{"id":"34ca0e27-c7f8-45fb-981b-48d648e0d753","arxiv_id":"2412.08322","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Quantum reservoir filters are injective if the state update is input-invertible at reachable states, reducible to a rank condition in affine quantum systems.","lead":"This paper proves conditions under which a quantum reservoir computer assigns different internal states to different input histories, a property called filter injectivity. It gives a rank condition that is easy to check on standard quantum reservoir models and illustrates how losing injectivity degrades short-term memory performance.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 5's rank condition only yields local injectivity; a periodic SAS counterexample meets all hypotheses yet has a non-injective filter, so the main sufficient criterion is false as stated.","rationale":"The reader's weakest assumption concerned the imported ESP/FMP characterization from Ref. [32] and its applicability to the filter expansion (19). The concern identified here is different and more direct: Proposition 5, the main sufficient condition for global filter injectivity, is mathematically false as stated. Full rank of the linearized maps at every point guarantees only local injectivity of each F_x; it does not make F_x globally injective, which is what Lemma 4 needs. The periodic SAS counterexample satisfies all stated hypotheses of Proposition 5 — rank condition, uniform contractivity of p, compact input space, and hence ESP — while the corresponding filter is not injective. This is not a matter of an extra technical assumption imported from another paper; it is an internal logical gap in the central theorem. The paper's own Example 24 demonstrates the same periodicity phenomenon and avoids Proposition 5 for exactly that reason, which underscores the issue. Lemma 4 is correct, and the local injectivity results in Propositions 11 and 15 may be salvageable, but the advertised 'condition that guarantees injectivity' fails. Because this is the central claim of the paper, the current version cannot be accepted as is; a major reformulation of the sufficient condition, likely to a much weaker or more complex one, is required. Therefore the verdict should move from CONDITIONAL to REJECT, while acknowledging that parts of the paper, especially the local analysis and the constant-filter characterization, retain value.","tokens_in":25842,"tokens_out":14122,"duration_ms":165773,"concrete_test":"Implement the SAS map with r=0.5, V a large ball, D=[0,4π]×[0,1], and F(x,z)=0.5 x+(cos z1, sin z1, z2). Use expansion (19) to evaluate the filter U for the constant input sequences (0,0) and (2π,0); verify that U is identical for both while the input sequences differ. Also evaluate the derivative Dq(z) at, say, z=(π/2, 0.5) to confirm it has rank 2, and confirm ||p(z)||<1 so the ESP hypothesis holds. If all checks pass, Proposition 5's hypotheses are satisfied but its conclusion fails, so an additional global-injectivity hypothesis on F_x or a restriction on D^n is required.","verdict_should_be":"REJECT","load_bearing_attack":"Proposition 5's rank condition (22) is not sufficient for the SI-invertibility required by Lemma 4. The Local Injectivity Theorem cited in (20) only gives, for each fixed x and each z, a neighborhood on which F_x is injective; full rank at every point does not imply global injectivity of F_x. A concrete SAS counterexample: take N=3, n=2, D=[0,4π]×[0,1], and F(x,z)=r x+(cos z1, sin z1, z2) with 0<r<1 and V a sufficiently large closed ball. Then Dp=0 and Dq(z) has rank 2 for all z, so (22) holds for every x and z; p is uniformly contractive, so the ESP holds under the paper's own criteria. Yet F_x(z1,z2)=F_x(z1+2π,z2) for every x, so SI-invertibility fails and the filter is not injective: the constant sequences (0,0) and (2π,0) have the same image. The paper itself encounters this obstruction in Example 24, where periodicity forces a retreat to local injectivity via Proposition 15, and Example 10 shows q can be an essentially arbitrary ergodic fixed-point function, including periodic ones. Since Proposition 5 is the paper's headline design criterion, this gap invalidates the central claim as stated.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies input-dependence of quantum reservoir computing filters. It builds on the state-affine system (SAS) representation of finite-dimensional quantum channels and proposes conditions under which the reservoir filter is injective as a map from input sequences to reservoir state sequences. The main results are Lemma 4 (SI-invertibility plus ESP implies filter injectivity), Proposition 5 (a rank condition on the SAS derivative that is claimed to guarantee SI-invertibility), a local injectivity theorem (Proposition 11) and its SAS specialization (Proposition 15, Corollary 16), and a classification of constant filters for contracted-encoding quantum channels (Theorem 19). The paper also contains several examples and a numerical demonstration that breaking global injectivity degrades a short-term memory task.","tokens_in":26044,"tokens_out":7388,"duration_ms":82887,"significance":"The motivation is sound: filter injectivity is a natural and practically relevant design goal for reservoir computing, and the connection to generalized synchronization and learnability in Remark 8 is useful. The paper has clear strengths: Lemma 4 is correct and concise; Proposition 12 correctly characterizes preimages of constant output sequences; the contracted-encoding family is well motivated; and the examples are concrete, with reproducible code for the numerical experiment. If a correct sufficient condition in the spirit of Proposition 5 existed, the paper would make a valuable design contribution. However, the central rank criterion is false as stated, and the local injectivity theorem has a serious proof gap, so the headline results cannot be accepted in their current form.","major_comments":[{"comment":"The claimed sufficient condition is false. The rank condition (22) only gives, via the Local Injectivity Theorem, that for each fixed x and z there is a neighborhood on which F_x is injective; full rank at every point does not imply global injectivity of F_x, and SI-invertibility (Definition 3) is a global property. Concretely, take N=3, n=2, D=[0,4π]×[0,1], V a sufficiently large closed ball, and F(x,z) = r x + (cos z1, sin z1, z2) with 0<r<1. Then p(z)=rI and q(z)=(cos z1, sin z1, z2); Dp=0 and Dq has rank 2, so (22) holds for every x and z. The uniform contraction of p gives the ESP. Yet F_x(z1,z2)=F_x(z1+2π,z2) for every x, so F is not SI-invertible and the filter is not injective: the constant input sequences (0,0) and (2π,0) have identical fixed-point images. This invalidates the central design criterion of the paper, and the same issue propagates to the contracted-encoding sufficient conditions in Section III.B, in particular Eq. (48) and Eq. (50).","section":"III.A, Proposition 5 and Eq. (22)"},{"comment":"The compactness argument in the proof of Proposition 11 is not valid. The function f(x) = inradius(L_{z0}(x), z0) is asserted to attain its minimum because there is a sequence x_n with f(x_n) -> a and a convergent subsequence x_{n_k} -> l; the proof then says f(l) would otherwise contradict that convergence. This would require f to be lower semicontinuous, or at least that the neighborhoods L_{z0}(x) vary continuously in x, which is neither established nor generally true. An infimum of positive numbers over a compact set need not be attained for an arbitrary positive function, and inradius can drop at a limit point. Consequently the existence of the uniform neighborhood K_{z0} in Eq. (27), and with it Proposition 11(i) and the local injectivity results Proposition 15 and Corollary 16, is not proved as written.","section":"III.A, Proposition 11 proof"}],"minor_comments":[{"comment":"The notation B0(H) for the space of traceless operators and B0 for the Gell-Mann basis of su(d) is confusing; please use different symbols for the basis and the subspace.","section":"Section II"},{"comment":"The zero block in the block matrix is written as 0_{d^2-1}; it should be identified explicitly as a row vector of length d^2-1 so that the dimensions of the four blocks are unambiguous.","section":"Eq. (14)"},{"comment":"The filter expansion is written for left-infinite sequences but the product order and the indices assume a specified time origin; please state explicitly whether the domain is (D^n)^{Z-} or (D^n)^Z and how the origin is chosen.","section":"Eq. (19)"},{"comment":"The statement 'Let F=V×D^n→V' is missing a colon after the domain, and the phrase 'have all rank n and hence are injective' should read 'have rank n for all ...' for clarity.","section":"Proposition 5 statement"},{"comment":"The sentence 'showing U(z)_0 ≠ 0 for a given epsilon seems an unfeasible task' is vague; please state precisely which claim is being discarded and why the positivity of the filter output is relevant to the rank condition.","section":"Example 24"}],"recommendation":"reject","confidential_remarks":"The counterexample to Proposition 5 is simple and directly hits the paper's central claim, so I cannot see how a minor revision could repair it without replacing the main sufficient condition with a substantially weaker or more restrictive one. The local injectivity results may be salvageable with additional hypotheses, and the examples are useful, but the current manuscript's headline contribution is not correct. I would note that the paper's reliance on the authors' own earlier results is not problematic; the issue is the new injectivity criterion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my private take: the paper's headline global injectivity criterion (Proposition 5) is false as stated. The rank condition (22) only gives local injectivity, not global state-input invertibility. A simple SAS counterexample: N=3, n=2, D=[0,4π]×[0,1], F(x,z)=r x+(cos z1, sin z1, z2) with 0<r<1. Dp=0, Dq has rank 2 everywhere, so (22) holds for all x,z; p is uniformly contractive so ESP holds; yet the filter maps (0,0) and (2π,0) to the same output. The paper's own Example 24 shows the same periodic obstruction and retreats to local injectivity via Proposition 15, so the authors were aware but did not correct the global claim. What is actually new and good: Lemma 4 is correct and clean — SI-invertibility plus ESP implies filter injectivity. The specialization to contracted-encoding channels, especially Theorem 19 on constant filters, is a useful design tool and appears original. The worked examples are thorough, and the code/notebooks are available, which is real evidence. The local criteria — Corollary 16 in particular — are practically checkable and on much firmer ground. Soft spots: the proof of Proposition 11 has a flawed compactness argument; the inradius function need not be continuous, so the existence of a uniform neighborhood is not proved as written. This is likely repairable but needs a fix. The numerical experiment (Fig. 2) is presented as evidence that injectivity is fundamental, but it does not isolate injectivity from other effects like input scaling or nonlinearity. Also, the novelty boundary with Refs. [57,59] should be clarified; Lemma 4 is largely a reformulation of known state-input invertibility ideas. Who this is for: QRC theorists designing input encodings and checking whether a given quantum reservoir can distinguish input sequences. The local results are useful, the global criterion is not. Recommendation: this deserves a serious referee, but the referee should be alerted to the false global claim. I would not cite Proposition 5 in its current form. With a revision that replaces or repairs the global criterion, tightens Proposition 11, and reframes the numerics as suggestive, the paper would be a solid addition.","headline":"The local injectivity results and contracted-encoding analysis are salvageable, but the headline global criterion (Prop. 5) is false as stated, so the paper needs major revision before it can be relied on.","tokens_in":680,"tokens_out":2133,"would_cite":false,"duration_ms":81006,"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":"Distinct input sequences reach distinct reservoir states exactly when the reservoir map is state-input invertible and has the echo state property; for finite-dimensional quantum reservoirs this becomes a checkable rank condition on the…","keywords":["quantum reservoir computing","filter injectivity","echo state property","fading memory property","state-affine systems","contracted-encoding quantum channels","generalized synchronization","input dependence"],"falsifier":"Symbolically evaluate the rank condition (22) for a state-affine quantum reservoir that has the echo state property, then compute the filter U_F on two distinct input sequences whose entries all lie in the input domain. Equality of the two state histories for any such pair would refute Proposition 5, since the proposition predicts no collision can occur when the rank condition holds at every reachable state; a random search over contractive p(z) and differentiable q(z) with a symbolic rank check would settle the sufficient condition directly.","tokens_in":25565,"feed_emoji":"⚛️","tokens_out":6558,"duration_ms":64878,"temperature":0.7,"pith_summary":"This paper addresses a gap in quantum reservoir computing: even when a reservoir has the echo state property (each input has a unique reservoir history) and fading memory, its filter may still fail to tell different input sequences apart. The authors establish that injectivity of the filter—the one-to-one correspondence between input histories and reservoir state histories—is guaranteed once the state-update map is state-input invertible and the system has ESP. For finite-dimensional quantum reservoirs, which admit an affine (state-affine) representation in a Gell-Mann basis, this reduces to a rank condition on the derivatives of p and q that is readily checkable in many models. The relevance is that injective filters are a prerequisite for learning deterministic dynamical systems through injective generalized synchronization, and the paper shows numerically that losing injectivity, for example through periodic input encodings, degrades temporal tasks such as short-term memory.","feed_headline":"Rank condition guarantees a quantum reservoir tells inputs apart","feed_subtitle":"Even with the echo state property, quantum reservoirs can be input-blind; this paper pins down when filters are injective.","key_machinery":"The load-bearing object is the state-affine system (SAS) representation of a reservoir in a generalized Gell-Mann basis, written x_t = p(z_t)x_{t-1} + q(z_t), which turns the quantum channel into an affine map on the Bloch vector; contractivity of p guarantees the echo state property and the absolutely convergent filter expansion U(z)_t = \\sum_{j\\geq 0} \\left(\\prod_{k=0}^{j-1} p(z_{t-k})\\right) q(z_{t-j}). The argument is carried by Lemma 4, which reduces filter injectivity to state-input invertibility, and by the local injectivity theorem, which converts the rank condition DF_x(z) rank n into that invertibility. For contracted-encoding channels, the composition structure lets the rank condition be expressed through p_E, p_J(z), q_E, and q_J(z), with direct simplifications when E is invertible or J is unitary.","core_discovery":"On its own terms, the paper's central discovery is that filter injectivity for reservoir computers, quantum or not, follows from a local, finite-dimensional condition. Lemma 4 proves the general statement: if F is state-input invertible (each map F_x(z)=F(x,z) is injective) and has the echo state property, then the filter U_F is injective. In the state-affine form that every finite-dimensional quantum reservoir takes in a generalized Gell-Mann basis, state-input invertibility is implied by the rank condition that the linear maps (Dp(z)(·))x + Dq(z) have rank n for all reachable states x and all inputs z (Proposition 5). The paper extends the same logic to the widely used contracted-encoding architecture, where an input-encoding quantum channel is followed by a strictly contractive channel; there the injectivity conditions factor through the affine data of the encoding channel J and the contractive channel E, with simplifications when E is invertible or J is unitary. It also gives local injectivity results around constant sequences, characteizes when the filter is constant through the fixed-point map, and shows numerically that a periodic rotation encoding that breaks global injectivity at g=2π also destroys short-term memory capacity.","pith_inferences":["Because Lemma 4 is purely set-theoretic and the SAS rank condition is not quantum-specific, the same injectivity criterion should apply to classical state-affine reservoirs; the paper's examples are quantum, but the mechanism is agnostic.","The rank condition is sufficient rather than necessary, so there is likely a larger class of injective reservoirs whose filters remain injective even where the Jacobian drops rank; mapping the boundary of the sufficient region could reveal cheaper design criteria.","Sampling the rank condition over the full state space instead of the reachable set can produce false negatives, since states that are never visited may violate the condition without affecting injectivity; automatic differentiation over simulated trajectories could estimate the reachable set in practice.","The connection to nonstationary echo state properties noted in the paper suggests that injectivity of the fixed-point map, together with contraction, may provide a route to formalizing when time-varying or nonstationary reservoirs can process sequences, going beyond the constant-input analysis."],"forward_implications":["Designers of quantum reservoir computers can certify that their reservoir distinguishes input histories by checking a finite-dimensional rank condition on the affine maps p(z) and q(z) over the reachable set, rather than comparing arbitrary input sequences.","For contracted-encoding channels, the injectivity check splits into the encoding channel and the contractive channel: an invertible E reduces it to the input-encoding map's derivatives, and a unitary encoding reduces it to p_E Dp_J(z)(\\cdot)x with x\\neq 0.","Local injectivity around constant sequences gives an easy-to-verify criterion: if (Dp(z)(\\cdot))x^*(z)+Dq(z) has rank n at the fixed point x^*(z)=(I-p(z))^{-1}q(z), the filter is injective on a neighborhood of that constant input.","Non-injective filters cannot support injective generalized synchronization, so the rank conditions are prerequisites for learning deterministic dynamical systems by reservoir forecasting.","Periodic input encodings destroy global injectivity, and the paper's numerical experiment ties that loss to a drop in short-term memory capacity, making injectivity an explicit design constraint for temporal tasks."],"supporting_citations":[{"why":"Supplies the state-affine representation of finite-dimensional quantum reservoir systems, the contractivity characterization of ESP and FMP, the constant-filter Theorem 2, and the master-equation examples used throughout.","marker":"[32]"},{"why":"Establishes the universality of non-homogeneous state-affine systems and provides the filter expansion (19) on which the injectivity analysis relies.","marker":"[42]"},{"why":"Provides the local injectivity theorem that converts the rank condition DF_x(z) rank n into state-input invertibility.","marker":"[58]"},{"why":"Shows that injective generalized synchronization is a sufficient route to learning deterministic dynamical systems, which motivates filter injectivity as an architectural requirement.","marker":"[35, 36]"},{"why":"Introduced the notion of state-invertibility in the context of causal embedding for forecasting, which the paper adapts to reservoir filters.","marker":"[57]"},{"why":"Supplies the classical reservoir computing framework: causality, time-invariance, weighted norms, fading memory continuity, and the relevant product topology on sequence spaces.","marker":"[43]"},{"why":"Proposes the nonstationary echo state property, which the paper connects to injectivity of the fixed-point map in Remark 13.","marker":"[33]"}],"fun_headline_variants":["Rank condition ensures quantum reservoir input injectivity","Local rank test guarantees reservoir filters tell inputs apart","Quantum reservoir injectivity from a finite rank check","Echo state alone insufficient; rank condition fills gap","Finite-dimensional rank criterion for reservoir input separation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's main theorems rely on the imported characterization from Ref. [32] that, for compact input spaces, uniform operator-norm contraction of the reservoir map on traceless operators is equivalent to the echo state and fading memory properties; if that contractivity characterization fails, or the uniform epsilon-contraction needed for the series expansion does not hold, the filter expansion and the injectivity results built on it do not apply.","fun_headline_variants_meta":{"raw":{"variants":["Rank condition ensures quantum reservoir input injectivity","Local rank test guarantees reservoir filters tell inputs apart","Quantum reservoir injectivity from a finite rank check","Echo state alone insufficient; rank condition fills gap","Finite-dimensional rank criterion for reservoir input separation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1197,"prompt_tokens":939,"completion_tokens":258,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":187}},"tokens_in":555,"tokens_out":258,"duration_ms":3008,"temperature":1.0,"reasoning_tokens":187,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:58:22.904747+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Symbolically evaluate the rank condition (22) for a state-affine quantum reservoir that has the echo state property, then compute the filter U_F on two distinct input sequences whose entries all lie in the input domain. Equality of the two state histories for any such pair would refute Proposition 5, since the proposition predicts no collision can occur when the rank condition holds at every reachable state; a random search over contractive p(z) and differentiable q(z) with a symbolic rank check would settle the sufficient condition directly.","supporting_citations":[{"cited_title":"Mart ´ ınez-Pe˜ na and J.-P","cited_arxiv_id":null,"evidence_quote":"Supplies the state-affine representation of finite-dimensional quantum reservoir systems, the contractivity characterization of ESP and FMP, the constant-filter Theorem 2, and the master-equation examples used throughout."},{"cited_title":"Grigoryeva and J.-P","cited_arxiv_id":null,"evidence_quote":"Establishes the universality of non-homogeneous state-affine systems and provides the filter expansion (19) on which the injectivity analysis relies."},{"cited_title":"Abraham, J","cited_arxiv_id":null,"evidence_quote":"Provides the local injectivity theorem that converts the rank condition DF_x(z) rank n into state-input invertibility."},{"cited_title":"Universal set of Observables for Forecasting Physical Systems through Causal Embedding","cited_arxiv_id":"2105.10759","evidence_quote":"Introduced the notion of state-invertibility in the context of causal embedding for forecasting, which the paper adapts to reservoir filters."},{"cited_title":"Grigoryeva and J.-P","cited_arxiv_id":null,"evidence_quote":"Supplies the classical reservoir computing framework: causality, time-invariance, weighted norms, fading memory continuity, and the relevant product topology on sequence spaces."},{"cited_title":"Kobayashi, Q","cited_arxiv_id":null,"evidence_quote":"Proposes the nonstationary echo state property, which the paper connects to injectivity of the fixed-point map in Remark 13."}],"review_version":1}