{"id":"33159f94-d87f-4185-aa84-c69c18e764ea","arxiv_id":"2412.16191","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A cubic-bounded nonlinearity, derived from a continued fraction of lines, is used as the score function in logistic regression and claims improved convergence and accuracy on Fashion-MNIST.","lead":"A single-author preprint proposes replacing the linear score in logistic regression with the real root of a cubic equation, expressed as a continued fraction of straight lines. It reports lower-variance weights and higher test accuracy on Fashion-MNIST than its linear baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed accuracy advantage rests on an unfair baseline: Section 3 Step 6e gives the linear model a unit-step-size, untuned mini-batch update while the nonlinear model gets an implicit adaptive step, so the comparison conflates the cubic nonlinearity with proper optimization.","rationale":"Both the reader and I identify the same critical weak point: the empirical comparison is not fair. The mathematical derivation of the cubic-root solution and the continued-fraction series appears internally consistent, and I do not object to the boundedness/variance claims insofar as they describe weight magnitudes. But the strongest claim is 'more accurate than the linear counterpart', and that claim is only as strong as the baseline. Step 6e's linear baseline is constructed from the nonlinear update by setting m=1,a=0, which removes the nonlinear method's implicit adaptive denominator. A unit-step-size averaged gradient update is not a standard way to train logistic regression and would be expected to be suboptimal. A tuned logistic regression is therefore the necessary control. Without it, and without repeated trials and error bars, the central claim is unsupported. The absence of code further impedes verification. I therefore agree with the reader's REJECT verdict.","tokens_in":564,"tokens_out":3412,"duration_ms":71008,"concrete_test":"Run a standard multinomial logistic regression on the same 60k/10k Fashion-MNIST split with the same pixel normalization and no data augmentation, tuning only the learning rate (e.g., grid-search η in {0.01, 0.05, 0.1, 0.5} with a decaying schedule, or using scikit-learn's LogisticRegression with default L2 and solver). Compare test accuracy after 50 epochs against the y-form results in Figs. 7 and 8. If tuned linear accuracy is within or above the y-form accuracy, the central accuracy claim collapses. Repeat at least 5 random seeds and report mean ± std for both methods.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is empirical superiority over the 'linear counterpart'. The comparison in Section 3 is not apples-to-apples. In Step 6e, the linear case is defined by substituting m=1 and a=0 into Eq. (14), which yields wi ← wi − (1/n) Σ (σj − p*_j) x_ij, i.e., a mini-batch gradient step with fixed step size 1, no learning-rate schedule, no regularization, and no tuning. In contrast, the nonlinear updates in Eqs. (12)-(14) include the factor 1/(1+3ay^2), which acts as an adaptive, state-dependent step size. Thus the reported faster convergence and higher test accuracy could be due to this implicit adaptive scaling rather than to the continued-fraction representation itself. A properly tuned linear logistic regression on Fashion-MNIST is known to reach high accuracy and may match or exceed the reported y-form accuracy. The paper reports only single trajectories without error bars or code, so the headline 'more accurate than the linear counterpart' is not established. The variance claim also concerns magnitudes of learned weights, not repeated-run variance, and is not by itself evidence of better classification performance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a parametric nonlinear transformation of a linear regression score, defined by the cubic equation (1+ay^2)y = mx, whose real root is expressed as a continued fraction of straight lines and then as an infinite series following Euler's method. The authors apply this transformation to Fashion-MNIST image classification by replacing the linear predictor w^Tx in logistic regression with the real root y(w^Tx) of the cubic equation, estimating per-class parameters w, a, and m by mini-batch gradient descent with implicit gradients. The central claim is that the continued-fraction-based model yields parameters with lower variance, faster convergence, and higher test accuracy than the plain linear logistic regression baseline.","tokens_in":10933,"tokens_out":9508,"duration_ms":87769,"significance":"If the empirical claims were established, the work would demonstrate that a simple bounded nonlinear transformation of the linear predictor can improve optimization stability and classification accuracy over ordinary logistic regression. The closed-form expression for the cubic root and the implicit-gradient derivation are internally consistent and could be useful for other models. However, the central empirical claim is not currently supported: the linear baseline is given an untuned, unit-step-size update, no error bars or multiple seeds are reported, and the convergence of the continued fraction is asserted without conditions. The theoretical contribution is also limited because the continued fraction is an equivalent rewriting of the cubic equation rather than a new mathematical object. The manuscript contains no code or data availability statement, so the results are not reproducible as presented.","major_comments":[{"comment":"The linear baseline is obtained by substituting m=1 and a=0 into the nonlinear update of Eq. (14), which yields the raw mini-batch update wi <- wi - (1/n) sum_j (sigma_j - p*_j) x_ij, i.e., gradient descent with a fixed step size of 1 and no learning-rate schedule, no regularization, no momentum, and no hyperparameter tuning. In contrast, the nonlinear model's update includes the factor 1/(1+3ay^2), which acts as a state-dependent damping that effectively reduces the step size as y grows. The accuracy gap reported in Figs. 7(c) and 8(c) therefore conflates the model change with a difference in optimization protocols. The authors must compare against a properly tuned linear logistic regression (for example, one with a chosen learning rate, regularization, or a standard solver) under identical training epochs and initialization conditions.","section":"Section 3, Step 6e, Eq. (14)"},{"comment":"All reported accuracy and loss curves are single trajectories: no error bars, no multiple random seeds, and no statistical significance test are provided. The Abstract's claim that the continued-fraction model is 'more accurate than the linear counterpart' is a point estimate from one run for each of two initial conditions. The authors should report means and standard deviations over at least several independent initializations, and ideally a paired test across the same initial weights, before the superiority claim can be evaluated.","section":"Section 3, Figs. 7 and 8"},{"comment":"The convergence of the continued fraction to the real root of Eq. (1) is asserted without any conditions or proof. The derived infinite series in Eq. (4) is not well formed (the third term has unbalanced parentheses), and the convergence properties are not analyzed. For large values of the reduced parameter ahat = a m^2 x^2, a naive fixed-point iteration of the truncations can oscillate rather than converge, so a precise convergence theorem with a stated domain of validity is required. If the continued-fraction representation is only motivational, the claim should be removed and the paper reframed around the closed-form root in Eq. (3), which is correct.","section":"Section 2, Eq. (4)"},{"comment":"The 'less variance' claim is based on comparing the raw magnitudes of the learned weight vectors (thousands for the linear model versus about 2 for the nonlinear model). This is not a statistical variance over repeated runs, and it is not scale-invariant: because the nonlinear model has an additional per-class scale parameter m, the magnitude of w is not directly comparable to the magnitude of the linear model's w. The authors should instead report repeated-run variances of test accuracy, of predicted probabilities, or of the resulting decision boundaries.","section":"Section 4.1 and Fig. 9"},{"comment":"The statement 'Once a converges, the other parameters converge' is used as a justification for the optimization procedure, but no argument or theorem is provided. The plotted trajectories of a and m in Figs. 7(a)-(b) and 8(a)-(b) show that a and m settle to nearly constant values in the particular runs, but this does not imply that the full gradient-descent dynamics converge, nor that convergence of a is a general condition for convergence of w. This should be presented as an empirical observation, not as a property of the model.","section":"Section 2, Property 2 and Section 4.2"}],"minor_comments":[{"comment":"There is a typo 'infnite' in the sentence preceding Eq. (4), and the continued-fraction notation with nested exponents is difficult to parse; please rewrite the derivation more clearly.","section":"Section 2, around Eq. (4)"},{"comment":"The loss function in Eq. (9) is written for a scalar output p*_j, but p*_j is a one-hot vector and the model is a multi-class classifier. Please specify how the loss is summed over classes (e.g., one-vs-rest with sigmoid, as implied by the text, or a softmax variant).","section":"Section 3, Eq. (9)"},{"comment":"The xy-plane plots in Fig. 10 are simply the learned curves for each class; calling this a representation of a multi-dimensional problem on a plane is an overstatement, since the plot shows the class-specific scalar function y(x) after training, not the full decision boundary. Consider rewording.","section":"Section 4.3 and Fig. 10"},{"comment":"Reference [3] (Ruder's overview of gradient descent) is a survey, not a primary source for the specific mini-batch updates; please cite a more standard optimization text. Reference [1] is a translation of Euler's work; please provide the original publication details or a more accessible citation.","section":"References"},{"comment":"Several figure captions contain garbled or duplicated text (e.g., Fig. 1 and Fig. 9), making it hard to interpret the markers and shaded regions. Please revise all captions for clarity.","section":"Figures"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is self-contained and the algebra of the cubic root is correct, but the contributions are modest: the continued fraction is an equivalent rewriting of the cubic equation, and the classifier is essentially a generalized linear model with a parametric link function. The experimental evidence is the main weakness; the baseline is untuned and no repeated trials are reported. These issues are fixable within the scope of a revision, but the authors will need to redo the experiments substantially and add a proper convergence analysis. I also note the absence of code and the very thin reference list, which may require editorial attention."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper takes a classical cubic equation, solves it with Cardano's formula, rewrites the solution as a continued fraction, and then uses the real root as a bounded score inside logistic regression for Fashion-MNIST. The forward and backward equations are internally consistent, and the implicit differentiation is right. But the paper's central claim—that this y-form is more accurate than the linear counterpart—rests on a baseline that is optimized with a fixed unit step size and no tuning, while the y-form gets an implicit adaptive step from the 1/(1+3ay^2) factor. That is not an apples-to-apples comparison.\n\nWhat is actually new: applying this specific cubic nonlinearity as a trainable score in logistic regression on one dataset. That is a narrow extension of learnable activations, not a new theory. The continued fraction rewriting is a known technique, and the paper uses the Cardano root for training anyway; the series is only motivational.\n\nWhat the paper does well: the derivations in Section 2 are careful and the plots of the curves are clear. The implicit gradient computations in Eq. (11) are correct. The observation that the factor 1/(1+3ay^2) gives bounded, adaptive step sizes is mildly interesting, though it is not a substitute for a fair baseline.\n\nThe soft spots are substantial. First, the linear comparator in Step 6e is a raw unit-step-size mini-batch update with no learning rate schedule, no regularization, and no tuning. Standard logistic regression on Fashion-MNIST, properly optimized, should be at least as accurate as the reported y-form accuracy. Second, there are no error bars, no multiple seeds, and no code, so we cannot assess run-to-run variance or reproducibility. Third, the 'variance' claim is about the magnitude of learned weights, not statistical variance across runs, and is largely a consequence of the m scaling. Fourth, the convergence of the continued fraction is asserted without conditions, and the initial step size alpha is not reported. These are not minor quibbles; the first one directly undermines the paper's headline result.\n\nWho is this for? Someone interested in bounded score functions might find the idea worth a glance, but the evaluation as it stands does not establish an advantage. A serious referee would need a properly tuned baseline, repeated runs, and code.\n\nMy recommendation: this deserves a desk reject in its current form. The mathematical core is sound but classical, and the empirical claim is not supported by the evidence presented. If the author returns with a fair comparison and error bars, it might be worth another look.","headline":"A correct but classical cubic-root score function, presented with an unfair linear baseline that sinks the only empirical claim.","tokens_in":11438,"tokens_out":5199,"would_cite":false,"duration_ms":46390,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that using the real root of $ay^3 + y = m w^T x$ as the logistic-regression score, computed by a real-valued continued fraction, yields more stable, faster-converging, and more accurate image classifiers on Fashion-MNIST…","keywords":["continued fraction","bounded nonlinearity","logistic regression","Fashion-MNIST","gradient descent","cubic equation","parameter estimation"],"falsifier":"Run a tuned linear logistic regression on Fashion-MNIST with a chosen learning rate, mini-batch schedule, and optional L2 regularization, and compare test accuracy and weight variance with the $y$-form. If the tuned linear model matches or exceeds the $y$-form accuracy, the central empirical claim would fail.","tokens_in":10438,"feed_emoji":"📈","tokens_out":9121,"duration_ms":70371,"temperature":0.7,"pith_summary":"This paper tries to establish that a simple bounded nonlinearity, obtained by replacing a straight-line score with the real root of a cubic equation and expressing that root as a real-valued continued fraction, improves logistic regression for image classification. The proposed score $y$ solves $ay^3 + y = m w^T x$, where $a>0$ and $m$ are learned parameters, and the continued fraction is shown to converge to this real root. On Fashion-MNIST, the paper reports that parameters estimated through the $y$-score have less variance, converge faster and more smoothly, and give higher test accuracy than the paper's linear logistic regression baseline. The motivation is that straight lines with large slope grow faster than the independent variable, while the bounded $y$-curves grow more slowly and therefore produce more stable estimates.","feed_headline":"Curved score beats plain logistic regression on Fashion-MNIST","feed_subtitle":"A continued-fraction root shrinks weight variance and converges more smoothly than untuned linear baseline.","key_machinery":"The load-bearing object is the real root of the cubic equation $ay^3 + y = m w^T x$, expressed as $y = -\\frac{1}{3}t^{1/3} + \\frac{1}{a}t^{-1/3}$ with $t = -\\frac{27mw^Tx}{2a} + \\sqrt{\\left(\\frac{27mw^Tx}{2a}\\right)^2 + \\frac{27}{a^3}}$. This root is also produced by the continued fraction $y = \\frac{m w^T x}{1 + ay^2}$, whose truncated forms all remain bounded by the linear line $y = m w^T x$. The continued-fraction representation is what makes the update rules special: the derivative denominators $1 + 3ay^2$ automatically shrink the gradient steps as $y$ grows, so the optimization is self-bounding. The two extra parameters $a$ and $m$ carry the nonlinearity and the overall scale, and once $a$ converges the other parameters converge.","core_discovery":"The central discovery is that the standard linear score $w^T x$ of logistic regression can be replaced by the real solution $y$ of $ay^3 + y = m w^T x$ without introducing a step-size hyperparameter for $w$ and $m$. The paper derives this solution through a continued fraction of straight lines, $y = mx/(1 + ay^2)$, and reduces the continued fraction to an infinite series by the standard transformation for continued fractions. Because $1 + 3ay^2$ appears in the denominators of the derivatives, gradient updates for $w$ and $m$ remain bounded and adaptive; only the positivity constraint on $a$ requires a step size. Empirically, after 50 mini-batch iterations on Fashion-MNIST, the $y$-form reaches higher test accuracy and lower loss than the linear form, while the learned weights stay in a small range rather than growing into the hundreds or thousands.","pith_inferences":["The comparison that would most directly test the paper's claim is against a tuned linear logistic regression with an optimized learning rate and L2 regularization; the paper's baseline is a single raw gradient substitution, so the reported accuracy gap may shrink under a fair baseline.","The same bounded score could be plugged into other linear classifiers, such as softmax regression or linear support vector machines, since the construction only changes the scalar score before the loss.","The $(i, ii)$-plane view suggests a diagnostic: if categories form separated clusters for particular initial conditions, that could indicate sensitivity of the learned model to initialization.","Because $a \\to 0$ recovers the linear case, the parameter $a$ can be read as a measure of how much nonlinear bounding the data require; classes needing larger $a$ are more nonlinear."],"forward_implications":["On Fashion-MNIST, the $y$-form score should reach higher or comparable test accuracy with smaller weight magnitudes than a raw linear logistic regression run with the same schedule.","Because $w$ and $m$ updates need no learning rate, the method reduces the number of tuning hyperparameters to one, namely the step size for $a$.","The learned $(a, m)$ pairs are stable per class, so a multi-dimensional classifier can be summarized by ten curves on the $xy$-plane or on the $(i, ii)$-plane.","The smooth, near-monotonic convergence of the weights is a direct consequence of the bounded gradient denominators and does not require early stopping or regularization."],"supporting_citations":[{"why":"supplies the transformation used to reduce the continued fraction into the infinite series that defines the approximating score.","marker":"[1]"},{"why":"provides the regression parameter estimation setup that the classification section follows.","marker":"[2]"},{"why":"supplies the mini-batch gradient descent update procedure used to estimate the parameters.","marker":"[3]"},{"why":"provides the Fashion-MNIST dataset and the benchmarking context for the comparison.","marker":"[4]"}],"fun_headline_variants":["Cubic score from continued fractions wins on Fashion-MNIST","Bounded curved score outperforms linear logistic regression","Continued-fraction root improves logistic regression","Cubic curve beats linear in logistic regression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the linear baseline was optimized fairly; in the paper it is obtained by setting $m=1$ and $a=0$ inside the nonlinear update, which yields an untuned raw gradient step with unit step size and no regularization.","fun_headline_variants_meta":{"raw":{"variants":["Cubic score from continued fractions wins on Fashion-MNIST","Bounded curved score outperforms linear logistic regression","Continued-fraction root improves logistic regression","Cubic curve beats linear in logistic regression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000608,"raw_usage":{"total_tokens":2815,"prompt_tokens":914,"completion_tokens":1901,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":1842}},"tokens_in":530,"tokens_out":1901,"duration_ms":12296,"temperature":1.0,"reasoning_tokens":1842,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:51:24.067580+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a tuned linear logistic regression on Fashion-MNIST with a chosen learning rate, mini-batch schedule, and optional L2 regularization, and compare test accuracy and weight variance with the $y$-form. If the tuned linear model matches or exceeds the $y$-form accuracy, the central empirical claim would fail.","supporting_citations":[{"cited_title":"(2004) On the Transformation of Infinite Series to Continued Fractions (D","cited_arxiv_id":null,"evidence_quote":"supplies the transformation used to reduce the continued fraction into the infinite series that defines the approximating score."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the regression parameter estimation setup that the classification section follows."}],"review_version":1}