{"id":"fc42a35f-e27f-4f3c-bbc7-7f12456e1115","arxiv_id":"2412.16023","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The optimal Gaussian probe for homodyne phase estimation switches discontinuously from displaced coherent-like states to squeezed vacuum as the prior phase uncertainty shrinks.","lead":"This paper determines which single-mode Gaussian probe state gives the most precise phase estimate with homodyne detection, given your current uncertainty about the phase. It shows the optimum switches abruptly from coherent-like displaced states to squeezed vacuum as the prior uncertainty decreases, a jump that matters for adaptive phase estimation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Global optimality rests on an unproven two-minima assumption; a third basin would break the discontinuous-jump claim.","rationale":"The reader's weakest assumption correctly identifies the absence of an exhaustive optimality analysis over the full parameter space as the main gap. My reading of the paper confirms this: the analytic Fisher-information results (Eq. 17, Eq. 18) are derived carefully and are internally consistent, but the leap from 'numerical minimization consistently results in one of two strategies' to 'we can identify the optimal probe state' is not logically secured. The paper itself provides only a single energy value (E = 0.5) in the representative basin plot (Fig. 7), and the text explicitly notes that which strategy is the global minimum depends on energy and prior, yet no systematic scan is reported. A third local minimum would not necessarily invalidate all results—the FI analysis and the qualitative HUS/LUS comparison could survive—but it would invalidate the headline claims of having identified the optimal states and of a sharp discontinuous transition. The proposed grid-plus-refinement test directly addresses this gap. The secondary concern about the adaptive-strategy optimality in Section V is real but less central to the paper's main contribution; it would require a separate dynamic-programming or counterexample analysis. Overall, the conditional verdict is appropriate: the core analytic and numerical results are plausible, but the global-optimality claim requires stronger evidence or a proof of the stationary-point structure.","tokens_in":15258,"tokens_out":6105,"duration_ms":59244,"concrete_test":"For each E in {0.5, 1, 2, 5} and a dense set of prior variances σ² in (0, 0.2] (e.g., 20 values), perform a global search over (|α|, τ, φ): first evaluate APV on a regular grid of 40×40×40 points in the compact domain, then run local minimization from every grid point whose APV is within the lowest 5% of grid values; finally, cluster the converged minima and record all distinct local minima with APV within 0.1% of the best. If any converged minimum is not well approximated by HUS or LUS (e.g., |α|²/E differs by more than 0.01 from both families, or τ/φ deviate significantly), or if the lowest APV at some σ² belongs to a third family, the global-optimality and discontinuity claims are falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim in Section IV.C—that the optimal pure single-mode Gaussian probe state is always either the HUS family (τ = θ0 - π/2, φ = -2θ0) or the LUS family (α = 0), with a discontinuous switch between them—depends on the assertion that no other local or global minima of the average posterior variance exist in the full (|α|, τ, φ) space. The only evidence is numerical multistart minimization, and the representative results in Fig. 7 are shown for E = 0.5 only. For E = 1, 2, 5, the claim rests on the statement that minimization 'consistently results in one of the two strategies,' with no basin analysis, grid scan, or uniqueness proof. The APV is a non-convex integral over a compact 3D domain (|α| ∈ [0,√E], τ, φ ∈ [0,2π)), so additional stationary points are plausible, especially near the claimed transition where the objective may be nearly degenerate. If a third family of states achieves a lower APV in some prior-variance range, both the 'two strategies' identification and the abrupt-jump conclusion would need revision. The Section V claim that per-round optimal adaptation 'will on average outperform any other strategy' is also asserted without proof for finite horizons, but the exhaustiveness of the single-round optimality landscape is the more load-bearing issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript addresses the question of which pure single-mode Gaussian probe state is optimal for estimating an unknown phase rotation under homodyne detection, as a function of the prior knowledge of the phase. Working in a Bayesian framework with a Gaussian prior of variance σ² and fixed energy E, the authors numerically minimize the average posterior variance (APV) over (|α|, τ, φ). They identify two candidate optimal families: a high-uncertainty strategy (HUS) with τ = θ0 − π/2 and φ = −2θ0, in which energy is split between displacement and squeezing, and a low-uncertainty strategy (LUS) with α = 0, i.e., squeezed vacuum. The global optimum is claimed to switch discontinuously from HUS to LUS as σ² decreases. The paper also provides analytic Fisher information results (Appendix A), including I = 4E(E+1) for the optimal HUS state and I = 8E(E+1) for the LUS state, applies the results to repeated measurements with several adaptive strategies, and draws a qualitative analogy to noisy phase estimation.","tokens_in":15515,"tokens_out":10651,"duration_ms":82541,"significance":"If correct, the paper's main contribution is a clear map of knowledge-dependent optimal Gaussian probe states for homodyne phase estimation, connecting the known local (frequentist) optimality of squeezed vacuum to the Bayesian optimality of coherent-like states at large prior uncertainty. The analytic FI derivation is transparent and the limiting cases reproduce established results (the local FI of the LUS, 8E(E+1), and the constrained HUS FI, 4E(E+1)). The adaptive strategies for repeated measurements are practically motivated and the comparison in Fig. 4 is informative. The key weakness is that the global optimality of the two families is inferred from numerical minimization without exhaustive search or proof; this is the load-bearing element of the paper's central claim.","major_comments":[{"comment":"The central claim that the global minimum of the APV over the full (|α|,τ,φ) space is always attained by either the HUS or LUS family is not established. The evidence consists of multistart local minimization; Fig. 7 shows results only for E = 0.5, and for E = 1, 2, 5 the claim rests on the statement that minimization 'consistently results in one of the two strategies.' Since the APV is a non-convex integral over a compact 3D domain, other stationary points cannot be excluded. If a third family of states achieved a lower APV in some region of σ², both the identification of 'two distinct optimal strategies' and the 'clear jump' conclusion would need revision. Please provide a more exhaustive global optimization (e.g., dense grid combined with multistart or a global optimizer) or an analytic argument that every stationary point belongs to one of the two families.","section":"IV.C, Appendix C"},{"comment":"The claim that 'a strategy that calculates the optimal probe state in each round will on average outperform any other strategy' is not justified. Myopic single-round optimality does not generally imply finite-horizon optimality in adaptive estimation problems. The numerical comparison in Fig. 4 is limited to fixed, angle-adaptive, predetermined, and a constrained adaptive strategy; the true fully optimal adaptive policy is not computed. Either prove the claim (e.g., by dynamic programming or by a dominance argument) or restate it as 'outperforms the strategies considered here.'","section":"V"},{"comment":"The repeated-measurement analysis approximates the posterior after each round as Gaussian with the same mean and variance, and asserts that the optimal probe state is 'nearly identical' to that for the true posterior. Because this approximation underlies the adaptive results in Fig. 4, a quantitative error estimate is needed (for example, by comparing the true and approximated posteriors for representative parameter values, or by comparing the resulting APVs). Without such a check, the performance ranking of the adaptive strategies may be affected.","section":"V"}],"minor_comments":[{"comment":"The figures use specific parameters (e.g., φ = −π, τ = 0) that correspond to θ0 = π/2; please state this convention explicitly in the figure captions.","section":"IV.B, Figs. 2 and 3"},{"comment":"The numerical details are not specified: integration grid for the APV, number of starting points for the minimization, and convergence tolerances. Please include these for reproducibility.","section":"Appendix C"},{"comment":"The notation switches between θ (local true value) and θ0 (prior mean). In Eq. (17) and the discussion of the HUS, clarify that the expressions are evaluated at the prior mean when used in the Bayesian context.","section":"IV.A and IV.B"},{"comment":"The Van Trees bound could be compared with the numerically computed APV in a figure; this would show how tight the bound is for the parameter range studied.","section":"Appendix B"},{"comment":"Ref. [29] is cited as the source of the uniform-prior result; because the present second author is a coauthor of that reference, the relationship should be acknowledged explicitly when the connection is made.","section":"I and VI"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is well within the scope of the journal. The overlap with Ref. [29] is a potential perception issue rather than a scientific one; the authors should cite and contextualize it explicitly. If the authors can strengthen the global-optimality evidence, I would support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the quick read on this one. The paper does something genuinely useful: it maps out the optimal pure single-mode Gaussian probe state for homodyne phase estimation as a function of the prior variance, and finds a discontinuous transition between displaced squeezed states (HUS) and squeezed vacuum (LUS). The analytic Fisher information derivation in Appendix A is clean, and the special cases check out (HUS local FI 4E(E+1), LUS 8E(E+1)). Connecting the known coherent-state result for broad priors to squeezed-vacuum optimality in the local limit is a nice bridge.\n\nThe soft spots are real, though. The central 'optimal probe state' claim is based on numerical multistart minimization that consistently lands in one of two basins. That is evidence, not proof. The APV is a non-convex function on a 3D compact domain, and the paper shows no grid scan, no basin analysis, and no argument that a third family cannot be globally optimal in some variance range. If such a family exists, both the 'two strategies' identification and the discontinuous-jump conclusion would need revision. The authors should either add an exhaustive scan or prove the absence of other stationary points, or at minimum temper the abstract's global-optimality language.\n\nSection V has a similar overreach. The claim that per-round optimal adaptation 'will on average outperform any other strategy' is asserted without proof. In adaptive estimation, greedy per-round optimization is not guaranteed to be globally optimal over a finite horizon. It may be true in this setting, but that needs an argument or a comparison against a broader class of policies.\n\nMinor: the numerical results are not accompanied by code or data, so the APV minimization is not independently checkable. The full-landscape figure is only shown for E=0.5; other energies rely on a sentence in the text.\n\nOverall, the paper is worth engaging with. The analytic core is solid, the question is relevant, and the qualitative conclusion (more squeezing as knowledge improves) is well supported. The load-bearing overclaim is about global optimality, and that is fixable. I'd send it to peer review, with the request that the authors either prove or substantially tone down the optimality claims and provide reproducible numerics.","headline":"Useful numerical map of optimal Gaussian probes vs prior variance, but the 'optimal' claim rests on an unproven two-minima assumption.","tokens_in":16033,"tokens_out":4855,"would_cite":true,"duration_ms":38355,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"For homodyne phase estimation with fixed probe energy, the state that minimizes the average posterior variance switches discontinuously between a displaced squeezed state and a squeezed vacuum as the prior variance decreases.","keywords":["phase estimation","quantum metrology","Gaussian states","homodyne detection","Bayesian estimation","average posterior variance","Fisher information","squeezed vacuum"],"falsifier":"Run a global optimization of the average posterior variance over $(|\\alpha|, \\tau, \\varphi)$ at fixed energy $E$ and a prior variance $\\sigma^2 \\in (0, 0.2]$, using a dense multi-start grid or a certified branch-and-bound method; a single point with lower average posterior variance than both the HUS and LUS optima would falsify the claim that those two families exhaust the optima. A cheaper partial check is to evaluate the Hessian at both putative optima and look for additional stationary points in the domain.","tokens_in":15019,"feed_emoji":"⚛️","tokens_out":13566,"duration_ms":102381,"temperature":0.7,"pith_summary":"The paper asks how to choose a probe state—a single-mode light state from the Gaussian family—when estimating an unknown phase rotation by homodyne detection, given a fixed energy budget and a Gaussian prior on the phase. It claims the best choice depends sharply on how much is already known: with a broad prior, the optimal states are close to coherent states, with most energy in displacement; as the prior narrows, the optimum moves continuously through displaced squeezed states, and then jumps abruptly to a squeezed vacuum state at a critical prior variance that depends on the energy. The result connects two previously separate findings: squeezed vacuum is asymptotically optimal in the local, many-measurement limit, while coherent states win in a single-shot Bayesian setting with little prior knowledge. If the claim is right, adaptive phase-estimation protocols should expect a discontinuous switch in the optimal probe, not a smooth trade-off, and can choose their probe state from precomputed maps of the two optimal families.","feed_headline":"Optimal probe jumps from displaced to squeezed vacuum as prior narrows","feed_subtitle":"Bayesian analysis: the best Gaussian probe for homodyne phase estimation switches discontinuously as uncertainty narrows","key_machinery":"The argument is carried by two one-parameter families of probe states. The high-uncertainty strategy (HUS) fixes $\\tau = \\theta_0 - \\pi/2$ and $\\varphi = -2\\theta_0$, so that after the unknown rotation the state's mean is on the $p$-axis and its squeezing is along the measured $q$-quadrature; the remaining parameter $|\\alpha|$ is chosen to minimize the average posterior variance at fixed energy $E = |\\alpha|^2 + \\sinh^2 r$. The low-uncertainty strategy (LUS) sets $\\alpha = 0$, putting all energy into squeezing, with the squeezing angle $\\varphi$ tuned to the prior; it gives Fisher information $8E(E+1)$ and locally saturates the quantum Cramér–Rao bound. The key calculation is the homodyne likelihood, a Gaussian in $q$ with mean $\\sqrt{2}|\\alpha|\\cos(\\tau-\\theta)$ and variance $[\\cosh 2r - \\cos(\\varphi+2\\theta)\\sinh 2r]/2$; from it the Fisher information is computed in closed form, while the average posterior variance is minimized numerically over the three parameters.","core_discovery":"Within the three-dimensional parameter space $(|\\alpha|, \\tau, \\varphi)$ of pure single-mode Gaussian states of fixed energy $E$, the paper finds that the average posterior variance under homodyne detection has two local-minimum families for Gaussian priors with variance $\\sigma^2 \\in (0, 0.2]$. The first, called the high-uncertainty strategy (HUS), sets $\\tau = \\theta_0 - \\pi/2$ and $\\varphi = -2\\theta_0$ and splits the energy between displacement and squeezing; the second, called the low-uncertainty strategy (LUS), sets $\\alpha = 0$ and uses a squeezed vacuum whose angle is tuned to the prior. Which family is globally optimal depends on $\\sigma^2$ and $E$: for large prior variance the HUS wins, for small prior variance the LUS wins, and the global optimum switches discontinuously between the two—there is no continuous path in parameter space connecting them. Within each family the parameters vary smoothly with the prior variance, so the discontinuous switch between families is the central non-smooth feature. The paper also computes the Fisher information for both families ($4E(E+1)$ for the HUS optimum, $8E(E+1)$ for the LUS) and shows that a purely local analysis misleads at finite uncertainty because the squeezed-vacuum Fisher information has a spurious second peak corresponding to a mirrored state homodyne detection cannot distinguish.","pith_inferences":["A natural extension the paper does not pursue is an analytical proof that the average posterior variance has no stationary points beyond the HUS and LUS families; if such a proof fails, the abrupt-jump claim would need to be weakened to a statement about the numerically accessible landscape.","The two-family structure was established for Gaussian priors and homodyne detection; for heavier-tailed priors or other Gaussian measurements, the discontinuity may become a smooth crossover, which would be a testable prediction of the same optimization machinery.","The near-optimality of the predetermined-rule strategy suggests a practical controller that stores the precomputed maps of $|\\alpha|^2/E$ and $\\varphi$ as functions of $\\sigma^2$ and $E$ and updates the probe from the current posterior mean and variance; its regret relative to the fully adaptive strategy is an experimentally measurable quantity.","If the discontinuous jump is generic, then any continuous feedback policy that interpolates smoothly between coherent and squeezed probes will pay a systematic penalty near the critical variance; this is a concrete design constraint for real-time adaptive interferometry."],"forward_implications":["In repeated measurements, updating the probe state from one round to the next beats any fixed probe state; the paper's numerical comparison shows that adapting only the measurement angle already gives most of the benefit, and updating the energy split according to the current variance matches the fully adaptive strategy.","The prior variance at which the switch to squeezed vacuum occurs grows as the probe energy shrinks, so low-energy probes should abandon displaced states at larger prior uncertainty than high-energy probes.","The Bayesian optimum within the HUS approaches the local frequentist optimum $|\\alpha|^2/E = (E+1)/(2E+1)$ as the prior variance tends to zero, but the global optimum in that limit is the LUS, not the HUS.","Because the squeezed-vacuum Fisher information has a second, misleading peak at a mirrored phase, any Fisher-information-only search for optimal probes at finite prior uncertainty can pick a state that systematically estimates the wrong angle.","The same displacement-versus-squeezing tradeoff appears in noisy phase estimation with phase-diffusive noise, suggesting that noise and prior uncertainty can be treated as interchangeable resources in probe design."],"supporting_citations":[{"why":"Supplies the local-analysis optimum: squeezed vacuum with the adapted angle saturates the quantum Cramér–Rao bound with Fisher information $8E(E+1)$, the anchor of the LUS family.","marker":"[15]"},{"why":"Gives the opposite extreme used as the paper's starting point: with a uniform prior in a single round, coherent states outperform squeezed states of equal energy.","marker":"[29]"},{"why":"Establishes that optimal phase measurements with pure Gaussian states use squeezed vacuum, supporting the identification of $|\\alpha|=0$ as the LUS optimum.","marker":"[14]"},{"why":"Provides the general single-mode Gaussian quantum Fisher information formula the paper invokes to show the QFI is maximized at zero displacement.","marker":"[32]"},{"why":"Supplies the squeezed-vacuum-as-universal-probe intuition for why the LUS is locally powerful but fragile away from the estimate.","marker":"[16]"},{"why":"Defines the phase-diffusion noisy estimation model whose optimal-probe behavior the paper compares with its own knowledge-dependent transition.","marker":"[17]"},{"why":"Provides the efficient Bayesian phase-estimation framework used to motivate and compare the repeated-measurement adaptive strategies.","marker":"[28]"}],"fun_headline_variants":["Phase probe jumps discontinuously as prior narrows","Optimal probe for phase estimation abruptly becomes squeezed vacuum","From displaced to squeezed: optimal probe jumps as prior tightens","Best Gaussian probe for phase: squeezed vacuum wins at low uncertainty","Discontinuous optimal probe: coherent to squeezed vacuum as prior shrinks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central claim rests on the numerical observation that every minimization of the average posterior variance over the full $(|\\alpha|, \\tau, \\varphi)$ parameter space lands on one of just two local-minimum families; if a third family were optimal in some part of the scanned energy and variance range, the abrupt-jump picture would collapse.","fun_headline_variants_meta":{"raw":{"variants":["Phase probe jumps discontinuously as prior narrows","Optimal probe for phase estimation abruptly becomes squeezed vacuum","From displaced to squeezed: optimal probe jumps as prior tightens","Best Gaussian probe for phase: squeezed vacuum wins at low uncertainty","Discontinuous optimal probe: coherent to squeezed vacuum as prior shrinks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001273,"raw_usage":{"total_tokens":5231,"prompt_tokens":993,"completion_tokens":4238,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":4156}},"tokens_in":609,"tokens_out":4238,"duration_ms":26972,"temperature":1.0,"reasoning_tokens":4156,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:52:30.931709+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a global optimization of the average posterior variance over $(|\\alpha|, \\tau, \\varphi)$ at fixed energy $E$ and a prior variance $\\sigma^2 \\in (0, 0.2]$, using a dense multi-start grid or a certified branch-and-bound method; a single point with lower average posterior variance than both the HUS and LUS optima would falsify the claim that those two families exhaust the optima. A cheaper partial check is to evaluate the Hessian at both putative optima and look for additional stationary points in the domain.","supporting_citations":[{"cited_title":"Monras, Optimal phase measurements with pure Gaussian states, Physical Review A 73 (2006)","cited_arxiv_id":null,"evidence_quote":"Supplies the local-analysis optimum: squeezed vacuum with the adapted angle saturates the quantum Cramér–Rao bound with Fisher information $8E(E+1)$, the anchor of the LUS family."},{"cited_title":"Wiebe and C","cited_arxiv_id":null,"evidence_quote":"Gives the opposite extreme used as the paper's starting point: with a uniform prior in a single round, coherent states outperform squeezed states of equal energy."},{"cited_title":"Friis, D","cited_arxiv_id":null,"evidence_quote":"Establishes that optimal phase measurements with pure Gaussian states use squeezed vacuum, supporting the identification of $|\\alpha|=0$ as the LUS optimum."},{"cited_title":"Pezz´ e and A","cited_arxiv_id":null,"evidence_quote":"Provides the general single-mode Gaussian quantum Fisher information formula the paper invokes to show the QFI is maximized at zero displacement."},{"cited_title":"Olivares and M","cited_arxiv_id":null,"evidence_quote":"Supplies the squeezed-vacuum-as-universal-probe intuition for why the LUS is locally powerful but fragile away from the estimate."},{"cited_title":"Gaiba and M","cited_arxiv_id":null,"evidence_quote":"Defines the phase-diffusion noisy estimation model whose optimal-probe behavior the paper compares with its own knowledge-dependent transition."},{"cited_title":"Teklu, S","cited_arxiv_id":null,"evidence_quote":"Provides the efficient Bayesian phase-estimation framework used to motivate and compare the repeated-measurement adaptive strategies."}],"review_version":1}