Several recent cosmology papers argue that the largest-scale patterns in the cosmic microwave background (CMB), the relic radiation from the early universe, carry a physical signature. On these scales the CMB shows less variation than the standard model predicts, and it may also show faint ringing. The proposed cause is a brief early phase in which the field driving inflation moved fast, before the long, gentle expansion began. The evidence for this story rests on a small number of long-wavelength modes, each computed by a numerical solver. That solver is then called millions of times inside a Bayesian sampler. Nested sampling, the sampler these papers use, keeps whichever points score highest. So a single silently wrong solver output that happens to score well can be retained and can shift the evidence. The source could be a hardware bit-flip or a numerical failure. The hardware fault-tolerance literature has shown how to catch such errors cheaply, but only for neural networks and matrix codes. Nobody has asked how large a fault rate a cosmological claim can survive.
This note proposes a cheap error detector built into one fast solver. The idea is that the solver's own physics supplies free checksums, so no computation needs to be duplicated. The target is the transfer-matrix Mukhanov–Sasaki solver of Rapid numerical solutions for the Mukhanov-Sasaki equation, which computes how each primordial fluctuation mode evolves. Every step of that solver must preserve an exact conserved quantity. The solver also already computes each segment twice while deciding how finely to subdivide time, and that duplicate can serve as a redundancy check at no extra cost. We use these two properties to measure the "fault budget" of published large-scale-suppression results: the corruption rate at which a Bayesian conclusion such as a 100:1 preference for a closed universe would flip. We also use them to guard those results at near-zero overhead.
Core Claim
The Mukhanov–Sasaki equation governs how one primordial fluctuation mode evolves. It can be written as a time-dependent oscillator, $\ddot x + \omega^2(t)x = 0$. Here $x$ is the rescaled mode amplitude and $\omega(t)$ is a frequency set by the wavenumber $k$ and the background expansion. The solver of Rapid numerical solutions for the Mukhanov-Sasaki equation splits time into segments. On each segment it approximates $\omega^2$ by a linear or exponential piece and solves that piece exactly with Airy or Bessel functions. It then multiplies the resulting $2\times 2$ matrices $T_j$ along the chain. It reports under 0.1% fractional error and constant runtime at high $k$.
Because the equation has no friction term, each exact segment map preserves phase-space area. This gives two exact identities:
$$ \det T_j = 1, \qquad W = x\,\dot x^{*} - x^{*}\dot x = i , $$
Here $W$ is the Wronskian of the complex mode function, whose value is fixed by the quantum normalization of the initial state. The first identity checks every segment matrix at the cost of a few flops. The second checks the whole chain once per mode. A bit-flip in any single matrix entry changes the determinant by the flipped amount times that entry's cofactor. It is therefore visible unless it is too small to matter.
These identities are blind to one class of fault. A corrupted $\omega^2$ value produces a perfectly area-preserving matrix that describes the wrong physics. For this case we reuse the solver's adaptive bisection, which already compares one segment against its two halves. Our change is to also flag a mismatch whenever it is inconsistent with the local smoothness of $\omega^2$. This follows the logic of Piggyback-DMR in SpareTrain, which turns an existing recomputation (the activation-checkpoint replay) into the redundant verification pass.
The claim has two parts. First, these checks should detect nearly all consequential faults in the power-spectrum stage at a few percent overhead. Second, detection matters: unprotected nested sampling should show evidence inflation from faults at rates far below those needed to perturb ordinary parameter means.
Research Program
Each prior line of work plays a different role here. The physics motivation comes from three papers. Finite inflation in curved space reports odds above 100:1 for a closed universe and $\Omega_{K,0} = -0.051 \pm 0.017$ from Planck 2018, using primpy, oscode, and PolyChord inside Cobaya. Constraining the kinetically dominated universe finds that kinetic-dominance models tie standard slow-roll models in evidence. Improved cosmological fits with quantized primordial power spectra reports $\Delta\chi^2 = -8.55$ for a discretely quantized spectrum.
These claims are marginal, and that is what makes them useful here. They sit where a few corrupted low-$k$ evaluations could plausibly matter. None of these papers audits its pipeline for silent errors.
The method borrows from three hardware papers. SpareTrain supplies the idea of reusing an existing recomputation as the redundant check. 3D Coded SUMMA and ReaLM supply the broader stance that algebraic structure can replace duplication. ReaLM adds a specific lesson: check only when an error could change the outcome. We use this by setting check tolerances relative to how sensitive the likelihood is to each $k$-band.
The range-supervision work in FT-ClipAct and Geissler et al. constrains the design. Their success means a simple bounds check on the output spectrum is the cheapest serious competitor, so our design has to beat it on the faults bounds cannot see.
The work has three parts. We first implement the checks in the transfer-matrix solver for flat kinetic-dominance-plus-inflation models, where the solver is native.
We then extend the solver to closed geometries. Analytical approximations for curved primordial power spectra shows that curvature acts largely as a shift in the participating wavenumbers. We will test whether the curved equation keeps oscillator form after a variable change, since only then do the invariants carry over. If it does not, the curved study falls back to oscode with a Wronskian-only check.
Finally, we build a fault-injection harness around the full Cobaya/PolyChord pipeline. It flips bits at controlled rates in three places: the segment matrices, the $\omega^2$ tables, and the output $C_\ell$ vectors handed to the likelihood. It records which injected evaluations the sampler actually keeps. We cannot cite measured silent-error rates for CPU cosmology clusters (no reported result found on Lacuna for silent data corruption rates in cosmological pipelines). The study therefore sweeps the fault rate and reports the threshold at which each published conclusion breaks, rather than assuming a rate.
Baselines
The first reference point is doing nothing: the unprotected pipeline, run with the same injection schedule. It fixes the scale of the problem, namely how much evidence inflation or posterior shift appears at each fault rate.
The strongest competitor is full dual modular redundancy: every mode or every likelihood call is computed twice and discarded on disagreement. It catches every fault and needs no physics insight. SpareTrain reports that naive redundancy costs roughly 100% overhead, and that its piggybacked version brings this to 3–14% with 100% coverage on LLM training. We run duplication on the same machines and the same PolyChord settings (live points, seeds, prior). We compare wall-clock time per converged run and detection coverage. Beating it means matching its coverage of consequential faults at under 10% overhead.
The second competitor is output range supervision in the style of Geissler et al.. It rejects any power spectrum or $C_\ell$ vector that falls outside bounds profiled from fault-free runs. Their Clipper and Backflip methods reduced residual silent corruption from over 50% to at most 0.01 in CNNs. We tune its bounds on the same fault-free profiling budget we use to set our own tolerances.
Two ablations isolate our key ingredients. One uses only the determinant/Wronskian checks, which should miss $\omega^2$ faults. The other uses only the bisection-reuse check, which should miss faults in the matrix multiplications themselves. Together they show whether both components earn their place.
Why not just…?
Why not just duplicate every likelihood call? It works, and it is the coverage standard we compare against. For a weeks-long nested sampling run, though, doubling compute is exactly the tax the hardware literature has spent a decade avoiding. The comparison above measures the overhead directly. If the power-spectrum solver turns out to be a negligible fraction of each likelihood call compared with the Boltzmann code, the savings shrink. In that case we will report that the checks matter only in pipelines where the spectrum solve dominates, such as the quantized discrete-$k$ sums in the Bartlett et al. setup.
Why not just bound the output spectrum? This is the real threat. Exponent and high-order bit flips produce absurd spectra that any bound catches. If those are the only faults that shift posteriors, range supervision wins on simplicity. The case for our checks is the other kind of fault: mid-mantissa or $\omega^2$ corruptions that produce a plausible extra dip or ring at low $k$, which is exactly the feature these models are fitted to. We stratify injections by bit position and location and report, for each stratum, which detector catches the faults that actually changed the evidence. If bounding matches our coverage there, we will say so and recommend it.
Why not just run independent chains and compare evidences? Repeat runs catch faults that happen to differ between runs. They do not catch a fault class that persists across runs, such as a systematic numerical failure in the solver's segment switching. Repeats also cost at least double, and they only detect large discrepancies after the fact. We include repeated runs in the unprotected arm to measure how often run-to-run scatter would have flagged the injected bias.
Proposed Experiments
All inference uses Planck 2018 plik TTTEEE+lowl+lowE+lensing, the likelihood of Improved cosmological fits with quantized primordial power spectra. We reproduce three targets without injection before running any fault study. The first is the closed-versus-flat odds of Finite inflation in curved space, above 100:1. The second is the $\Delta\chi^2 = -8.55$ of the quantized spectrum. The third is the tie between kinetic-dominance and slow-roll evidence from Constraining the kinetically dominated universe, rerun on Planck 2018.
The first decisive check is detection. For each injection location and bit class, we measure the fraction of consequential faults caught by each detector. A consequential fault is one that raised the likelihood enough to enter the live-point set. We also measure the per-call overhead.
The second check is scientific. We sweep the injection rate and report the rate at which the unprotected pipeline flips each conclusion: the log-odds shift by more than one unit, or $N_\dagger$ and $\Omega_{K,0}$ move by more than half a standard deviation. We then confirm that the protected pipeline holds these quantities fixed up to rates at least two orders of magnitude higher.
We predict that evidence is far more fragile than posterior means, because a single spuriously high point enters the evidence sum directly. If instead the unprotected fault budget sits far above any plausible hardware rate and bounding catches everything that matters, that is a clean negative result worth publishing. It would tell the field that these low-$\ell$ claims are computationally sound as they stand.
References
- Rapid numerical solutions for the Mukhanov-Sasaki equation
- SpareTrain: Fault-Tolerant LLM Training via Low-Cost Dual Modular Redundancy
- Finite inflation in curved space
- Constraining the kinetically dominated universe
- Improved cosmological fits with quantized primordial power spectra
- 3D Coded SUMMA: Communication-Efficient and Robust Parallel Matrix Multiplication
- ReaLM: Reliable and Efficient Large Language Model Inference with Statistical Algorithm-Based Fault Tolerance
- FT-ClipAct: Resilience Analysis of Deep Neural Networks and Improving their Fault Tolerance using Clipped Activation
- Towards a Safety Case for Hardware Fault Tolerance in Convolutional Neural Networks Using Activation Range Supervision
- Analytical approximations for curved primordial power spectra