{"id":"dc6d4df0-728c-4398-b363-8f363a908434","arxiv_id":"2412.00295","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"At spectral radius 1, the feature-space motifs of a linear Simple Cycle Reservoir are exactly the sampled Fourier basis vectors.","lead":"This paper proves that a linear Simple Cycle Reservoir, a minimal recurrent neural network, uses the Fourier basis as its internal feature representation when run at the edge of stability. The result explains a previously observed collapse in motif diversity and links reservoir computing to classical signal processing.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.4 overcounts symmetric and skew-symmetric motifs: it holds only when the input coupling w has nonzero projection on every Fourier mode, and fails for degenerate w.","rationale":"The reader's weakest assumption identifies exactly the gap I consider most load-bearing: Theorem 5.4's unconditional count of symmetric and skew-symmetric motifs assumes nonzero projection of w on every Fourier mode, which is not guaranteed and frequently fails for ±1 input vectors. The proof's use of [CB76] gives a basis of the whole space, not of the positive-eigenvalue subspace, so the theorem is mathematically false without a genericity assumption. This is not merely a cosmetic issue: the paper's abstract and Section 6 claim that linear SCR at the edge of stability performs Fourier decomposition, which suggests the full real Fourier basis appears as motifs. For degenerate w, the reservoir's motif space is a strict subspace, so the 'Fourier decomposition' is only partial. My concrete test with n=2, w=(1,1) settles the matter and should be included in any revision. I also note the explicit real Fourier basis matrix in Eq. (6.1) is mis-indexed (zero sine column, wrong count of columns, missing Nyquist for even n), which is a separate but also concrete issue; however, the degeneracy of w is the more fundamental mathematical flaw because it invalidates a stated theorem even when Eq. (6.1) is corrected. The complex-domain Theorem 4.4 appears sound, and the numerical experiments are consistent with the generically correct behavior, so the appropriate disposition remains CONDITIONAL: the paper needs a genericity hypothesis and corrected statements, but the central idea that eigenvectors are Fourier basis vectors is supported.","tokens_in":13214,"tokens_out":10286,"duration_ms":92180,"concrete_test":"Set n=2, w=(1,1), and C=[[0,1],[1,0]]. Compute Q with entries Q_{i,j}=w^T (C^T)^{i-1} C^{j-1} w for i,j∈{1,2}. Diagonalize Q and list the positive eigenvalues and the corresponding eigenvectors. If the positive-eigenvalue eigenspace does not contain both a symmetric and a skew-symmetric vector, Theorem 5.4 is false as stated. Repeat with n=4 and w=(1,1,1,1) (or w=(1,-1,1,-1)) to verify which Fourier modes vanish and whether the motif counts deviate from ⌈n/2⌉ and ⌊n/2⌋.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central real-domain claim (Section 5, Theorem 5.4) states that at ρ=1, a linear SCR with cyclic permutation has exactly ⌈n/2⌉ symmetric and ⌊n/2⌋ skew-symmetric motifs. The proof invokes the Cantoni–Butler theorem [CB76], which provides an orthonormal eigenbasis of the whole space R^n with those counts. However, motifs are defined (Section 2) as eigenvectors of Q with positive eigenvalues. The proof never establishes that every vector in that [CB76] basis has positive eigenvalue; it only shows the full space admits such a basis. When |d_j|^2 = |ξ_j^* w|^2 = 0 for some Fourier mode, the corresponding eigenvector lies in the null space of Q and is not a motif, so the positive-eigenvalue subspace loses that direction. Thus the counts of positive-eigenvalue symmetric and skew-symmetric motifs can drop below ⌈n/2⌉ and ⌊n/2⌋. A concrete counterexample: n=2, w=(1,1). Then d_0=√2 and d_1=0, so Q has a single positive eigenvalue with eigenvector (1,1)/√2, which is symmetric. There are 1 symmetric motif and 0 skew-symmetric motifs, contradicting Theorem 5.4's prediction of 1 and 1. The paper does not state any genericity assumption on w; its experimental w is likely generic, but the theorem as written is false. Because Section 6 and the abstract's 'Fourier decomposition' claim rely on the full real Fourier basis being present, this gap is load-bearing: for degenerate w the reservoir only projects onto a proper subset of the Fourier modes. The complex-domain Theorem 4.4 is not affected, as it correctly identifies eigenvectors (including zero-eigenvalue ones) with the Fourier basis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes the reservoir kernel induced by a linear Simple Cycle Reservoir (SCR) at unit spectral radius. For a full-cycle permutation coupling C and look-back window τ = n, the authors show in the complex domain that the kernel metric tensor Q shares its eigenvectors with the discrete Fourier basis, with eigenvalues equal to |ξ_j^* w|^2 (Theorem 4.4). In the real domain they argue that Q is circulant and use the Cantoni–Butler theorem to claim exactly ⌈n/2⌉ symmetric and ⌊n/2⌋ skew-symmetric motifs (Theorem 5.4). Section 6 combines these results to state that the motifs are exactly the real Fourier basis columns, and Section 7 reports numerical experiments comparing an SCR-motif reservoir model with a Fourier-basis model, finding near-identical behavior. The paper concludes that a linear SCR at the edge of stability performs a weighted projection of the input onto the Fourier basis.","tokens_in":13466,"tokens_out":3317,"duration_ms":30984,"significance":"If the main claim holds, the paper provides a clean, fully analytic explanation of the previously observed collapse of motif richness at ρ = 1, and it bridges reservoir computing with classical Fourier analysis. The derivation is self-contained and free of fitted parameters: the eigenvectors are forced by the cyclic permutation structure and the definition of Q, and the numerical experiments are reproducible with publicly available code. These are genuine strengths. The central real-domain claim, however, needs a genericity condition on the input vector w; without it, the stated motif counts and the exact identification with the Fourier basis in Section 6 fail for degenerate inputs.","major_comments":[{"comment":"Theorem 5.4 overclaims the number of motifs. The proof invokes the Cantoni–Butler theorem [CB76], which provides an orthonormal eigenbasis of the whole space R^n with ⌈n/2⌉ symmetric and ⌊n/2⌋ skew-symmetric eigenvectors, but the paper defines motifs as eigenvectors of Q with positive eigenvalues (Section 2, Eq. (2.3)). The proof never shows that every vector of that basis has a positive eigenvalue. When |d_j|^2 = |ξ_j^* w|^2 = 0 for some Fourier mode j, the corresponding eigenvector lies in the null space of Q and is not a motif, so the numbers of symmetric and skew-symmetric motifs can drop below ⌈n/2⌉ and ⌊n/2⌋. A concrete counterexample is n = 2, w = (1,1): d_0 = √2, d_1 = 0, so Q has rank 1 with only one symmetric motif (1,1)/√2 and zero skew-symmetric motifs, contradicting the theorem's prediction of one symmetric and one skew-symmetric motif. The theorem should either be restated with a genericity condition such as |d_j|^2 > 0 for all j, or its conclusion should be weakened to the existence of an orthonormal eigenbasis of R^n with those symmetry counts, not of the positive-eigenvalue motif subspace.","section":"Section 5.1, Theorem 5.4"},{"comment":"The claim that the motifs of a linear SCR over R are exactly the first n columns of the real Fourier basis F depends on the degenerate-case issue in Theorem 5.4 and also requires an explicit correspondence between the complex eigenvectors ξ_j of Theorem 4.4 and the real cosine/sine columns of F. For example, the case n even includes the Nyquist frequency, whose eigenvector is real and not paired with a sine vector. The text currently asserts this correspondence without proving that each column of F has positive eigenvalue and without addressing the fact that vanishing d_j removes entire frequencies. The identification should be re-stated under the genericity condition, and the indexing of F, which uses k = 0,...,⌈n/2⌉ with 'even columns' and 'odd columns', should be made unambiguous regarding column numbering.","section":"Section 6, Eq. (6.1) and items R.1–R.2"}],"minor_comments":[{"comment":"In the proof of Lemma 4.3, the summations are written as ∑_{k=0}^{n-1} ρ^{2k} when they should be ∑_{k=0}^{τ-1} ρ^{2k}, since λ was defined with τ terms. The same index error appears in the displayed expansion of X_ρ ξ̂_j. This does not affect Theorem 4.4, which uses τ = n, but it contradicts the statement on page 10 that the results of the section hold for general τ.","section":"Section 4, Lemma 4.3"},{"comment":"The paper's theoretical framework assumes strictly contractive W with ∥W∥<1, but Theorem 4.4 and all results at ρ=1 take W=C with ∥C∥=1, which is not contractive. This is presumably intentional as an edge-of-stability idealization, but it should be stated explicitly that the fading-memory assumption is relaxed in the limit ρ→1 and that the analysis applies to finite look-back windows.","section":"Section 2, Definition 2.1 and Section 4"},{"comment":"There are several typos and stylistic issues: 'W e' at the start of the introduction, 'This decline is also observed in is also shown in Figure 2' on page 6, 'space of space of motif' in the proof of Theorem 5.4, and 'While this function may not visually resemble' in Remark 6.1. A careful proofreading pass is needed.","section":"Section 3 and throughout"},{"comment":"The figure captions are terse and do not specify the normalization used for the 'column-wise FFT' or the shuffling procedure in panels (c) and (d). Please add enough detail in the captions or text so that the numerical verification can be reproduced without guessing.","section":"Section 6, Figures 3 and 4"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—here's the short version: the paper's main result is a clean linear-algebra fact, and it's correct. At spectral radius 1, the kernel of a linear simple cycle reservoir is diagonalized by the discrete Fourier basis, with eigenvalue |ξ_j^* w|^2 for input coupling w. That turns Tino's earlier empirical observation about motif richness collapse into a precise statement. The complex-domain proof (Theorem 4.4) is sound, and the numerical comparison between SCR motifs and a hand-built Fourier basis is convincing.\n\nThe real-domain counting theorem, however, is stated too strongly. Theorem 5.4 claims exactly ceil(n/2) symmetric and floor(n/2) skew-symmetric motifs for every cyclic permutation, but that holds only when w has nonzero projection on every Fourier mode. The authors never state this genericity condition. Concrete counterexample: n=2, w=(1,1). Then Q has a single positive eigenvalue with symmetric eigenvector (1,1)/√2 and a zero eigenvalue with skew-symmetric eigenvector (1,-1)/√2. So you get one symmetric motif and zero skew-symmetric motifs, not one and one. The proof invokes the Cantoni–Butler theorem for an orthonormal eigenbasis of the whole space, but the definition of motif requires a positive eigenvalue; zero-eigenvalue vectors drop out. This is a load-bearing gap only in the sense that the exact counts in Section 6 are wrong for degenerate w. The central claim that the kernel projects onto the Fourier basis at ρ=1 still stands; it may simply project onto a proper subset. That is still a Fourier decomposition.\n\nOther soft spots are minor: Eq. (6.1) has the wrong DC normalization and indexing, Lemma 4.3 has an index typo, and the code URL is malformed. The paper also quietly drops the strict contractivity assumption at ρ=1, which is fine for an edge-of-stability analysis but deserves a sentence. None of this threatens the main theorem.\n\nThis is a paper that deserves a serious referee. The authors should add a genericity assumption (or the case analysis for degenerate w), fix the basis formula, and make the code link work. For the reservoir-computing community, the result is genuinely useful: it connects SCRs to classical signal processing and explains why the edge of stability collapses motif diversity. I'd take it for review, with the understanding that Theorem 5.4 needs a patch.","headline":"Main theorem is right; the real-domain motif counts need a genericity condition.","tokens_in":14102,"tokens_out":3595,"would_cite":true,"duration_ms":31376,"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":"A linear Simple Cycle Reservoir at the edge of stability projects its input onto the Fourier basis.","keywords":["simple cycle reservoir","reservoir computing","edge of stability","Fourier basis","kernel motif","time series analysis","spectral radius"],"falsifier":"For $n = 4$ with constant input weights $w = (1,1,1,1)$, the metric tensor $Q$ is the all-ones matrix with rank 1, giving exactly one positive-eigenvalue motif rather than the two symmetric and two skew-symmetric motifs claimed by Theorem 5.4.","tokens_in":12904,"feed_emoji":"📈","tokens_out":10689,"duration_ms":86366,"temperature":0.7,"pith_summary":"This paper establishes that a linear Simple Cycle Reservoir (SCR) placed exactly at the edge of stability — its recurrent coupling a full-cycle permutation with spectral radius one — induces a time-series kernel whose eigenvectors are precisely the discrete Fourier basis. In the complex domain the motifs are the Fourier columns with weights equal to the projections of the input coupling vector onto each mode; in the real domain they split into cosine and sine motifs. The result means such a reservoir is effectively a Fourier analyzer: its state-space features carry a weighted projection of the input window onto harmonic functions. This connection between reservoir computing and classical signal processing is confirmed by numerical experiments in which the motif spectra match the Fourier matrix and forecasting with SCR motifs matches forecasting with Fourier motifs.","feed_headline":"Reservoir at the edge of stability Fourier-decomposes its input","feed_subtitle":"Its feature-space motifs match the discrete cosine and sine basis, linking reservoir computing to signal processing.","key_machinery":"The central object is the reservoir kernel's metric tensor $Q$, defined by $Q_{ij} = w^\\top (W^\\top)^{i-1} W^{j-1} w$, whose eigen-decomposition yields the motifs (eigenvectors with positive eigenvalue). The argument runs on two algebraic facts: when $W$ is unitary at $\\rho = 1$, $Q$ factors as $A^* A$ with $A = D F$, where $F$ is the Fourier matrix, which forces the eigenvectors to be the Fourier basis; and when $W$ is a cyclic permutation over $\\mathbb{R}$, $Q$ is symmetric circulant — hence centrosymmetric — so its eigenbasis has a fixed count of symmetric and skew-symmetric vectors. The edge of stability is what makes the powers of $W$ reduce to roots of unity.","core_discovery":"The paper proves two structural facts about the metric tensor $Q$ of the reservoir kernel at unit spectral radius. Over $\\mathbb{C}$, with $W$ equal to a full-cycle permutation $C$, the eigenvectors of $Q$ are exactly the Fourier basis vectors $\\xi_j$, with eigenvalues $|d_j|^2$, where $d_j = \\xi_j^* w$ is the projection of the input coupling vector onto the $j$-th Fourier mode (Theorem 4.4). Over $\\mathbb{R}$, $Q$ is symmetric and circulant, hence centrosymmetric, so it admits an orthonormal basis with $\\lceil n/2 \\rceil$ symmetric and $\\lfloor n/2 \\rfloor$ skew-symmetric eigenvectors; combined with the complex-domain result, these are the columns of the real Fourier basis matrix, cosines in even columns and sines in odd columns. The paper concludes that a linear SCR at the edge of stability performs a weighted projection of the input time series onto the discrete Fourier basis, and supports this with numerical experiments showing the Fourier spectra of the motifs coincide with those of the Fourier matrix.","pith_inferences":["A practical design rule follows: for linear reservoirs one can bypass training and use the Fourier basis directly as the feature map, which is cheaper and more transparent.","By choosing $w$ orthogonal to selected Fourier modes, the reservoir becomes a spectral filter (for instance, a band-pass or notch filter) for time-series preprocessing.","The proof strategy — recognizing the kernel matrix as a Gram matrix of a unitary group action — likely extends to other structured couplings such as block-circulant or multi-cycle reservoirs, where motifs would become generalized harmonics.","The stated motif counts presuppose a generic input coupling; for degenerate $w$ the reservoir's effective dimensionality drops, a caveat that matters for interpretability of the feature space."],"forward_implications":["At $\\rho = 1$, the motif matrix of a linear SCR is, up to phase and column ordering, the discrete Fourier transform matrix, so the reservoir's feature space is the space of Fourier coefficients of the input window.","Because each mode's eigenvalue is $\\left|\\langle w, \\xi_j \\rangle\\right|^2$, the input coupling vector $w$ acts as a spectral filter: it decides which Fourier modes the reservoir keeps and how strongly.","The real-domain motifs alternate between symmetric cosine vectors and skew-symmetric sine vectors, giving an explicit real Fourier basis for the motif space.","Since a Reservoir Motif Machine trained with unit-radius SCR motifs and one trained with the Fourier basis give virtually identical forecasts, the SCR representation is numerically the Fourier representation."],"supporting_citations":[{"why":"Introduces the kernel view of linear reservoirs and defines motifs as eigenvectors of the metric tensor, the framework the paper reverse-engineers.","marker":"[Tin20]"},{"why":"Defines Simple Cycle Reservoirs, the architecture whose motif structure is analyzed.","marker":"[RT10]"},{"why":"Supplies the approach of analyzing unitary dynamical couplings first; the complex-domain proof follows this template.","marker":"[LFT24]"},{"why":"Supplies the orthogonal-coupling approach used in the real-domain analysis and the universality context for SCRs.","marker":"[FLT24]"},{"why":"Provides the theorem that symmetric centrosymmetric matrices have $\\lceil n/2 \\rceil$ symmetric and $\\lfloor n/2 \\rfloor$ skew-symmetric eigenvectors, used for the motif counts.","marker":"[CB76]"},{"why":"Defines Reservoir Motif Machines, the forecasting model used in experiments to compare SCR motifs with Fourier motifs.","marker":"[TFL24]"}],"fun_headline_variants":["Reservoir at stability edge acts as Fourier transformer","Stable-edge reservoir computes Fourier transform","Edge-of-stability reservoir reveals Fourier basis","Reservoir at critical stability aligns with Fourier modes","Linear reservoir at edge of stability is a Fourier projector"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed numbers of cosine and sine motifs assume the input weight vector is generic, meaning it has nonzero projection on every Fourier mode; a weight vector that misses a mode produces fewer motifs than stated.","fun_headline_variants_meta":{"raw":{"variants":["Reservoir at stability edge acts as Fourier transformer","Stable-edge reservoir computes Fourier transform","Edge-of-stability reservoir reveals Fourier basis","Reservoir at critical stability aligns with Fourier modes","Linear reservoir at edge of stability is a Fourier projector"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000744,"raw_usage":{"total_tokens":3283,"prompt_tokens":878,"completion_tokens":2405,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":2334}},"tokens_in":494,"tokens_out":2405,"duration_ms":15779,"temperature":1.0,"reasoning_tokens":2334,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:32:10.048548+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For $n = 4$ with constant input weights $w = (1,1,1,1)$, the metric tensor $Q$ is the all-ones matrix with rank 1, giving exactly one positive-eigenvalue motif rather than the two symmetric and two skew-symmetric motifs claimed by Theorem 5.4.","supporting_citations":[],"review_version":1}