REVIEW 2 major objections 5 minor
Learning Asymptotics with Convergence-Rate Guarantees using Linear Least Squares
T0 review · 2 major / 5 minor · reviewed 2026-07-31 · deepseek-v4-flash
Pith's one-line read Two sliding least-squares estimators compute unknown constants inside proven asymptotic expansions with explicit convergence-rate guarantees.
desk verdict A genuinely rigorous convergence analysis for sliding least squares, but the advertised learning of DLW-type constants works only when you already have the hard part (high-order expansions); sT-LLSQ only certifies the dominant parameter. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the exponential-product asymptotic form bf(n;x) = exp(Σ φ_j(n) u_j(x_j)), whose logarithm is linear in the transformed parameters γ_j = u_j(x_j). The sliding estimator solves min_y ||A y − b||², with A the windowed matrix of the φ_j's and b the log-data vector; its closed-form solution A†b (and Tikhonov variant Cb, with C = (AᵀA + μI)⁻¹Aᵀ) is what carries the argument. The key identities are the decomposition b = Aγ + log(1 + Σβ_l g_l) + z, which splits the estimate into the true parameter plus a controllable perturbation, and the asymptotic invariants that transfer Θ-equivalent error rates through the inverse transformation x = u^{-1}(y).
What would settle it
For a known sequence such as a_n = 7 n^{1/3} (5/2)^n (1 + (−1)^n / √n), compute the sLLSQ estimate y*_n = A†b with m = k = 3 and compare ||y*_n − γ|| to the theorem's O(n^{3/2} log n) bound: if the error does not grow like Θ(n^{3/2} log n), or if it converges despite ϑ_j(n) not being o(1), the claimed sufficient conditions would be wrong.
Extended reading notes
Core claim
The discovery is a reduction: after taking logarithms, the exponential-product asymptotic form becomes linear in the unknown parameters, so estimating them is exactly an ordinary least-squares problem with a design matrix built from the known basis functions φ_j. The main theorems give closed-form estimators and matching error bounds: for sLLSQ, ||y*_n − γ|| = O(η(n)), where η(n) is a maximum of two terms involving the pseudoinverse of the design matrix; for sT-LLSQ, ||y*_n − γ|| = O(1) for all components and O(φ_{k−1}(n)/φ_k(n)) = o(1) for the dominant one. Since the map u is bijective with nonzero derivative, these rates transfer to the original parameters α via the convergence-rate invari
Load-bearing premise
The entire guarantee rests on the assumption that an asymptotic expansion of sufficiently high order, with the exact basis functions φ_j and error scale g_l, has already been proven for the sequence; without those higher-order terms, sLLSQ's bound can diverge.
Editorial extensions
If this is right
- Any asymptotics of the form (2.1) with the exponential-product structure can be numerically inverted with certified rates, complementing the ratio method where only rough estimates were available.
- The convergence-rate invariant means one proven expansion yields simultaneous estimates of all parameters in a single vector, rather than separate per-parameter sequences.
- The sT-LLSQ bound O(φ_{k−1}/φ_k) = o(1) gives an unconditional, if slow, way to compute the dominant parameter even when the asymptotics order is low.
- Higher-order asymptotic knowledge strictly improves or preserves the sLLSQ error bound (η(n;p) = O(η(n;p−1))), so investing in deeper expansions is provably rewarded.
- The counterexample shows that sLLSQ's global-minimum value is not a reliable indicator of success; zero residuals can coexist with divergence, shifting practical attention to the ϑ_j conditions.
Reading between the lines
- A natural extension is to other loss functions or regularization penalties (e.g., ℓ1/LASSO) that preserve the sliding-window idea; the paper's universal dominant rate suggests the method is robust to the choice of penalty.
- The approach may adapt to asymptotics with unknown basis functions, first estimating the φ_j themselves, though the current proof assumes the basis is known a priori.
- Since the estimators are closed-form, the theory could combine with exact arithmetic to certify constants to arbitrary precision, turning the asymptotic bound into a finite-time guarantee once explicit constants are extracted.
- The structure of the error bound suggests the same technique may work for multivariate or multi-scale expansions where the ratio method breaks down.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces a framework called Asymptotics Learning Theory (ALT) for computing the finite parameter vector α in a proven asymptotic expansion f(n) = bf(n;α)(1 + Σ_{l=1}^p β_l g_l(n) + O(g_{p+1}(n))) from exact values of f. The theory specializes to exponential-product forms bf(n;x) = exp(Σ_j φ_j(n)u_j(x_j)), covering α_1 n^{α_2} α_3^n, stretched-exponential, and power-law asymptotics. Two sliding-window estimators are studied in closed form: sLLSQ, y*_n = (A^T A)^{-1}A^T b = A†b, and sT-LLSQ, y*_n = (A^T A + μI)^{-1}A^T b, where A is the windowed matrix of the φ_j. Theorem 3.14 establishes uniqueness, residual decay F*_n=O(||g_1||^2), and the bound ||y*_n − γ||=O(η(n)) with per-component rates ϑ_j(n); convergence is certified exactly when ϑ_j(n)=o(1), and Proposition 3.7 transfers rate bounds from the transformed parameters γ to the original α. Theorem 3.27 gives ||y*_n − γ||=O(χ(n))=O(1) and an unconditional dominant-component rate y*_{k,n} − γ_k = O(φ_{k−1}(n)/φ_k(n)) = o(1). Applications to analytic combinatorics (Fibonacci, Catalan, factorials, and an eleven-form family in Tables 4–6) are compared with the ratio method (Prop. 5.1) and tested numerically. Section 6.4 constructs a p=0 example with a hidden alternating term where sLLSQ diverges despite zero residuals, showing that the sufficiency conditions are genuinely necessary.
Significance. If the results hold up, the paper contributes a fairly general, closed-form family of estimators for the parameters of exponential-product asymptotics, with explicitly stated sufficient conditions, per-component rate bounds, and a Θ-transfer theorem (Prop. 3.7) that lets rates for y* convert to rates for the original parameters. The appendix proofs are coherent, and I independently verified several representative computations: det(A^T A)=6 for m=3 and 20 for m=4 in Remark 3.23, the Θ(1/n), Θ(1/n^2) vs Θ(1), Θ(1/n) components of A† log(1+β_1 g_1), and the p=0 AF-1 bound ||y*_n − γ||=O(n^2 log n) of Remark 4.1. The framework is not circular: γ enters only through the assumed expansion (2.1); the bounds are functions of A† and the decreasing scale g_l, with β_l as symbolic constants; there are no self-citations. The strongest clean result is the unconditional dominant-component rate y*_{k,n} − γ_k = O(φ_{k−1}/φ_k) (Lemma 3.26/Theorem 3.27), valid for any μ>0, any norm, and any ν≥1 for the general regularized problem. The §6.4 counterexample (zero residuals, diverging sLLSQ) is a valuable negative result that correctly identifies the sufficiency conditions as the load-bearing checkpoin
major comments (2)
- [Abstract; §1.1; Theorems 3.14/3.27; Remarks 4.1–4.3] The paper's motivating settings (DLW theorem, catalytic equations, chordal graphs; §1.1 and eq. (1.3)) provide only leading-order asymptotics, i.e., AF-2 with p=1. For that order, Table 4 gives ||y*_n − γ||=O(log n) for sLLSQ, so the condition ϑ_j(n)=o(1) in Theorem 3.14 fails and full-vector convergence is not certified. Theorem 3.27 certifies only the dominant component γ_k (Table 6, Remark 3.34); for j<k it gives only y*_{j,n} − γ_j = O(1), and the o(1) in the Table 6 footnote describes convergence to the regularized bias 0, not to γ_j. Thus, in the very settings that motivate the paper, certified recovery is limited to the leading exponential parameter (via sT-LLSQ) unless higher-order terms have been proven in advance — precisely the 'hard part' flagged in Remarks 4.1–4.3. The abstract's 'convergence (to the correct values of unknown parameters)' should be qualified. This is a scope
- [Section 4, Tables 4–6] Tables 4–6 carry the applications and operationalize the ϑ_j(n)=o(1) checkpoints, but the Puiseux estimates are asserted without derivations or scripts. Remark 3.23 shows the rates are fragile: with the same φ_j and g_l, m=3 gives η(n)=Θ(1/n) while m=4 gives η(n)=Θ(1). With the many symbolic rate entries (11 forms × orders p=0..3, plus det(A^T A) values and the §6.4 Θ-estimates), the reader cannot audit those entries. Please provide the CAS notebook/script (e.g., the MATLAB symbolic commands) or a supplementary derivation for each row family. Without such artifacts, the rate tables function as trusted assertions rather than verified results.
minor comments (5)
- [Title] Typo in the manuscript header: 'CONVERGENCE-RA TE GUARANTEES' should be 'CONVERGENCE-RATE GUARANTEES'.
- [Abstract] The phrase 'introduce a new research area' is stronger than this first paper supports; 'framework' or 'methodology' would be more defensible and would not undercut the technical contribution.
- [Appendix B.3, proof of Lemma 3.26] The step following (B.4) moves from entrywise O(1) and O(φ_{k−1}/φ_k) estimates, valid for arbitrary (eγ, w), to uniform bounds on the matrices C and D. Since the O-constants may in principle depend on eγ and w, the argument should restrict to unit basis vectors and use finiteness of the test set; the conclusion is correct, but this step is compressed.
- [Section 4, Tables 4–6] All rate tables use m=k. Given Remark 3.23 (increasing m from 3 to 4 flips η from Θ(1/n) to Θ(1)), practitioners need explicit guidance on whether and when m>k is safe; Proposition 3.22 covers only the regime where convergence already holds. A sentence in Section 4 would prevent misreading the tables as valid for all m.
- [Table 7] For the rows 'AF-2 with known α_2 / α_3', the sT-LLSQ rate columns follow from the transformations noted in parentheses, but the dimension change of the transformed problem is not shown; a pointer to the corresponding AF-6/AF-10 rows would make the comparison self-contained.
Circularity Check
No significant circularity: the estimators are derived from the assumed asymptotic expansion and their error bounds follow from that expansion; the target γ is never used to construct the estimators.
full rationale
The paper's derivation chain is self-contained in the sense required by the circularity criteria. The input is a proven asymptotic expansion (2.1) with unknown α (equivalently γ) plus a known asymptotic scale {g_l} and known basis functions {φ_j}. The estimators y*_n = A†b (sLLSQ, eq. (3.17)) and y*_n = Cb (sT-LLSQ, eq. (3.34)) depend only on the data vector b and the design matrix A built from φ_j(n); they do not use the target γ. The main error estimates are obtained by substituting the assumed expansion into b, yielding b = Aγ + log(1 + Σ β_l g_l) + z, and then bounding A† times the remainder terms (Theorem 3.14, eqs. (B.1), (3.20)) or C times the remainder plus the regularization bias -μDγ (Theorem 3.27, eqs. (B.5), (3.36), (3.40)). The β_l are treated as symbolic constants whose existence is assumed, not fitted values, so there is no fitted-input-called-prediction pattern. The sufficient convergence conditions ϑ_j(n)=o(1) and ψ_k(n)=o(1) are expressed in terms of the known matrix A† or C and the known error scale {g_l}, not in terms of the unknown γ. The paper explicitly documents where convergence fails (Remark 4.1, Section 6.4), which further confirms that the guarantees are not vacuously forced by construction. There are no self-citations: the reference list is entirely external, and no uniqueness theorem or ansatz is imported from the author's prior work. The only substantive limitation is that sLLSQ's guarantee requires sufficiently high-order proven asymptotics, and sT-LLSQ's unconditional guarantee covers only the dominant component; this is an applicability caveat, not circular reasoning. The numerical examples use known constants for verification, which is an external consistency check rather than a circular construction.
Assumptions & free parameters
free parameters (2)
- Tikhonov regularization parameter µ =
1 in all numerical experiments; any fixed µ>0 allowed
- Sliding window length m =
m=k in Tables 4-6 and most examples; m=2,3,4,5,10 in figures
assumptions (5)
- domain assumption The data sequence {f(n)} is exact or of sufficiently high accuracy
- domain assumption Assumptions 3.1–3.3: f has the proven expansion (2.1) with bf of exponential-product form (3.1); φ_j form an increasing asymptotic scale with φ_1=Ω(1); u_j are bijective differentiable with nonzero derivative
- domain assumption Assumption 3.10/3.11: the design matrix A(n) has full column rank, det(A^TA)=Ω(h(n))
- domain assumption Assumption 3.24: φ_1(n)→∞ when k=1; for k≥2, φ₁=Ω(1) suffices
- ad hoc to paper Tables 4–6 symbolic Puiseux estimates are correct as stated
Cite this review
Pith. "Pith review of Learning Asymptotics with Convergence-Rate Guarantees using Linear Least Squares." pith.science (2026). https://pith.science/paper/AJ3GEI4F
@misc{pith2026260723287,
author = {Pith},
title = {Pith review of: Learning Asymptotics with Convergence-Rate Guarantees using Linear Least Squares},
year = {2026},
howpublished = {\url{https://pith.science/paper/AJ3GEI4F}},
note = {Machine review of arXiv:2607.23287}
}
read the original abstract
We introduce a new research area that is called Asymptotics Learning Theory (ALT) and combines optimization with asymptotic analysis. In particular, ALT provides a unified approach for computing unknown constants/parameters in proven asymptotic expansions using optimization theory. In this paper, we focus on a general asymptotic form which includes a broad class of asymptotics. Furthermore, we study two powerful numerical methods, namely, sliding Linear Least Squares (sLLSQ) and sliding Tikhonov Linear Least Squares (sT-LLSQ). For these techniques we rigorously prove asymptotic estimates that lead to sufficient conditions for convergence (to the correct values of unknown parameters) and convergence-rate guarantees. Despite their strengths, both methods have also limitations, e.g., slow convergence---or even, counterintuitively, divergence---in some cases. Moreover, we present fundamental applications in analytic combinatorics, a beautiful field of mathematics that deals with asymptotic enumeration of discrete structures using complex analysis. The proposed techniques complement existing approaches, such as the ratio method and its variants. Numerical examples also verify the theoretical results. Finally, we discuss interesting research directions in ALT.
Figures
Figures from the paper (2 more)
Reviewed July 31, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.