Predicates resolved by code, and the model's own readings and reading list.
The hypothesis ledger: predicates the planning model committed to before the data could answer them, each resolved by code. Open a row for the statement, the proposed mechanism and the control that would have shown the mechanism wrong. Below the ledger are the model's readings of the archive and its literature digests.
Each row pairs a machine-checkable predicate over per-cell statistics, for example slow.p3s4.heldout < slow.p4s4.heldout, with the mechanism and control the model recorded before the data could answer. p and s name an (order, stages) cell, the leading word picks the cost model, and heldout in a predicate is that cell elite's error on the held-out set, which selects elites and so is not a clean test set.
Verdicts come from code, never from the model. Once every cell a predicate names holds min_samples records, the predicate is evaluated against those cells' running statistics. d is Cohen's d; below 0.2 the verdict is inconclusive whichever way the comparison went, so a weak effect cannot be claimed. A predicate that names an empty cell is inconclusive too, because absence of evidence never refutes. n is the smallest record count among the named cells, and c is the cycle the verdict landed on (c123+ means proposed at cycle 123 and still open).
×N means the model posed that predicate N times over the run. The row shows the newest verdict and lists the id of every posing; mixed beside the verdict means those posings did not all get the same one.
Source: rk-work/hypotheses.jsonl; exact counts.
| verdict group | distinct predicates |
|---|---|
| supported | 47 |
| refuted | 162 |
| inconclusive | 20 |
| open | 160 |
27 older predicates are not shown here. The whole log is in rk-work/hypotheses.jsonl.
142 older predicates are not shown here. The whole log is in rk-work/hypotheses.jsonl.
140 older predicates are not shown here. The whole log is in rk-work/hypotheses.jsonl.
Model-written text, in this section and the next. The sources are citations the model collected itself, so check them before relying on them.
Readings of the archive the model wrote during the run, newest at the top. Each one comments on the numbers as they stood at its cycle, so where the text and the tables disagree, trust the tables. Unlike hypotheses these readings carry no predicate, so code cannot catch them being wrong.
The archive remains sparse despite 2,599 cycles and a stall counter of 569. Order 2 has the broadest coverage, with 8 of 40 cells filled, compared with 6 of 32 for order 3 and 3 of 24 for order 4; no searchable order-1 cell is filled. Most occupied cells lie in buckets 1, 3, leaving the higher buckets unexplored. The lowest recorded held-out error is 2.862e-02 for the order-2, three-stage cell in bucket 1, followed by 3.766e-02 for the order-3, six-stage cell in bucket 3. Search error and held-out error are only weakly aligned: the latter cell has the best search error, 4.851e-03, while several candidates with similarly small search errors generalize substantially worse.
This mismatch is consistent with coefficient-specific quantization effects rather than a simple advantage from adding stages or increasing classical order. Floor rounding can accumulate directional bias, and individual tableaux can amplify or suppress that bias differently across trajectories. Extra stages also introduce more rounded intermediate values and more opportunities for error accumulation, so their theoretical approximation benefit need not survive fixed-point execution. The archive reflects this tension: the order-3, six-stage result is strong, but the corresponding five-stage cells are not clearly better than order-2 alternatives, and the available order-4 cells have held-out errors between roughly 6.0e-02 and 8.5e-02.
The recent sequence of observations also provides little support for clean cost-model crossovers based on order and stage count. Several comparisons involving order-2 and order-3 methods were resolved as refuted, including both error reversals and cycle-cost reversals across fast, slow, and AVR-approximate models. Some cycle comparisons have large effect sizes, while error comparisons are more moderate. This fits the literature digest: arithmetic precision, stage placement, rounding behavior, compiled instruction paths, memory traffic, and right-hand-side evaluation costs can vary independently. The mixed fixed-reference result and the different RK4 coefficient fractions under fast and slow models likewise suggest that hardware assumptions materially alter which coefficient patterns are favored.
Substantial uncertainty remains because most grid cells are empty, only best-in-cell summaries are shown, and the current search is stalled at the order-5-stage, bucket-2 cell. The next useful examinations are repeated held-out evaluation of the strongest order-2 three-stage and order-3 six-stage candidates, coefficient-level measures of rounding sensitivity and directional drift, and invariant or pseudo-energy drift over longer trajectories. Cycle comparisons would also benefit from compiled, hardware-appropriate measurements rather than stage-count proxies. Expanding coverage around the empty neighboring buckets and the entirely empty order-4 five-stage row would help distinguish genuine structural patterns from archive selection effects.
The archive remains sparse and strongly concentrated in low-to-middle buckets: 17 of 136 searchable cells are filled, with no searchable order-1 entries and no entries above bucket 3. Order 2 has the broadest coverage, while orders 3 and 4 occupy progressively fewer stage, bucket combinations. The lowest held-out error is 2.862e-02 for the order-2, three-stage entry in bucket 1; the order-3, six-stage bucket-3 entry follows at 3.766e-02. More stages therefore do not produce a monotone improvement, and the current five-stage bucket-3 target remains empty despite the long run and a stall counter of 544.
The populated cells show a persistent gap between search and held-out error. One of the sharpest examples is the order-3, six-stage bucket-3 entry, whose search error is 4.851e-03 but held-out error is 3.766e-02; the order-2, four-stage bucket-2 entry similarly moves from 7.239e-03 to 5.865e-02. This is consistent with coefficient choices fitting the search cases more closely than unseen cases, with quantization buckets, stability constraints, and stage interactions limiting transfer. Cycle cost also generally rises with stage count, although its dependence on order and cost model is not simple enough to support model-independent rankings.
The recently resolved hypotheses reinforce that qualification. Several proposed reversals between fast, slow, and AVR-approximate cost models were not observed, including comparisons across two versus three stages and across four, five, and six stages. Effect sizes range from modest to very large, but the sample counts vary from 114 to 4,933, so the archive supports some contrasts much more densely than others. This pattern is compatible with the literature digest’s general emphasis on hardware-aware Runge, Kutta tradeoffs: arithmetic structure, coefficient representation, and target architecture can matter alongside formal order and stage count, while simple expectations about higher order or additional stages need not predict realized error or cost.
Substantial uncertainty remains because most cells are empty, order 1 appears only outside the searchable stage range, and the search is currently stalled. The fixed-reference result is mixed, with the RK4 coefficient fraction changing from about 0.559 under the fast model to 0.362 under the slow model and a reported crossover at h = 0.15625; this suggests that conclusions may depend on workload scale as well as cost assumptions. Useful next examinations include repeated held-out evaluation of the best occupied cells, analysis of why the search-to-held-out gap is so large, targeted attempts to populate the empty five-stage and higher-bucket cells, and sensitivity checks around the crossover step size and the coefficient quantization boundaries.
The archive is sparse and strongly concentrated in low-to-middle cost buckets: 17 of 136 searchable cells are filled, with no searchable order-1 entries and no occupied buckets above b3. Order 2 has the broadest coverage, while orders 3 and 4 become progressively thinner. The best held-out error is 2.862e-02 for the three-stage, order-2 entry, followed by 3.766e-02 for the six-stage, order-3 entry. Increasing stage count or formal order does not produce a monotone improvement: some candidates have low search error but substantially worse held-out error, especially the order-3 six-stage b3 entry, whose search error is 4.851e-03 but held-out error is 3.766e-02.
This pattern is consistent with a tradeoff among approximation quality, quantized coefficients, and execution cost. Additional stages enlarge the coefficient search space and can reduce error on the search workload, but they also add cycles and more opportunities for fixed-point rounding or workload-specific fitting. The gap between search and held-out error in several cells suggests that search optimization is not always transferring cleanly. Cycle counts generally rise with stage count, although bucket placement and cost-model details complicate direct comparisons.
The recent hypothesis observations reinforce that these tradeoffs are cost-model dependent rather than simple stage-count rules. Several proposed reversals between fast, slow, and AVR-approximate models were refuted, with effect sizes ranging from modest to very large. Together with the literature-digest reference result, mixed falsification, RK4 coefficient fractions of about 0.56 under the fast model and 0.36 under the slow model, and a crossover near h = 0.15625, the archive fits the familiar expectation that implementation details and step size can change which method looks preferable. It does not support treating formal order, stage count, or one cost model as a universal proxy for practical performance.
The main uncertainty is coverage. The stall counter is 519, the current five-stage b3 region remains empty for orders 1 and 4, and most buckets have no representative, so the visible frontier may reflect search reachability as much as underlying method quality. Next examinations should distinguish genuine infeasibility from search failure, repeat the strongest entries across workloads and fixed-point formats, and map error around the reported step-size crossover. It would also be useful to investigate why the order-2 three-stage candidate generalizes best despite its small stage count, and why some very low search-error candidates show comparatively large held-out degradation.
The archive remains sparse and strongly concentrated in low-to-middle cost buckets. Order two occupies 8 of 40 searchable cells, order three 6 of 32, and order four only 3 of 24; order one has no searchable entries, although one result exists outside the stage range. The best held-out value is 2.862e-02 for the three-stage, order-two entry, followed by 3.766e-02 for the six-stage, order-three entry. Additional stages therefore do not produce a monotone improvement: they sometimes reduce search error, but held-out performance varies irregularly and can worsen substantially.
A persistent gap between search and held-out error suggests that selection on the search workload is a major mechanism shaping the archive. The six-stage, order-three entry is the clearest example, with search error 4.851e-03 but held-out error 3.766e-02; several other cells show still larger proportional gaps. Stage count also raises cycle cost broadly, while fixed-point quantization and coefficient structure can change both approximation error and execution cost enough to disrupt simple order-based rankings. The recent hypothesis outcomes reinforce this picture: several proposed reversals between cost models or neighboring order-stage combinations were not observed, including some with large standardized effects.
This pattern accords with the literature digest’s general expectation that classical Runge, Kutta order alone is not a reliable proxy for performance after coefficients are quantized and implementation costs are included. It also fits the fixed-reference result, whose mixed verdict, differing RK4-coefficient fractions under fast and slow models, and crossover at h = 0.15625 indicate workload-dependent tradeoffs rather than a single dominant construction. The current archive offers more support for interactions among stage count, cost model, step size, and generalization than for a uniform rule favoring higher order or more stages.
Substantial uncertainty remains because most cells are empty, the search has stalled for 494 cycles, and the current five-stage bucket-three cell is itself unfilled. Comparisons may consequently reflect uneven discovery as much as underlying feasibility, especially for order four and higher buckets. The next useful examinations are targeted attempts to populate nearby empty cells, repeated held-out evaluation of entries with unusually low search error, and direct coefficient-level comparisons around the h = 0.15625 crossover. It would also help to separate infeasible cells from cells that are merely hard to reach and to test whether the observed search-to-held-out gaps persist across additional workloads and fixed-point formats.
The archive remains sparse and strongly concentrated in low-numbered buckets: 17 of 136 searchable cells are filled, with no searchable order-1 cells, eight order-2 cells, six order-3 cells, and three order-4 cells. The lowest held-out error is 2.862e-02 for the order-2, three-stage entry in bucket 1, followed by 3.766e-02 for order 3 with six stages in bucket 3. Search error and held-out error are not closely aligned: the latter entry has the best search error, 4.851e-03, but not the best held-out result, while several other cells show much larger gaps. This suggests that optimizing the search workload alone does not reliably predict performance on held-out cases.
More stages generally carry higher recorded cycle costs, but accuracy does not improve monotonically with either stage count or formal order. Order 2 with three stages does better than the populated two-, four-, five-, and six-stage order-2 cells on held-out error, whereas the strongest populated order-3 result occurs at six stages. The recent hypothesis records also show that several proposed cost-model crossovers were refuted, sometimes with large effect sizes, indicating that simple stage-count comparisons do not capture the observed ordering across fast, slow, and AVR-approximate multiplication models. Coefficient arithmetic, derivative work, controller overhead, and model-specific multiply latency can all change the balance.
The literature digest provides mechanisms for both the cost variability and the search-to-held-out gaps. Embedded pairs share stages, and FSAL methods can reuse a derivative after accepted steps, so nominal stage count can overstate steady-state work; rejected steps interrupt that reuse and make realized cost dependent on controller behavior. PI-style control can reduce step-size oscillation, while Q15 floor rounding introduces a systematic negative bias and small updates can disappear entirely through finite-precision stagnation. These effects plausibly produce method-specific failure regimes that are not summarized by formal order, stage count, or a single aggregate search score. The fixed reference crossover at h = 0.15625 and the differing RK4 coefficient fractions under fast and slow models further illustrate the sensitivity to arithmetic-cost assumptions.
Considerable uncertainty remains because most cells are empty, the search has stalled for 469 cycles, and the current order-1 target cell has no searchable representative. Comparisons are therefore conditional on a narrow and unevenly sampled archive, and some hypothesis records have far fewer samples than others. The most useful next examination would separate accepted and rejected steps, FSAL reuse rates, saturation events, vanished updates, and controller arithmetic by cell and cost model. Re-evaluating leading candidates across step sizes near the reference crossover, with alternative rounding modes and matched workloads, would help determine whether the observed rankings arise from truncation behavior, fixed-point bias, controller dynamics, or hardware cost assumptions.
35 older entries are not shown here. The whole log is in rk-work/interpretation/interpretations.jsonl.
Digests of published work the model looked up on the web during the run, newest at the top. They feed the directive and hypothesis prompts. Nothing here is a measurement from this project, and no measured page depends on it.
Bogacki, Shampine introduced an embedded 3(2) Runge, Kutta pair whose shared stages provide a third-order update and a second-order error estimate. The commonly used scheme has four stages and the earliest-same-as-last property, allowing the final derivative of an accepted step to be reused as the earliest derivative of the next step. Dormand and Prince constructed embedded 5(4) pairs with seven stages, emphasizing small principal fifth-order truncation terms and extended absolute-stability regions. These results show that nominal stage count alone does not determine recurring cost: derivative reuse, accepted-step rate, coefficient arithmetic, and stability-limited step size also matter. They therefore give no general ordering between the archive cells in H-703 through H-709 without cell-specific tableaux and cycle measurements.
sources:
Recent mixed-precision Runge, Kutta research treats arithmetic precision and stage evaluation placement as separate design variables. A 2026 study of Bogacki, Shampine 3(2) found that mixed precision can retain much of the high-precision accuracy, but low precision did not reduce the number of function evaluations enough to offset slower operations in the tested setting. Work on reduced-precision fixed-point ODE solvers also shows that rounding mode materially affects trajectory error: stochastic rounding outperformed deterministic round-to-nearest in the reported neuron-model experiments. Because floor rounding is directional rather than unbiased, classical order and stage count alone do not characterize its accumulated error.
Structure preservation introduces another axis beyond classical order. Explicit pseudo-energy-preserving RK schemes have been constructed with energy-preservation order higher than their classical order; tests reported smaller errors than same-order RK methods over moderately long Hamiltonian simulations. Thus, comparisons among four-, five-, and six-stage archive cells require coefficient-level measures of quantization sensitivity and invariant drift, not only nominal order. For cycle hypotheses, Arm identifies the Cortex-M0+ technical reference manual as the source for instruction timing, while Zephyr exposes direct cycle-difference measurements; measured compiled paths are therefore needed because equal stage counts need not imply equal instruction, memory-access, or right-hand-side costs.
sources:
Arm documents two implementation-time multiplier choices for Cortex-M0+: a fast single-cycle array multiplier and a smaller 32-cycle iterative multiplier. Both execute MULS, which returns the low 32 bits of a 32-by-32-bit product. The core timing table lists register addition, subtraction, and immediate or register shifts as one-cycle instructions, while loads can take one or two cycles depending on the implementation. Consequently, a cost model that assigns one universal price to multiplication does not represent all Cortex-M0+ devices. For signed Q15 operands, the full product fits within 31 significant bits, so MULS followed by ASRS #15 can implement the usual arithmetic-shift floor scaling without needing a 64-bit product.
Constant multiplication can instead be expressed as shifts plus additions or subtractions. Canonical signed-digit encoding minimizes the number of nonzero signed digits, but it does not necessarily produce the shortest possible addition chain; adder-graph methods can reuse intermediate expressions and sometimes improve on direct CSD expansion. GCC explicitly distinguishes Cortex-M0/M0+/M1 targets with the small iterative multiplier and may replace multiplication with add/subtract/shift sequences. These findings support treating fast-multiplier, slow-multiplier, and constant-specific shift-add costs separately when comparing Runge-Kutta cells: stage count alone does not determine cycles, because coefficient encodings, reused partial products, constant materialization, scaling shifts, and register spills can materially change the arithmetic cost.
sources:
Runge, Kutta design is naturally a constrained optimization problem: the Butcher coefficients must satisfy polynomial order conditions, while remaining degrees of freedom can optimize leading truncation-error norm, stability, SSP behavior, or storage. RK-Opt supports searches indexed by stage count and order, including minimization of the principal error norm. Its documentation stresses that coefficient optimization is highly nonconvex, so multistart searches can converge to different local solutions and do not automatically certify global optimality. Classical constructions likewise use free parameters and numerical searches to reduce error coefficients; Dormand and Prince’s search over a parameterized family is a documented example.
Extra stages can therefore improve a method even without raising its formal order, but neither order nor stage count alone determines realized error. Higher order suppresses truncation error asymptotically, whereas principal-error optimization addresses only the leading exact-arithmetic term. Internal-error research shows that roundoff or inexact stage computations can be amplified and that amplification depends strongly on the implementation. Consequently, the literature makes an order-4 five-stage advantage over order-2 or order-3 cells plausible in a truncation-dominated regime, but it does not establish H-682 or H-684 through H-687 under Q15 floor rounding and fixed cycle budgets; those comparisons require objectives or validation that explicitly include quantized coefficients, floor-rounded stage arithmetic, implementation form, and the applicable cost model.
sources:
For an explicit Runge, Kutta method, linear stability is governed by a polynomial whose earliest coefficients are fixed by the requested order and whose remaining coefficients can enlarge the stability region. Recent work formulates method selection as maximizing the admissible step for the actual eigenvalue spectrum. Consequently, stage count and order alone do not establish superiority: a five-stage fourth-order method has fewer free stability-polynomial coefficients than a six-stage second-order method, while a four-stage third-order method can devote its extra coefficient to stability. Runge, Kutta, Chebyshev results show the limiting alternative for diffusion-like spectra: low-order stabilized explicit methods can obtain a negative-real stability interval proportional to the square of the stage count, although their applicability assumes spectra concentrated near the negative real axis. Internal amplification of roundoff is a separate constraint and becomes important in low precision.
Under a fixed cycle budget, the relevant comparison is stable simulated time and achieved error per cycle, not error per step. Extra stages pay when their larger stable step or higher-order accuracy compensates for fewer completed steps; optimized stability must also match whether the problem spectrum is real-negative or substantially complex. Implicit DIRK and ESDIRK methods provide stiffly accurate, L-stable formulas, but each implicit stage requires solving an algebraic system. They pay for themselves only when explicit stability restrictions would otherwise require enough additional steps to exceed those solve costs. That crossover moves against implicit methods on Cortex-M0+, where hardware division is absent and multiplication latency is implementation-dependent; fixed-point nonlinear solves also require convergence tests, scaling, and safeguards. For the stated comparisons, the literature therefore supports measuring spectrum-specific stability extent, internal roundoff amplification under Q15 floor rounding, and total cycles per accepted unit of simulated time alongside held-out error.
sources:
Runge, Kutta error in low precision contains both discretization error and propagated arithmetic perturbations. General RK analysis shows that roundoff introduced at successive steps can accumulate globally, while reduced-precision analyses give a deterministic roundoff contribution proportional to u/h in addition to the method term h^p. Consequently, decreasing the step size or increasing formal order does not necessarily reduce total error once Q15 quantization dominates. Floor rounding is directed and introduces a non-positive error at every nonexact quantization, so its errors cannot be modeled as unbiased independent noise; repeated stage multiplications and state updates can create systematic drift.
Experiments with fixed-point neural ODE solvers found stochastic rounding more accurate than round-to-nearest fixed point and even binary32 for the studied neuron models and solvers. That work used a fixed-step, second-order Runge, Kutta variant optimized for fixed-point execution and reported that a small random residual was sufficient for stochastic rounding. Related Runge, Kutta heat-equation analysis found that implementation details affect error accumulation and that stochastic rounding can avoid stagnation, although random-number generation adds work and conflicts with a tightly deterministic cycle budget.
For Cortex-M implementations, Q15 stores 15 fractional bits in a signed 16-bit value and Q31 uses a signed 32-bit representation. Arm CMSIS-DSP supplies wider intermediate storage and both saturating and nonsaturating operations, making accumulator width, rescaling points, and saturation behavior part of the numerical method. A five- or six-stage RK cell has more potential quantization sites than a two-stage cell, but the resulting error depends on where values are narrowed and whether products are accumulated at wider precision. Thus formal order and stage count alone do not determine the comparisons in H-668, H-669, H-671, H-672, or H-673; they require evaluation with the exact floor-rounding schedule, scaling, overflow policy, and target instruction costs.
sources:
SciPy’s default solve_ivp method, RK45, stores the state as a one-dimensional NumPy array and the stage derivatives in a contiguous matrix shaped (stages + 1, state dimension). Its embedded error vector is divided componentwise by atol + rtol·max(|y_old|, |y_new|), then reduced with an RMS norm. Defaults are rtol=1e-3 and atol=1e-6. The proportional step controller uses safety 0.9, permits factors from 0.2 to 10, and prevents immediate growth following a rejected step. Boost.Odeint instead templates its steppers over the state container, algebra, operations, resizer, error checker, and step adjuster. Its default checker scales each component by eps_abs + eps_rel·(a_x|x| + a_dxdt|dt·dx/dt|), with eps_abs=eps_rel=1e-6 and both scale coefficients equal to 1; controlled Runge, Kutta steppers own reusable derivative, output, and error buffers and reduce the scaled error through the selected algebra.
CVODE is not a Runge, Kutta implementation: it uses variable-order Adams, Moulton formulas for nonstiff systems or BDF formulas for stiff systems, together with solution-history vectors, predictor-corrector data, and abstract N_Vector storage supporting multiple backends. It applies a weighted RMS norm with weights 1/(rtol·|y_i|+atol_i). Its controller jointly selects step size and order, begins at order 1, uses a 1/6 safety factor in the local-error step calculation, normally caps growth at 10, and retries rejected steps with additional failure-dependent limits. These libraries therefore demonstrate that production adaptive solvers incur costs beyond tableau stages: full-vector stage or history storage, component scaling, squared-error accumulation, square roots, controller powers, rejection paths, and sometimes nonlinear solves. Their designs do not establish any of H-662 through H-666, which compare fixed-cycle, fixed-point archive cells; they instead identify operations that should be excluded, fixed, or explicitly charged when relating floating-point library behavior to a Q15 Cortex-M0+ cost model.
sources:
Bogacki, Shampine 3(2) uses four stages and embeds a second-order estimate in a third-order result; its FSAL structure can reduce subsequent accepted steps to three new right-hand-side evaluations. Dormand, Prince 5(4) was constructed as a higher-order embedded family with shared stages and favorable principal truncation errors. These results show that formal order alone does not determine cycle cost: stage count, coefficient arithmetic, retained-stage reuse, rejected steps, and the target processor’s instruction costs all matter. They therefore do not establish either proposed ordering between the archived order-4/five-stage and order-2/four-stage cells, although they support separately measuring evaluations and coefficient operations under each cost model.
sources:
ODE error analyses predict a crossover as the step size decreases: discretization error falls with the formal order, while accumulated arithmetic error grows because more steps are taken. Kehlet and Logg derive a worst-case roundoff contribution proportional to the inverse step size and, under independent zero-mean errors, an expected contribution proportional to the inverse square root of the step size. Their residual-and-dual formulation also shows that problem-dependent stability factors govern amplification. Mosbach and Turner similarly model deterministic Euler and RK4 trajectories probabilistically and show that unstable structures such as saddle points can amplify tiny arithmetic perturbations into distributions of long-run trajectories.
These probabilistic cancellation results do not directly describe Q15 floor rounding. Floor is directed rounding and therefore introduces negative bias rather than zero-mean perturbations; the assumptions behind square-root accumulation and stochastic-rounding backward-error bounds are consequently absent. Fixed-point ODE experiments report the characteristic error minimum where truncation and quantization balance, along with accumulated errors from truncated constants and operations. Backward error analysis interprets a numerical trajectory as the exact solution of a modified differential equation, but arithmetic quantization adds implementation-dependent residuals to that modification. Thus, a fourth-order five-stage method can retain an error advantage over lower-order cells only while its smaller discretization error exceeds neither the Q15 quantization floor nor systematic floor-rounding drift; formal order alone does not establish the comparisons in H-647 and H-649 through H-651, and these error analyses provide no evidence for the cycle-cost claim H-648.
sources:
Recent reduced-precision Runge, Kutta literature shows that formal order alone does not predict realized accuracy or cost. Naive low-precision implementations can lose convergence as the step size decreases, while compensated accumulation and selective higher-precision operations can recover accuracy. Mixed-precision stabilized RK methods further demonstrate that stages serving stability and stages serving accuracy need not use identical arithmetic. For a Q15 implementation with mandatory floor rounding, these results support evaluating each complete arithmetic schedule: floor is biased rather than unbiased, and mitigation operations would consume part of the fixed cycle budget.
Hard-real-time literature continues to favor fixed-step methods because they provide a constant operation count and analyzable worst-case execution time; adaptive schemes introduce variable work unless adaptation and rejection are explicitly bounded. Research on FPGA hardware-in-the-loop simulation likewise evaluates integration methods jointly by stability, model bandwidth, accuracy, and execution time. Consequently, equal-stage RK formulas can have different measured costs because coefficient representation, multiplication, accumulation, reuse, and storage differ; higher formal order does not by itself establish lower cost or lower finite-precision error.
Energy-preserving collocation methods offer a different objective from minimizing trajectory error: for port-Hamiltonian systems they can impose an exact discrete energy-balance relation and retain high formal accuracy, with reported advantages at larger sampling intervals. These methods are generally implicit and therefore bring solve work that is difficult to reconcile with a small fixed cycle allowance. The literature nevertheless suggests adding invariant or energy-drift measurements to held-out evaluation, since a conventional explicit RK method may have small short-horizon state error while introducing undesirable long-horizon energy behavior.
sources:
47 older digests are not shown here. The whole log is in rk-work/literature/digests.jsonl.