{"id":"f7e6504b-61ca-4ae0-81dd-3121685f8632","arxiv_id":"2607.06271","paper_version":1,"verdict":"ACCEPT","confidence":"UNKNOWN","novelty_score":8.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A unique normal form for tensor trains over arbitrary fields is defined using a novel LDPU matrix decomposition, with a polynomial-time reduction algorithm and direct construction from full tensors.","lead":"The paper defines a unique normal form for tensor trains (a compressed tensor representation) that works over any field, including finite fields like F2. This matters because it enables canonical, minimal-size representations analogous to reduced binary decision diagrams, but for a strictly more general data structure.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"No significant objection identified. The P-conditioning uniqueness argument in Appendix C is subtle but sound: REF zero-structure before leading indices enables an inductive peeling argument that forces σ₁(k)=σ₂(k) and L₁=L₂ entry by entry.","rationale":"The reader correctly identified the load-bearing concern: the P-conditioning's sufficiency for LDPU uniqueness. Having traced through the proof in detail, the argument is sound. The REF zero-structure (entries before the leading index are zero) is the key property that makes the inductive peeling argument work, forcing permutation agreement and then entry-by-entry agreement of L. The extension to tensor train normal form uniqueness (Theorem 16) is a straightforward induction using LDPU uniqueness at each core boundary, supported by the REF/CEF preservation lemmas (Lemmas 9–13), whose proofs are also correct. The algorithmic claims (O(nm min(n,m)) for LDPU, O(rs) for reduction, 8/3 size bound) are all justified by the complexity analysis in the proof of Theorem 19 and Proposition 21. The paper has no formal verification or reproducible code, but the mathematical arguments are self-contained and verifiable. No adjustment to the ACCEPT verdict is warranted.","tokens_in":22260,"tokens_out":26824,"duration_ms":1103803,"concrete_test":"Implement the LDPU decomposition (Algorithm 4) over F₂ and over Q, then stress-test uniqueness by brute force: for random matrices A of sizes up to 5×5 with ranks 0 through 5, enumerate all valid (L,D,P,U) tuples satisfying the LDPU constraints and verify exactly one exists. Pay particular attention to rank-deficient matrices and matrices where multiple pivot orderings are a priori possible. Additionally, implement the full two-sweep reduction (Algorithms 2–3) and verify that two randomly generated tensor trains representing the same tensor (via random gauge transformations A·A⁻¹) reduce to identical normal forms.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader correctly identifies the P-conditioning (Definition 4) as the most delicate assumption. Having worked through the proof of Proposition 7 in Appendix C carefully, I believe the argument holds. The key step is the inductive proof that σ₁(k)=σ₂(k) for the k-th pivot. Suppose for contradiction σ₁(k)<σ₂(k). The equation A[i_k] = Σ_{j∈S₁} L₁[i_k,j]D₁[j,j]U₁[σ₁(j)] + D₁[k,k]U₁[σ₁(k)] = Σ_{j∈S₂} L₂[i_k,j]D₁[j,j]U₁[σ₁(j)] + D₂[k,k]U₂[σ₂(k)] where S₁={jσ₁(k), and U₂[σ₂(k)][σ₁(k)]=0 since σ₂(k)>σ₁(k). By examining the equation at position σ₁(j₁) (smallest leading index in S₁), only U₁[σ₁(j₁)] contributes, forcing L₁[i_k,j₁]=L₂[i_k,j₁]. Peeling inductively through all j∈S₁ by ascending σ₁(j), we get L₁=L₂ on S₁. Then at position σ₁(k): D₁[k,k]=0, contradicting D invertible. So σ₁(k)=σ₂(k), and the rest follows. The extension to Theorem 16 (tensor train uniqueness) is a clean induction using LDPU uniqueness at each core cut, supported by Lemmas 9–13. I found no gap. The notation in Appendix C is occasionally confusing (σ_i denotes the inverse of the permutation from Definition 4, and there is a likely typo 'Ui[σi(k)]' for 'U₁[σ₁(k)]'), but these are presentation issues, not correctness issues.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper introduces the LDPU decomposition—a variant of LU decomposition that is rank-revealing and, crucially, unique over arbitrary fields—and uses it to define the first unique normal form for tensor trains. The main results are: (1) the LDPU decomposition exists, is unique, and is computable in O(nm min(n,m)) time (Proposition 7); (2) a normal form for tensor trains exists and is unique (Theorem 16); (3) a two-sweep reduction algorithm reaches this normal form in O(rs) time without increasing the tensor train size (Theorem 19), implying minimality of TT-ranks and size (Corollary 20); and (4) the size of a fully reduced tensor train is at most 8/3 times the full tensor size, which is tight (Proposition 21). The paper also provides algorithms for direct construction from a full tensor (Algorithm 1), nullity testing (Proposition 17), and leading-index extraction (Proposition 18). All proofs are given in full in the appendix.","tokens_in":22666,"tokens_out":3577,"duration_ms":153805,"significance":"The paper addresses a genuine gap: existing SVD-based tensor train reductions (HSVD) are not unique and are restricted to R and C. The LDPU decomposition and its uniqueness proof are the core technical contributions. The extension to tensor trains via a clean inductive argument at each core cut is well-executed. The connection to BDDs over F2 provides concrete motivation, and the size bound (Proposition 21) with a matching lower bound is a nice addition. The author is appropriately candid about practical limitations (numerical instability, exponential bit complexity for infinite fields), positioning the contribution as primarily relevant to finite fields. The string-diagrammatic framework, while adding length, does make the reshaping arguments more transparent.","major_comments":[],"minor_comments":[{"comment":"Appendix C, proof of Proposition 7, induction step: the text states 'we get σ₁(1) = σ₂(1)' where it should read 'σ₁(k) = σ₂(k)'. The same typo repeats for D₁[1,1] = D₂[1,1] and U₁[σ₁(1)] = U₂[σ₂(1)], which should reference index k. This is a notational error in the induction step, not a gap in the argument.","section":null},{"comment":"Appendix C, proof of Proposition 7: the phrase 'By taking each U₁[σ₁(j)] in order by ascending leading coefficient' should read 'ascending leading index', since all leading coefficients are 1 in unit REF. The peeling argument proceeds by ascending leading index, not coefficient.","section":null},{"comment":"Appendix C, proof of Proposition 7: the notation σᵢ is introduced as 'the inverse of the permutation induced by Pᵢ', which is somewhat unusual. A brief clarifying remark that σᵢ maps column-pivot indices to their positions would help readers follow the inductive argument.","section":null},{"comment":"Algorithm 4 (LDPU): the return statement 'return L^⊺D⁻¹, D, P, P⁻¹A' could benefit from a brief comment explaining why the transpose appears (converting the row-operation tracking matrix to column-echelon form) and why D⁻¹ normalizes L to unit form. Without this, the connection between the algorithm's internal bookkeeping and the LDPU definition (Definition 5) is not immediately obvious.","section":null},{"comment":"Section 3.1, Example 6: the example is helpful, but the permutation σ is described only through its inverse values. Stating σ explicitly (σ(0)=1, σ(1)=2, σ(2)=0) would make the P-conditioning verification easier to follow.","section":null},{"comment":"Definition 14: the two conditions reference Pᵢ-conditioning, but the definition of P-conditioning (Definition 4) is stated for column-echelon form matrices. A brief remark confirming that intertwine(Cᵢ, nᵢ, Pᵢ₋₁)Pᵢ⁻¹ is indeed in the form required by Definition 4 would improve readability.","section":null},{"comment":"Theorem 19: the complexity is stated as O(rs) where r is the max TT-rank and s is the size. It would help to explicitly note that this assumes naive matrix multiplication, as mentioned later in the text, since the O(rs) bound depends on the LDPU cost being O(nm min(n,m)) with naive arithmetic.","section":null},{"comment":"Proposition 21, proof: the derivation uses the bound nᵢ ≥ 2 to replace geometric sums with 4/3. This assumption should be stated explicitly in the proposition or its proof, since modes of dimension 0 or 1 are allowed elsewhere in the paper.","section":null},{"comment":"The paper would benefit from a brief discussion of how the normal form interacts with core reordering (permuting the modes of the tensor train), given that the author notes in Section 2.1 that core ordering can dramatically change size. Even a sentence acknowledging that the normal form is defined for a fixed mode ordering would suffice.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is a solid contribution that fills a clear gap in the tensor train literature. The LDPU decomposition appears to be a novel or at least underused tool, and its uniqueness proof is the key technical insight. The referee checks (reader report and stress-test) both confirm the correctness of the central arguments. The presentation issues are real but local—none affect the load-bearing claims. I would encourage the editor to accept after minor revision, as the fixes are straightforward. The paper fits well within the scope of a data structures / formal methods venue."},"author_rebuttal":{"model":"glm-5.2","summary":"The referee report is uniformly positive, recommending minor revision, with no major comments listed. We thank the referee for the careful reading and note that we will address minor typographical issues in the next version.","responses":[],"tokens_in":21724,"tokens_out":115,"duration_ms":7294,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Short version: this paper gives the first unique normal form for tensor trains over arbitrary fields, and the proofs hold up. The core construction — the LDPU decomposition — is a genuine new contribution, and the uniqueness argument is clean once you sit with it. This deserves a serious referee and likely belongs in a good venue. The reader's assessment is on target; the stress-test concern about P-conditioning is real but resolves on careful reading. The paper's main achievement is defining a rank-revealing matrix decomposition (LDPU) that is both unique and field-agnostic, then using it to build a canonical tensor train form. The P-conditioning constraint (Definition 4) is the key insight: it adds just enough structure to force uniqueness without sacrificing existence. Proposition 7's proof by inductive peeling of pivot rows is correct — the REF zero-structure before leading indices does force sigma_1(k) = sigma_2(k) and then L entry-by-entry. Theorem 16 extends this to tensor trains by cutting at each core and applying LDPU uniqueness inductively, supported by Lemmas 9–13. The two-sweep reduction algorithm (Theorem 19) is polynomial, and the size bound (Proposition 21, factor 8/3) is tight with a matching lower bound. The string-diagram notation is a real help here — it makes the reshape arguments readable rather than index-heavy. The BDD connection over F_2 is well-motivated and not oversold. Soft spots are minor. The notation in Appendix C is occasionally confusing — sigma_i denotes the inverse of the permutation from Definition 4, and there is a likely typo where U_i[sigma_i(k)] should be U_1[sigma_1(k)]. These are presentation issues, not correctness issues. The paper honestly flags that the approach suffers from numerical instability and exponential bit complexity over infinite fields — the construction is really best suited for finite fields. This is fine and appropriately scoped. One thing I'd want a referee to check: whether the LDPU can be computed in matrix-multiplication time O(n^omega), which the author leaves open. This affects practicality but not correctness. This paper is for readers in formal methods, knowledge compilation, and tensor network theory who need exact equivalence checking or canonical representations over finite fields. It bridges BDDs and tensor networks in a concrete way.","headline":"First unique normal form for tensor trains over arbitrary fields, built on a novel LDPU decomposition. Proofs check out.","tokens_in":23381,"tokens_out":551,"would_cite":true,"duration_ms":83140,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Unique normal form found for tensor trains over any field","keywords":[],"falsifier":"A matrix over some field for which two distinct LDPU decompositions exist — i.e., two different sets of (L, D, P, U) all satisfying the P-conditioning and echelon-form constraints — which would break the uniqueness claim of Proposition 7 and cascade to destroy Theorem 16.","tokens_in":22562,"feed_emoji":"🧮","tokens_out":1044,"duration_ms":238723,"temperature":0.7,"pith_summary":"The paper proves that every tensor train — a chain of small tensors whose product represents a large multidimensional array — can be brought to a canonical minimal form that is unique, and this form can be reached in polynomial time. The key innovation is the LDPU decomposition: a variant of LU matrix factorization that, unlike the singular value decomposition used in prior approaches, works over arbitrary fields (including finite fields like the 2-element Galois field) and is provably unique. The author defines the normal form through two structural conditions on the cores of the tensor train (row-echelon form on one axis, column-echelon form with a permutation-conditioning constraint on the other), shows that a deterministic algorithm produces it, and then proves no other tensor train with the same modes can satisfy both conditions simultaneously. A two-sweep reduction algorithm, running in time proportional to the product of the maximum rank and the tensor train size, brings any tensor train into this form without ever increasing its size. The paper also provides a direct construction algorithm that builds the normal form from a full tensor, a method to extract the leading nonzero index and value, and a tight worst-case size bound showing the normal form uses at most 8/3 times the space of naive storage.","feed_headline":"Unique normal form found for tensor trains over any field","feed_subtitle":"An LDPU matrix decomposition replaces SVD to give tensor trains a canonical form over finite fields, with polynomial-time reduction and a 8/","key_machinery":"The LDPU decomposition (Definition 5): a matrix factorization A = LDPU where L is unit column-echelon and P-conditioned, D is diagonal, P is a permutation, and U is unit row-echelon. Uniqueness is proved by an inductive argument showing pivot rows, column pivots, diagonal entries, and all remaining entries of L are forced. The tensor train normal form (Definition 14) requires two conditions on each core: the block matrix of slices is in full row-rank row-echelon form, and intertwine(C_i, P_i) is in full column-rank column-echelon form and P-conditioned. The two-sweep reduction (Theorem 19) enforces these conditions sequentially.","core_discovery":"The central object is the LDPU decomposition, which factors any matrix A as L·D·P·U where L is in unit column-echelon form and is P-conditioned (a constraint tying the zero-pattern of L to the permutation P), D is diagonal, P is a permutation matrix, and U is in unit row-echelon form. The author proves this decomposition exists, is unique for any matrix over any field, and is rank-revealing — the three properties needed to serve as the replacement for SVD in the tensor train setting. Uniqueness of the tensor train normal form (Theorem 16) is then established by showing that each cut between consecutive cores corresponds exactly to an LDPU decomposition of a reshaped partial tensor, so the LD","pith_inferences":[],"forward_implications":["Tensor trains over finite fields can now serve as a canonical data structure with the same uniqueness guarantee that BDDs enjoy, but with provably better compactness for some function families — enabling direct substitution in model checking, satisfiability, and circuit synthesis pipelines.","Equality testing between two tensor trains reduces to subtracting one from the other (polynomial time), applying the first sweep, and checking whether the first core is null — a consequence of Proposition 17 that avoids full tensor reconstruction.","The leading nonzero index and value of a tensor train can be extracted in O(dnr) after the first sweep (Proposition 18), which is faster than the O(dr^2) needed for an arbitrary index, exploiting the echelon structure of the cores.","The worst-case size bound of 8/3 times naive storage (Proposition 21) means tensor trains in normal form are never catastrophically larger than the full tensor, matching the constant-factor overhead known for BDDs."],"fun_headline_variants":["First canonical form for tensor trains over arbitrary fields","LDPU decomposition gives tensor trains a unique normal form","Tensor trains get provable reduction over any field via LDPU","Unique tensor train normal form extends beyond real and complex","Polynomial-time reduction to a unique tensor train normal form"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The uniqueness of the LDPU decomposition — and hence of the entire normal form — rests on the P-conditioning constraint on the lower triangular factor L being sufficient to force every entry of L to be uniquely determined. The proof argues inductively that pivot rows and column pivots are shared between any two decompositions and that P-conditioning pins down all remaining entries, but if some edge case involving specific field characteristics or degenerate rank profiles were","fun_headline_variants_meta":{"raw":{"variants":["First canonical form for tensor trains over arbitrary fields","LDPU decomposition gives tensor trains a unique normal form","Tensor trains get provable reduction over any field via LDPU","Unique tensor train normal form extends beyond real and complex","Polynomial-time reduction to a unique tensor train normal form","LDPU replaces SVD to yield unique tensor train canonical form","Canonical tensor train form proven over arbitrary fields","Rank-revealing LDPU decomposition enables unique tensor trains"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":673,"prompt_tokens":530,"completion_tokens":143,"prompt_tokens_details":null},"tokens_in":530,"tokens_out":143,"duration_ms":11987,"temperature":1.0,"reasoning_tokens":29,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T11:14:50.111345+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"A matrix over some field for which two distinct LDPU decompositions exist — i.e., two different sets of (L, D, P, U) all satisfying the P-conditioning and echelon-form constraints — which would break the uniqueness claim of Proposition 7 and cascade to destroy Theorem 16.","supporting_citations":[],"review_version":1}