{"id":"75921569-814c-4d4b-bbb2-b0f80f1739b2","arxiv_id":"2412.19670","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Conjugation invariants of the path signature are exactly the cyclic-rotation sums of words; loop and closure invariants are characterized through a Lie-algebraic condition.","lead":"This paper characterizes which iterated-integral features of a path's signature are unaffected by conjugation, by moving the start point of a closed loop, and by closing an open path. The main theorem shows conjugation invariants are exactly the sums of cyclic rotations of words, giving a complete, explicit description of a natural class of stable features.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.10's key equivalence is asserted via an unproven 'derivation and linearization' step, and Lemma 3.5's R_n definition is internally inconsistent as written; the loop-invariant characterization needs a complete proof.","rationale":"The reader's CONDITIONAL verdict is appropriate: Theorem 3.10 is load-bearing and its proof is too compressed to verify, but the identified DLPR21 dependency is less direct than stated. The theorem can likely be repaired with a short Jacobi-identity argument, and Theorem 2.11 (ConjInv = im rot) is proved cleanly and is correct over R. The closure section has a concrete algebraic operator with explicit proofs, and the dimension tables are consistent with the stated formulas. The open conjectures are honestly flagged. The main weakness is the unexpanded derivation/linearization step and the inconsistent definition of R_n in Lemma 3.5; both are fixable but must be addressed before the loop-invariant characterization can be regarded as fully established. Hence no change to the reader's conditional verdict is needed.","tokens_in":16953,"tokens_out":37066,"duration_ms":296075,"concrete_test":"Independently complete the missing proof of Thm 3.10: (a) show, via Chen–Chow, that any grouplike h is a finite product of exponentials of linear elements, so invariance under conjugation by exp(a) suffices; (b) prove by induction on Lie length that, if ⟨[v,i],φ⟩=0 for all v∈V and all letters i, then ⟨[L,v],φ⟩=0 for every Lie polynomial L and v∈V, using Lemma 3.13 and the Jacobi identity without invoking Lemma 3.5/Prop 3.6 or [DLPR21]. If the induction goes through, Thm 3.10 is independent of the shuffle-generation input; if it fails, the theorem's characterization must be revised. As a complementary computational check, recompute the dimensions of the solution spaces {φ∈T_n : ⟨[v,i],φ⟩=0 for all v∈V_{n−1}, i} for d=2,3 and n≤8 using a PBW basis of V, and compare with the loop-invariant dimensions implied by the tables.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central loop-invariant theorem (Thm 3.10) rests on two compressed steps. After reducing arbitrary grouplikes h to linear exponentials exp(a), the paper states 'by derivation and linearization' that the resulting condition is equivalent to ⟨[v,i],φ⟩=0 for all v∈V and letters i. Neither the reduction to linear paths nor the derivation/linearization is written out, and the proof contains a typo ('Prop 3.2 and Prop 3.2'). To make the equivalence rigorous one must show, from the letter-level conditions, that ⟨[L,v],φ⟩=0 for every Lie polynomial L and v∈V, using the Jacobi identity and Lemma 3.13; this induction is absent. The reader's related concern about Lemma 3.5/Prop 3.6 and [DLPR21, Cor. 4.5] is partly real: those lemmas are needed for the dimension identities V_n=S_n^⊥ and for Prop 3.6, and Lemma 3.5's proof defines R_n first as 'shuffle monomials in non-letter words' and then identifies it with 'non-letter coordinates of the first kind'. For d=2,n=2 these spaces differ (span{12,21} versus span{12−21}), so the proof as written is not internally consistent. However, Thm 3.10 itself may be provable without Lemma 3.5, using only Prop 3.3, Lemma 3.13 and the Jacobi identity; the paper does not exhibit that proof, leaving the main characterization only partially verified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies three classes of signature invariants for paths in R^d: conjugation invariants, loop invariants, and closure invariants. Section 2 proves that the space of conjugation invariants is exactly the image of the cyclic rotation operator rot on words (Theorem 2.11), with a clean proof via Chen's identity and a characterization in terms of annihilation of Lie brackets with letters (Proposition 2.6). Section 3 characterizes loop invariants: Theorem 3.10 asserts that an element is a loop invariant if and only if it annihilates all brackets [v,i] with v in the span V of zero-increment grouplikes and i a letter. Section 4 introduces a right-closure operator rcl and proves that right-closure invariants form the image of rcl, with dimension counts given by the generating function of V; it also relates loop-and-closure invariants to letter-reduced loop invariants. Several conjectures are clearly stated as such, including the description LoopInv = S + AreaConj and related dimension conjectures.","tokens_in":17225,"tokens_out":7866,"duration_ms":70754,"significance":"If the main theorems are correct, the paper gives a complete and explicit algebraic classification of conjugation invariants as necklace words, and a tractable linear criterion for loop invariants, together with a concrete operator formulation for closure invariants. Theorem 2.11 is proved in full and is a satisfying result with direct computational consequences. The loop-invariant characterization in Theorem 3.10 is the central new result and is potentially very useful, but its proof is incomplete. The closure-invariant section is largely self-contained and contains useful explicit identities and dimension data. The paper is honest about its conjectural parts and provides reproducible-looking tables and OEIS links, which strengthens its value as a reference.","major_comments":[{"comment":"The proof of Theorem 3.10 is only a sketch and omits the load-bearing reduction from Lie brackets to letter brackets. After Proposition 3.3 one obtains <[L,g],phi>=0 for all Lie polynomials L and all zero-increment grouplikes g, hence <[L,v],phi>=0 for all v in V. To pass from this to condition (7) one must prove that [FL(R^d), V] is generated, as a linear span or closure, by brackets [i,v] with i a letter and v in V; this is not immediate and requires an induction using the Jacobi identity and Lemma 3.13. Conversely, to recover loop invariance from (7) one must show that the letter-level conditions imply <[L,v],phi>=0 for every Lie polynomial L; this reverse induction is also absent. The sentence 'by derivation and linearization' is therefore not a complete proof of the central characterization and must be expanded into a full argument.","section":"Section 3.1, Theorem 3.10"},{"comment":"The definition of R_n inside the proof of Lemma 3.5 is internally inconsistent. The proof first defines R_n as the span of shuffle monomials in non-letter words and then identifies it with the span of shuffle monomials in non-letter coordinates of the first kind. These two spaces differ; for d=2 and n=2 they are respectively span{12,21} and span{12-21}. With the first choice the direct sum T_n = S_n + R_n fails because S_2 and span{12,21} intersect nontrivially. Consequently the argument for the dimension identity V_n = S_n^\\perp is not established as written. Since Lemma 3.5 is used in Theorem 3.10, Proposition 3.6, Corollary 3.7, and the dimension counts in Section 4, this proof needs to be repaired or replaced with a consistent definition of R_n.","section":"Section 3.1, Lemma 3.5"},{"comment":"The proof of Proposition 3.6 relies on [DLPR21, Corollary 4.5] for the assertion that the dual elements f_i freely shuffle-generate T(R^d). This same result is also used in Lemma 3.5 to construct R_n. Because the cited item is a self-cited preprint and is load-bearing for the loop-invariant characterization, the statement should be reproduced in the paper, or the proof should be given directly. As it stands, the reader cannot verify from the manuscript that the particular dual basis constructed here has the required free shuffle-generation property.","section":"Section 3.1, Proposition 3.6"}],"minor_comments":[{"comment":"The proof says 'By Proposition 3.2 and Proposition 3.2'; the second reference should presumably be Proposition 3.3.","section":"Section 3.1, Theorem 3.10 proof"},{"comment":"The unnumbered display contains 'dim(V_n/ dim[V_{n-1}, R^d])'; the stray 'dim' in the denominator should be removed so the expression reads 'dim(V_n/[V_{n-1},R^d])'.","section":"Section 3.1, after Corollary 3.7"},{"comment":"The abstract contains the typo 'interated integrals', and the title page has spacing errors such as 'TU Berli n' and 'a natura l'; these should be corrected.","section":"Abstract and title page"},{"comment":"The level-12 row is missing entries for [V_{n-1},R^2] and for the letter-reduced loop invariants; the missing values should be supplied or the row should be annotated as incomplete.","section":"Table 3.1"},{"comment":"The reference [DLPR21] should include the arXiv identifier or a journal/digital object identifier if one is available.","section":"References"},{"comment":"The phrase 'For dimension 2 and up' should be 'For dimensions 2 and up'.","section":"Remark 2.14"}],"recommendation":"major_revision","confidential_remarks":"The central theorem of Section 3 depends essentially on the self-cited preprint [DLPR21] for shuffle-generation of the tensor algebra. I would ask the editor to ensure that this result is publicly available and correct. The remainder of the paper, especially Section 2 and the closure-invariant part, is in good shape; the main work for revision is to complete the proof of Theorem 3.10 and fix Lemma 3.5."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my honest take. The paper's headline result is Theorem 2.11: conjugation invariants of the iterated-integrals signature are exactly the image of the cyclic rotation operator rot. The proof is short and I could not find a gap: Proposition 2.6 reduces to a Lie-bracket condition via Chen-Chow, Lemma 2.7 handles the right-comb induction, and Proposition 2.10 plus the averaging argument gives the classification. That settles a natural question cleanly.\n\nThe other main contribution is the framework for closure invariants, with the rcl operator and the projection property. Lemma 4.7 and the isomorphisms in Section 4 are genuinely useful; I have not seen this way of separating the trivial loop invariants from the letter-reduced ones before.\n\nThe soft spots are in Section 3. Theorem 3.10, which characterizes loop invariants as [V, R^d]^{\\\\perp}, is only sketched. The proof says 'by derivation and linearization' and moves from conditions on linear paths to the Lie-bracket condition on all of V. That step is doing real work: one has to show that the letter-level conditions propagate to all Lie polynomials, via the Jacobi identity and Lemma 3.13. The induction is absent, and there is also a duplicate reference typo in the proof. I believe the theorem is likely true — I can see the induction one would write — but the paper as published does not actually prove it. Lemma 3.5, which identifies the orthogonal complement of V with the shuffle ideal S, also has an internally inconsistent definition of R_n: first it is 'shuffle monomials in non-letter words', then immediately it is 'shuffle monomials in non-letter coordinates of the first kind'. Those spaces differ (e.g., for d=2,n=2: span{12,21} vs span{12−21}). The lemma's statement is probably correct and the proof can be fixed, but as written it is not right.\n\nThe reliance on [DLPR21, Cor. 4.5] for free shuffle-generation of coordinates of the first kind is real, but that is an external published result, and not obviously suspect; the dependency is at least explicit. The dimension tables and the conjectures are honest and look consistent.\n\nNet: this is a paper worth publishing after revision. The conjugation-invariant classification and the rcl machinery are solid, and the loop-invariant theorem is worth having even in its current sketched form, but it should not go to press without a complete proof and a repaired Lemma 3.5. I would send it to a serious referee, and I would cite the conjugation and closure results in my own work.\n\nWho for: anyone working on signature invariants, rough paths, or combinatorial Hopf algebras. The reading group would enjoy the clean parts and could spend a session reconstructing the missing induction.","headline":"Conjugation invariants are completely nailed down by a short, clean theorem; the loop-invariant half is an attractive but under-proved sketch.","tokens_in":17795,"tokens_out":6579,"would_cite":true,"duration_ms":133952,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["17B01","16T30","68R15"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that the conjugation-invariant features of the iterated-integrals signature are exactly the cyclic rotations of words, characterizes loop invariants by a bracket annihilation condition, and constructs a closure projection…","keywords":["iterated-integrals signature","conjugation invariants","loop invariants","closure invariants","shuffle algebra","free Lie algebra","combinatorial necklaces","signed area"],"falsifier":"Compile the pairing matrix between the space $V_n$ (spanned by grouplike elements with zero increment) and the span of shuffle monomials in non-letter coordinates of the first kind, and check that its rank equals $\\dim V_n$ predicted by the generating function $(1-q)^d/(1-dq)$. A rank mismatch gives a concrete counterexample to Lemma 3.5 and hence to the loop-invariant theorem; a concrete element satisfying the bracket condition but failing loop invariance would also falsify Theorem 3.10 directly.","tokens_in":16710,"feed_emoji":"🔁","tokens_out":11751,"duration_ms":94713,"temperature":0.7,"pith_summary":"This paper gives algebraic characterizations of three natural invariance classes for the iterated-integrals signature of a path: features unchanged by path conjugation, features of closed loops unchanged by moving the starting point, and features unchanged by closing a path with a straight segment. It proves that the conjugation invariants are exactly the sums of cyclic rotations of words, $\\operatorname{rot}(w)$, so they correspond level by level to combinatorial necklaces. For loops it proves that a feature $\\varphi$ is invariant exactly when $\\langle[v,i],\\varphi\\rangle=0$ for every zero-increment grouplike element $v\\in V$ and every coordinate direction $i$, and it constructs a projection $\\mathrm{rcl}$ whose image is the right-closure invariants and whose kernel is the shuffle ideal $S$. Loop-and-closure invariants then coincide, as a graded algebra, with the letter-reduced loop invariants, and explicit dimension formulas follow from the generating function $(1-q)^d/(1-dq)$.","feed_headline":"Conjugation invariants equal cyclic rotations of words","feed_subtitle":"Loop invariants are the annihilator of a bracket space, with dimension formulas at every level.","key_machinery":"The central objects are the rotation operator $\\operatorname{rot}(w)$, which sums the cyclic rotations of a word; the shuffle ideal $S$ generated by the letters; and the space $V$ spanned by grouplike elements whose signature has zero increment. A further load-bearing piece is the coordinates of the first kind indexed by Lyndon words, which freely shuffle-generate the tensor algebra and identify $V_n^\\perp$ with $S_n$. The closure results are carried by the explicit operator $\\mathrm{rcl} = \\mathrm{sh}\\circ(\\mathrm{id}\\otimes H)\\circ\\Delta_\\bullet$, where $\\mathrm{sh}$ is the shuffle product and $H$ is the signed shuffle-antipode map; its kernel is exactly $S$, making it a projection onto the right-closure invariants.","core_discovery":"The paper's central discovery is that three a priori geometric invariance questions about paths reduce to finite algebraic conditions on words. Theorem 2.11 classifies conjugation invariants completely: $\\mathrm{ConjInv} = \\operatorname{im}\\operatorname{rot}$, the image of the operator that maps a word to the sum of its cyclic rotations, with the repetition multiplicity counted. Theorem 3.10 characterizes loop invariants as $[V,\\mathbb{R}^d]^\\perp$, where $V$ is the span of grouplike elements with vanishing increment; equivalently, $\\varphi$ is a loop invariant iff $\\langle[v,i],\\varphi\\rangle=0$ for all $v\\in V$ and all letters $i$. The paper further proves that $\\varphi$ is a right-closure invariant iff it lies in the image of the explicit operator $\\mathrm{rcl} = \\mathrm{sh}\\circ(\\mathrm{id}\\otimes H)\\circ\\Delta_\\bullet$, and that the kernel of $\\mathrm{rcl}$ is the shuffle ideal generated by the letters. It leaves as conjectures the precise structure of the loop invariants in low dimensions, in particular whether they are generated by signed areas and conjugation invariants.","pith_inferences":["A practical consequence the paper does not spell out: because $\\operatorname{rot}$ is a linear operator on words, one can precompute a basis of conjugation invariants for each fixed degree once, then project any signature word vector onto this basis to obtain a canonical, start-point-independent feature set.","The closure projection $\\mathrm{rcl}$ suggests a data-cleaning step: replacing a raw signature feature $\\varphi$ by $\\mathrm{rcl}(\\varphi)$ removes exactly the component that changes when the path is closed by a straight segment, which is testable on piecewise-linear trajectory data.","The conjectured low-dimensional structure $\\mathrm{LoopInv}=S+\\mathrm{AreaConj}$ in dimensions 2 and 3 is a strong falsifiable prediction: a computer search through the letter-reduced loop-invariant tables should either find a first counterexample or support the conjecture, effectively settling Conjecture 3.12."],"forward_implications":["At each word length $m$, the conjugation invariants form a vector space of dimension equal to the number of $m$-bead necklaces, with a canonical basis given by the rotations $\\operatorname{rot}(w)$ over necklace representatives.","A signature element is a loop invariant exactly when it is orthogonal to every bracket $[v,i]$ with $v$ in the zero-increment grouplike span $V$; level by level this is a finite-dimensional linear condition.","Right-closure invariants are exactly the image of the explicit operator $\\mathrm{rcl}$, and since $\\ker\\mathrm{rcl}=S$, the loop-and-closure invariants are isomorphic as graded algebras to the letter-reduced loop invariants.","Dimension formulas follow: the loop-and-closure invariants at level $n$ have dimension $\\dim V_n - \\dim[V_{n-1},\\mathbb{R}^d]$, computed from the generating function $(1-q)^d/(1-dq)$.","For $d\\ge 2$ there are infinitely many shuffle-algebraically independent conjugation invariants and infinitely many shuffle-independent loop-and-closure invariants, so the invariant feature sets are large enough for practical use."],"supporting_citations":[{"why":"Supplies Chen's identity and the Chow/Chen theorem that connect signature values to grouplike elements, the bridge from path-level invariance to algebraic bracket conditions.","marker":"[FV10]"},{"why":"Corollary 4.5 gives the free shuffle-generation of the tensor algebra by coordinates of the first kind, the load-bearing input for Lemma 3.5 and Proposition 3.6.","marker":"[DLPR21]"},{"why":"Provides the lemma that grouplike elements span the relevant tensor spaces and the signed-area and signed-volume examples used throughout.","marker":"[DR19]"},{"why":"Supplies the Lyndon basis of the free Lie algebra, shuffle algebra facts, and the necklace correspondence underlying rot and the dimension counts.","marker":"[Reu93]"},{"why":"Establishes that the signature is well-defined up to reparametrization and tree-like equivalence, the path-equivalence setting used in the loop-invariant arguments.","marker":"[BGLY16]"},{"why":"Introduces the iterated-integrals signature of paths, the object whose invariant features the paper classifies.","marker":"[Che58]"}],"fun_headline_variants":["Cyclic rotations capture conjugation invariants","Loop invariants: annihilator of a bracket space","Path invariants boil down to word algebra","Three invariance questions, one algebraic result"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The loop-invariant theorem rests on the assumption that a certain set of non-letter coordinates freely generates the whole tensor algebra under the shuffle product; if that generation statement is false or incomplete, the key identification of the zero-increment grouplike span with the orthogonal complement of the shuffle ideal, and with it the characterization of loop invariants, collapses.","fun_headline_variants_meta":{"raw":{"variants":["Cyclic rotations capture conjugation invariants","Loop invariants: annihilator of a bracket space","Path invariants boil down to word algebra","Three invariance questions, one algebraic result"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1434,"prompt_tokens":941,"completion_tokens":493,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":437}},"tokens_in":557,"tokens_out":493,"duration_ms":10475,"temperature":1.0,"reasoning_tokens":437,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:59:59.752371+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compile the pairing matrix between the space $V_n$ (spanned by grouplike elements with zero increment) and the span of shuffle monomials in non-letter coordinates of the first kind, and check that its rank equals $\\dim V_n$ predicted by the generating function $(1-q)^d/(1-dq)$. A rank mismatch gives a concrete counterexample to Lemma 3.5 and hence to the loop-invariant theorem; a concrete element satisfying the bracket condition but failing loop invariance would also falsify Theorem 3.10 directly.","supporting_citations":[],"review_version":1}