{"id":"e34ab0a0-e6e4-4b9d-93ab-a556e91c7f30","arxiv_id":"2607.06330","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"A spectral algorithm with minimax-optimal guarantees is proposed for directly estimating bilinear forms in Tucker2 tensor panels with staggered-adoption missingness.","lead":"This paper develops a spectral algorithm to directly estimate bilinear forms (like average treatment effects) from noisy, partially-observed tensor data under staggered adoption designs. It provides non-asymptotic error bounds showing a phase transition in the number of tensor layers and matches these with minimax lower bounds.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"No significant objection identified. The proof structure is sound: the Haar compression argument (Lemma 6) correctly yields the projection factor enabling K-dependent SNR improvement, the clipped inverse is inactive on the high-probability event (Eq. 24), and the minimax lower bounds match the upper","rationale":"The reader correctly identified A1 (c_l > 0) as the most restrictive assumption and correctly noted that the paper provides hardness results justifying it. The reader's verdict of ACCEPT with HIGH confidence is appropriate. The central claim—Theorem 1's MSE bound with phase transition in K, matched by minimax lower bounds—is supported by careful technical work. The Haar compression argument is the key innovation enabling the improved SNR dependence on K, and it is correctly constructed. The clipped spectral inverse introduces no bias on the high-probability event. The gap between upper and lower bounds (rank and log factors) is standard and acknowledged. The real-data applications, while on small datasets, demonstrate practical relevance. No adjustment to the verdict is needed.","tokens_in":74667,"tokens_out":6422,"duration_ms":474726,"concrete_test":"Independently verify the Haar compression bound in Lemma 6, Eq. (20): specifically, confirm that the conditional distribution of Q given F = σ(G_1, Σ_2, V_2) is Haar on St(N−r, q) and independent of F, by checking that (i) the reduced block in the eigendecomposition of Ŝ is measurable with respect to F, and (ii) the left-orthogonal invariance of G_2 implies left-invariance of Q conditional on (Σ_2, V_2). If this conditional independence fails, the projection factor √((p+r+ζ_T)/N) would not hold, weakening the K-dependent SNR improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"After careful review, I do not identify a load-bearing concern that would undermine the central claim. The key technical innovation—the Haar compression argument in Lemma 6—is correctly constructed: the conditional Haar distribution of Q on the Stiefel manifold (given F = σ(G_1, Σ_2, V_2)) is justified by the rotational invariance of Gaussian matrices, and the projection factor √((p+r+ζ_T)/N) in Eq. (20) is the correct compression that replaces the ambient dimension N. The reduction P_⊥_2 S = 0 (Eq. 15) is valid because Λ_c + σ²T_{1,p}I_r is invertible whenever Λ_c ≻ 0, which is established in part (i) via Weyl's inequality and the SNR condition (A3). The clipped spectral inverse is inactive on the good event G_1 (Eq. 24), so no bias is introduced with high probability; the bad-event contribution is controlled by condition (4), which is satisfiable because the LHS involves exponentially small tail probabilities while the RHS involves polynomial terms. The matching between upper and lower bounds holds up to acknowledged rank and logarithmic factors: the gap of ~(r+ζ_N) between the upper bound's first term and the lower bound's corresponding term is standard in the literature. The reader's identified weakest assumption (A1 with c_l > 0) is indeed the most restrictive condition, but the paper provides hardness results (Propositions 18–19) and sufficient conditions (Lemmas 16–17) that justify it. The extension to staggered adoption (Corollary 4) follows cleanly by reduction to the four-block setting with substituted dimensions.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"This paper studies the estimation of bilinear forms from noisy, partially observed tensor data under a Tucker2 model with structured missingness motivated by staggered adoption designs. The authors propose a spectral algorithm that pools information across tensor layers and targets the functional directly, rather than completing the entire tensor. They prove a non-asymptotic mean squared error bound (Theorem 1) exhibiting a phase transition in the number of layers K, and match it with local minimax lower bounds (Theorems 2 and 3). The four-block missingness pattern is extended to general staggered adoption via an anchored reduction (Corollary 4). Simulations and two real-data applications validate the theory.","tokens_in":75505,"tokens_out":1375,"duration_ms":256711,"significance":"The paper addresses a well-motivated problem at the intersection of causal inference and tensor estimation. The direct targeting of bilinear forms, rather than full tensor completion, is a meaningful contribution with clear practical relevance. The key technical innovation is the Haar compression argument (Lemma 6), which yields a projected subspace error bound with a dimension-dependent compression factor that replaces the ambient dimension N; this is what enables the SNR condition to improve with K. The matching lower bounds via van Trees inequality are correctly constructed and establish rate-optimality up to constants, rank, and logarithmic factors. The extension to staggered adoption (Corollary 4) via the anchored four-block reduction is clean. The reproducible code and real-data applications add value. The paper ships falsifiable predictions (the phase transition in K) that are confirmed in simulation.","major_comments":[{"comment":"Theorem 1, condition (4): This condition requires that a certain tail-probability term is bounded by c_0 * Upsilon_xy. The authors argue this is satisfiable because the LHS involves exponentially small tail probabilities while the RHS involves polynomial terms. However, the condition couples tau, the bad-event probability, and Upsilon_xy in a non-trivial way. A brief explicit verification that (4) is compatible with the regime where the SNR condition (A3) holds and tau is chosen as in Theorem 1 (tau <= c_l * N_{1k} / (2N)) would strengthen the result. Currently the reader must take on faith that the polynomial-in-K Upsilon_xy always dominates the exponentially small tail terms; a short remark or a line in the proof of Theorem 1 checking this inequality case by case would help.","section":null}],"minor_comments":[{"comment":"Section 2.1, Assumption A1: The discussion of c_l > 0 is well-supported by Propositions 18-19, but the reader would benefit from a brief note in the main text (not just the appendix) on how restrictive this is in practice. Lemma 16 shows the missing block must be small relative to the observed block; a one-sentence pointer to this in Section 2.3 would contextualize the assumption.","section":null},{"comment":"Table 1: The notation table is comprehensive but dense. Consider splitting into two tables (one for block/signal notation, one for empirical/SVD notation) or adding a brief reading guide so readers can quickly find definitions when reading the algorithm description.","section":null},{"comment":"Section 5.1, Algorithm 2: The construction of anchor sets RowAnc_k(j) and ColAnc_k(j) is clearly defined, but a small worked example showing the sets for a concrete staircase (e.g., o_k = 4, a = 3, b = 3) would help readers verify their understanding of the index set computations.","section":null},{"comment":"Section 6.2: The Castle Doctrine application uses r = 3. It would be useful to report sensitivity to the choice of r (e.g., r = 2, 4) for at least one functional, to assess robustness of the conclusions to rank misspecification in a real-data setting.","section":null},{"comment":"Section 6.3: The COVID-19 application has N = 18, T = 32, K = 2. With only 2 fully observed rows in the target layer (Norway, Iceland), Assumption A1 requires c_l > 0 for the restricted Gram matrix of 2 rows. For r = 3, this is rank-deficient. The authors should clarify whether the auxiliary four-block construction resolves this (by pooling anchor rows from the other layer), or whether the theoretical guarantees are not plausibly satisfied here.","section":null},{"comment":"Lemma 6, proof: The Haar-measure argument is the key technical step. While correct, the proof is quite involved. A brief high-level summary at the start of the proof (2-3 sentences) explaining the strategy—(1) reduce to controlling S = U_perp^T hat_U_left H_U, (2) show S lives in col(G_2), (3) apply Haar compression—would make the result more accessible.","section":null},{"comment":"References: Several references cite 2026 dates (e.g., Agterberg 2026, Choi and Yuan 2026, Agarwal et al. 2026, Ma and Ma 2026, Liu et al. 2026). If these are forthcoming or revised versions, please update to the correct final citations.","section":null},{"comment":"Appendix B.1, equation (7): The table of functionals is informative but the Trend row uses a linearity assumption on the row-averaged trajectory. This assumption is stated but not discussed in the main text; a brief remark in Section 6.2 on whether it is plausible for the Castle Doctrine data would improve readability.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is technically strong and the central claims are well-supported. The Haar compression argument is a genuine technical contribution that enables the K-dependent improvement. The main substantive question is whether condition (4) in Theorem 1 is automatically satisfied under the other assumptions or requires an additional mild regularity condition; the authors' argument is plausible but should be made explicit. The COVID-19 application (Section 6.3) is borderline in terms of whether the theoretical assumptions are met (N=18, only 2 fully observed rows, r=3), but the authors frame it as illustrative rather than a rigorous test of the theory, so this does not block publication. I recommend minor revision."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. We address the single major comment below.","responses":[{"response":"We agree that an explicit verification of condition (4) would strengthen the presentation, and we will add it to the revised manuscript. Below we provide the argument that will appear as a remark following the statement of Theorem 1, with the detailed case-by-case verification in the proof. Recall that condition (4) requires two inequalities to hold simultaneously: (I) gamma_max^2 * tau^{-1} * (N_{1k}/N) * (p_N^{-10} + p_T^{-10}) <= c_0 * Upsilon_xy, and (II) sigma^2 * tau^{-1} * (N_{1k} + T) * (p_N^{-5} + p_T^{-5}) <= c_0 * Upsilon_xy. We verify each in turn under the regime where (A3) holds and tau = c_l * N_{1k} / (2N). For inequality (I): Substituting tau gives a LHS of order gamma_max^2 * (N/N_{1k}) * (p_N^{-10} + p_T^{-10}). The RHS contains the term sigma^2 * (r + zeta_N) / rho_N * ||U_{2k}^T x||^2. Using (A1), we have rho_N = N_{1,p}/N >= c_l * N_{1k}/N, so (N/N_{1k}) <= c_l^{-1} * rho_N^{-1} * N_{1k}/N * rho_N is not directly the right path; instead, note that (N/N_{1k}) * p_N^{-10} <= (N/N_{1k}) * (max{N_{1,p}, T})^{-10}. Since N_{1,p} >= N_{1k} (as it is a sum including N_{1k}), we have N/N_{1k} <= N/N_{1k} and p_N >= N_{1k}, so (N/N_{1k}) * p_N^{-10} <= N_{1k}^{-1} * N * N_{1k}^{-10} which is exponentially small. More precisely, under (A2) we have N - r >= c_blk * N so N is of the same order as the ambient dimension, while p_N = max{N_{1,p}, T} >= N_{1k}. The term (N/N_{1k}) * p_N^{-10} is thus of order N * N_{1k}^{-11}, which is polynomially (in fact, nearly exponentially) small in the dimensions. The RHS term sigma^2 * (r + zeta_N) / rho_N * ||U_{2k}^T x||^2 is of order sigma^2 * r * (N/N_{1,p}) * (r/N) (using (A4)) = sigma^2 * r^2 / N_{1,p}, which is polynomial in the dimensions and the SNR parameters. Under (A3), sigma is bounded by c_0 * gamma_min / max","revision_made":"no","referee_comment":"Theorem 1, condition (4): The condition requires that a certain tail-probability term is bounded by c_0 * Upsilon_xy. The authors argue this is satisfiable because the LHS involves exponentially small tail probabilities while the RHS involves polynomial terms. However, the condition couples tau, the bad-event probability, and Upsilon_xy in a non-trivial way. A brief explicit verification that (4) is compatible with the regime where the SNR condition (A3) holds and tau is chosen as in Theorem 1 (tau <= c_l * N_{1k} / (2N)) would strengthen the result. Currently the reader must take on faith that the polynomial-in-K Upsilon_xy always dominates the exponentially small tail terms; a short remark or a line in the proof of Theorem 1 checking this inequality case by case would help."},{"response":"Continuing the verification: the LHS of (II) is of order sigma^2 * (N/N_{1k}) * (N_{1k} + T) * p_N^{-5}, which simplifies to sigma^2 * N * (1 + T/N_{1k}) * p_N^{-5}. Since p_N = max{N_{1,p}, T} and N_{1,p} >= N_{1k}, we have p_N >= N_{1k}, so the LHS is at most sigma^2 * N * (1 + T/N_{1k}) * N_{1k}^{-5}. Under (A2), N, T, N_{1k} are all polynomially related, so the factor N * N_{1k}^{-5} decays at least as fast as N_{1k}^{-4} (using N <= c * N_{1k} * (N/N_{1k}) and the dimensional regularity conditions). Meanwhile, the RHS term sigma^2 * (r + zeta_N) / rho_N * ||U_{2k}^T x||^2 is of order sigma^2 * r * log(p_N) * (N/N_{1,p}) * (r/N) = sigma^2 * r^2 * log(p_N) / N_{1,p}, which is polynomially bounded away from zero under (A3) and (A4). The ratio LHS/RHS is thus of order (N * N_{1k}^{-5}) / (r^2 * log(p_N) / N_{1,p}) ~ N * N_{1,p} / (N_{1k}^5 * r^2 * log(p_N)), which goes to zero as the dimensions grow, since N_{1k}^5 grows polynomially while the numerator grows only linearly. The same argument applies to the p_T terms by symmetry. Therefore, both inequalities (I) and (II) are satisfied for sufficiently large dimensions, and the constant c_0 can be chosen to absorb the finite-dimensional cases. In summary: the key observation is that the LHS of (4) always contains a factor of the form p_N^{-5} or p_N^{-10} (or the analogous T terms), which decays polynomially in the dimensions, while the RHS Upsilon_xy contains terms that are polynomial in the SNR and dimensional parameters under (A3) and (A4). The polynomial decay of the tail probabilities always dominates the polynomial terms in Upsilon_xy for sufficiently large dimensions, and the constant c_0 > 0 can be chosen small enough to handle the transition regime. We will add a remark after Theorem 1 summarizing this argument,","revision_made":"yes","referee_comment":"sqrt(N), sqrt(T), sqrt(N/rho_T), sqrt(N*T/N_{1k})}, so sigma^2 is at most of order gamma_min^2 / (N + T + N/rho_T + N*T/N_{1k}), which is polynomial. The exponentially decaying p_N^{-10} factor ensures that the LHS is negligible compared to any polynomial term appearing in Upsilon_xy. For inequality (II): Substituting tau gives a LHS of order sigma^2 * (N/N_{1k}) * (N_{1k} + T) * (p_N^{-5} + p_T^{-5}). Again, p_N >= N_{1k} and p_T >= T_{1k}, and under (A2) the dimensions N, T, N_{1k}, T_{1k} are all of the same polynomial order. The factor (p_N^{-5} + p_T^{-5}) decays polynomially (with exponent 5) in the dimensions, while the RHS of (4) contains terms like sigma^2 * (r + zeta_N) / rho_N * ||U_{2k}^T x||^2, which is of order sigma^2 * r * log(N_{1,p} + T) * (N/N_{1,p}) * (r/N) = sigma^2 * r^2 * log(N_{1,p} + T) / N_{1,p}. Under (A3), this is a polynomial quantity in the problemparameters"}],"tokens_in":74230,"tokens_out":1768,"duration_ms":731110,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"This paper does something genuinely new: it estimates bilinear forms of missing tensor blocks directly, without completing the full tensor, under a Tucker2 model with MNAR staggered-adoption missingness. The key result is Theorem 1, which gives a non-asymptotic MSE bound exhibiting a phase transition in K — pooling terms decay as 1/K while an irreducible layer-specific term remains. The matching local minimax lower bounds (Theorems 2–3) are constructed via clean one-dimensional parametric submodels and van Trees, and they match the upper bound up to constants and log factors. The extension to general staggered adoption via the anchored four-block reduction (Corollary 4) is clean and practical. Code is available on GitHub. The Haar compression argument in Lemma 6 is the technical highlight — it replaces the ambient dimension N with a projected factor sqrt((p+r+ζ_T)/N), which is what makes the SNR condition improve with K. This is a real improvement over the leave-one-block-out approach in Yan and Wainwright (2024), whose remainder control deteriorates with K. The clipped spectral inverse is a nice practical touch: it's inactive on the high-probability event (Eq. 24), so it introduces no bias but stabilizes the algorithm on bad events. The simulations confirm the predicted phase transition, including under rank misspecification and incoherence violation. The main soft spot is Assumption A1 (c_l > 0). The reader flags this as the weakest assumption, and I agree it's the most restrictive condition. Lemmas 16–17 show it essentially requires the missing block to be small relative to the observed block — in Lemma 16, N_1 must be nearly N when r is nontrivial. The paper is transparent about this and provides hardness results (Propositions 18–19) showing c_l = 0 makes the problem unidentifiable, which is the right response. But the practical reach is still limited to settings with small missing fractions. This is a structural limitation of the problem, not a flaw in the paper. The real-data applications are small (50×11×4 and 18×32×2) and serve as illustrations rather than stress tests. The COVID-19 example, where only two rows are fully observed in the target layer, is a nice showcase for the tensor advantage. I don't find load-bearing flaws. The proof structure is sound, the lower bounds are correctly constructed, and the upper bound follows from first principles. The gap between upper and lower bounds (rank and log factors) is standard. This paper is for researchers working on tensor completion, causal panel data, or high-dimensional statistics. It deserves a serious referee — the technical content is substantial and the contribution is real.","headline":"Direct estimation of bilinear forms in Tucker2 tensor panels with staggered-adoption missingness — solid theory, clean phase transition result, one genuinely restrictive assumption.","tokens_in":75497,"tokens_out":644,"would_cite":true,"duration_ms":112950,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62H12","62H25","62F12"],"pacs":[],"model":"glm-5.2","headline":"Spectral algorithm estimates causal bilinear forms directly from staggered tensor panels","keywords":["bilinear forms","Tucker2 decomposition","staggered adoption","tensor completion","missing not at random","spectral methods","minimax lower bounds","causal inference"],"falsifier":"If the restricted Gram matrices U_1j^T U_1j and V_1j^T V_1j become singular or near-singular—which happens when the observed block is too small or the factor structure is poorly conditioned within a layer—the target bilinear form is not identifiable and no estimator can consistently recover it.","tokens_in":74898,"feed_emoji":"📊","tokens_out":824,"duration_ms":193024,"temperature":0.7,"pith_summary":"This paper addresses the problem of estimating specific quantities—bilinear forms—from partially observed, noisy tensor data where the missingness follows a staggered adoption pattern, as arises in causal inference when units adopt a treatment at different times. The authors model the data using a Tucker2 decomposition, which assumes shared latent factors across tensor layers (e.g., multiple outcomes or policies) while allowing each layer to have its own core structure. Rather than completing the entire missing tensor and then computing the target quantity—a plug-in approach that may be statistically and computationally wasteful—the authors propose a spectral algorithm that targets the bilinear form directly. The key methodological idea is a regress-then-denoise procedure: the algorithm learns shared subspaces by pooling information across layers via stacked singular value decompositions, then uses a spectral inverse to predict the missing block and evaluate the desired functional without ever materializing the full missing matrix. The central theoretical result is a non-asymptotic mean squared error bound that exhibits a phase transition in the number of layers K. When K is small, the error decreases as 1/K because pooling improves estimation of the shared factors. When K is large, the error saturates at an irreducible layer-specific term that reflects the difficulty of estimating the slice-specific core. This elbow behavior is proven to be rate-optimal via local minimax lower bounds that match the upper bound up to constants. The construction is extended from a basic four-block missingness pattern to general staggered adoption designs through an anchored reduction that decomposes arbitrary staircase patterns into auxiliary four-block problems.","feed_headline":"Spectral algorithm estimates causal bilinear forms directly from staggered tensor panels","feed_subtitle":"By pooling across tensor layers and targeting functionals directly, the method achieves rate-optimal error with a provable phase transition—","key_machinery":"The Tucker2 tensor model with shared unit and time factors across layers and slice-specific cores; the four-block missingness pattern as the building block for staggered adoption; the Stack-SVD procedure for pooling shared subspaces across layers; the clipped spectral inverse for stabilizing the regression operator; the anchored four-block reduction for extending to general staircase missingness; and the Haar compression bounds on the Stiefel manifold that yield sharper projected subspace perturbation estimates than standard Wedin-type bounds.","core_discovery":"The central discovery is that direct estimation of bilinear forms from partially observed tensor panels achieves a mean squared error rate that exhibits a provable phase transition in the number of layers K: pooling across layers yields 1/K decay for the shared-subspace estimation components, while an irreducible layer-specific term remains that cannot be reduced by pooling. This rate is matched by local minimax lower bounds, establishing optimality. The algorithm achieves this by combining stacked SVDs for subspace learning with a clipped spectral inverse for block prediction, targeting the functional directly rather than completing the tensor.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Direct estimation of bilinear forms in staggered tensor panels shows layer-pooling phase t","Spectral method targets bilinear forms in tensor panels without full completion","Phase transition in layer count governs pooling benefit for tensor bilinear estimation","Minimax-optimal bilinear form estimation for partially observed tensor panels","Staggered adoption designs enable direct bilinear estimation with stacked SVDs"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The algorithm requires that the observed portion of the data within each layer is sufficiently well-conditioned relative to the latent factor structure—specifically, that the restricted Gram matrices of the shared factors over observed entries have eigenvalues bounded away from zero. This effectively requires the missing block to be small relative to the observed block, limiting applicability when the fraction of missing entries is large.","fun_headline_variants_meta":{"raw":{"variants":["Direct estimation of bilinear forms in staggered tensor panels shows layer-pooling phase transition","Spectral method targets bilinear forms in tensor panels without full completion","Phase transition in layer count governs pooling benefit for tensor bilinear estimation","Minimax-optimal bilinear form estimation for partially observed tensor panels","Staggered adoption designs enable direct bilinear estimation with stacked SVDs"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":599,"prompt_tokens":501,"completion_tokens":98,"prompt_tokens_details":null},"tokens_in":501,"tokens_out":98,"duration_ms":39544,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T09:27:15.242672+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If the restricted Gram matrices U_1j^T U_1j and V_1j^T V_1j become singular or near-singular—which happens when the observed block is too small or the factor structure is poorly conditioned within a layer—the target bilinear form is not identifiable and no estimator can consistently recover it.","supporting_citations":[],"review_version":1}