{"id":"61acc25e-07d8-48c3-afd8-7b5899854a3b","arxiv_id":"1908.03090","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"TDVP is extended from matrix product states to arbitrary tree tensor networks, and used to time-evolve Fork Tensor Product States with off-diagonal hybridizations.","lead":"This paper generalizes the time-dependent variational principle, an algorithm for simulating quantum time evolution, to any finite loop-free tensor network. The authors show it can handle Hamiltonians with off-diagonal hybridizations relevant to spin-orbit coupling in impurity models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 9's vertical-subspace parameterization appears false already for a two-site MPS: the contraction requires an index transposition at one end of each link, so the derivation of Eq. 15 is unsupported as written.","rationale":"The reader's weakest assumption is Eq. 9's parameterization of the vertical subspace. My check confirms that this is the load-bearing point, but it identifies a more concrete defect than mere lack of proof: the parameterization is algebraically wrong for a two-site MPS because it uses right multiplication by X at both ends of a link instead of right multiplication at one end and the transposed action at the other. This invalidates the derivation of the gauge-fixing constraints, the least-squares solution, and hence the proof that Eq. 15 is the tangent-space projector. The final projector Eq. 15 itself appears to be correct in the two-site limit, and the numerical demonstrations are consistent with the algorithm working, so the manuscript may be fixable by correcting Eq. 9 and re-deriving the projector. The verdict should remain CONDITIONAL: the central claim is plausible and empirically supported, but the written derivation is not reliable as it stands.","tokens_in":16940,"tokens_out":34113,"duration_ms":376638,"concrete_test":"Specialize Eq. 9 to a two-site MPS with bond dimension D=2, physical dimension 2, and canonical tensors A = C = I_2. Choose X with a single off-diagonal entry X_{12}=1 and all other entries zero; compute |Theta[B]> from Eq. 8. If |Theta[B]> = |1,2> - |2,1> != 0, Eq. 9 is false. Then replace the second-site term with B2 = -C X^T and confirm the cancellation, and check whether the corrected derivation still reproduces Eq. 15 and the standard MPS tangent-space projector.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The derivation of Eq. 15 rests on Eq. 9, which asserts that the vertical subspace is parameterized by one matrix X per link. As written, Eq. 9 does not give the vertical subspace even in the MPS limit. For a two-site network (N=2, bond q, endpoint N=2) with canonical tensors A_{s1,q} and C_{s2,q}, Eq. 9 gives B1 = A X and B2 = -C X (right multiplication by X on both sites). The zero-state condition following from Eq. 8 is sum_q B1_{s1,q} C_{s2,q} + sum_q A_{s1,q} B2_{s2,q} = 0, i.e. B1 C^T + A B2^T = 0. Substituting B1 = A X and B2 = -C X yields A X C^T - A X^T C^T, which vanishes only for symmetric X. Taking A = C = I_2 and X with a single off-diagonal entry X_{12}=1 gives |Theta[B]> = |1,2> - |2,1> != 0, so Eq. 9 maps a nonsymmetric X to a nonzero tangent vector. The gauge variation of the second tensor must instead be delta C = -C X^T, because the contraction is A C^T, not A C; Eq. 9 omits this transposition on one side of each link. Consequently, the counting argument after Eq. 10 and the minimization leading to Eq. 15 are not valid as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper generalizes the time-dependent variational principle (TDVP) to finite loop-free tensor networks (TTNs). The derivation follows the MPS route of Haegeman et al.: it defines a tangent-space representation, parameterizes the vertical subspace by one matrix per link (Eq. 9), imposes gauge-fixing constraints (Eq. 10), and solves a least-squares problem to obtain a tangent-space projector (Eq. 15). The projector is then integrated by sequential local site and link updates, yielding first- and second-order single-site and two-site sweeping algorithms. As an application, the authors implement TDVP for Fork Tensor Product States (FTPS) and use it to compute impurity Green's functions for multi-orbital Anderson impurity models with off-diagonal hybridizations, benchmarking against TEBD and exact diagonalization. The numerical results in Figs. 10–12 show good agreement and convergence.","tokens_in":17261,"tokens_out":19692,"duration_ms":190429,"significance":"If the central derivation is correct, the paper provides a genuinely useful framework: TDVP for arbitrary tree tensor networks, beyond the MPS and binary-tree cases, with the practical advantage that only a Hamiltonian representation in the same tensor-network structure is needed. The FTPS application targets a real problem in DMFT, namely off-diagonal hybridizations relevant for spin-orbit coupling and lattice distortions. The numerical tests are appropriate and include external benchmarks (TEBD, exact diagonalization) rather than self-referential checks, and the convergence study in Fig. 12 is a strength. The presentation is clear and the claimed generalizations are concrete. However, the derivation of the central projector contains a load-bearing error in the parameterization of the vertical subspace, which currently prevents the main theoretical claim from being accepted as stated. The algorithm may well be correct, but the paper's proof is not.","major_comments":[{"comment":"The proposed parameterization of the vertical subspace is incorrect, because it uses right multiplication by the same matrix X on both sides of each link. For a two-site TTN with endpoint N=2, tensors A_{s1,q} and C_{s2,q}, Eq. (9) gives B1 = A X and B2 = -C X. The zero-state condition following from Eq. (8) is B1 C^T + A B2^T = 0. Substituting yields A X C^T - A X^T C^T = A (X - X^T) C^T, which vanishes only for symmetric X. Taking A = C = I_2 and X with the single off-diagonal entry X_{12} = 1 gives |Theta[B]> = |1,2> - |2,1> != 0, so a non-symmetric X is mapped to a nonzero tangent vector. The correct gauge variation on the side of the link away from the endpoint requires the transpose: B2 = -C X^T (equivalently, left multiplication by X). Consequently, Eq. (9) does not reduce to the MPS kernel of Ref. [29], and the derivation of Eqs. (10)–(15) as written is unsupported. The kernel parameterization and the subsequent counting argument must be re-derived with the correct orientation of the X matrices on each side of every link; the final projector may be correct, but the proof needs to be fixed.","section":"Sec. 3.1, Eq. (9)"},{"comment":"The Hamiltonian representation in the FTPS network is never constructed explicitly. The effective one-site and link Hamiltonians in Eqs. (16) and the two-site Hamiltonian in Eq. (18) presuppose a tensor-network (MPO-like) representation of H in the same FTPS geometry. For the off-diagonal hybridization Eq. (22), the paper only states that the matrix V can be chosen lower-triangular; it does not show how the terms c^†_{l0} c_{l'k} are encoded as operators on the TTN. Since the abstract and introduction emphasize that the method works whenever such a representation is available, and since the FTPS demonstration is a central part of the paper, the explicit construction (or a precise pointer to it) is needed for reproducibility and to substantiate the claim that the off-diagonal case is within the method's scope.","section":"Sec. 4, Eqs. (16)–(22)"}],"minor_comments":[{"comment":"The notation T_N^{[q_l]} in Eq. (9) is confusing: the subscript N suggests the endpoint of the tree, but the text says this is the tensor on site i orthogonalized towards the neighbor along q_l. It should read T_i^{[q_l]} (or an equivalent site label). The same ambiguity appears in Sec. 2.2 when tensor 4 is denoted as T_N^{[q2]}.","section":"Sec. 3.1, Eq. (9) and Sec. 2.2"},{"comment":"There are several typos and formatting issues: 'coefficientcs1···sN' in Sec. 2.1, 'Where∑<i,j>qk' after Eq. (15), and '10 −9' in the captions of Figs. 10 and 11 should be corrected to proper mathematical notation.","section":"Throughout"},{"comment":"In the algorithm description, 'for k = Nb : 1' should be written as a decreasing loop (e.g., 'k = N_b, N_b - 1, ..., 1'), and 'impurity tensors itself' should be 'impurity tensors themselves'.","section":"Sec. 4"},{"comment":"The caption reports TDVP with dt = 0.1 and TEBD with dt = 0.01 and notes that TDVP generally allows larger steps. A brief explanation of why this is the case for the present model, beyond citing Ref. [50], would help the reader.","section":"Fig. 10 caption"}],"recommendation":"major_revision","confidential_remarks":"The error in Eq. (9) is serious and currently blocks acceptance, but it appears to be a fixable derivation issue rather than a fundamental flaw: the final projector Eq. (15) is structurally consistent with known formulas for MPS and hierarchical Tucker, and the numerical demonstrations are credible. I recommend asking the authors to re-derive the tangent-space projector with the correct orientation of gauge transformations on each link and to provide the missing FTPS Hamiltonian representation. The authors' acknowledgment of the parallel work by Kohn et al. is appropriate and not a concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the result is useful and the numerics are convincing, but the derivation of the tangent-space projector in Sec. 3.1 has a genuine index mistake. The final formula is very likely right; the proof as printed is not.\n\nWhat is new: previous TDVP derivations covered MPS and binary hierarchical Tucker/tensor-train trees, while this paper extends the construction to arbitrary finite loop-free tensor networks, including non-binary nodes. That is the right level of generality and a real step beyond the existing literature. The FTPS application is also genuinely new: off-diagonal hybridizations are awkward for TEBD, and the exact noninteracting benchmark in Fig. 11 is a meaningful test, not just a repeat of a known comparison. The convergence checks in Fig. 12 (the dt^2 scaling and the bond-dimension dependence) are solid.\n\nWhere it is soft: the load-bearing vertical-subspace parameterization, Eq. (9), is asserted without proof, and on inspection it is not correct as written. Even for a two-site MPS—the limit the authors claim it reduces to—Eq. (9) requires the second tensor to be multiplied by X on the left, or equivalently by X^T on the right, not by X on the right. Taking A = C = I and X with a single off-diagonal entry gives a nonzero tangent vector, so Eq. (9) does not parameterize the kernel. The counting and gauge arguments around Eq. (10) therefore do not establish the projector in Eq. (15). That said, Eq. (15) has the expected MPS form (site projectors minus link projectors), and the numerical results are consistent with it. I read this as a fixable derivation bug rather than a wrong algorithm. The absence of released code or data is a minor concern; the benchmarks are clear enough to give moderate confidence, but independent implementation checks would help.\n\nCitation pattern looks clean: the authors cite Haegeman, Lubich, Kohn, and note the independent concurrent work. The self-citations are to their own FTPS papers and are appropriate.\n\nBottom line: this deserves serious refereeing, not a desk rejection. The referee should ask for a corrected treatment of the vertical subspace—specifically, fix Eq. (9) so the contraction transposition is handled properly—and then the paper should be a solid methods contribution.","headline":"A useful TDVP generalization for arbitrary loop-free tensor networks, with convincing numerics and a real but fixable index error in the vertical-subspace derivation that should be corrected before publication.","tokens_in":17759,"tokens_out":6877,"would_cite":true,"duration_ms":76691,"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 single projector generalizes TDVP to every loop-free tensor network, not just matrix product states.","keywords":["time-dependent variational principle","tree tensor networks","tangent space projector","Fork Tensor Product States","Anderson impurity model","off-diagonal hybridization","real-time evolution","Trotter decomposition"],"falsifier":"On a small non-binary tree (for example a star graph with three leaves), compute the dimension of the kernel of the linear map from tangent tensors $\\{B^{s_i}_{Q_i}\\}$ to the state $|\\Theta[B]\\rangle$. If this dimension is larger than the number of free parameters in the $X$-matrices allowed by Eq. (9) under Eq. (10) — or if a single gauge-equivalent-to-zero perturbation cannot be written in that form — then Eq. (15) is not the full tangent-space projector, and TDVP on such a tree would not be the exact manifold projection.","tokens_in":16758,"feed_emoji":"🌳","tokens_out":7079,"duration_ms":72500,"temperature":0.7,"pith_summary":"The paper extends the time-dependent variational principle (TDVP) from matrix product states to any finite, loop-free tensor network, which it calls a tree tensor network (TTN). Its central claim is that Eq. (15) gives the orthogonal projector onto the tangent space at any TTN state, so that time evolution stays on the tensor-network manifold as long as the Hamiltonian can be written in the same network structure. This matters because TDVP then offers a time-evolution method for Hamiltonians with long-range or off-diagonal couplings, where TEBD becomes difficult to implement. The authors demonstrate the method on Fork Tensor Product States for multi-orbital Anderson impurity models, showing that off-diagonal hybridizations relevant for spin-orbit coupling and lattice distortions can be treated.","feed_headline":"A single projector generalizes TDVP to every tree tensor network","feed_subtitle":"The method time-evolves impurity models with off-diagonal hybridizations that TEBD cannot easily reach.","key_machinery":"The central object is the tangent-space projector of Eq. (15), built from site tensors $T^{s_i}_{Q_i}$ orthogonalized toward each neighbor (written $T_N^{[q_k]}$) and from the mutually orthogonal link states $|q^{(i)}_k\\rangle$ that each cut of the tree defines. This projector decomposes the global projected Schrödinger equation into one-site effective Hamiltonians $H_{(s_i Q_i),(s'_i Q'_i)}$ and link effective Hamiltonians $K_{(q^{(i)}_k q^{(j)}_k),(q^{(i)'}_k q^{(j)'}_k)}$; each local piece is integrated by exponentiating the effective Hamiltonian with Krylov methods. The sweeping order, from a chosen start leaf to a chosen end leaf, turns these local updates into a first-order TDVP step, and reversing the sweep with half time steps gives a second-order integrator. The two-site variant replaces the site and link updates by a two-site effective Hamiltonian followed by an SVD, allowing adaptive bond-dimension growth.","core_discovery":"The paper establishes that for any finite loop-free tensor network, the tangent space at a state $|\\psi[T]\\rangle$ is spanned by local tensor variations, and the orthogonal projector onto it has the explicit form $P_{T|\\psi[T]\\rangle} = \\sum_i \\mathbb{1}_{s_i}\\otimes\\sum_{Q_i}|q^{(i)}_1\\cdots q^{(i)}_{r_i}\\rangle\\langle q^{(i)}_1\\cdots q^{(i)}_{r_i}| - \\sum_{\\langle i,j\\rangle_{q_k}}\\sum_{q_k q'_k} |q^{(j)'}_k\\rangle\\langle q^{(j)'}_k|\\otimes |q^{(i)}_k\\rangle\\langle q^{(i)}_k|$ (Eq. 15). The derivation generalizes the MPS construction by parameterizing the vertical subspace with one matrix $X$ per link, fixing the gauge with $N-1$ constraints, and choosing any leaf as end point. Integrating the projected Schrödinger equation term by term with a Suzuki-Trotter breakup yields single-site and two-site TDVP sweeps; the paper verifies the scheme on impurity models, including a non-interacting case with off-diagonal hybridizations where it matches the exact Green's functions.","pith_inferences":["Beyond the paper: a direct count of the kernel of the map $B \\mapsto |\\Theta[B]\\rangle$ on a small non-binary tree would test whether Eq. (9) really captures all gauge directions; if it does not, the projector needs extra null vectors for high-degree nodes, though the FTPS demonstration would still stand.","Beyond the paper: because the projector is constructed from the current orthogonalized state, the same local update structure could be combined with a Hamiltonian-adapted sweeping order, possibly reducing the number of long-range terms each sweep must cover.","Beyond the paper: the non-interacting off-diagonal benchmark is the cleanest reported test; repeating it at finite interaction with a controlled impurity-impurity bond dimension would show whether the method's accuracy survives the interacting regime."],"forward_implications":["Any Hamiltonian that admits a representation in the same TTN structure as the state can be time-evolved with TDVP, including long-range couplings that are hard for TEBD.","The single-site TDVP variant preserves exactly the conserved quantities of the Hamiltonian, because the evolution never leaves the tensor-network manifold.","The two-site variant provides a controlled truncation and can grow bond dimensions dynamically during the evolution.","For FTPS impurity solvers, off-diagonal hybridizations (for example from spin-orbit coupling or lattice distortions) can be included without reformulating TEBD.","Second-order convergence in the time step $\\Delta t$ is observed, matching the numerical $\\sim \\Delta t^2$ scaling reported in the paper."],"supporting_citations":[{"why":"Supplies the MPS tangent-space projector and gauge-fixing derivation that this paper generalizes to arbitrary TTNs.","marker":"[29]"},{"why":"Defines TDVP as the projected Schrödinger equation and establishes the conservation properties that single-site TDVP inherits.","marker":"[30]"},{"why":"Provides the tensor-train time-integration scheme whose sweeping and Trotter splitting are adapted for TTNs.","marker":"[31]"},{"why":"Introduces TDVP for binary TTNs, the earlier special case that this paper extends to general loop-free networks.","marker":"[47]"},{"why":"Introduces the FTPS tensor network and the TEBD-based impurity solver that TDVP is compared against and extends.","marker":"[18]"},{"why":"Gives the MPO-based alternative for long-range time evolution that TDVP avoids by using the Hamiltonian in the same network structure.","marker":"[44]"},{"why":"Compares real-time evolution algorithms for Anderson impurity models, showing where TEBD outperforms TDVP and motivating the hybrid scheme.","marker":"[50]"},{"why":"Supplies the Suzuki-Trotter decomposition used to break the tangent-space projector into local updates for first- and second-order integration.","marker":"[39]"}],"fun_headline_variants":["TDVP now covers every tree tensor network","One projector opens TDVP to all tree tensor networks","TDVP extends beyond MPS to every tree tensor network","Generalized TDVP handles off-diagonal impurity hybridizations","TDVP sweeps any tree tensor network with one projector"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument depends on the assumption that every redundant direction of the tensor network is caught by exactly one matrix per link, with signs set by the chosen end point, and that the $N-1$ gauge conditions fix these matrices uniquely.","fun_headline_variants_meta":{"raw":{"variants":["TDVP now covers every tree tensor network","One projector opens TDVP to all tree tensor networks","TDVP extends beyond MPS to every tree tensor network","Generalized TDVP handles off-diagonal impurity hybridizations","TDVP sweeps any tree tensor network with one projector"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001998,"raw_usage":{"total_tokens":7779,"prompt_tokens":907,"completion_tokens":6872,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":6795}},"tokens_in":523,"tokens_out":6872,"duration_ms":47799,"temperature":1.0,"reasoning_tokens":6795,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:25:20.473722+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a small non-binary tree (for example a star graph with three leaves), compute the dimension of the kernel of the linear map from tangent tensors $\\{B^{s_i}_{Q_i}\\}$ to the state $|\\Theta[B]\\rangle$. If this dimension is larger than the number of free parameters in the $X$-matrices allowed by Eq. (9) under Eq. (10) — or if a single gauge-equivalent-to-zero perturbation cannot be written in that form — then Eq. (15) is not the full tangent-space projector, and TDVP on such a tree would not be the exact manifold projection.","supporting_citations":[{"cited_title":"Lubich, T","cited_arxiv_id":null,"evidence_quote":"Introduces TDVP for binary TTNs, the earlier special case that this paper extends to general loop-free networks."},{"cited_title":"Comparison of MPS based real time evolution algorithms for Anderson Impurity Models","cited_arxiv_id":"1906.09077","evidence_quote":"Compares real-time evolution algorithms for Anderson impurity models, showing where TEBD outperforms TDVP and motivating the hybrid scheme."},{"cited_title":"Suzuki, Fractal decomposition of exponential operators with applications to many-body theories and Monte Carlo simulations , Physics Letters A 146(6), 319 (1990)","cited_arxiv_id":null,"evidence_quote":"Supplies the Suzuki-Trotter decomposition used to break the tangent-space projector into local updates for first- and second-order integration."}],"review_version":1}