{"id":"1e2d20d9-6a17-408d-9fbf-5a3236b2ee70","arxiv_id":"2505.02491","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Non-Markovian quantum reservoir computers can revive memory of past inputs, while Markovian reservoirs are claimed to obey an exponential memory decay bound.","lead":"The paper argues that every Markovian quantum reservoir computer loses access to old inputs exponentially fast, and that non-Markovian dynamics can revive that memory. It demonstrates the effect in a residual-style quantum reservoir and in an embedded auxiliary-qubit design, with numerical gains on short-term and chaotic time-series tasks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (6) does not follow from Eq. (4): exponential decay of Volterra coefficients does not imply exponential decay of capacity, so the central memory theorem is unproven.","rationale":"I agree with the reader's conditional verdict but identify a different weakest point than the one formally listed. The reader's 'weakest_assumption' concerns the echo-state and contractivity assumptions restricting theorem scope; the more load-bearing gap is the derivation of Eq. (6) from Eq. (4). This gap affects the main theorem's conclusion, not just its scope: even under all stated assumptions (unique stationary state, ||T||<1, fading memory, Volterra expansion), exponential decay of Volterra coefficients does not imply exponential decay of capacity, because capacity is a normalized squared projection and the optimal readout can amplify arbitrarily small amplitudes. The reader's rationale does flag that 'the central analytical step from Volterra coefficient decay to capacity decay is not justified as written,' so the concern is partially acknowledged, but the formal weakest-assumption field points elsewhere. The concrete test would settle whether Eq. (6) is quantitatively correct or whether a corrected bound (e.g., O(e^{-2aτ}) or a cutoff behavior) is needed. The numerical figures are consistent with exponential memory loss in the specific unoptimized models used, but they do not test the claimed rate or the theorem's generality. Thus the verdict remains CONDITIONAL: the qualitative resource claim about non-Markovian enhancement is plausible and numerically supported, but the central analytical memory bound must be corrected or re-derived before acceptance.","tokens_in":13676,"tokens_out":11209,"duration_ms":133548,"concrete_test":"On the Markovian model of Eqs. (9)-(10) with N=1 (or N=3 matching Fig. 2), set λ=1 and compute the optimal linear readout capacity C for targets ŷ_k = s_{k-τ} and ŷ_k = s²_{k-τ}, using all independent single-qubit observables, a long training set (e.g., 10^4 steps), and no regularization. Extract the decay rate r from C ∝ e^{-rτ} and compare with a = -ln max_s ||T(s)||, computed by diagonalizing L(s) on a grid of s values. If r ≈ 2a, the rate in Eq. (6) is wrong; if C saturates at 1 for small τ and then drops sharply, the functional form is not a simple exponential; if C remains O(1) for all τ, the central memory theorem is false. Also verify the counterexample y_k = e^{-aτ}s_{k-τ} gives C=1 under Eq. (5).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim is Eq. (6), C[ŷ(τ)] = O(e^{-aτ}), stated as 'combining Eqs. (4) and (5)'. This combination is a non-sequitur. Eq. (4) bounds the magnitude of each Volterra kernel f_n by O(e^{-aτ_n}). Eq. (5) defines capacity as 1 - MSE_T/⟨y²⟩_T, i.e., a normalized squared-projection of the target onto the linear span of the output features after optimal linear readout. The optimal readout can rescale features arbitrarily; a target that lies in the span is reconstructed exactly (C=1) no matter how small its coefficient in any single Volterra term is. Even in a noisy or finite-sample setting, the variance explained scales with the square of the projection, so if the projection is dominated by a kernel of amplitude e^{-aτ}, the capacity decays as e^{-2aτ}, not e^{-aτ}. The paper provides no argument that the projection magnitude is O(e^{-aτ}); it simply asserts the combination. A concrete counterexample to the implication: a reservoir output y_k = e^{-aτ} s_{k-τ} satisfies |f_1(τ)| = O(e^{-aτ}) but has capacity C=1 for the target s_{k-τ} for every τ under Eq. (5). Thus the claimed exponential decay of memory capacity is not established by the given proof, and the stated rate in Eq. (6) is likely incorrect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims a general theorem that Markovian quantum reservoir computers (QRC) have exponentially decaying memory capacity, and argues that quantum non-Markovianity can overcome this limitation. It introduces a 'Quantum Residual Reservoir' with an explicit delayed-state recurrence and an embedding scheme based on partial swaps and depolarizing channels, and supports the claims with numerical experiments on short-term memory tasks and Mackey-Glass chaotic time-series forecasting.","tokens_in":13973,"tokens_out":10206,"duration_ms":137359,"significance":"If the central memory bound were established, the paper would be an important contribution: it points to a largely underexplored design axis for QRC and gives a concrete, tunable embedding that interpolates between Markovian and non-Markovian operation. The numerical work is substantial, with 100 random realizations, washout/training/test protocol, and multiple benchmarks. However, the main analytical claim is currently under-supported, so the significance is conditional on a repair or careful restatement of the theorem.","major_comments":[{"comment":"The inference from Eq. (4) to Eq. (6) is not valid as written. Eq. (4) bounds the absolute value of each Volterra kernel f_n, whereas Eq. (5) defines capacity as a normalized squared error after an unconstrained linear readout. The optimal readout can rescale features by arbitrary factors, so an exponentially small feature can be amplified to achieve unit capacity. Concretely, if a reservoir observable has expectation e^{-aτ}s_{k-τ}, the linear readout can choose a weight of order e^{aτ} and obtain C=1 for the target s_{k-τ} for every τ, consistently with |f_1(τ)|=O(e^{-aτ}). Thus the claimed exponential decay of memory capacity is not established; a proof would need to bound the projection of the target onto the readout feature space, or impose a bound on the readout weights, a noise model, or regularization. Since this is the paper's central theoretical result, it must be fixed or explicitly scoped down.","section":"Memory bounds of Markovian reservoirs, Eq. (6)"},{"comment":"The memory revival at τ=τ_E is put into the evolution rule by hand. The recurrence ρ_{k+1}=e^{L(s_{k+1})Δt}[λρ_k+(1-λ)ρ_{k-τ_E}] contains an explicit ρ_{k-τ_E} term, so the numerical revival in Fig. 2(a) is a direct consequence of the recursion rather than an emergent prediction of non-Markovian dynamics. The manuscript should present Eq. (8) as a designed or toy model, and the analysis of how non-Markovianity enhances memory should be tied to the embedded model, where the memory of the auxiliary system arises from the partial-swap interaction.","section":"Quantum Residual Reservoir, Eq. (8)"},{"comment":"The statement that the result applies to 'all reservoir models proposed so far' is broader than the assumptions used in the proof. The derivation in the Supplemental assumes a unique stationary state for every input, contractivity ∥T(s)∥<1, and a fading-memory property guaranteeing convergence of the Volterra expansion. These assumptions are asserted, not verified for the published QRC models cited in the paper. Please either verify the assumptions for the cited models or explicitly scope the theorem to the class of Markovian QRCs satisfying those assumptions.","section":"Introduction and Supplemental Sec. I"}],"minor_comments":[{"comment":"Eq. (8) states λ∈[0,1), but Fig. 2(a) labels the curve λ=1 as the Markovian case; please clarify how the boundary value is treated.","section":"Eq. (8) and Fig. 2"},{"comment":"The formula for the non-Markovianity measure appears to contain a typo: D(ρ_{k,1},ρ_{k,1}) should presumably be D(ρ_{k,1},ρ_{k,2}).","section":"Supplemental Eq. (S4)"},{"comment":"The phrase 'in descending order with respect to Ω' is ambiguous; the listed MSE values are 1.8×10⁻², 6.8×10⁻³, 2×10⁻² for Ω=1, 0.5, 0, so 'in order of decreasing Ω' would be clearer.","section":"Fig. 3 caption"},{"comment":"There are several typos: 'Sante F e' should be 'Santa Fe', and 'non-Markocian' in the reference to the Supplemental material should be 'non-Markovian'.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The gap in the derivation of Eq. (6) is substantial and may require the authors to weaken or reframe the central claim. The numerical and modeling contributions are still valuable, so I would not reject outright, but the revision must address the proof of the memory bound directly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what you should know: this paper claims to prove that all Markovian quantum reservoir computers have exponentially decaying memory, and that non-Markovian dynamics can revive information at a chosen delay. The numerical revival is real, and the embedding scheme is a useful tool. But the central analytical step, Eq. (6), is a non-sequitur. The fact that Volterra coefficients decay exponentially does not imply that memory capacity decays exponentially. The optimal linear readout can rescale features arbitrarily, so a target that lies in the span of an exponentially small component can be reconstructed perfectly. A concrete counterexample: a reservoir output y_k = e^{-aτ} s_{k-τ} has |f_1(τ)| = O(e^{-aτ}) but capacity C = 1 for the target s_{k-τ} for every τ. The proof needs an additional argument—noise, weight regularization, or a bound on the readout norm—to go from coefficient decay to capacity decay. Without that, the paper's main theorem is unproven.\n\nWhat is genuinely new and good: the paper is, as far as I know, the first to frame non-Markovianity as a tunable resource specifically for quantum reservoir computing. The embedding method—coupling reservoir qubits to auxiliaries via partial swaps and tuning the interaction with depolarizing channels—is concrete and experimentally plausible. The numerical evidence for memory revival in the embedded model (Fig. 2 and S3) is independent of the by-construction residual reservoir of Eq. (8), and that matters. The authors also honestly report that non-Markovianity does not always help (Santa Fe task), which is the right kind of nuance. The Volterra coefficient decay proof in the Supplemental Material is fine as far as it goes.\n\nSoft spots, in proportion: the proof gap in Eq. (6) is load-bearing, not minor. The theorem's scope is also narrower than the abstract suggests: it requires a unique stationary state, a contractive map, and fading memory for every input, which are not verified for all previously proposed QRC architectures. The residual reservoir of Eq. (8) revives memory by construction—there is a literal ρ_{k−τ_E} term in the update—so that demonstration is not an emergent effect; the embedding model carries the weight. No code or data are released, so the numerics are not independently reproducible as-is.\n\nWho is this for? The quantum reservoir computing and quantum machine learning community. It is a paper a serious referee should see, not a desk reject. The right outcome is likely major revision: either fix the derivation with a regularized setting or soften the theorem to a conjecture, verify the assumptions on published models, and release the code. I would bring it to a reading group and would probably cite the embedding construction even while being skeptical of the proof.\n\nBottom line: worth refereeing, but the authors need to do real work on the central claim before it is sound.","headline":"A genuinely interesting non-Markovian reservoir construction, but the paper's central memory theorem is not proven as written—the capacity bound does not follow from the Volterra coefficient decay.","tokens_in":14541,"tokens_out":2541,"would_cite":true,"duration_ms":34833,"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":"Markovian quantum reservoir computers are proven to lose past-input memory exponentially; switching to non-Markovian dynamics revives it and improves forecasting.","keywords":["quantum reservoir computing","non-Markovian dynamics","memory capacity","Volterra series","fading memory","echo state property","chaotic time-series forecasting","open quantum systems"],"falsifier":"Measure the short-term memory capacity $C[\\hat{y}(\\tau)]$ of any Markovian reservoir satisfying the paper's assumptions (unique stationary state and $\\|T(s)\\|<1$) out to large delays $\\tau$; if the capacity decays polynomially or stays flat instead of exponentially, the bound in Eq. (6) is false. A second check is to exhibit a Markovian reservoir from the published literature that violates the unique-stationary-state assumption yet still shows exponential decay, which would narrow the claimed universality.","tokens_in":13454,"feed_emoji":"🔄","tokens_out":10189,"duration_ms":107538,"temperature":0.7,"pith_summary":"This paper targets a fundamental limit of quantum reservoir computers: any reservoir whose update is Markovian—the rule used by every architecture proposed so far—loses information about past inputs exponentially with the delay. The authors prove that the Volterra coefficients of a Markovian reservoir output scale as $O(e^{-a\\tau_n})$ and that the capacity to reconstruct a target depending on an input $\\tau$ steps back scales as $O(e^{-a\\tau})$. They then show that non-Markovian update rules, in which the new reservoir state also depends on an older state $\\rho_{k-\\tau_E}$, revive the memory at the delay $\\tau_E$ and can outperform Markovian reservoirs on tasks that need both short- and long-term correlations. A concrete embedding scheme using auxiliary qubits and partial swaps gives a tunable knob between Markovian and non-Markovian operation, and the simulations show intermediate non-Markovianity gives the best chaotic forecasting.","feed_headline":"Exponential memory loss in quantum reservoirs can be reversed","feed_subtitle":"Tuning non-Markovianity lets quantum reservoirs recall both recent and long-past inputs, improving chaotic forecasts.","key_machinery":"The machine doing the work is the decomposition of the Markovian propagator into a stationary-state projector and a contraction: $e^{\\mathcal{L}(s)\\Delta t}=S(s)+T(s)$, with $S(s)S(w)=S(s)$ and $S(s)T(w)=0$. That identity forces every term in the Volterra expansion that reaches back $\\tau_n$ steps to contain $\\tau_n$ factors of $T$, each bounded by $\\max_s\\|T(s)\\|<1$, hence the exponential bound. On the non-Markovian side, the load-bearing object is the residual update rule $\\rho_{k+1}=e^{\\mathcal{L}(s_{k+1})\\Delta t}[\\lambda\\rho_k+(1-\\lambda)\\rho_{k-\\tau_E}]$, a quantum analogue of residual and dilated skip connections, with $\\lambda$ controlling how much stale state $\\rho_{k-\\tau_E}$ is mixed back; the embedding variant uses partial swaps with auxiliary qubits and depolarizing noise $\\Omega$ to realize the same effect in a larger Markovian model, giving a physical tuning knob for non-Markovianity.","core_discovery":"On the paper's own terms, the central discovery is that memory in quantum reservoir computing is not a fixed resource but is controlled by the dynamical character of the reservoir. For a Markovian reservoir with update rule $\\rho_{k+1}=e^{\\mathcal{L}(s_{k+1})\\Delta t}\\rho_k$, the evolution map splits into a stationary-state projector $S(s)$ and a contracting part $T(s)$ with $\\|T(s)\\|<1$; because $S(s)T(w)=0$, any memory term involving an input $\\tau_n$ steps in the past must carry a product of $T$'s, forcing $|f_n(\\ldots)|=O(e^{-a\\tau_n})$ and, for delay-$\\tau$ targets, capacity $C[\\hat{y}(\\tau)]=O(e^{-a\\tau})$. Replacing the update by a non-Markovian rule such as $\\rho_{k+1}=e^{\\mathcal{L}(s_{k+1})\\Delta t}[\\lambda\\rho_k+(1-\\lambda)\\rho_{k-\\tau_E}]$ introduces a controlled revival of old states at delay $\\tau_E$, so information that would be untraceable in the Markovian window becomes retrievable again. The same tunability is realized in an embedding with auxiliary qubits (partial swap followed by depolarizing noise of strength $\\Omega$), and the numerics show capacity revivals and superior chaotic forecasting at intermediate $\\Omega$.","pith_inferences":["Beyond the paper: the proof's engine—the $S(s)T(w)=0$ orthogonality—does not use the details of the Lindblad structure, so the same exponential-memory wall should apply to any input-driven contracting map with a unique fixed point, quantum or classical; that would make the Markovian memory decay a general property of contractive reservoir dynamics.","Beyond the paper: the $\\tau_E$-delayed state mixing is formally the quantum analogue of dilated skip connections; choosing $\\tau_E$ and $\\lambda$ to match a task's correlation timescales could turn non-Markovianity into a design parameter for long-range temporal tasks.","Beyond the paper: a clean experimental test of the revival mechanism would measure the capacity $C[\\hat{y}(\\tau)]$ for the embedded model as a function of $\\Omega$, predicting that the revival peak at $\\tau=\\tau_E$ grows as $\\Omega\\to0$ while the short-delay capacity shrinks; the authors show this trend but do not derive its scaling.","Beyond the paper: the contrast between the autonomous forecasting task (gain from non-Markovianity) and the externally driven benchmark (no gain) suggests that non-Markovian memory matters most in closed-loop operation, where prediction errors feed back as inputs; quantifying this regime would connect quantum reservoir memory to control and forecasting theory."],"forward_implications":["For any Markovian quantum reservoir satisfying a unique stationary state and fading memory, the paper's bound makes exponential decay of long-term memory unavoidable; no amount of Hamiltonian or dissipation tuning within the Markovian class can produce a polynomial or flat memory profile.","The Volterra and capacity bounds apply also when inputs are quantum states, so the limitation is not an artifact of classical encoding.","Non-Markovian updates with delay $\\tau_E$ create a memory revival near $\\tau_E$, giving reservoirs the coexistence of fresh and distant past information that tasks requiring both can exploit.","The proposed embedding with auxiliary qubits gives an experimentally accessible control parameter $\\Omega$ interpolating from Markovian ($\\Omega=1$) to non-Markovian ($\\Omega\\to0$), with optimal forecasting at intermediate values.","In autonomous forecasting of the chaotic benchmark series, the non-Markovian reservoir at $\\Omega=0.5$ outperforms the Markovian one, while maximum non-Markovianity ($\\Omega=0$) degrades performance, meaning the resource must be tuned, not maximized."],"supporting_citations":[{"why":"Supplies the dissipation-based quantum reservoir model and the echo-state condition $\\|T(s)\\|<1$ that the proof relies on.","marker":"[18]"},{"why":"Contains the proofs of the Volterra-coefficient bound and the numerical details; the main text's Eqs. (4) and (6) rest on it.","marker":"[48]"},{"why":"Establishes that fading-memory systems admit Volterra expansions, the framework for the bound.","marker":"[50]"},{"why":"Defines the information processing capacity used to quantify memory in the tasks.","marker":"[51]"},{"why":"Introduces residual echo state networks, the classical analogue the Quantum Residual Reservoir is built on.","marker":"[52]"},{"why":"Provides dilated skip connections, the design pattern behind the $\\tau_E$-delayed state mixing.","marker":"[53]"},{"why":"Describes the embedding method for realizing non-Markovian dynamics from Markovian evolutions in a larger space.","marker":"[31]"},{"why":"Presents the auxiliary-qubit collision model with tunable non-Markovianity that the embedded reservoir adapts.","marker":"[62]"},{"why":"Defines the chaotic delayed-differential equation used as the forecasting benchmark.","marker":"[63]"},{"why":"Provides the trace-distance based non-Markovianity measure used to characterize the embedded model.","marker":"[34]"}],"fun_headline_variants":["Non-Markovian reservoirs boost quantum memory retention","Quantum reservoir memory revival via non-Markovian dynamics","Tunable non-Markovianity extends quantum reservoir memory","Controlled non-Markovianity recovers quantum reservoir memory","Quantum reservoirs: non-Markovianity as a memory resource"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The memory-decay proof assumes that every input maps the reservoir to a unique stationary state and that the update map is strictly contracting with fading memory, so that the Volterra expansion converges; reservoir models with multiple stationary states or non-contracting updates are outside the proof.","fun_headline_variants_meta":{"raw":{"variants":["Non-Markovian reservoirs boost quantum memory retention","Quantum reservoir memory revival via non-Markovian dynamics","Tunable non-Markovianity extends quantum reservoir memory","Controlled non-Markovianity recovers quantum reservoir memory","Quantum reservoirs: non-Markovianity as a memory resource"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1460,"prompt_tokens":991,"completion_tokens":469,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":383}},"tokens_in":607,"tokens_out":469,"duration_ms":5702,"temperature":1.0,"reasoning_tokens":383,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:50:33.726565+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the short-term memory capacity $C[\\hat{y}(\\tau)]$ of any Markovian reservoir satisfying the paper's assumptions (unique stationary state and $\\|T(s)\\|<1$) out to large delays $\\tau$; if the capacity decays polynomially or stays flat instead of exponentially, the bound in Eq. (6) is false. A second check is to exhibit a Markovian reservoir from the published literature that violates the unique-stationary-state assumption yet still shows exponential decay, which would narrow the claimed universality.","supporting_citations":[{"cited_title":"Sannia, R","cited_arxiv_id":null,"evidence_quote":"Supplies the dissipation-based quantum reservoir model and the echo-state condition $\\|T(s)\\|<1$ that the proof relies on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contains the proofs of the Volterra-coefficient bound and the numerical details; the main text's Eqs. (4) and (6) rest on it."},{"cited_title":"Boyd and L","cited_arxiv_id":null,"evidence_quote":"Establishes that fading-memory systems admit Volterra expansions, the framework for the bound."},{"cited_title":"Dambre, D","cited_arxiv_id":null,"evidence_quote":"Defines the information processing capacity used to quantify memory in the tasks."},{"cited_title":"Ceni and C","cited_arxiv_id":null,"evidence_quote":"Introduces residual echo state networks, the classical analogue the Quantum Residual Reservoir is built on."},{"cited_title":"De Vega and D","cited_arxiv_id":null,"evidence_quote":"Describes the embedding method for realizing non-Markovian dynamics from Markovian evolutions in a larger space."},{"cited_title":"Rijavec and G","cited_arxiv_id":null,"evidence_quote":"Presents the auxiliary-qubit collision model with tunable non-Markovianity that the embedded reservoir adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the chaotic delayed-differential equation used as the forecasting benchmark."}],"review_version":1}