{"id":"e125ed42-0243-4c01-b1ad-c45aa564534b","arxiv_id":"2509.07815","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For quadratic truncation, the minimal number of segments needed to recover a path from the barycenter of signature matrices is min(d, total segments) if all segment counts are even, and min(d, total segments minus the number of odd counts plus one) otherwise.","lead":"This paper finds exactly how many line segments are needed to draw a path representing the barycenter, a special group average, of several signature matrices. It provides a closed formula and explicit matrix transformations, which matters for learning representative paths from signature data in machine learning and finance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 7.2 proof asserts without verification that the congruence transform sends 1_m to e1; Lemma 7.6's explicit P for odd m maps 1_m to √2 e2, so the simultaneous vector-matrix reduction is unproven.","rationale":"The reader's weakest_assumption is precisely the simultaneous vector-matrix compatibility. Our independent check shows the concern is concrete: the explicit transform in Lemma 7.6 does not by itself map 1_m to e_1, so the proof must invoke an additional block permutation or scaling that is not stated. However, Example 7.13 demonstrates that a compatible transform exists for at least one case, so the theorem is likely correct but the proof needs a detailed verification. This supports the reader's CONDITIONAL verdict rather than a rejection.","tokens_in":24674,"tokens_out":20454,"duration_ms":195191,"concrete_test":"Using the explicit formulas in Lemma 7.6, Theorem 7.3, and the alternative proof of Theorem 7.10, compute the composite P for odd α such as (3,2) in R^5, (5,2,2) in R^9, and (1,2,2) in R^5, and verify directly whether P W_α P^T = C_r ⊕ 0 and P (1/N)1_m = [1_r;0], with r = m - #odd + 1. Do this symbolically in OSCAR/Sage. If the equality holds for all tested cases, the concern is resolved; if it fails for any, the proof of Theorem 7.2 is incomplete and the formula may be only a lower bound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 7.2 (odd case) reduces B_{d,2}(α) to finding one congruence transform P such that P W_α P^T = C_r ⊕ 0 and P (1/N) 1_m = [1_r; 0]. It asserts: 'since the matrix from Theorem 7.6 transforms 1_{α1} to e_1 ...', but this property is not demonstrated and does not hold for the explicit P in Lemma 7.6. For m=3, the given P = √2 T_{12}(I_3 - E_{21} - E_{31} + E_{13} - E_{12}) sends 1_3 to √2 e_2, not e_1; for m=5 it sends 1_5 to √2 e_2 as well (after T_{12}). The 'inverted transform' used in Example 7.13 employs a different matrix (without T_{12}) that does map 1_5 to e_1, but the proof does not reconcile this discrepancy. If the composite transform does not send the vector exactly to [1_r; 0], then the recovered path would require more segments than rk(W_α), and Theorem 7.2 would hold only as a lower bound. The issue is load-bearing because the upper bound B ≤ rk(W_α) depends on this simultaneous normalization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the Buser–Karcher Lie group barycenter on free nilpotent Lie groups of path signatures. It proves that the barycenter map is an epimorphism of algebraic varieties (Theorem 4.7), is a non-commutative polynomial map (Theorem 4.9), and gives a closed-form expression at quadratic truncation (Theorem 4.11). For piecewise linear paths, it defines the barycenter recovery order B_{d,k}(α) and, for k=2, claims an exact formula B_{d,2}(α)=min(d, α+) if all α_i are even, and min(d, α+−#odd_α+1) otherwise (Theorem 7.2). The proof proceeds via explicit congruence normal forms of the signature matrix. The paper also provides an OSCAR implementation and several worked examples.","tokens_in":24904,"tokens_out":25149,"duration_ms":229545,"significance":"The main contribution is the exact recovery order for quadratic truncation and the explicit transformation to congruence normal forms, connecting signature varieties with classical matrix congruence theory. If the proof gap identified below is repaired, this is a valuable and citable result. The paper ships executable code and several examples that verify special cases, which strengthens reproducibility. The clean statement of Theorem 7.2 will likely be useful in further work on path recovery and signature geometry.","major_comments":[{"comment":"The simultaneous vector–matrix normalization is asserted without proof and is false for the displayed transform in Lemma 7.6. The proof states \"since the matrix from Theorem 7.6 transforms 1_{α1} to e1\", but the explicit P in Lemma 7.6 (with T_{12}) sends 1_m to √2 e_2 for m≥2, e.g. m=3. Example 7.13 instead uses a different, untwisted transform that sends 1_5 to e_1. The upper bound B ≤ rk(W_α) depends on this compatibility; without it, Theorem 7.2 would give only a lower bound. Please state and prove a lemma that the congruence normal-form transform can be chosen to also map 1_m to the prescribed vector (or prove the consistency of the two-step inverted transform).","section":"§7, proof of Theorem 7.2 and Lemma 7.6"},{"comment":"The proof concludes B_{d,2}(α)=rk(W_α), but the theorem states min(d, rk(W_α)). For α=(2,2) and d=2, one has rk(W_α)=4 while the theorem gives B=2. The proof does not address the dimensional cap d. A separate argument is needed for the case rk(W_α)>d: show that recovery with d segments is possible (e.g. by the general upper bound B≤d from [1, Theorem 3.4]) and that fewer than min(d,rk) segments are impossible, e.g. via the rank of the skew-symmetric part.","section":"§7, proof of Theorem 7.2"},{"comment":"The proof of Lemma 7.6 says \"by construction\" and does not verify that the displayed P realizes the stated normal form, nor does it reconcile the block ordering with the T_{12} permutation. Since Theorem 7.2 relies on the vector action of this same matrix, the explicit expression should be checked and the version used in Example 7.13 (without T_{12}) should be reconciled with the lemma as stated.","section":"§7, Lemma 7.6"}],"minor_comments":[{"comment":"The notation T_{m+1}^{12} is undefined; presumably T_{12}. Also check the indices in the expression E_{j,m-2(i-1)-1}.","section":"§7, Lemma 7.6"},{"comment":"The proof writes B_{d,k}(α) where the theorem concerns B_{d,2}(α); correct the symbol.","section":"§7, proof of Theorem 7.2"},{"comment":"Example 8.4 refers to \"Theorem 6.2\", but the statement being verified is Example 6.2, not a theorem.","section":"§8, Example 8.4"},{"comment":"The text says 'In Theorem 2.2 we recall...' but the relevant statement is Example 2.2.","section":"§4, before Theorem 4.11"},{"comment":"The final transformation matrix is very large and hard to verify visually; consider splitting the computation or providing the corresponding OSCAR output in the repository.","section":"§7, Example 7.13"}],"recommendation":"major_revision","confidential_remarks":"The two self-citations [1] and [8] are to published peer-reviewed works, so they are not a concern. The main issue is the unproven simultaneous normalization in Theorem 7.2; based on Example 7.13 it appears repairable, but the manuscript must be revised carefully. The topic and methods fit the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi — this one is worth engaging with. The main new result, Theorem 7.2, gives an exact formula for the minimal number of segments needed to realize the quadratic-truncation barycenter of signature matrices. It looks correct, and the paper has a genuine proof gap in that theorem that the authors need to fix.\n\nWhat is actually new: the closed-form barycenter expression in Theorem 4.11 generalizes the known Heisenberg formula to all dimensions, and the explicit congruence transformations in Section 7 are new. The proof of Theorem 4.7 that the barycenter is an epimorphism is clean, and the accompanying OSCAR code is a real plus. The worked examples check out; I verified Example 6.2 and the parity cases in Figure 1. The self-citations to [1] and [8] are to published work and not a problem.\n\nThe soft spot is in the proof of Theorem 7.2, odd case. The text says 'since the matrix from Theorem 7.6 transforms 1_α1 to e1', but the explicit P from Lemma 7.6 does not do that. For m=3, that P sends 1_3 to √2 e2, not e1. This matters: the upper bound B ≤ rk(W_α) depends on simultaneously normalizing the vector and matrix levels, and the given proof does not establish it. Example 7.13 uses a different transformation that does map 1_5 to e1, so the gap is likely fixable — the formula is not in doubt — but as written the proof is incomplete. The two-segment impossibility in Lemma 8.5 is also asserted through a computational ideal rather than a displayed argument; that is minor since the code is provided.\n\nBottom line: the paper deserves a serious referee. Send it to review, and the referee should ask the authors to repair the simultaneous normalization step and reconcile the two different transformations. Once that is patched, it is a solid, citable contribution to signature-based path recovery.","headline":"Solid signature-recovery paper with a real proof gap in the odd-case simultaneous normalization; the formula likely holds, but the proof needs a patch before publication.","tokens_in":25489,"tokens_out":4230,"would_cite":true,"duration_ms":43329,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60L10","22E25","15A21","14Q15"],"pacs":[],"model":"deepseek-v4-flash","headline":"At matrix truncation level, this paper determines the exact minimal segment count for recovering a path from the barycenter of path signatures, and shows every group element is a barycenter of a sample with all but one element fixed.","keywords":["Lie group barycenters","free nilpotent Lie groups","iterated-integral signatures","piecewise linear paths","path recovery","matrix congruence normal forms","non-commutative polynomials","signature matrices"],"falsifier":"Take d=4 and α=(2,1), where the formula predicts B_{4,2}(2,1)=3. Let X_1 be the two-segment axis path in R^4 and X_2 a one-segment path along e_1, compute the quadratic barycenter with the paper's closed form (Theorem 4.11), and write its matrix projection W_α. Then apply the explicit congruence transform P constructed in the alternative proof of Theorem 7.10 and check whether P W_α P^T = C_3 and P(1_3/2) = e_1 simultaneously. If no simultaneous transform exists, the true minimal count is 4 and the formula fails; the check is a finite symbolic computation.","tokens_in":24480,"feed_emoji":"📐","tokens_out":12428,"duration_ms":121749,"temperature":0.7,"pith_summary":"This paper asks how to average a collection of paths so that the average is still a path signature. The standard expected signature of a random path generally leaves the signature space, so the authors work with the Lie-group barycenter, a group-valued mean, and study when that barycenter is the signature of a piecewise linear path and with how few segments. They prove that every element of the free nilpotent Lie group is a barycenter of some N-tuple of group elements, with any N−1 of them fixed in advance, and they give a closed-form expression for the barycenter at matrix truncation level. The central result is an exact formula for the minimal segment count: for segment counts α, B_{d,2}(α) = min(d, α_+) if every α_i is even, and B_{d,2}(α) = min(d, α_+ − #odd_α + 1) otherwise. If correct, this makes barycenter recovery constructive and low-dimensional: a matrix rank decides the segment count, and explicit matrix congruence transformations recover a realizing path.","feed_headline":"Exact formula finds the simplest path behind averaged signatures","feed_subtitle":"For matrix-level signatures, the minimum segment count depends only on the sample's total and odd counts.","key_machinery":"The load-bearing object is W_α, the matrix-level (quadratic) projection of the barycenter of the sample signatures after reducing to axis subpaths. It is a block-direct sum of skew-symmetric matrices U_{α_i} − U_{α_i}^T, where U_{α_i} is the strictly upper triangular matrix with ones above the diagonal, plus a rank-one all-ones term. The argument turns on the congruence normal form of W_α: it is congruent to Γ2⊕H2(−1)^{⊕...} when all α_i are even, and to 1⊕H2(−1)^{⊕...}⊕0^{⊕...} otherwise, so its rank is m in the first case and m−#odd_α+1 in the second. The proof uses these normal forms to construct explicit matrices P that send W_α to C_r = 1/2 I_r + U_r, the signature matrix of the r-segme","core_discovery":"On its own terms, the paper's claim is twofold. First, the barycenter map on the free nilpotent Lie group is an epimorphism of algebraic varieties: every group element occurs as the barycenter of some sample, and fixing N−1 of the N sample elements makes the map an isomorphism (Theorem 4.7). Second, at quadratic truncation the minimal number of segments needed to realize the barycenter of signatures of piecewise linear paths is exactly the value given in Theorem 7.2. The proof exhibits the matrix projection of the barycenter, W_α = (1/2N) ⊕_i (U_{α_i} − U_{α_i}^T) + (1/2N^2) 1_{m×m}, computes its rank and congruence normal form, and gives explicit transformations sending W_α to the signature","pith_inferences":["Editorial inference: the rank formula for W_α points to a plausible higher-level analogue in which B_{d,k}(α) would be governed by ranks of higher-order tensor congruence orbits; that analogue is not established here and could be tested computationally at k=3 for simple α.","Editorial inference: the compatibility issue flagged by the one-sentence vector–matrix normalization suggests that for truncation level k≥3, simultaneous normalization across tensor levels will become a genuine obstruction; the cubic two-segment example B_{d,3}(1,1)=3 may be the first instance of such behavior.","Editorial inference: because the authors note that all results lift to weighted distributions, the explicit quadratic barycenter formula can be turned into a practical weighted-average path estimator for ensembles of piecewise linear data streams, a step the paper does not develop.","Editorial inference: a randomized numerical check—draw many α_i-segment paths, compute W_α, and compare its rank with the minimal segment count found by solving the recovery system—would localize any failure of the compatibility assumption to particular α."],"forward_implications":["For k=2, barycenter recovery becomes an exact, constructive procedure: the rank of an explicitly written matrix gives the number of segments, and the paper's explicit congruence transformations produce a realizing path.","The barycenter of any number of one-segment path signatures is itself a one-segment signature, so averaging straight-line paths in signature space preserves straight-line structure.","No quadratic signature barycenter ever needs more than d segments, so in fixed dimension the barycenter of paths with many segments has a bounded-complexity representation.","The cubic case already differs: the barycenter of two one-segment signatures in R^2 needs three segments (B_{d,3}(1,1)=3), so segment-count behavior depends essentially on truncation level."],"supporting_citations":[{"why":"supplies existence and uniqueness of the barycenter plus the polynomial-map and Lyndon-basis techniques used to prove Theorem 4.7.","marker":"[8]"},{"why":"defines the Lie group barycenter equation that the paper adopts as its notion of group-valued mean.","marker":"[6]"},{"why":"provides equivariance properties of barycenters and the Heisenberg-group closed form that Theorem 4.11 generalizes to all d.","marker":"[22]"},{"why":"sets up learning paths from signature tensors via canonical axis paths and the recovery system s = A·σ(Axis_m).","marker":"[24]"},{"why":"describes the variety of signature tensors for piecewise linear paths and supplies the bound B_{d,2}(α)≤d and the closed-form axis signature.","marker":"[1]"},{"why":"gives Chen's identity and Chen-Chow's theorem, which yield the upper bound for recovery and the group structure of signatures.","marker":"[14]"},{"why":"supplies the canonical building blocks and congruence normal forms for matrices that structure the proof of Theorem 7.10.","marker":"[15]"},{"why":"guarantees real congruence transformations can be obtained from cosquare similarity, a step used in the normal-form arguments.","marker":"[12]"},{"why":"characterizes one-segment paths by their log-signature, used to prove Proposition 7.1 that all-one-segment samples have one-segment barycenter.","marker":"[13]"}],"fun_headline_variants":["Every signature is a barycenter of a sample","Minimal segments for signature barycenters","Exact path recovery from signature averages","Barycenter on signatures yields shortest path","Signature barycenter theorem: minimal segments"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The exact formula rests on the premise that one and the same linear change of coordinates can simultaneously put the matrix part of the averaged signature into its standard normal form and align its vector part with the first axis; the paper asserts this compatibility in a single sentence without a detailed verification. If that premise fails, the theorem's value would be only a lower bound on the true segment count.","fun_headline_variants_meta":{"raw":{"variants":["Every signature is a barycenter of a sample","Minimal segments for signature barycenters","Exact path recovery from signature averages","Barycenter on signatures yields shortest path","Signature barycenter theorem: minimal segments"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000664,"raw_usage":{"total_tokens":2827,"prompt_tokens":661,"completion_tokens":2166,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":405,"completion_tokens_details":{"reasoning_tokens":2111}},"tokens_in":405,"tokens_out":2166,"duration_ms":20586,"temperature":1.0,"reasoning_tokens":2111,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:41:44.754895+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take d=4 and α=(2,1), where the formula predicts B_{4,2}(2,1)=3. Let X_1 be the two-segment axis path in R^4 and X_2 a one-segment path along e_1, compute the quadratic barycenter with the paper's closed form (Theorem 4.11), and write its matrix projection W_α. Then apply the explicit congruence transform P constructed in the alternative proof of Theorem 7.10 and check whether P W_α P^T = C_3 and P(1_3/2) = e_1 simultaneously. If no simultaneous transform exists, the true minimal count is 4 and the formula fails; the check is a finite symbolic computation.","supporting_citations":[{"cited_title":"Clausel, J","cited_arxiv_id":null,"evidence_quote":"supplies existence and uniqueness of the barycenter plus the polynomial-map and Lyndon-basis techniques used to prove Theorem 4.7."},{"cited_title":"Buser and H","cited_arxiv_id":null,"evidence_quote":"defines the Lie group barycenter equation that the paper adopts as its notion of group-valued mean."},{"cited_title":"Pennec and M","cited_arxiv_id":null,"evidence_quote":"provides equivariance properties of barycenters and the Heisenberg-group closed form that Theorem 4.11 generalizes to all d."},{"cited_title":"Pfeffer, A","cited_arxiv_id":null,"evidence_quote":"sets up learning paths from signature tensors via canonical axis paths and the recovery system s = A·σ(Axis_m)."},{"cited_title":"Améndola, P","cited_arxiv_id":null,"evidence_quote":"describes the variety of signature tensors for piecewise linear paths and supplies the bound B_{d,2}(α)≤d and the closed-form axis signature."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives Chen's identity and Chen-Chow's theorem, which yield the upper bound for recovery and the group structure of signatures."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the canonical building blocks and congruence normal forms for matrices that structure the proof of Theorem 7.10."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"guarantees real congruence transformations can be obtained from cosquare similarity, a step used in the normal-form arguments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"characterizes one-segment paths by their log-signature, used to prove Proposition 7.1 that all-one-segment samples have one-segment barycenter."}],"review_version":1}