η≈N, and post-medium shaping reaches up to about 4.6N numerically.","lead":"This paper derives and simulates the maximum enhancement of two-photon correlations that wavefront shaping can provide when entangled photon pairs pass through a thick scattering medium. It finds that the optimal enhancement depends on where the spatial light modulator sits and on whether both photons are detected in the same mode, with some configurations beating classical shaping.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"2P-DS enhancement values 1.91N and 4.6N are labeled optimal but rest only on local L-BFGS optimization of a non-convex problem; without a global-optimality check they are lower bounds, not fundamental bounds.","rationale":"The reader's weakest_assumption was the absence of a global-optimality certificate for the 2P-DS configurations, and my reading agrees that this is the load-bearing point. The abstract and Section II.C present η ≈ 1.91N and η ≈ 4.6N as optimal enhancements, and the introduction asserts that the results are determined by the optimal shaping phases. The only evidence for these values is a local optimizer on a non-convex unit-modulus quadratic objective, with SI S3 explicitly deferring formal complexity analysis. This makes the 2P-DS portion of the 'fundamental bounds' claim an empirical observation, not a proven bound. I checked the analytical 1P-S, 2P-IS, and 2P-IS(OPC) derivations: they provide explicit phase constructions and closed-form enhancements, so they carry independent support. The same-mode detection prescription is consistently handled through the beamsplitter factors in Eq. (S3), and I do not see an internal inconsistency that would justify REJECT. The missing global-optimality certificate is exactly what the reader's CONDITIONAL verdict should require, so no verdict change is needed.","tokens_in":22440,"tokens_out":11719,"duration_ms":160276,"concrete_test":"For 20 Gaussian IID disorder realizations at N = 8, 12, 16, and 24, compute the certified global maximum of |Σ_{n,m} f_{βn}(T^T T)_{nm} f_{αm} s_n s_m|^2 over |s_n|=1 using a rigorous global optimizer (e.g., interval branch-and-bound on the unit-modulus quadratic form). Compare each global optimum with the PyTorch L-BFGS result on the same realization. If L-BFGS systematically falls below the certified global optimum, or if the exact optimum's η/N extrapolation toward N = 512 disagrees with the reported 1.91 and 4.6 prefactors, then the 2P-DS claims should be downgraded from 'optimal enhancements' to 'achieved enhancements'.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the 2P-DS enhancement is optimally ~1.91N (non-symmetric) and ~4.6N (same-mode, Gaussian IID) requires that the numerically optimized phase masks are global maximizers of P_{αβ}^{(2P-DS)} = (1/(2N)) |(F S T^T T S F)_{βα}|^2. The manuscript supports this only with PyTorch's L-BFGS on the non-convex objective |Σ_{n,m} A_{nm} s_n s_m|^2 with |s_n|=1 (main text §II.C; SI §S4). No global-optimality certificate is given; the paper's own SI §S3 states that \"a formal complexity analysis for this specific problem ... is beyond the scope of this work\" and only cites structural resemblance to NP-hard MAXQP / XY spin-glass problems. Because L-BFGS is a local method and the objective is known to have many stationary points, the reported prefactors are achieved values, i.e., rigorous lower bounds on the maximum, rather than proven optimal enhancements. Consequently the abstract's and title's \"fundamental bounds\" claim is not established for the 2P-DS configurations; this is not a disagreement with the numerics but a missing certification of optimality. The analytical 1P-S, 2P-IS, and 2P-IS(OPC) results are unaffected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies optimal wavefront shaping of spatially entangled photon pairs that both propagate through a thick scattering medium, modeled by either a random unitary or a Gaussian IID transmission matrix. Three configurations are analyzed: one-photon shaping (1P-S), two-photon illumination shaping (2P-IS), and two-photon detection shaping (2P-DS), each with an optional symmetric detection variant labeled OPC. The main results are analytical expressions for the enhancement in 1P-S and 2P-IS, η=(π/4)N and η=(π/4)^2N respectively, an exact N prefactor for the symmetric 2P-IS(OPC) case, and numerical claims of η≈1.91N (non-symmetric) and η≈4.6N (symmetric) for 2P-DS in the Gaussian IID model. The paper argues that the 2P-DS optimization problem resembles NP-hard MAXQP/XY spin-glass problems, and that the numerically obtained phases achieve enhancements that can exceed the total pre-optimization coincidence rate.","tokens_in":22712,"tokens_out":8459,"duration_ms":108537,"significance":"If the results hold, this is a useful contribution to quantum wavefront shaping: it identifies configuration-dependent enhancement prefactors, demonstrates that phase-only control can restore perfect two-photon correlations in the OPC variants, and highlights a physically interesting self-consistent optimization problem in the 2P-DS geometry. The analytical derivations in SI S2 are explicit and self-contained, and the code and data are provided, which supports reproducibility. For the 1P-S, 2P-IS, and 2P-IS(OPC) configurations, the results are clean and convincing. However, the paper's headline claim of 'fundamental bounds' for 2P-DS rests on numerical results from a local optimizer on a non-convex landscape, and the paper's own SI states that a formal complexity analysis is beyond scope. This means the 2P-DS prefactors are achieved enhancements (lower bounds), not proven optimal enhancements, which weakens the central claim as currently worded.","major_comments":[{"comment":"The reported 2P-DS enhancements η≈1.91N and η≈4.6N are described in the abstract and main text as optimal and as 'fundamental bounds,' but they are obtained with PyTorch's L-BFGS, a local optimizer, applied to the non-convex objective P_{αβ}^{(2P-DS)} = (1/(2N)) |(F S T^T T S F)_{βα}|^2. SI §S3 explicitly states that 'a formal complexity analysis for this specific problem ... is beyond the scope of this work.' Consequently, these numerical values are rigorous lower bounds on the maximum enhancement, not certificates of the global maximum. This does not invalidate the numerical results, but it does invalidate the claim that the prefactors 1.91 and 4.6 are the optimal (maximal) enhancements. The authors should either (i) reframe the 2P-DS claims as achieved enhancements or lower bounds throughout the abstract, main text, and title, or (ii) provide a global-optimality certificate, for example by exhaustive search or rigorous upper bounds for small N that match the numerical values.","section":"§II.C, SI §S3–S4"},{"comment":"The finite-size status of the 2P-DS numbers is also not addressed. SI §S4.E states that 'there does seem to be a continued non-negligible dependence on N ... even for a few thousand modes,' and Fig. S6 shows η/N still increasing at N values beyond 512. The abstract quotes η≈1.91N and η≈4.6N without this qualification, so these are finite-N, local-optimizer values, not established asymptotic prefactors. The sentence in the Introduction that 'these results are determined by the optimal shaping phases, regardless of the method used to obtain them' is therefore too strong; the method (local optimization) is exactly what prevents the conclusion that the phases are optimal. Please qualify these claims, for example by reporting 2P-DS values as numerical achievements at N=512 and by presenting the observed N-dependence as an open issue.","section":"Intro, §II.C, SI §S4.E"}],"minor_comments":[{"comment":"The heading '2P-DS (OPC) configurtation' contains a typo; it should be 'configuration'.","section":"SI §S4.D"},{"comment":"The notation '1-PS' is used once in the text, while the rest of the paper uses '1P-S'; please make the notation uniform.","section":"SI §S6.A"},{"comment":"Reference 23 contains 'V os' which should be 'Vos', and Reference 43 contains 'F oundations' which should be 'Foundations'.","section":"References"},{"comment":"The sentence 'we note that it is bounded by σ1^2, where σ1 is the maximal singular value of TT^T' is ambiguous: the antecedent of 'it' could be the enhancement or the total optimized coincidence counts. Since η/N ≈ 4.6 exceeds typical values of σ1^2 ≈ 4 in the Gaussian IID model, the bound presumably applies to the total optimized coincidence counts; please clarify.","section":"§II.C"},{"comment":"Please change 'errorbars' to 'error bars' for consistency.","section":"Fig. 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The analytical 1P-S, 2P-IS, and 2P-IS(OPC) results are solid and likely publishable. The main concern is the mismatch between the 'fundamental bounds' language and the lack of a global-optimality certificate for the 2P-DS numerical results. If the authors are willing to reframe the 2P-DS claims as achieved enhancements rather than proven optimal bounds, and to add an explicit caveat about the finite-N, local-optimizer status of those numbers, a revised version could be acceptable. Requiring a full global-optimality proof for a non-convex NP-hard-like problem may be too demanding for this paper; the reframing route seems the most practical path."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things worth knowing. First, the analytic derivations for one-photon shaping and two-photon illumination shaping are clean and correct: the paper constructs the optimal phase mask explicitly, and the enhancements (π/4)N and (π/4)^2 N follow from Gaussian moment statistics. The symmetric same-mode case, where phase-only control restores perfect correlations and gives N, is a genuinely nice result with a persuasive digital phase-conjugation interpretation. Second, the 2P-DS numbers in the title's promise of 'fundamental bounds' are not bounds. They are achieved values from local L-BFGS optimization of a non-convex quadratic form, with no global optimality certificate, and the SI explicitly punts on complexity analysis. That does not make the numerics wrong, but it makes them lower bounds, and the abstract's 'up to 4.6N' should be labeled as such.\n\nWhat is new: the analytic scalings, the incomplete-control slopes (π/2 vs π/4 for the symmetric versus non-symmetric illumination cases), and the OPC interpretation are not in the prior literature. The paper also ships code and data, and the simulation methodology is described well enough to reproduce. Credit where due.\n\nThe soft spots are proportionate. The 2P-DS section is the only load-bearing weakness. For a 'fundamental bound' claim, you need either exhaustive small-N verification, many random restarts with a report of the distribution, or a rigorous upper bound. The same-mode detection (both photons into one spatial mode) is an idealization; the paper mentions it but does not discuss how an experiment would implement two detectors on the same mode. The N-dependence of the 2P-DS prefactor is still slowly varying at N=512, which undercuts the claim that 1.91N and 4.6N are asymptotic constants. These are fixable with caveats or more numerics, not fatal.\n\nThe citation pattern is fine; they cite Courme et al. for the NP-hard mapping, and the self-citations are to their own relevant prior work.\n\nBottom line: this deserves a serious referee. Send it to review, and ask the authors to either prove or explicitly soften the 2P-DS 'optimal' language. The analytical core holds up and is worth publishing.","headline":"Solid analytical core for 1P-S and 2P-IS, but 'fundamental bounds' overstates the 2P-DS numerical results, which are lower bounds from local optimization.","tokens_in":23248,"tokens_out":2302,"would_cite":true,"duration_ms":26877,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["42.50.Dv","42.25.Dd"],"model":"deepseek-v4-flash","headline":"For entangled photon pairs, the best wavefront-shaping gain depends on modulator placement—from about 0.62N to 4.6N.","keywords":["wavefront shaping","spatially entangled photons","two-photon correlations","scattering media","transmission matrix","enhancement factor","optical phase conjugation","NP-hard optimization"],"falsifier":"Run an independent search—many random restarts with local refinement, or exhaustive branch-and-bound at small $N$—on a fixed Gaussian IID transmission matrix realization and the same same-mode target; if any phase pattern yields $\\eta/N > 1.91$ for distinct detectors or $> 4.6$ for same-mode detection, the reported values are lower bounds, not fundamental bounds.","tokens_in":22249,"feed_emoji":"⚛️","tokens_out":14644,"duration_ms":156155,"temperature":0.7,"pith_summary":"This paper asks a sharp question: when both photons of a spatially entangled pair scatter through a thick random medium, what is the best that a phase-only spatial light modulator can do, and where should the modulator sit? The answer is that the optimal enhancement of two-photon coincidences is configuration-dependent and differs from the classical wavefront-shaping limit. Shaping one photon after the sample reproduces the classical gain $\\eta \\approx (\\pi/4)N$, while shaping both photons before the sample lowers it to $\\eta \\approx (\\pi/4)^2 N$; detecting both shaped photons in the same spatial mode restores near-perfect correlations, $\\eta \\approx N$, with phases alone. Shaping both photons after the sample is computationally hard but numerically gives the largest gains, up to about $4.6N$ in the same-mode Gaussian case. This matters for quantum imaging and communication through scattering media, where a classical beacon cannot provide the correct feedback.","feed_headline":"Shaping both photons beats classical wavefront shaping—up to 4.6N","feed_subtitle":"For N spatial modes, the best two-photon gain shifts with modulator placement, from (π/4)²N to 4.6N.","key_machinery":"The central object is the two-photon coincidence probability written as a squared matrix element, $P_{\\alpha\\beta} = (2/N)|(H_2 H_1^T)_{\\beta\\alpha}|^2$, where $H_1,H_2$ are the single-photon transmission matrices from the crystal plane to the two detectors. In the three configurations this becomes $|(F T T^T S F)_{\\beta\\alpha}|^2$, $|(F T S S T^T F)_{\\beta\\alpha}|^2$, and $|(F S T T^T S F)_{\\beta\\alpha}|^2$, where $T$ is the scattering matrix, $S$ the diagonal phase-only SLM matrix, and $F$ the discrete Fourier transform to the far-field detection plane. The advanced wave picture—replacing one detector by a source and the crystal by a mirror—turns each configuration into a classical double-pass propagation problem and supplies the physical intuition for the formulas. The analytic prefactors come from averaging Gaussian moments of $T$, using the identity $\\langle |t| \\rangle^2/\\langle |t|^2 \\rangle = \\pi/4$. In the after-medium configuration, $S$ appears on both sides of $T T^T$, so every SLM pixel influences the light incident on every other pixel on the second pass; this makes the phase pattern a complex quadratic form in unit-modulus variables, which is the structure the paper identifies as NP-hard-like.","core_discovery":"The central claim is that the fundamental limits of two-photon wavefront shaping are set by a short list of formulas, each tied to a specific modulator and detector geometry. For $N$ spatial modes with phase-only control, one-photon shaping gives $\\eta = 1 + (N-1)\\pi/4 \\approx (\\pi/4)N$; shaping both photons before the medium gives $\\eta = 1 + (N-1)(\\pi/4)^2 \\approx (\\pi/4)^2 N$; and in the symmetric case where both shaped photons are detected in the same mode, perfect correlations are restored, $\\eta = N$ for a unitary medium and $\\eta = N+1$ for a Gaussian IID medium. When the modulator shapes both photons after the medium, the paper reports numerically obtained optimal enhancements of $\\eta \\approx 0.89N$ (unitary) and $\\eta \\approx 1.91N$ (Gaussian IID) for distinct detectors, and $\\eta \\approx N$ and $\\eta \\approx 4.6N$ respectively when both photons land in the same mode. The after-medium gains can exceed the total coincidence rate summed over all modes before shaping, and the paper notes they are bounded above by the largest singular value of the transmission-matrix product. The analytic prefactors follow from Gaussian statistics of the transmission matrix; the after-medium numbers come from numerical optimization of a problem the paper identifies as NP-hard-like.","pith_inferences":["The prefactor pattern hints at a compounding rule the paper does not state: every pass of a phase pattern illuminated by a speckle field may reduce the gain by a factor $\\pi/4$, so a chain of $k$ such passes would give $\\eta \\approx (\\pi/4)^k N$; the 1P-S and 2P-IS formulas are consistent with this.","The Gaussian-IID after-medium gain rising above the pre-optimization total suggests the optimizer is recruiting open transmission eigenchannels, so the enhancement per realization should correlate with the largest singular value of $TT^\\dagger$; that correlation could be tested from the paper's code and data.","Because the after-medium problem is NP-hard-like, the reported $1.91N$ and $4.6N$ are best read as benchmarks of the local gradient optimizer on a non-convex landscape; a dedicated global optimizer at small $N$ could either confirm tightness or find better patterns.","The same-mode 'OPC' configurations offer a self-aligned way to restore correlations without ever characterizing the medium, which could simplify entanglement distribution through fibers or tissue if feedback is taken directly from coincidence counts."],"forward_implications":["When both photons are shaped before the medium and detected in the same spatial mode, phase-only control can fully undo strong scattering, $\\eta \\approx N$, without amplitude control, which the paper identifies as digital optical phase conjugation.","In the after-medium configuration, the optimized coincidence count at one output mode can exceed the total coincidence count summed over all modes before shaping, so scattering alone no longer sets the rate ceiling.","Because every SLM pixel affects the light that reaches all other pixels on the second pass, the after-medium optimization is self-consistent and NP-hard-like; standard pixel-by-pixel iterative algorithms fail there.","Under incomplete control, enhancement still grows linearly with the number of controlled modes at low degree of control, and the symmetric same-mode cases start at twice the classical slope ($\\pi/2$ versus $\\pi/4$), a signature the paper traces to coherent backscattering.","Classical beacon feedback is not directly usable for both-photon shaping, because the phases that optimize two-photon correlations differ from those that focus a classical beam; the paper presents this as the reason new feedback strategies are needed."],"supporting_citations":[{"why":"Introduces wavefront shaping of light through scattering media, the classical capability this paper extends to spatially entangled photon pairs.","marker":"[1]"},{"why":"Supplies the classical benchmark $\\eta \\approx (\\pi/4)N$ for phase-only control that the one-photon shaping configuration reproduces.","marker":"[33]"},{"why":"Provides the circular Gaussian statistics and moment theorem used to evaluate the disorder averages that produce the $\\pi/4$ prefactors.","marker":"[36]"},{"why":"Supplies the Gaussian IID transmission-matrix model for a strongly scattering medium and the bound on maximal intensity transmission used for the after-medium enhancements.","marker":"[37]"},{"why":"Introduces the advanced wave picture that reduces two-photon coincidence probability to a classical double-pass intensity, the paper's main interpretive machinery.","marker":"[38]"},{"why":"Provides the mapping of two-photon correlation optimization to non-convex, NP-hard-like problems used to characterize the after-medium configuration.","marker":"[31]"},{"why":"Defines digital optical phase conjugation, the concept used to explain perfect same-mode restoration with phase-only control.","marker":"[41]"}],"fun_headline_variants":["Two-photon shaping: after-medium beats classical, up to 4.6N","Shaping both before medium drops gain to (π/4)²N","Same-mode detection restores perfect correlations: gain N","Quantum wavefront shaping: 4.6N max, placement decides"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The after-medium enhancements of $1.91N$ and $4.6N$ are treated as the optimal values, but they come from a local optimizer on a non-convex problem, so no proof yet rules out a better phase pattern.","fun_headline_variants_meta":{"raw":{"variants":["Two-photon shaping: after-medium beats classical, up to 4.6N","Shaping both before medium drops gain to (π/4)²N","Same-mode detection restores perfect correlations: gain N","Quantum wavefront shaping: 4.6N max, placement decides"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000378,"raw_usage":{"total_tokens":2070,"prompt_tokens":1066,"completion_tokens":1004,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":682,"completion_tokens_details":{"reasoning_tokens":934}},"tokens_in":682,"tokens_out":1004,"duration_ms":8997,"temperature":1.0,"reasoning_tokens":934,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:59:11.601315+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run an independent search—many random restarts with local refinement, or exhaustive branch-and-bound at small $N$—on a fixed Gaussian IID transmission matrix realization and the same same-mode target; if any phase pattern yields $\\eta/N > 1.91$ for distinct detectors or $> 4.6$ for same-mode detection, the reported values are lower bounds, not fundamental bounds.","supporting_citations":[{"cited_title":"Soro , author E","cited_arxiv_id":null,"evidence_quote":"Supplies the classical benchmark $\\eta \\approx (\\pi/4)N$ for phase-only control that the one-photon shaping configuration reproduces."},{"cited_title":"Shekel , author O","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian IID transmission-matrix model for a strongly scattering medium and the bound on maximal intensity transmission used for the after-medium enhancements."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the advanced wave picture that reduces two-photon coincidence probability to a classical double-pass intensity, the paper's main interpretive machinery."},{"cited_title":"Lib , author G","cited_arxiv_id":null,"evidence_quote":"Provides the mapping of two-photon correlation optimization to non-convex, NP-hard-like problems used to characterize the after-medium configuration."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines digital optical phase conjugation, the concept used to explain perfect same-mode restoration with phase-only control."}],"review_version":1}