{"id":"6cb3ff32-e6e1-4a7c-8b87-867b7fe04fd2","arxiv_id":"2411.17522","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Conditional diffusion transformers and their latent variants receive score approximation and estimation rates, but the minimax optimality headline rests on setting a derived constant equal to 1/2.","lead":"This paper derives approximation, estimation, and distribution recovery rates for conditional diffusion transformers under smooth-data assumptions, and claims their latent versions reach minimax optimality. A generalist should read it because these rates would say how many samples transformer-based conditional generation needs before it hits fundamental statistical limits.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Minimax optimality rests on the bare assertion 'By setting 2*nu3=1'; nu3 is a fixed function of already-chosen constants and is generally far from 1/2, so the upper bound does not match the lower bound.","rationale":"The reader's verdict of REJECT is correct, and the central flaw is the minimax optimality claim. I agree with the reader's rationale but not with their chosen 'weakest_assumption' label: Assumption 3.2 is not itself a flaw, since the lower bound in Corollary 3.4.2 is stated for exactly that class (f >= C > 0). The real load-bearing concern is that the upper bound exponent contains 1/(2*nu3) and the paper asserts 2*nu3=1 without any justification, even though nu3 is determined by constants that were already fixed in Theorem 3.4's parameter choices. Checking the definition shows nu3 is generally much larger than 1/2 for typical DiT architectures (d>=2, L>=2), so the upper bound is strictly looser than the minimax lower bound. The paper never explains how the equality could be achieved; there is no free parameter left to tune once C_alpha and C_sigma are set. This is not a question of failing to match the state of the art; it is an internal inconsistency in the main text. The approximation and estimation results for conditional DiTs may be salvageable if the minimax section is removed or corrected, but as stated the abstract's central claim of 'minimax optimality of unconditional DiTs under identified settings' is not established. Hence the verdict should remain REJECT.","tokens_in":76158,"tokens_out":7280,"duration_ms":59716,"concrete_test":"Evaluate nu3 as defined in Theorem 3.3/3.4 under Assumption 3.2 with the theorem's own choices C_alpha = 2*beta/(dx+dy+2*beta), C_sigma = beta/(dx+dy+beta) and dy=0, for a grid of allowed architecture parameters (e.g., d=2..64, L=2..16, dx=d*L, beta in {0.5,1,2}). If nu3 != 1/2 for all entries, the sentence 'By setting 2*nu3=1' cannot be satisfied and the minimax claim fails. Also search the proof of Theorem 3.4 for any step that derives nu3=1/2; if none exists, the equality is an unsupported postulate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim of exact minimax optimality for unconditional DiTs is the only place where the upper and lower bounds are asserted to match. In Section 3.4 the authors write: 'By setting 2*nu3=1, we show that the unconditional DiT is the minimax optimal distribution estimator.' The constant nu3 is defined in Theorem 3.3 (Assumption 3.2) as nu3 = 4(12*beta*dx + 31*beta*d + 6*beta)/(d*(dx+dy)) + 12*(12*C_alpha*dx + 25*C_alpha*d + 6*C_alpha)/d + 72*C_sigma. In Theorem 3.4 the choices C_alpha = 2*beta/(dx+dy+2*beta) and C_sigma = beta/(dx+dy+beta) are already fixed. With these values and dy=0, nu3 is a concrete function of beta, dx, and d; for any allowed DiT architecture (d>=2, L>=2, dx=d*L) it is typically much larger than 1/2. No derivation or parameter-selection argument is given for the equality 2*nu3=1; it is simply asserted. Moreover, even if nu3 were a free constant, imposing nu3=1/2 after the fact would change the early-stopping time and the covering-number bounds used to derive Theorem 3.4, so the resulting rate would not be the one stated. Thus the claimed matching with the lower bound n^(-beta/(dx+2*beta)) is unsupported. This is an internal inconsistency visible without checking the lengthy appendix: the upper bound has an extra factor 1/(2*nu3) in the exponent, and the paper provides no reason this factor equals 1.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives approximation, score-estimation, and distribution-estimation rates for conditional diffusion transformers (DiTs) in the in-context conditioning architecture, under generic and stronger Hölder smoothness assumptions, and extends the analysis to latent conditional DiTs under a linear subspace assumption. The main technical engine is a term-by-term Taylor expansion of the conditional score on a discretized grid, combined with a universal-approximation theorem for one-layer transformers. The headline claim is that, by reducing the conditional analysis to the unconditional setting, unconditional DiTs achieve the minimax optimal total-variation rate n^{-\\beta/(dx+2\\beta)}, with a parallel claim for latent unconditional DiTs.","tokens_in":76446,"tokens_out":4674,"duration_ms":48137,"significance":"If the central claims were fully established, the paper would make a substantial contribution: it would provide the first exact minimax optimality result for diffusion transformers, together with explicit parameter-norm bounds and sample-complexity rates for conditional and latent DiTs. The appendix contains an extensive proof apparatus, including transformer universal approximation lemmas with quantitative parameter bounds, covering-number arguments, and a careful treatment of score blow-up via truncation. These are genuine strengths. However, the minimax optimality step rests on an unsupported constant-setting assertion, and this is load-bearing for the paper's headline contribution. The estimation-rate theorems may still be of interest, but the exact-optimality claim is not established as written.","major_comments":[{"comment":"The minimax matching is asserted rather than derived. The text states: 'By setting 2*nu3 = 1, we show that the unconditional DiT is the minimax optimal distribution estimator.' But nu3 is defined in Theorem 3.3 as nu3 = 4(12*beta*dx + 31*beta*d + 6*beta)/(d*(dx+dy)) + 12*(12*C_alpha*dx + 25*C_alpha*d + 6*C_alpha)/d + 72*C_sigma, and Theorem 3.4 has already fixed C_alpha = 2*beta/(dx+dy+2*beta) and C_sigma = beta/(dx+dy+beta). With dy = 0, nu3 is a concrete function of beta, dx, and d (with d = dx/L and L >= 2), and it is generally much larger than 1/2. The upper bound in Theorem 3.4 has exponent -(1/(2*nu3))*beta/(dx+dy+2*beta); matching the lower bound n^{-beta/(dx+2*beta)} requires nu3 = 1/2. No argument is given for this equality, and the paper's own definitions make it false in typical regimes. This is not a presentation issue: it is the only step that supports the abstract's and Section 3.4's minimax-optimality claim.","section":"Section 3.4"},{"comment":"The same flaw appears in the latent result. Remark 4.3 states 'Setting 2*e_nu3 = 1, we obtain the minimax optimality of latent unconditional DiT.' The constant e_nu3 in Theorem 4.2 is again a fixed function of beta0, d0, ed, dy, C_alpha, and C_sigma, with C_alpha and C_sigma already chosen. The equality 2*e_nu3 = 1 is not derived from the preceding bounds; it is imposed to force the upper bound to meet the lower bound. Consequently, the latent minimax claim is unsupported for the same reason as the unconditional claim.","section":"Remark 4.3 / Section 4"},{"comment":"Even if nu3 were treated as a free constant, imposing nu3 = 1/2 after the fact is circular. The quantity nu3 enters the proof through the choices t0 = N^{-C_sigma}, T = C_alpha log n, and through the covering-number and approximation bounds used to obtain Theorem 3.4. Changing nu3 to force the matching would change those choices and therefore change the very bound that is being matched. A valid minimax argument must either prove that nu3 = 1/2 under the stated assumptions or show that the exponents can be optimized independently; the paper does neither.","section":"Section 3.4 and Corollary 3.4.2"},{"comment":"The lower-bound class in Corollary 3.4.2 is stated as P = {p(x) = f(x) exp(-C2 ||x||^2/2) : f in H^beta(R^dx, B), f(x) >= C >= 0}. The upper bound in Theorem 3.4 is established under Assumption 3.2, which requires f(x,y) >= C for a positive constant C. Since the lower bound is stated for a potentially larger class (with f >= C >= 0, allowing C = 0), a lower bound on that larger class does not by itself transfer to a lower bound on the smaller class used by the upper bound. The relation between the two classes needs to be stated precisely; as written, the exact minimax conclusion does not follow from the cited lower bound.","section":"Corollary 3.4.2"}],"minor_comments":[{"comment":"The 'Minimax Optimality' column contains the string '/reve' in the first and third rows; this appears to be a typo and should be fixed.","section":"Table 1"},{"comment":"The text says 'the distribution estimation error becomes eO(epsilon^{-1/(2*nu3)*beta/(dx+2*beta)})'; since the theorem is stated in terms of sample size n, this should likely read n, not epsilon.","section":"Section 3.4"},{"comment":"The lemma statement begins 'Suppose Assumption Assumption D.3 holds'; the duplicated word should be removed.","section":"Lemma D.3"},{"comment":"After equation (G.21), the text begins with 'dk Applying the light tail property'; the stray 'dk' appears to be an editing artifact.","section":"Appendix G"},{"comment":"The theorem title reads 'under Generic Hölder Assumption' but the theorem is stated under Assumption 3.2 (the stronger Hölder assumption); the title appears mismatched.","section":"Theorem C.2"}],"recommendation":"reject","confidential_remarks":"The paper's headline result is the exact minimax optimality of unconditional DiTs, and that result is not established: the 'By setting 2*nu3 = 1' step is an unsupported constant-fitting assertion, not a consequence of the derived bounds, and it is repeated for the latent case. This is a load-bearing error that cannot be repaired by local edits; the authors would need to either prove that nu3 = 1/2 under the stated assumptions, identify a different parameter regime where matching holds, or substantially weaken the claims to non-minimax rates. Given the many typos and the scope of the incomplete step, the manuscript is not publishable in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe non-optimality part of this paper is real work: conditional DiT score approximation and estimation rates under two Hölder assumptions, plus latent variants, with explicit transformer parameter bounds. That is a solid chunk of analysis, and the rates improve on prior ReLU-based results. The authors also avoid the double-exponential covering-number hit from Hu et al. by using the single-layer attention universal approximation. Those parts deserve a careful referee.\n\nThe soft spot is exactly where the abstract aims highest. Section 3.4 says \"By setting 2*nu3=1\" and then claims minimax optimality. But nu3 is defined in Theorem 3.3 as a fixed expression in beta, dx, d, C_alpha, C_sigma, and in Theorem 3.4 the authors have already fixed C_alpha and C_sigma. With dy=0, nu3 is generally not 1/2; it is a concrete number much larger than 1/2 for typical DiT widths. No argument is given for why this equality should hold, and \"setting\" it changes the early-stopping time and covering-number bounds used to get the upper bound. So the upper bound does not match the lower bound, and the minimax claim is unsupported. The same issue appears in Remark 4.3 for latent DiTs.\n\nI am not going to pretend I verified every appendix step; the appendices are long. But the flaw is in the main text, and it is load-bearing for the headline. The honest fix is to either drop the minimax claims or derive the condition on the constants properly and show it is satisfiable with the theorem's own choices. If the condition cannot be satisfied, the non-optimality results still stand on their own.\n\nOther concerns are minor by comparison. Assumption 3.2's lower bound f(x0,y) >= C is restrictive, but the authors flag it and it is a standard workaround for score explosion. There are typos and some informal theorem statements in the main text (e.g., Theorem 3.2 is \"informal\" with the formal version deferred to the appendix), which is annoying but not fatal.\n\nBottom line: the approximation and estimation analysis is substantial and likely correct after repair, but the minimax optimality as stated is not established. A serious editor should send this to referees, with instructions to focus on Section 3.4 and the nu3 condition. I would cite the non-optimality rate results once they are cleaned up. I would not bring it to reading group until the minimax business is resolved.\n\nRecommendation: send to peer review, but expect heavy revision or removal of the minimax claims.","headline":"Solid approximation and estimation rates for conditional DiTs, but the minimax optimality claim rests on an unjustified constant-setting and should not survive as stated.","tokens_in":77079,"tokens_out":1994,"would_cite":false,"duration_ms":23677,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G20","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Conditional diffusion transformers can be minimax-optimal estimators of smooth densities.","keywords":["diffusion transformers","conditional diffusion models","classifier-free guidance","minimax optimality","score approximation","score estimation","total variation distance","Hölder smoothness"],"falsifier":"Take a smooth density of the form $e^{-C_2\\|x\\|^2/2}f(x)$ with $f\\in H^\\beta$ but $f=0$ on a small ball, sample $n$ points, and measure the total-variation error of a trained unconditional DiT; if it still achieves $n^{-\\beta/(d_x+2\\beta)}$, the positivity lower bound is unnecessary, while a degraded error exponent would confirm that the minimax claim relies on it.","tokens_in":75836,"feed_emoji":"📊","tokens_out":8173,"duration_ms":67132,"temperature":0.7,"pith_summary":"This paper derives finite-sample statistical rates for conditional diffusion transformers (DiTs) trained with classifier-free guidance. Working under Hölder-smooth conditional densities, the authors prove bounds for score approximation, score estimation, and distribution estimation in total variation, for both the standard and the latent linear-subspace setting. The central payoff is that, under a stronger Hölder assumption, the upper bound for unconditional DiTs matches the known lower bound $n^{-\\beta/(d_x+2\\beta)}$, making the unconditional DiT a minimax-optimal distribution estimator. These rates matter because they say when transformer-based diffusion models are statistically efficient, and they make the dependence of sample complexity on data smoothness and intrinsic dimension explicit.","feed_headline":"Diffusion transformers hit the minimax rate for smooth densities","feed_subtitle":"With enough samples, a tuned DiT matches the best achievable total-variation error on smooth data.","key_machinery":"The engine is a two-stage approximation: discretize the input and label domains into $N^{d_x+d_y}$ infinitesimal hypercubes and perform a term-by-term Taylor expansion of the Gaussian-smoothed conditional density $p_t(x|y)=\\int p(x_0|y)\\exp(-\\|x-\\alpha_t x_0\\|^2/(2\\sigma_t^2))\\,dx_0$. This turns the score $\\nabla p_t/p_t$ into polynomial pieces and algebraic operators (product, inverse, square root of the noise schedule) that single-layer self-attention transformers can approximate with controlled error. Under the stronger assumption that $p(x_0|y)=e^{-C_2\\|x_0\\|^2/2}f(x_0,y)$ with $f\\ge C>0$, the score splits into a linear term plus a well-behaved ratio $\\nabla h/h$, removing the threshold that causes extra log factors in the generic case.","core_discovery":"Under the paper's assumptions, a conditional DiT score network can approximate the conditional score $\\nabla\\log p_t(x|y)$ with error scaling like $N^{-\\beta/(d_x+d_y)}$ under generic Hölder smoothness and $N^{-2\\beta/(d_x+d_y)}$ under the stronger factored-density assumption; converting the grid resolution $N$ into sample size $n$ gives score-estimation and total-variation distribution-estimation bounds. When the conditioning label is removed ($d_y=0$) and the constants are tuned, the total-variation rate becomes $n^{-\\beta/(d_x+2\\beta)}$, exactly matching the minimax lower bound for estimating a smooth density. The latent version, with intrinsic dimension $d_0 \\le d_x$, replaces $d_x$ by $d_0$ in these rates, so the latent unconditional DiT is also minimax optimal under the same assumptions.","pith_inferences":["The conditional model's own minimax optimality is not established; if the same lower-bound technique were extended to the conditional problem, the conditional DiT would likely be minimax as well, but that step is not in the paper.","The grid-and-Taylor technique should transfer to other transformer conditioning mechanisms, such as cross-attention, once a universal approximation theorem for those architectures is available; the paper explicitly leaves cross-attention and adaptive layer-norm variants unanalyzed.","A testable prediction is that datasets with low-density holes or rough conditional densities will need noticeably more samples than the Hölder-exponent formula suggests, since the tighter rates rely on $f$ staying bounded below by a positive constant.","The score decomposition that isolates the linear term also explains why Gaussian-like tails stabilize score matching: the remaining ratio term is bounded, so the network does not need to fit near-singular denominators."],"forward_implications":["Score approximation error improves from $O((\\log(1/\\epsilon))^{d_x}/\\sigma_t^4)$ under generic Hölder smoothness to $(\\log(1/\\epsilon))^{O(1)}/\\sigma_t^2$ under the stronger factored-density assumption.","A tuned unconditional DiT is minimax-optimal in total variation for densities of the form $e^{-C_2\\|x\\|^2/2}f(x)$ with $f\\in H^\\beta$ and $f\\ge C$, achieving the rate $n^{-\\beta/(d_x+2\\beta)}$.","Latent conditional DiTs inherit the same rates with the latent dimension $d_0$ in place of $d_x$, and the latent unconditional DiT matches the corresponding lower bound $n^{-\\beta_0/(d_0+2\\beta_0)}$.","The recommended early-stopping time $t_0=n^{-\\beta/(d_x+d_y+\\beta)}$ and terminal time $T\\sim \\log n$ are explicit enough to serve as practical guidance for choosing diffusion schedules."],"supporting_citations":[{"why":"Supplies the Hölder-smooth data assumptions and the diffused local polynomial decomposition that the rates build on.","marker":"[Fu et al., 2024b]"},{"why":"Provides the single-layer self-attention universal approximation result used to build the score networks.","marker":"[Kajitsuka and Sato, 2024]"},{"why":"Establishes transformer universal approximation, the baseline whose coarse DiT bounds are refined here.","marker":"[Yun et al., 2020]"},{"why":"Gives near-minimax diffusion estimation rates for ReLU networks, the comparison point for the minimax claim.","marker":"[Oko et al., 2023]"},{"why":"Provides the prior DiT approximation and estimation analysis that this paper extends to conditional and latent settings.","marker":"[Hu et al., 2024b]"},{"why":"Supplies the low-dimensional linear latent subspace score decomposition used for latent DiTs.","marker":"[Chen et al., 2023c]"},{"why":"Defines the classifier-free guidance objective that the training loss is based on.","marker":"[Ho and Salimans, 2022]"}],"fun_headline_variants":["DiTs hit minimax rates for smooth densities","Conditional DiTs match minimax optimality bounds","Diffusion transformers achieve minimax rates","Latent DiTs reach minimax rates faster","Optimal rates proved for conditional diffusion transformers"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is Assumption 3.2: after factoring out a Gaussian, the conditional density must stay bounded below by $C>0$ everywhere; if the data has empty or low-density holes, the score function can blow up and the tight rates and minimax matching no longer follow.","fun_headline_variants_meta":{"raw":{"variants":["DiTs hit minimax rates for smooth densities","Conditional DiTs match minimax optimality bounds","Diffusion transformers achieve minimax rates","Latent DiTs reach minimax rates faster","Optimal rates proved for conditional diffusion transformers"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000149,"raw_usage":{"total_tokens":1174,"prompt_tokens":908,"completion_tokens":266,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":197}},"tokens_in":524,"tokens_out":266,"duration_ms":3455,"temperature":1.0,"reasoning_tokens":197,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:01:53.034124+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a smooth density of the form $e^{-C_2\\|x\\|^2/2}f(x)$ with $f\\in H^\\beta$ but $f=0$ on a small ball, sample $n$ points, and measure the total-variation error of a trained unconditional DiT; if it still achieves $n^{-\\beta/(d_x+2\\beta)}$, the positivity lower bound is unnecessary, while a degraded error exponent would confirm that the minimax claim relies on it.","supporting_citations":[],"review_version":1}