{"id":"4e37e584-3064-44fa-9ceb-237fb71e4831","arxiv_id":"2605.24628","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Presents a memory uncertainty relation bounding short-term memory from below in dynamical systems, with harmonic memory as the lower bound, and analyzes equality cases plus noise-induced memory under regularization in noisy reservoir systems.","lead":"The paper derives an inequality bounding short-term memory in dynamical systems from below, framed as an uncertainty relation with input-induced state fluctuations, and defines harmonic memory as the achievable lower bound via readout weights. It also identifies noise-induced memory under regularization in recurrent networks, which may inform limits and design in reservoir computing.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"The inequality's validity as a general bound hinges on the precise (but unspecified in abstract) definitions of short-term memory capacity and input-induced state fluctuation size.","rationale":"The reader's weakest_assumption correctly isolates the missing quantitative definitions as the point that prevents verification of the bounding claim. Because the full text is now referenced but the concern is precisely about whether those definitions support a non-trivial general result, the UNVERDICTED status is unchanged pending that check.","tokens_in":1728,"tokens_out":323,"duration_ms":32443,"concrete_test":"From the theory section, extract the exact expressions for the memory capacity C and fluctuation measure F; substitute into the claimed inequality and verify whether it follows by direct algebraic manipulation (e.g., Cauchy-Schwarz or eigenvalue bound) with no additional assumptions on the dynamics or readout.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is an inequality that lower-bounds a memory measure by a fluctuation measure, achieved by a particular readout (harmonic memory). This requires explicit quantitative definitions of both quantities (e.g., whether memory capacity is the standard linear readout sum-of-squares or a different functional, and whether fluctuation size is variance, L2 norm of the response operator, or another norm). Without those, one cannot check whether the derivation is an identity, a Cauchy-Schwarz application, or an assumption-dependent result, nor whether equality cases are non-trivial. The paper states it holds for general dynamical systems but only examines it in reservoir networks; the definitions are therefore the least secure link.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript presents an inequality that lower-bounds a quantitative measure of short-term memory capacity in dynamical systems by a measure of the size of state fluctuations induced by input signals. This is interpreted as an uncertainty relation, with the bound attained exactly by a specific linear readout weight vector termed 'harmonic memory.' The authors derive and examine the relation analytically and numerically in random recurrent reservoir networks driven by input noise, documenting cases of exact equality, asymptotic equality, and strict inequality. They further analyze the effects of state-space regularization, identifying a 'noise-induced memory' phenomenon under certain regularization strengths, and note that the uncertainty relation does not hold in general for the regularized memory measure.","tokens_in":1859,"tokens_out":807,"duration_ms":20211,"significance":"If the inequality is non-tautological and holds under the stated definitions for a reasonably broad class of systems, the result would supply a concrete lower bound on memory performance in recurrent dynamical systems and a mechanistic explanation for suboptimal but analytically tractable memory (harmonic memory). The explicit treatment of noise-induced memory under regularization and the classification of equality cases add concrete value for reservoir-computing applications. The work is strongest where it supplies reproducible numerical protocols and explicit constructions; its broader impact hinges on whether the definitions of memory capacity and fluctuation size are standard or ad-hoc and on the scope of the derivation beyond the reservoir networks studied.","major_comments":[{"comment":"The abstract and introduction claim the inequality applies to general dynamical systems, yet all analytic derivations and numerical tests are performed exclusively on random recurrent networks with linear readouts. The manuscript should clarify whether the derivation uses only properties common to all dynamical systems or relies on the specific structure of reservoir state updates (e.g., the echo-state property or the form of the recurrent weight matrix).","section":"Introduction and §3"},{"comment":"The central inequality is described as bounding memory capacity from below by a fluctuation measure, with equality achieved by harmonic memory. Without the explicit definitions of both quantities (e.g., whether memory capacity is the standard sum-of-squared correlations or a different functional, and whether fluctuation size is an L2 norm of the response operator or input-induced variance), it is impossible to determine whether the bound is a non-trivial consequence of Cauchy-Schwarz or an identity by construction. The manuscript must supply these definitions and the derivation steps in a dedicated section.","section":"§2 (Definitions) and §3 (Derivation)"},{"comment":"The claim that 'the memory uncertainty relation does not hold in general for the regularized memory and harmonic memory' is presented as a consequence of the noise-induced memory mechanism. This statement requires a precise counter-example or analytic condition under which the regularized quantities violate the inequality; the current discussion appears to rest on numerical observation alone.","section":"§5 (Regularization and noise-induced memory)"}],"minor_comments":[{"comment":"Notation for the memory capacity measure and the fluctuation measure should be introduced once with explicit symbols and then used consistently; several passages refer to 'the memory' without distinguishing the unregularized, regularized, and harmonic variants.","section":"Throughout"},{"comment":"Figure captions should state the precise parameter values (network size, spectral radius, noise variance, regularization strength) used in each panel so that the equality/asymptotic/strict cases can be reproduced from the caption alone.","section":"Figure captions"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is submitted to nlin.AO; the topic sits at the intersection of dynamical systems and reservoir computing, which is appropriate, but the absence of explicit definitions in the abstract raises a flag about whether the core contribution is a genuine bound or a re-expression of existing linear-algebra identities. I recommend requesting the definitions and derivation as a condition for further review."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments, which help clarify the scope and presentation of our results. We address each major comment below.","responses":[{"response":"The inequality is derived from the Cauchy-Schwarz inequality applied to the inner product between the readout vector and the state response operator, which is a general property once a well-defined state response to inputs is assumed. However, the explicit analytic expressions and all numerical tests rely on the linear state update and echo-state property of random recurrent networks. We will revise the introduction and §3 to state the assumptions explicitly and limit the general claim to systems admitting a linear readout and bounded state response.","revision_made":"yes","referee_comment":"[Introduction and §3] The abstract and introduction claim the inequality applies to general dynamical systems, yet all analytic derivations and numerical tests are performed exclusively on random recurrent networks with linear readouts. The manuscript should clarify whether the derivation uses only properties common to all dynamical systems or relies on the specific structure of reservoir state updates (e.g., the echo-state property or the form of the recurrent weight matrix)."},{"response":"We agree that the definitions and derivation must be stated explicitly. Memory capacity is the sum of squared correlations (standard definition), and the fluctuation measure is the squared L2 norm of the input-induced state deviation. The bound follows directly from Cauchy-Schwarz on these quantities, with equality when the readout equals the normalized fluctuation vector. We will insert a new dedicated subsection in §2 with the full definitions followed by the step-by-step derivation in §3.","revision_made":"yes","referee_comment":"[§2 (Definitions) and §3 (Derivation)] The central inequality is described as bounding memory capacity from below by a fluctuation measure, with equality achieved by harmonic memory. Without the explicit definitions of both quantities (e.g., whether memory capacity is the standard sum-of-squared correlations or a different functional, and whether fluctuation size is an L2 norm of the response operator or input-induced variance), it is impossible to determine whether the bound is a non-trivial consequence of Cauchy-Schwarz or an identity by construction. The manuscript must supply these definitions and the derivation steps in a dedicated section."},{"response":"The current manuscript presents the violation through numerical observation under varying regularization strengths. We will add an analytic derivation showing that the inequality fails when the regularized covariance matrix rotates the effective memory vector away from the harmonic direction by an angle whose cosine falls below the normalized fluctuation term; a concrete counter-example with explicit regularization parameter and noise variance will be included in the revised §5.","revision_made":"partial","referee_comment":"[§5 (Regularization and noise-induced memory)] The claim that 'the memory uncertainty relation does not hold in general for the regularized memory and harmonic memory' is presented as a consequence of the noise-induced memory mechanism. This statement requires a precise counter-example or analytic condition under which the regularized quantities violate the inequality; the current discussion appears to rest on numerical observation alone."}],"tokens_in":1541,"tokens_out":652,"duration_ms":22527,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The central claim is an inequality that lower-bounds a measure of short-term memory by a measure of input-driven state fluctuations in dynamical systems, with equality achieved by a specific readout they call harmonic memory. They also report that moderate input noise can increase memory capacity when state regularization is present, an effect they name noise-induced memory, and show that the uncertainty relation breaks under regularization.\n\nThe paper checks the inequality analytically and numerically across several reservoir networks, covering exact equality, asymptotic equality, and strict inequality cases. The regularization experiments illustrate how the fluctuation structure changes and how noise interacts with it. These concrete checks in random recurrent networks are the most useful part of the work.\n\nThe soft spot is that the abstract supplies no explicit formulas for the memory capacity or the fluctuation size. Without those, it is impossible to tell whether the bound is a non-trivial result or follows directly from the chosen definitions, for instance via Cauchy-Schwarz or a similar identity. The claim is stated for general dynamical systems, yet all analysis stays inside reservoir networks, so the generality is not demonstrated. The account of why the relation fails under regularization is plausible but would need to be tight to rule out artifacts from the regularization term.\n\nThis is for people working on reservoir computing and short-term memory bounds in nonlinear dynamics. A reader who already knows the standard memory capacity definitions in the field could extract the numerical patterns and the noise-regularization interaction.\n\nIt deserves peer review because the claims are specific enough to be tested and the topic is relevant to practical reservoir design, even if the definitions and scope need tightening.","headline":"The paper introduces a lower bound on short-term memory via an uncertainty relation with state fluctuations in recurrent networks, realized by harmonic memory, plus a noise-induced memory effect under regularization, but the abstract leaves the core definitions unspecified.","tokens_in":2385,"tokens_out":409,"would_cite":false,"duration_ms":21269,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"An inequality sets a lower bound on short-term memory capacity of dynamical systems using input-induced state fluctuations.","keywords":["memory uncertainty relation","harmonic memory","short-term memory capacity","reservoir computing","random recurrent networks","noise-induced memory","dynamical systems","input noise"],"falsifier":"A dynamical system in which measured short-term memory capacity falls below the numerical value of the lower bound computed from its observed state fluctuations under the same inputs.","tokens_in":2590,"feed_emoji":"","tokens_out":711,"duration_ms":31106,"temperature":0.7,"pith_summary":"The paper derives an inequality that places a minimum value on the short-term memory capability of any dynamical system, expressed in terms of the size of fluctuations in its internal states caused by input signals. This bound is framed as an uncertainty relation and is attained exactly by a particular choice of readout weights called harmonic memory. The authors analyze and simulate the inequality in several reservoir systems driven by input noise, identifying cases of exact equality, asymptotic equality, and strict inequality. They further show that state-space regularization can generate additional memory from noise, a phenomenon they term noise-induced memory, which causes the uncertainty relation to fail.","feed_headline":"Short-term memory in networks bounded below by state fluctuations","feed_subtitle":"An inequality gives the minimum memory capacity set by input-driven changes; harmonic memory reaches it and noise can exceed it under regula","key_machinery":"The memory uncertainty relation inequality, which lower-bounds short-term memory capacity by a measure of input-induced state fluctuation size, with the bound achieved by harmonic memory readout weights.","core_discovery":"We present an inequality that bounds the short-term memory capability of dynamical systems from below. It can be interpreted as an uncertainty relation between a measure of short-term memory and that of the size of state fluctuations induced by input signals. The lower bound can be achieved by a readout weight and thus represents a suboptimal memory called harmonic memory. We examine analytically and numerically the inequality in a number of reservoir systems subject to input noise. We illustrate cases in which equality is achieved exactly, equality holds asymptotically, and the inequality is strict. We also study the effect of a state-space regularization to elucidate the inequality in term","pith_inferences":["Network designers could deliberately adjust fluctuation levels to enforce or relax the memory lower bound.","The same fluctuation-memory trade-off might appear in biological or physical systems outside reservoir computing.","Identifying the precise conditions that trigger noise-induced memory could allow controlled memory enhancement."],"forward_implications":["The bound applies to the memory performance of any dynamical system used for temporal tasks.","Harmonic memory supplies an explicit readout construction that saturates the bound in some systems.","Input noise combined with regularization can produce memory capacity above the bound.","The uncertainty relation ceases to apply once regularization activates noise-induced memory."],"fun_headline_variants":["Memory uncertainty relation sets lower bound on short-term memory","Harmonic memory achieves fluctuation bound in random recurrent networks","State fluctuations impose minimum on network short-term memory capacity","Input noise induces extra memory in regularized reservoir systems"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The inequality and its interpretation depend on the particular quantitative definitions chosen for short-term memory capacity and the size of state fluctuations.","fun_headline_variants_meta":{"raw":{"variants":["Memory uncertainty relation sets lower bound on short-term memory","Harmonic memory achieves fluctuation bound in random recurrent networks","State fluctuations impose minimum on network short-term memory capacity","Input noise induces extra memory in regularized reservoir systems"]},"model":"grok-4.3","cost_usd":0.009259,"raw_usage":{"total_tokens":4059,"prompt_tokens":657,"num_sources_used":0,"completion_tokens":61,"cost_in_usd_ticks":92590500,"prompt_tokens_details":{"text_tokens":657,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3341,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":657,"tokens_out":61,"duration_ms":38203,"temperature":1.0,"reasoning_tokens":3341,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-30T12:11:33.535892+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A dynamical system in which measured short-term memory capacity falls below the numerical value of the lower bound computed from its observed state fluctuations under the same inputs.","supporting_citations":[],"review_version":1}