{"id":"49bb8449-6179-4879-8faa-83a16c1b5c4d","arxiv_id":"2411.15374","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Simple fitted initial-guess formulas improve solver convergence speed for Kepler's equation, especially for hyperbolic orbits.","lead":"This paper uses symbolic regression to discover simple formulas for the starting point of iterative solvers of Kepler's equation. The formulas give modest speedups for elliptical orbits and larger speedups for hyperbolic orbits in the paper's benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The hyperbolic initial guess (Eq. 4) saturates at min{M,e}; for M >> e it stays bounded while E ~ ln(2M/e), so the claimed major hyperbolic speedup likely reverses outside the benchmarked M <= 3*pi/4 range.","rationale":"The paper's elliptical contribution is separate and likely robust: the new elliptical guess reduces to the canonical form over much of parameter space and adds only a max operation, so the modest speedup near pericenter is plausible. The hyperbolic claim is the paper's main differentiator, and the concern above is an internal asymptotic mismatch between Eq. (4) and the hyperbolic Kepler equation, Eq. (3). The proposed test is decisive and does not depend on the unspecified Rust quartic solver: any reasonable iterative solver will require many extra iterations starting from E0 = 1.56 when the true solution is near 4.94. This is more fundamental than the reader's concern about missing code and solver representativeness, though that concern also remains. The verdict should stay conditional: the paper can be made correct by explicitly limiting the hyperbolic claim to small M, adding the large-M benchmark, and releasing the code. If the unqualified hyperbolic claim is retained, it should be rejected; conditional acceptance pending this clarification is the appropriate response.","tokens_in":3924,"tokens_out":6362,"duration_ms":58836,"concrete_test":"Use the same tolerance (e.g., 1e-14) and the same root-finding method (quartic or Newton) to compare Eq. (4) against Danby's ln(2M/e + 1.8) for e in {1.1, 1.5, 2, 4, 10} and M in {10, 100, 10^4}, plus the Fig. 2 range for calibration. Record iteration counts and wall time. If Eq. (4) requires more iterations or time than canonical for M >= 10 at any tested e, the major-hyperbolic-speedup claim must be restricted to small M and the abstract and conclusions revised accordingly.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that Eq. (4) gives 'major speed improvements for hyperbolic orbits' rests on benchmarks shown only for M in [0, 3*pi/4] and e in [1, 4] (Fig. 2). For fixed e > 1 and M > e, the proposed guess becomes E0 = e/(e - 0.538), which is bounded and approaches 1 as e grows, whereas the true hyperbolic anomaly grows logarithmically, E ~ ln(2M/e), as also encoded in the canonical Danby guess ln(2M/e + 1.8). Concretely, with e = 1.5 and M = 100, Eq. (4) gives E0 = 1.56, the canonical guess gives about 4.91, and the true solution is about 4.94; a root finder started at 1.56 must traverse a large interval, so iteration counts and wall time will be much worse than for the canonical guess. This is not a solver-implementation detail but a property of Eq. (4) at large M. Unless the intended application is explicitly restricted to the small-M range scanned in Fig. 2, the headline claim that Eq. (4) improves hyperbolic Kepler solves in general is not supported and would reverse in a physically relevant large-M regime.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper uses symbolic regression with a genetic algorithm to derive two new initial guesses for iterative solvers of Kepler's equation: Eq. (2) for elliptical orbits, E = e sin M + max{M, e(sin M + 0.591)}, and Eq. (4) for hyperbolic orbits, E = min{M,e}/(e - 0.538). Benchmarks using a quartic root-finding algorithm in Rust (Figs. 1 and 2) show modest speed improvements for elliptical orbits and larger speed improvements for hyperbolic orbits within the scanned parameter ranges. The author concludes that the new guesses improve on canonical initial guesses and that the symbolic-regression approach yields interpretable, low-cost expressions.","tokens_in":4204,"tokens_out":6305,"duration_ms":55994,"significance":"If the reported speedups hold, the results are practically useful for celestial mechanics software that solves Kepler's equation in high-throughput settings. The proposed guesses are simple, interpretable, and cheap to evaluate, and the paper is transparent that the constants 0.591 and 0.538 were found by fitting rather than derived, which avoids circularity in the benchmark comparison. The external comparison against standard guesses is a fair and relevant test of the central speed claim. However, the absence of released code and the incomplete specification of the root-finding algorithm weaken the reproducibility of the benchmark, and the hyperbolic guess is shown here to fail in a physically relevant large-mean-anomaly regime that the paper does not mention.","major_comments":[{"comment":"The hyperbolic initial guess is not uniformly better outside the benchmarked domain. For M >= e, Eq. (4) reduces to the constant e/(e - 0.538), independent of M, whereas the true hyperbolic anomaly grows logarithmically, E ~ ln(2M/e). For example, with e = 1.5 and M = 100, Eq. (4) gives E0 = 1.56, the canonical Danby guess gives about 4.91, and the true solution is about 4.94; a iterative solver started at 1.56 will require many more iterations than one started at 4.91. Since Figure 2 only scans M in [0, 3*pi/4] and e in [1, 4], the claimed 'major speed improvements for hyperbolic orbits' is not supported for larger M and in fact reverses. Please either restrict the claim to the benchmarked range or modify the guess to include an asymptotic regime, e.g., a switch to a logarithmic form for large M.","section":"Section 4, Eq. (4)"},{"comment":"The benchmarks rely on an unspecified 'quartic root-finding algorithm'; no source code or data are released. The central claim of this paper is a speed improvement, and that claim necessarily depends on the solver's update rule, stopping criterion, and implementation. Without stating which algorithm is used (or giving a reference) and ideally releasing the benchmark code, the results cannot be independently checked. Please specify the algorithm, report iteration counts and timing with error bars, and consider making the code available.","section":"Section 3, footnote 3"}],"minor_comments":[{"comment":"The timing panels show point values without error bars or confidence intervals; since criterion.rs can provide such intervals, please include them to quantify the 'some noise' mentioned in the hyperbolical figure caption.","section":"Section 3, Figures 1 and 2"},{"comment":"The details of the symbolic regression and genetic algorithm (fitness function, parsimony penalty strength, population size, number of generations) are not given, so the search procedure is not reproducible; adding a short description or reference would strengthen the methodological contribution.","section":"Section 2, paragraph 2"},{"comment":"As written, Eq. (4) has a singular denominator at e = 0.538. For hyperbolic orbits e > 1, so this is not a practical issue, but please state the domain of validity explicitly in the text.","section":"Section 4, Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The large-mean-anomaly failure of Eq. (4) is a substantive empirical issue rather than a style matter: it concerns the physical regime of hyperbolic orbits that are far from pericenter, which is common in applications such as interstellar object trajectory modeling. If the authors restrict the hyperbolic claim to the benchmarked range or add an asymptotic regime, the paper could be acceptable. The missing specification of the root-finding algorithm is also a reproducibility problem that should be fixed before publication. The elliptical part is sound, and the overall approach is promising, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the elliptic initial guess (Eq. 2) is a genuine, modest win, and the paper is transparent about how the constants were found. But the headline hyperbolic claim is over-sold: Eq. (4) caps at min{M,e} and is bad for M >> e, a regime the benchmarks don't cover. The paper deserves a referee, but not as is.\n\nThe genuinely new thing is the use of symbolic regression to fit initial guesses, yielding two closed-form expressions that are cheaper than Markley's and better than the classic E = M + e sin M near pericenter for high e. That part is plausible and useful. The author also deserves credit for noting the search recovered canonical guesses under parsimony, which is a nice sanity check.\n\nThe soft spots. First, the hyperbolic formula. For fixed e>1 and M >> e, E0 = e/(e-0.538) is roughly constant while true E grows as ln(2M/e). The benchmark only explores M up to 3π/4, so the claimed 'major speed improvements for hyperbolic orbits' does not extend to large M. A root finder starting at ~1.5 instead of ~5 will need extra iterations. That's a real limitation, not a quibble. The abstract and conclusions should be qualified, or the formula should be patched with a log term for large M. Second, the benchmarks rely on an unspecified 'quartic root-finding algorithm' in Rust, no code or data are released. That makes the speed claims hard to audit, though the iteration counts are at least reproducible in principle from the formulas. The elliptical formula is the more solid contribution; the hyperbolic one is promising in a limited domain.\n\nBottom line: this is a useful, honest little paper about a computational subroutine. It isn't a field changer, and it doesn't need to be. A serious referee should see it, primarily to push the author to fix the hyperbolic claim and release benchmarks. I would not cite it in my own work until the large-M behavior is addressed. Reading-group material? Maybe, as a cautionary example of extrapolating from a limited parameter scan.","headline":"Good elliptic initial guess, but the hyperbolic speedup is only shown in a small-mean-anomaly box and likely reverses at large M.","tokens_in":4730,"tokens_out":1841,"would_cite":false,"duration_ms":17936,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65H05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Symbolic regression discovers simple initial guesses that make iterative Kepler solvers faster, with the largest gains for hyperbolic orbits.","keywords":["Kepler's equation","initial guess","symbolic regression","genetic algorithm","elliptical orbit","hyperbolic orbit","numerical root finding","celestial mechanics"],"falsifier":"Benchmark both the new and canonical guesses using Newton's method and a secant method in a compiled language such as C or Fortran on a grid of $e\\in[0.9,1)$ and $M\\in[0,\\pi/4]$ for the elliptical case, and $e\\in(1,4]$, $M\\in[0,\\pi]$ for the hyperbolic case. If the new guesses do not reduce the median iteration count or wall-clock time to reach $10^{-15}$ accuracy in these settings, then the paper's central claim of a general speed improvement is falsified. A simpler check: verify that for hyperbolic orbits with $e$ just above 1 and small $M$, the new guess keeps the Newton iteration inside the basin of convergence; if it occasionally diverges, the speed gain is not real.","tokens_in":3727,"feed_emoji":"🛰️","tokens_out":7067,"duration_ms":59817,"temperature":0.7,"pith_summary":"Kepler's equation relates mean anomaly to eccentric anomaly and must be solved numerically billions of times in orbit simulations. This paper argues that the standard starting guesses for iterative solvers can be improved cheaply: a machine-learning search over simple arithmetic expressions produced the elliptical guess $E = e\\sin M + \\max\\{M, e(\\sin M + 0.591)\\}$ and the hyperbolic guess $E = \\min\\{M,e\\}/(e - 0.538)$. In benchmarks with a quartic root-finding algorithm, the new guesses match or beat the canonical guesses, giving modest time savings on elliptical orbits and large savings on hyperbolic orbits. The practical payoff is that any orbit-propagation code can adopt these two one-line formulas without adding memory lookups or branching.","feed_headline":"Two simple starting guesses speed up Kepler orbit solvers","feed_subtitle":"Elliptical orbits get a modest speedup; hyperbolic solvers finish in far fewer iterations.","key_machinery":"The load-bearing object is the initial-guess expression itself, produced by symbolic regression with a parsimony penalty. The algorithm explores the space of short formulas built from $e$, $M$, $\\sin M$, and constants of order unity using only $+$, $-$, $\\times$, $\\min$, and $\\max$, so the result is a closed-form, branch-free expression that costs almost nothing to evaluate. The crucial structural trick is the $\\max$ operation in the elliptical guess: it partitions the $(e,M)$ plane into two regions, applying the cheap canonical guess where it is already good and a corrected formula only where it is needed. For the hyperbolic guess, the operation $\\min\\{M,e\\}$ plays a similar role, approximating the behavior of the true solution as a function of whichever of $M$ and $e$ is limiting. The parsimony penalty is what keeps the expressions short enough to be evaluated faster than the extra iterations they save.","core_discovery":"The paper's central claim is that a genetic symbolic-regression search, restricted to the operations $+$, $-$, $\\times$, $\\min$, $\\max$ and the variables $e$, $M$, $\\sin M$, yields initial guesses that outperform the standard textbook guesses for both forms of Kepler's equation. For elliptical orbits the discovered guess coincides with the classic $M + e\\sin M$ in most of parameter space but uses a $\\max$ to switch to a better fit in the high-eccentricity, near-pericenter region, where the canonical guess is poor. For hyperbolic orbits the discovered guess $\\min\\{M,e\\}/(e-0.538)$ beats the usual $\\ln(2M/e+1.8)$ over most of the domain and appears to improve even at arbitrarily large eccentricity. The speed advantage is measured in iteration counts and wall-clock time at $10^{-15}$ (elliptical) and $10^{-14}$ (hyperbolic) precision. The paper presents these as practical improvements, not a theoretical ceiling: better guesses may exist, and the technique is offered as a template for finding them.","pith_inferences":["The machine-dependent $\\sin$ call in the elliptical guess could likely be replaced by a truncated Taylor series; the author flags this as a possibility but did not implement it, and a direct benchmark would show whether the trade-off is worthwhile.","The same genetic search over cheap arithmetic expressions could be applied to other transcendental equations in orbital mechanics, such as the Barker equation or universal-variable formulation, where analogous speedups may exist.","In very large simulations that solve Kepler's equation millions of times, even one saved iteration per solve can translate into substantial wall-clock savings; the paper's benchmark does not directly quantify this bulk-throughput effect."],"forward_implications":["Any celestial-mechanics code that solves Kepler's equation by iteration can replace the canonical initial guesses with Equations (2) and (4) and expect equal or faster convergence, with the largest gains for high-eccentricity elliptical orbits near pericenter and for hyperbolic orbits.","The hyperbolic guess $\\min\\{M,e\\}/(e-0.538)$ remains effective even as eccentricity grows without bound, making it relevant for unbound trajectories and interstellar objects.","Because the new guesses are plain analytic expressions, they carry no memory footprint and cause no cache misses, unlike grid-interpolation or neural-network initializers.","The symbolic-regression recipe can be rerun with custom cost weightings, such as emphasizing rare high-eccentricity orbits, to tailor initial guesses to specific applications."],"supporting_citations":[{"why":"Supplies the canonical initial guesses $E = M + e\\sin M$ and $E = \\ln(2M/e + 1.8)$, which are the baselines the new expressions are benchmarked against.","marker":"Danby 1992"},{"why":"Provides the highly accurate analytic expression whose evaluation is slower than iterative root-finding, motivating the search for cheaper initial guesses.","marker":"Markley 1995"},{"why":"Gives the contour-integral 'analytic' solution that is slower in practice than iterative methods, framing the speed problem the paper addresses.","marker":"Philcox et al. 2021"},{"why":"Supplies the micro-benchmarking harness used for the robust wall-clock time measurements reported in Figures 1 and 2.","marker":"Aparicio & Heisler 2014–2024"}],"fun_headline_variants":["Genetic algorithm finds faster initial guesses for Kepler's equation","Simple new guesses speed hyperbolic Kepler solvers dramatically","Kepler solver speedup: genetic discovery improves both orbit types","Symbolic regression yields better starting points for Kepler iterations","Hyperbolic orbits: new guess beats standard, major speedup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed speedups rest on timing measurements from a single, not fully specified quartic root-finding implementation written in Rust; if that solver, compiler, or machine is not representative of how most codes solve Kepler's equation, the improvements may shrink or vanish in practice.","fun_headline_variants_meta":{"raw":{"variants":["Genetic algorithm finds faster initial guesses for Kepler's equation","Simple new guesses speed hyperbolic Kepler solvers dramatically","Kepler solver speedup: genetic discovery improves both orbit types","Symbolic regression yields better starting points for Kepler iterations","Hyperbolic orbits: new guess beats standard, major speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1373,"prompt_tokens":811,"completion_tokens":562,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":427,"completion_tokens_details":{"reasoning_tokens":483}},"tokens_in":427,"tokens_out":562,"duration_ms":5660,"temperature":1.0,"reasoning_tokens":483,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:21:22.426390+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Benchmark both the new and canonical guesses using Newton's method and a secant method in a compiled language such as C or Fortran on a grid of $e\\in[0.9,1)$ and $M\\in[0,\\pi/4]$ for the elliptical case, and $e\\in(1,4]$, $M\\in[0,\\pi]$ for the hyperbolic case. If the new guesses do not reduce the median iteration count or wall-clock time to reach $10^{-15}$ accuracy in these settings, then the paper's central claim of a general speed improvement is falsified. A simpler check: verify that for hyperbolic orbits with $e$ just above 1 and small $M$, the new guess keeps the Newton iteration inside the basin of convergence; if it occasionally diverges, the speed gain is not real.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the canonical initial guesses $E = M + e\\sin M$ and $E = \\ln(2M/e + 1.8)$, which are the baselines the new expressions are benchmarked against."},{"cited_title":"2014–2024, criterion.rs: Statistics-driven micro-benchmarking library for Rust","cited_arxiv_id":null,"evidence_quote":"Supplies the micro-benchmarking harness used for the robust wall-clock time measurements reported in Figures 1 and 2."}],"review_version":1}