{"id":"da0925f0-a069-48b0-b6fe-577c3bc8b4da","arxiv_id":"2501.05410","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Second-order NQT functions, a C^1 piecewise quadratic approximation to log2, preserve second-order convergence of linear interpolation on logarithmic grids while running faster than true logarithms.","lead":"This paper introduces second-order not-quite-transcendental (NQTo2) functions, a faster, nearly exact replacement for logarithms when linearly interpolating tabulated data that spans many orders of magnitude. It demonstrates that NQTo2 keeps interpolation accuracy while speeding up equation-of-state lookups in astrophysical simulations, often by 10 to 30 percent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The drop-in accuracy claim conflates resampling onto an NQTo2 grid with replacing log in an existing log-grid table; the latter incurs a fixed coordinate error ~0.01 log2 units that does not vanish as the table is refined.","rationale":"The reader's CONDITIONAL verdict is appropriate, and this stress test sharpens the condition rather than overturning it. The key new observation is that the deviation between NQTo2 and log2 is a fixed, resolution-independent coordinate error of roughly 1% of a log2 unit. In a true drop-in replacement, where an existing log-spaced table is queried using lgo2, this error contributes a nonzero term to the interpolation error even as the table spacing goes to zero. That directly contradicts the unbounded phrasing of 'essentially no cost in accuracy.' The paper's own Section 4.3 concession, that the table density needed for matching eigenspectra has not been explored, is consistent with this concern. Credit is due for the open-source implementation, the cross-architecture benchmarks, and the clear demonstration that an NQTo2-resampled grid recovers second-order convergence; those parts are not in question. The unresolved issue is which usage mode the central claim refers to and whether the drop-in mode can be supported. The proposed test would settle this by separating the two modes and measuring the asymptotic behavior of the interpolation error. Since the paper's practical claims about typical EOS tables may still hold, the verdict remains CONDITIONAL rather than moving to rejection.","tokens_in":8956,"tokens_out":13803,"duration_ms":134914,"concrete_test":"Take a smooth analytic target g(u) (e.g., g(u) = exp(u) + u^2) and build a table sampled on a true uniform log2 grid of spacing delta. Perform linear lookup in two ways: (A) use log2 to locate the bracket, and (B) use lgo2 to locate the bracket on the same table. Sweep delta from 2^-2 down to 2^-12 and record the L-infinity error of both. If mode B's error plateaus near max_x |g'(log2 x)| * max|epsilon(x)| rather than tracking mode A's delta^2 curve, the drop-in accuracy claim fails in the dense-table regime. Repeat with the SFHo and Sesame tables at their native resolutions to see which regime realistic EOS tables occupy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, 'speedup at essentially no cost in accuracy' (Section 1), is not established for the drop-in mode. NQTo2 from Eq. (10) is not a perturbation of log2 that shrinks with grid spacing: over each dyadic mantissa interval the maximum deviation |lgo2(x) - log2(x)| is about 0.0097 log2 units, independent of table resolution. If a code keeps its existing log-spaced table and only replaces log10 with lgo2 when computing lookup coordinates, linear interpolation evaluates the table at u = lgo2(x) instead of u = log2(x). The resulting error contains a fixed term g'(u) * epsilon(x) ~ 0.01 g' that remains as delta -> 0, while the true-log interpolation error vanishes as delta^2. Thus, taken literally, 'essentially no cost in accuracy' is false asymptotically for the drop-in mode. If the intended use is instead to resample or respace the table on a uniform NQTo2 grid, second-order convergence is plausible (Section 3, Figure 1), but that is not a drop-in replacement for an existing log grid and requires an explicit resampling step that the paper never specifies. The finite EOS tests in Section 4.2 (SFHo and Sesame) cannot distinguish these two modes, and Section 4.3 explicitly defers the table-density comparison for eigenfrequencies, exactly the regime where the distinction matters.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript introduces 'second-order not-quite-transcendental' functions (NQTo2), a piecewise quadratic, exactly invertible approximation to log2 designed to replace logarithms when linearly interpolating tabulated data over many orders of magnitude. The authors derive the function, discuss portable and integer-aliased implementations, benchmark log and exp variants on several CPU and GPU architectures, and test the approach on a synthetic function, the SFHo nuclear EOS, a Sesame copper EOS, and an AthenaK neutron star simulation. The main claim is that NQTo2 is a faster drop-in replacement for linear interpolation on logarithmic grids with essentially no loss of accuracy.","tokens_in":24,"tokens_out":14526,"duration_ms":202884,"significance":"If the central claim is established, the method offers a simple, low-risk performance optimization for tabulated microphysics in astrophysical simulations, with open-source implementations and reported integrated speedups up to about 30%. The paper is also commendably candid: it corrects the prior NQTo1 claim of zero accuracy loss and explicitly flags the missing table-density/eigenspectra convergence study. However, the accuracy half of the claim is not yet rigorously supported: the results are mostly visual, no error norms as a function of resolution are reported, and the drop-in mode is not cleanly separated from a resampling mode.","major_comments":[{"comment":"Equation (9) is not satisfied by Eq. (10). For x = m 2^p with m in [1/2,1), Eq. (10) gives the derivative of lgo2 with respect to x at x = 2^p as 4/(3*2^p), and at x = 2^{p+1} also as 4/(3*2^p), so the ratio is 1 rather than the 2 required by Eq. (9). The stated derivation 'to fix these, we enforce ... (9)' is therefore internally inconsistent, even though the resulting function is indeed C^1. Please correct the condition or explain the actual constraint used, and adjust the surrounding convergence discussion accordingly.","section":"Section 3, Eq. (9)"},{"comment":"The statement that 'second order convergence of linear interpolation is only guaranteed for everywhere C^1 functions' is not correct. A C^1 function whose derivative has modulus of continuity h^alpha is only guaranteed O(h^{1+alpha}) error, so C^1 alone does not imply O(h^2). The convergence argument should cite a sufficient condition such as Lipschitz derivative or piecewise C^2 with bounded second derivative; NQTo2-transformed smooth data plausibly satisfies such a condition, but the printed justification needs revision.","section":"Section 2, convergence theorem"},{"comment":"The central 'drop-in replacement' claim conflates two different operations. If NQTo2 is substituted for log in an existing log-spaced table, the lookup coordinate u = lgo2(x) differs from the table coordinate u = log10(x) by a resolution-independent offset epsilon(x) with max |epsilon| approximately 9.7e-3 in log2 units. Linear interpolation at u + epsilon then carries an error term g'(u)*epsilon that remains as the log-grid spacing delta tends to zero, so the asymptotic error does not vanish. The visual comparisons in Figs. 3 and 4 cannot rule out a fixed relative error of order one percent, and Sec. 4.3 explicitly postpones the table-density comparison needed to assess this. The authors should either state clearly that the method requires resampling the table onto a uniform NQTo2 grid and specify that resampling procedure, or restrict the accuracy claim to finite resolution and quantify the error with norms as a function of delta.","section":"Section 1, Figs. 3-4, Sec. 4.3"}],"minor_comments":[{"comment":"The word 'eiqenspectra' should be 'eigenspectra', and 'the density of table required' should read 'the table density required'.","section":"Section 4.3"},{"comment":"The word 'efficated' should likely be 'effected' or 'implemented'.","section":"Section 4.2"},{"comment":"Two different works are cited as Miller et al. (2022): arXiv:2206.08957 and the Journal of Open Source Software paper. These should be disambiguated as 2022a and 2022b in the text and reference list.","section":"References"},{"comment":"The performance plots would benefit from a statement of the timing methodology, including the number of repetitions and any confidence intervals, and from a reconciliation of the 'indistinguishable on GPU' result in Sec. 4.2 with the up-to-30% speedup in Fig. 6.","section":"Figures 2 and 6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope well, and the public code release is a strength. The main concern is precision of the accuracy claim: the literal drop-in reading is misleading, and the Section 3 derivation contains a small but real inconsistency. I would be comfortable with a revision that adds error norms, specifies the intended resampling mode, and corrects the convergence discussion. No concerns about novelty or citation practice beyond the duplicate Miller et al. (2022) reference."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Josh, here's my read. The genuinely new thing is NQTo2, a C^1 member of the NQT family that fixes the convergence defect of NQTo1 in Lp norms for p>1. The derivation is simple, the formula in Equation (10) is correct, and the paper openly corrects Miller et al.'s earlier zero-loss claim. That honesty is worth something. The benchmark suite is broad, and the integrated AthenaK speedups (up to ~30% on an RTX 3070) are plausible given how much time EOS lookups take.\n\nWhere it gets soft is the drop-in claim. The stress-test note gets this right. If you keep an existing log-spaced table and simply swap log10 for lgo2 in the lookup coordinate, you're evaluating the table at u = lgo2(x) rather than u = log2(x), a coordinate offset that stays about 0.01 log2 units no matter how fine the table is. The interpolation error then has a fixed term that does not vanish. The 'essentially no cost in accuracy' sentence is literally false for that mode. If the intended use is to resample the table on a uniform NQTo2 grid, second-order convergence is plausible, but that is not a drop-in; it is a table regeneration, and the paper never says so.\n\nThat ambiguity is not resolved by the EOS tests in Section 4.2, which compare relative errors but do not say whether the tables were regridded. And Section 4.3 explicitly defers the table-density comparison for eigenspectra, exactly the regime where the two modes diverge.\n\nAlso, the accuracy comparisons are visual; there are no numeric error norms. The convergence discussion leans on a derivative condition (Equation 9) that is nonstandard and not fully justified. The ARM speedup is unexplained, though the authors flag it.\n\nNone of this breaks the core contribution. The construction is correct, the implementation is on GitHub, and the paper is useful for anyone doing tabulated EOS interpolation. But the revision needs to say precisely which mode the method targets, add error norms, and either drop the 'drop-in' phrasing or define the required resampling.\n\nWho's it for: computational astrophysics types doing tabulated EOS interpolation. It deserves a serious referee, with a request for revision before acceptance.","headline":"NQTo2 is a real improvement over NQTo1, but the paper needs to say exactly what 'drop-in' means before the accuracy claim holds.","tokens_in":9785,"tokens_out":2993,"would_cite":true,"duration_ms":27876,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a family of cheap, exactly invertible not-quite-transcendental functions can replace true logarithms in linear interpolation of logarithmically spaced tables, giving the same accuracy at lower cost.","keywords":["not-quite-transcendental functions","logarithmic interpolation","tabulated equations of state","floating-point bit manipulation","interpolation error","performance portability","neutron star simulations","second-order convergence"],"falsifier":"Take a smooth test function with known derivatives and build both a logarithmic and an NQTo2 table at a fixed number of points; if the NQTo2 grid's interpolation error does not converge at second order in the table spacing once the table is fine enough that the NQTo2 deviation dominates, the central claim fails. Alternatively, run the paper's neutron star simulation on tables of increasing density and compare the fundamental oscillation eigenfrequencies between NQTo2 and true log interpolation: if the frequencies do not converge to the same value as the table densifies, the claimed accuracy equivalence is false.","tokens_in":8745,"feed_emoji":"⚡","tokens_out":8039,"duration_ms":68625,"temperature":0.7,"pith_summary":"This paper claims that a family of cheap, exactly invertible 'not-quite-transcendental' functions can replace true logarithms in linear interpolation of logarithmically spaced tables, giving the same accuracy at lower cost. The specific new function, NQTo2, is everywhere smooth (C1) and fixes the slow-converging error spikes that hurt the earlier first-order version. On three test problems — a synthetic function, the SFHo nuclear equation of state, and a copper Sesame equation of state — NQTo2 interpolation is essentially indistinguishable from true log interpolation, while the underlying function evaluations are about twice as fast on x86 CPUs. Integrated into a neutron star simulation, the substitution speeds up the full calculation by as much as 30 percent. If the claim holds, simulation codes can gain this speedup by swapping a single function call.","feed_headline":"Not-quite log speeds up table interpolation by up to 30%","feed_subtitle":"A smooth, cheap stand-in for the logarithm gives the same accuracy in equation-of-state lookups while making the full simulation run faster.","key_machinery":"The central object is the second-order NQT pair, $\\mathrm{lgo}_2$ and its inverse, built from the floating-point decomposition $x = m \\cdot 2^{p}$. The forward map $\\mathrm{lgo}_2(x) = -\\frac{4}{3}(m-2)(m-1) + p$ is the unique quadratic that matches $\\log_2$ at the control points $x = 2^p$ and satisfies the recurrence $\\partial_x \\mathrm{lgo}_2|_{2^p} = 2 \\, \\partial_x \\mathrm{lgo}_2|_{2^{p+1}}$, which forces the derivative discontinuities of the first-order piecewise-linear version to vanish. The inverse requires extracting the bit-level mantissa, solving a quadratic, and recombining with the exponent. The whole method is carried by this $C^1$ property plus the integer-aliasing bit tricks that make the evaluation cheap.","core_discovery":"The paper introduces second-order not-quite-transcendental functions (NQTo2) as a drop-in replacement for logarithms in table interpolation. Since a positive float $x = m \\cdot 2^{p}$ with $m \\in [1/2,1)$, the function $\\mathrm{lgo}_2(x) = -\\frac{4}{3}(m-2)(m-1) + p$ is a Hermite interpolant of $\\log_2(x)$ whose control points are powers of two; continuity of the value and derivative at those control points makes it everywhere $C^1$. This smoothness recovers second-order convergence in all $L^p$ norms when linear interpolation is performed on an NQTo2 grid. The function is exactly invertible by solving a quadratic for the mantissa, and the integer-aliased implementation runs in a small number of bit operations. The paper argues that any code using linear interpolation on logarithmic data can replace $\\log$ with $\\mathrm{lgo}_2$ and get essentially unchanged accuracy with significant speedups.","pith_inferences":["The same Hermite-style smoothing could be applied to other piecewise-linear transforms (for example, logarithms in other bases, or single-precision floats), extending the speedup to any table lookup that currently pays for transcendental calls.","Because the integer-aliased implementation discards the lowest mantissa bits during squaring, there may be a precision floor in the transformed coordinate; production codes that need extreme dynamic range should test whether dropped bits affect quantities that depend on high-order derivatives.","The untested eigenspectrum agreement suggests a concrete design rule: when oscillation modes or other spectral quantities are the quantities of interest, the table density may need to be higher for NQTo2 than for true log interpolation, and that density should be measured before relying on the speedup in such simulations."],"forward_implications":["Any code currently doing linear interpolation on log-spaced tables can substitute NQTo2 for the logarithm with essentially no change in accuracy and, on x86 CPUs, about a 2x speedup in the function call.","An integrated neutron star simulation speeds up by up to 30 percent when NQTo2 replaces logarithmic interpolation in the equation-of-state lookups.","Higher-order interpolation schemes, such as biquintic interpolation, do not benefit from this replacement; the gain is specific to linear interpolation on logarithmic grids.","NQTo2 restores second-order convergence in $L^p$ norms for $p > 1$, which the first-order NQT lacked, making the error essentially indistinguishable from true log interpolation on the SFHo and copper equations of state."],"supporting_citations":[{"why":"Supplies the piecewise-linear interpolant of log2 with control points at powers of two that defines first-order NQT functions.","marker":"Hall et al. (1970)"},{"why":"Gives the interpolation form used to construct the everywhere-C1 second-order NQT function.","marker":"Hermite (1877)"},{"why":"Documents the series-expansion cost of true logarithms that NQT methods avoid.","marker":"Cody & Waite (1980)"},{"why":"Provides the integer-aliasing trick that makes NQT evaluation fast by reinterpreting float bits as integers.","marker":"Blinn (1997)"},{"why":"Supplies the Stellar Collapse table format used for the SFHo nuclear equation of state tests.","marker":"O'Connor & Ott (2010)"},{"why":"Provides the SFHo finite-temperature nuclear equation of state used in the interpolation and simulation tests.","marker":"Steiner et al. (2013)"},{"why":"The Singularity-EOS library in which the NQT methods are implemented for equation-of-state lookups.","marker":"Miller et al. (2024)"},{"why":"The numerical relativity code used for the integrated neutron star simulation and timing comparison.","marker":"Stone et al. (2024)"}],"fun_headline_variants":["Drop-in log fake cuts interpolation time 30%","New math function makes interpolation 30% faster, same accuracy","Pseudolog function speeds astrophysics tables by 30%","NQTo2 log stand-in keeps accuracy, boosts speed by 30%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The accuracy equivalence rests on the assumption that the deviation of NQTo2 from a true logarithm is small compared with the interpolation error of the table itself; the paper demonstrates this on only three test cases and explicitly leaves open the table density required for NQTo2 and logarithmic tables to produce matching neutron star oscillation eigenspectra.","fun_headline_variants_meta":{"raw":{"variants":["Drop-in log fake cuts interpolation time 30%","New math function makes interpolation 30% faster, same accuracy","Pseudolog function speeds astrophysics tables by 30%","NQTo2 log stand-in keeps accuracy, boosts speed by 30%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000911,"raw_usage":{"total_tokens":3870,"prompt_tokens":857,"completion_tokens":3013,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":2941}},"tokens_in":473,"tokens_out":3013,"duration_ms":20003,"temperature":1.0,"reasoning_tokens":2941,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:14:01.200023+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a smooth test function with known derivatives and build both a logarithmic and an NQTo2 table at a fixed number of points; if the NQTo2 grid's interpolation error does not converge at second order in the table spacing once the table is fine enough that the NQTo2 deviation dominates, the central claim fails. Alternatively, run the paper's neutron star simulation on tables of increasing density and compare the fundamental oscillation eigenfrequencies between NQTo2 and true log interpolation: if the frequencies do not converge to the same value as the table densifies, the claimed accuracy equivalence is false.","supporting_citations":[{"cited_title":"M., et al","cited_arxiv_id":null,"evidence_quote":"The Singularity-EOS library in which the NQT methods are implemented for equation-of-state lookups."},{"cited_title":"1970, IEEE Transactions on Computers, C-19, 97, 10.1109/T-C.1970.222874","cited_arxiv_id":null,"evidence_quote":"Supplies the piecewise-linear interpolant of log2 with control points at powers of two that defines first-order NQT functions."},{"cited_title":"1980, Software Manual for the Elementary Functions, Prentice-Hall Foundations of Earth Science Series (Prentice-Hall)","cited_arxiv_id":null,"evidence_quote":"Documents the series-expansion cost of true logarithms that NQT methods avoid."},{"cited_title":"1997, IEEE Computer Graphics and Applications, 17, 80, 10.1109/38.595279","cited_arxiv_id":null,"evidence_quote":"Provides the integer-aliasing trick that makes NQT evaluation fast by reinterpreting float bits as integers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Stellar Collapse table format used for the SFHo nuclear equation of state tests."}],"review_version":1}