{"id":"42da772c-39a6-40f5-b39b-0b7e37bdbe40","arxiv_id":"2511.10184","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For a minimal linear-Gaussian estimation model, optimal memory use is governed by dimensionless discriminants ΘQ and ΘM, yielding sudden, nonmonotonic, scale-invariant phase transitions.","lead":"This paper derives analytical conditions for when a resource-limited organism should use memory versus only current observation when estimating a noisy environment. It explains why the switch between strategies can be sudden, nonmonotonic in sensory noise, and governed by a simple scaling law.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The analytical bounds ΘQ/ΘM bracket stationary-point existence, not the optimal phase boundary; the paper itself notes zero remains optimal after nonzero solutions emerge, so the central claim overreaches.","rationale":"The reader's weakest assumption identified the same core issue: the analysis relies on simultaneous stationary points of J without proving global optimality. My stress-test sharpens this into a concrete, structurally identifiable gap: the paper's own numerical description shows zero remains optimal after the nonzero branch emerges, so the analytical discriminants cannot demarcate the optimal phase boundary. This is not a hypothetical boundary/non-stationary failure; it is a known feature of the model's first-order transition. The concern is load-bearing because the abstract and conclusion present the bounds as characterizing the resource-induced phase transition, whereas the derivations actually characterize the existence of non-memoryless stationary solutions. The paper still contains valid analytical results and a clear numerical picture; the gap is addressable by adding a comparison of J values or by carefully reframing the claims. Therefore the existing CONDITIONAL verdict is appropriate, and my analysis does not move the verdict.","tokens_in":21191,"tokens_out":11088,"duration_ms":121987,"concrete_test":"Fix β (e.g., β=1) and sweep δ through the predicted ΘM=1 threshold. Numerically solve ∂J/∂Φzy=0 and ∂J/∂Φzz=0 to obtain the nonzero stationary branch, then evaluate J_zero(δ) and J_nz(δ) using Eqs. (S30)-(S31). Locate δ* where J_zero=J_nz. If δ* is larger than the δ at ΘM=1—as the paper's own 'zero remains optimal after emergence' statement suggests—then ΘM≥1 is not a sufficient condition for optimal memory use, and the phase-boundary claim must be weakened to a statement about stationary-strategy existence. Repeat for β=0.1, 1, and 10 to test robustness across the nonmonotonic region.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the resource-induced phase boundary is analytically characterized by ΘQ=1 (necessary) and ΘM≈1 (sufficient). But ΘQ and ΘM are derived solely from intersections of the stationarity conditions ∂J/∂Φzy=0 and ∂J/∂Φzz=0 (Supp. Secs. IV C-D), i.e., from stationary points of a nonconvex objective. The main text states that as Q increases, nonzero Πzx,Πzz 'emerge discontinuously, while Πzx=Πzz=0 remains optimal' and only later does the optimal solution switch. This is the classic spinodal-versus-coexistence distinction: the nonzero branch is born at the spinodal, but it becomes globally optimal only at the J-crossing where J(nonzero)=J(zero). Equations (14), (15), and the exact ΘT defined after Eq. (15) locate the spinodal, not the coexistence curve. Therefore, ΘM≥1 is sufficient for the existence of nonzero memory control gains, but it is not sufficient for memory-based estimation to be the optimal strategy. The actual phase boundary can occur at a larger δ, and the paper gives no analytical control over that J-crossing. The quadratic approximation in Supp. Eq. (S73) is a secondary issue; even the exact ΘT would only give the spinodal. Thus the abstract's claim of analytically characterizing the phase transition, as opposed to the existence of alternative stationary strategies, is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a one-dimensional LQG estimation model with a scalar memory z and a quadratic control cost M v^2. The authors reformulate the controller as linear in observation and memory with gains Φzy and Φzz, derive the exact stationary cost J(Φzy, Φzz), and obtain the optimal Φzy for fixed Φzz in closed form (Eq. 9), with a discriminant Θ (Eq. 10). For simultaneous optimization, they introduce dimensionless variables β=E/D and δ=QD/MF and derive a necessary condition ΘQ≥1 (Eq. 14) and an approximate sufficient condition ΘM≥1 (Eq. 15) for the existence of nonzero stationary solutions, bracketing an exact but intractable threshold ΘT. They claim this analytically explains the discontinuous, nonmonotonic, and Q/MF-scaling features of the phase transition between memoryless and memory-based optimal estimation.","tokens_in":21628,"tokens_out":11584,"duration_ms":119122,"significance":"If the optimality interpretation were justified, the paper would be a valuable analytical counterpart to previous numerical work, providing a simple two-parameter description and falsifiable boundaries. Strengths include the exact derivation of J and Φzy*, the analytical proof of the Q/MF scaling in the simultaneous problem (Supp. IV A), and the clear mechanism for discontinuity via simultaneous stationarity. The numerical validation uses an independent Riccati solver from prior work, so the analytical expressions are not fitted. However, the paper's principal claim about optimal strategies currently overreaches what the stationary-point analysis proves.","major_comments":[{"comment":"The discriminants ΘQ, ΘM, and ΘT are conditions for the existence of simultaneous stationary points of the nonconvex cost J, not for the global optimality of the memory-based branch. The paper itself states in the Numerical Results that as Q increases, nonzero gains 'emerge discontinuously, while Πzx=Πzz=0 remains optimal' and only later does the optimal solution switch. Thus ΘQ/ΘM/ΘT locate the spinodal, not the coexistence curve where J(nonzero)=J(zero). The abstract's claim of characterizing 'phase transitions in optimal estimation strategies' is therefore not established. Please either re-scope the claims to stationary-strategy emergence or add a global-optimality analysis (e.g., comparing J on the two branches) to connect ΘT to the actual optimal boundary.","section":"Numerical Results; Theoretical Analysis after Eq. (10)"},{"comment":"The sufficient condition ΘM is based on a second-order Taylor expansion of the quartic G(Φzz) around Φzz=0. The approximation is uncontrolled; the paper only compares the approximate and exact discriminants for default parameters in Fig. S3(b). Since ΘM≥1 defines the orange 'memory-based' region in Fig. 5(b), a failure of the quadratic approximation elsewhere would invalidate this sufficiency claim. The authors should either provide a rigorous bound on the approximation error over the relevant β-δ domain or explicitly state the domain where the sufficient condition is proven.","section":"Supp. Sec. IV D, Eq. (S73)"}],"minor_comments":[{"comment":"The title contains a typo: 'Estimatio n Strategies' should be 'Estimation Strategies'.","section":"Title"},{"comment":"The caption states that black and red dots are zero and nonzero solutions of the Riccati equation, but it does not indicate which branch is optimal. Since the paper's central claim concerns optimal strategies, please mark optimality explicitly or clarify that the figure only shows stationary solutions.","section":"Fig. 5 caption"},{"comment":"The bracketing argument uses the inequality (S55) that the solution of ∂J/∂Φzz=0 lies between the solutions of ∂JQ/∂Φzz=0 and ∂JM/∂Φzz=0 for fixed Φzy. The step from this inequality to the claim that intersection with the middle curve is equivalent to intersection with the upper/lower curves (necessary/sufficient conditions) is not fully proven in the supplement. The numerical results support the claim, but a more explicit proof would strengthen the paper.","section":"Supp. Sec. IV B-C"},{"comment":"The text uses 'ΘM ≈' in Eq. (15) but later refers to 'ΘM ≥ 1' as a definitive sufficient condition. Please clearly state that the sufficient condition is approximate, and mark the orange region in Fig. 5(b) accordingly.","section":"Eq. (15) and main text"}],"recommendation":"major_revision","confidential_remarks":"The paper is heavily self-referential (Refs. [30-34]); the novelty relative to those PRR papers should be stated more plainly. The main issue is that the analytical thresholds describe stationary-point existence, not the optimal phase boundary; the authors should either re-scope the claims or add a global-optimality analysis. This is fixable within the manuscript's scope, hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid piece of analysis worth taking seriously, but the abstract overstates what the theory actually pins down. The authors derive exact expressions for the cost J and for the optimal Φzy at fixed Φzz, and they prove a useful scaling relation (phase behavior depends only on β = E/D and δ = QD/MF). The necessary condition ΘQ ≥ 1 and the approximate sufficient condition ΘM ≥ 1 are genuine analytical results, and the paper is honest about the exact boundary ΘT being intractable. That is real progress over the earlier numerical work, and the bracket ΘQ > ΘT > ΘM is a useful mechanistic statement: the discontinuity comes from the simultaneous optimization of two gains, and the nonmonotonicity in sensory noise is visible in the discriminants themselves.\n\nThe soft spot is exactly what the stress-test note says. The discriminants are conditions for the intersection of stationarity equations, i.e., for the existence of nonzero stationary solutions of the nonconvex objective. They are not conditions for those solutions to be optimal. The main text says it in passing: as Q increases, the nonzero branch emerges while the zero solution remains optimal, and only later does the optimum switch. So ΘQ and ΘM bracket a spinodal, not the coexistence curve. The numerical phase boundaries in Fig. 2(c) — the actual optimality switch — live inside the 'ambiguous' green region of Fig. 5(b), where the paper offers no analytical control. The quadratic approximation in Supp. Eq. (S73) is a secondary issue; even an exact ΘT would only give the spinodal.\n\nThis is a significant gap, but not a fatal one. The paper's mechanistic claims — discontinuity from simultaneous optimization, scaling, nonmonotonicity of the discriminants — survive. What does not survive is the abstract's phrasing that the conditions for resource limitations to alter estimation strategies have been identified. The authors identify conditions for memory-based strategies to exist as candidate optima. The actual switch requires a J-crossing analysis that is not here.\n\nMy advice: send it to peer review, but ask for a revised abstract and a clear statement distinguishing existence from optimality. If the authors can also show numerically that the J-crossing obeys the same scaling and nonmonotonicity (or prove it in a corner of parameter space), the paper would be much stronger. As it stands, this is a useful partial theory for people working on resource-rational analysis and information bottlenecks, and a good reading-group piece on how spinodals and coexistence curves can be conflated in optimization problems.","headline":"The analytical bounds are real, but they locate where nonzero strategies first appear, not where they become optimal — the paper's central claim outruns its own math.","tokens_in":21979,"tokens_out":3493,"would_cite":true,"duration_ms":38142,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The switch between memoryless and memory-based estimation is governed by two dimensionless parameters, β = E/D and δ = QD/MF, with a necessary condition Θ_Q ≥ 1 and a sufficient condition Θ_M ≥ 1.","keywords":["resource-limited estimation","phase transitions","optimal control theory","linear-quadratic-Gaussian","internal memory","sensory reliability","dimensionless parameters","Riccati equation"],"falsifier":"Solve the full observation-based Riccati equation over a dense grid of $(\\beta, \\delta)$, and check whether any point with $\\Theta_Q < 1$ still exhibits nonzero memory gains, or any point with $\\Theta_M \\geq 1$ still exhibits zero gains; either violation would falsify the bracketing claim. In the ambiguous middle band, a global minimization of $J$ over $(\\Phi_{zy}, \\Phi_{zz})$ should reveal whether the nonzero stationary point is ever beaten by the boundary $\\Phi_{zz} = 0$—if so, the stationarity criterion is the weak link.","tokens_in":21115,"feed_emoji":"🧠","tokens_out":8560,"duration_ms":82669,"temperature":0.7,"texified_at":"2026-08-05T20:38:28.831283+00:00","pith_summary":"Organisms estimating a fluctuating environment must decide whether to store past observations in an internal memory—a decision that costs energy and is subject to internal noise. This paper analyzes a minimal linear-Gaussian model of that decision and claims that the optimal strategy is determined by two dimensionless numbers: $\\beta = E/D$, the ratio of sensory noise to environment volatility, and $\\delta = QD/MF$, the ratio of energy availability to memory cost and memory noise. The paper derives a necessary condition ($\\Theta_Q \\geq 1$) and an approximate sufficient condition ($\\Theta_M \\geq 1$) for memory use, which bracket the exact phase boundary $\\Theta_T$ and reproduce three numerical features: the transition is discontinuous, it is nonmonotonic in sensory noise (memory is optimal only at intermediate noise), and the boundary obeys the scaling $Q/MF$. A sympathetic reader would care because these are parameter-free predictions that can be tested against biological experiments on when organisms switch between reactive and memory-based behavior.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":10937,"prompt_tokens":888,"completion_tokens":10049,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":888,"completion_tokens_details":{"reasoning_tokens":9141}},"feed_headline":"Two dimensionless ratios set the memory/no-memory switch","feed_subtitle":"Necessary and sufficient conditions bracket the exact boundary, explaining discontinuous, nonmonotonic onset.","key_machinery":"The central object is the reformulated memory control law $v_\\Phi(y,z) = -\\Phi_{zy} y - \\Phi_{zz} z$, where $\\Phi_{zy}$ is the gain that encodes the current observation into memory and $\\Phi_{zz}$ is a negative-feedback gain that stabilizes the memory state. The proof mechanism is the decomposition of the objective $J$ into estimation error $J_Q$ and control cost $J_M$, together with the observation that the stationary equations $\\partial J_Q / \\partial \\Phi_{zz} = 0$ and $\\partial J_M / \\partial \\Phi_{zz} = 0$ respectively provide upper and lower bounds for the true stationary equation $\\partial J / \\partial \\Phi_{zz} = 0$. Because the true equation is analytically intractable, these bounds convert the problem into two tractable intersection conditions, yielding the necessary discriminant $\\Theta_Q$ and t","core_discovery":"The paper establishes that resource-induced phase transitions in optimal estimation strategies are analytically tractable in a minimal linear-quadratic-Gaussian model. By rewriting the memory control as a linear function $v_\\Phi = -\\Phi_{zy} y - \\Phi_{zz} z$ and optimizing the two gains jointly, the paper shows that the emergence of nonzero memory gains is governed by the dimensionless parameters $\\beta = E/D$ and $\\delta = QD/MF$. The exact phase boundary $\\Theta_T$, defined by the simultaneous stationarity conditions $\\partial J / \\partial \\Phi_{zy} = 0$ and $\\partial J / \\partial \\Phi_{zz} = 0$, is bracketed by two explicit discriminants: $\\Theta_Q = \\frac{2\\beta^2 \\delta}{(1+2\\beta)^2 (1+4\\beta)}$ is a necessary condition (memory cannot appear when $\\Theta_Q < 1$), and $\\Theta_M \\approx \\frac{\\beta^2 \\delta}{(1+2\\beta)^2 \\left(1+6\\beta + \\sqrt{4\\beta(4+13\\beta)}\\right)}$ i","pith_inferences":["The discriminants could serve as a design rule for synthetic cellular circuits or neural controllers—choose parameters so that Θ_M ≥ 1 to favor memory integration, and stay below Θ_Q < 1 to keep the system reactive; this extends the paper's diagnostic to an engineering prescription.","The sufficient discriminant Θ_M rests on a quadratic truncation (Supp. Eq. S73); a natural test is to compute higher-order corrections or numerically map the exact Θ_T surface to see whether the approximation degrades for extreme β (very large or very small E/D), which the paper does not address.","The paper identifies memory use with a stationary point of J, but the LQG cost is nonconvex; if for some parameters the global optimum lies on the boundary Φ_zz = 0 or at a non-stationary point, the predicted phase boundary would shift, and a careful global optimization of (Φ_zy, Φ_zz) could detect this.","Because the model is LQG with quadratic costs, the same β–δ structure may extend to delayed prediction or tracking; replacing (x_t − \\hat{x}_t)² with a prediction error at lag τ might simply renormalize the effective δ, which would be a concrete testable extension."],"forward_implications":["The phase boundary is fully determined by β and δ; any change in Q, M, D, or F affects the boundary only through these two combinations, which is why the numerical boundaries collapse onto a single Q/MF curve.","Memory gains emerge discontinuously at the transition because the simultaneous stationarity condition for the two gains (Φ_zy, Φ_zz) has no continuous zero-crossing branch—the intersections appear at nonzero Φ_zy values.","The discriminants predict that memory is used only for intermediate sensory noise: when E is much smaller or much larger than D, Θ_Q and Θ_M fall below 1 and memoryless estimation wins.","Increasing environmental volatility ε is equivalent to replacing δ by δ/ε, so the framework predicts a monotonic transition from memory-based to memoryless as volatility increases, matching experiments.","The bracketing Θ_Q ≥ Θ_T ≥ Θ_M means that, for a given organism or circuit, measuring β and δ immediately tells whether memory cannot be optimal (Θ_Q < 1) or must be optimal (Θ_M ≥ 1); only in the narrow middle band is the exact answer ambiguous."],"fun_headline_variants":["Two ratios fix the memory/no-memory boundary","Resource limits set estimation phase switch","Exact phase boundary via two dimensionless parameters","Memory strategies emerge at a twin-parameter threshold","Estimation phase transition pinned by two ratios"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the onset of memory use coincides with the existence of a stationary point of the cost $J$ that is jointly optimal in both memory gains, and that the sufficient condition can be captured by a quadratic truncation of a quartic equation; if the true optimum lies at a boundary or the approximation fails, the claimed phase boundary shifts.","fun_headline_variants_meta":{"raw":{"variants":["Two ratios fix the memory/no-memory boundary","Resource limits set estimation phase switch","Exact phase boundary via two dimensionless parameters","Memory strategies emerge at a twin-parameter threshold","Estimation phase transition pinned by two ratios"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00014,"raw_usage":{"total_tokens":976,"prompt_tokens":698,"completion_tokens":278,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":442,"completion_tokens_details":{"reasoning_tokens":212}},"tokens_in":442,"tokens_out":278,"duration_ms":3726,"temperature":1.0,"reasoning_tokens":212,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T22:28:18.876597+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the full observation-based Riccati equation over a dense grid of $(\\beta, \\delta)$, and check whether any point with $\\Theta_Q < 1$ still exhibits nonzero memory gains, or any point with $\\Theta_M \\geq 1$ still exhibits zero gains; either violation would falsify the bracketing claim. In the ambiguous middle band, a global minimization of $J$ over $(\\Phi_{zy}, \\Phi_{zz})$ should reveal whether the nonzero stationary point is ever beaten by the boundary $\\Phi_{zz} = 0$—if so, the stationarity criterion is the weak link.","supporting_citations":[],"review_version":1}