Cosmology has two famous disagreements. The universe's present expansion rate, the Hubble constant $H_0$, comes out near 67 km/s/Mpc when inferred from early-universe light and near 73 km/s/Mpc when measured directly nearby. The clumpiness of matter today, summarized by a number called $S_8$, comes out lower in weak-lensing surveys than the early universe predicts. Weak lensing maps matter by the slight distortions it imprints on the shapes of background galaxies. These two problems are usually studied with separate pipelines. Yet in standard gravity they are tied together by a single piece of dynamics: the expansion rate acts as a friction term that slows the growth of matter clumps. If the hints of evolving dark energy seen in recent supernova and galaxy-survey data are real, they should also leave a predictable mark on $S_8$.
We propose to test that link without assuming any particular form for dark energy. The approach borrows from machine-learning work on uncertain dynamical systems. We place a Gaussian process (a probability distribution over smooth functions, which returns both a best guess and an error bar at every point) on the dark-energy density as a function of cosmic time. We then push every sampled history through the exact equations of expansion and of matter growth. A second Gaussian process, kept separate, captures any departure from standard gravity in the growth equation. The output answers two questions: whether some smooth dark-energy history consistent with the expansion data, plus ordinary gravity, can produce the low $S_8$ seen by lensing; and, if not, whether the data point to a change in the expansion history or in gravity itself.
Core Claim
The current evidence is split across incompatible representations. Flexible reconstructions of the dark-energy equation of state $w(a)$, where $a$ is the cosmic scale factor, use piecewise-linear "flexknots" fit only to distance data. In Non-parametric reconstructions of dynamical dark energy via flexknots, this approach found a W-shaped $w(a)$ and a localized clash between DESI galaxy-survey data and DES five-year supernovae near $a \approx 0.75$. Comparison of dynamical dark energy with ΛCDM in light of DESI DR2 found a Bayes factor of about 2.3 for the two-parameter CPL model over standard cosmology on DESI DR2 plus DES5Y. Neither paper asks what those histories imply for structure growth. The only Gaussian-process work in the file bearing directly on the tension, Kernel-, mean-, and noise-marginalized Gaussian processes for exoplanet transits and H0 inference, fits $H(z)$ as a curve against redshift $z$ and averages over kernel choices. It encodes no dynamics and is silent on $S_8$.
Our claim is that a Gaussian process with physical structure built in can turn the background data into an honest, parametrization-free prediction for $S_8$. Here "physical structure" means a prior on a physical quantity whose derivatives and integrals enter the equations exactly. That prediction can be checked against the Dark Energy Survey Year 3 value of $S_8 = 0.776 \pm 0.017$ reported in Dark Energy Survey Year 3 results: Cosmological constraints from galaxy clustering and weak lensing. Splitting the model into a nominal physics part and a learned residual tells us which equation, expansion or growth, has to absorb the discrepancy.
Model
We write the expansion rate with a Gaussian process on the logarithm of the dark-energy density:
$$H^2(a) = H_0^2\left[\Omega_m a^{-3} + (1-\Omega_m)\, e^{g(a)}\right], \qquad g \sim \mathcal{GP}(0, k_\theta), \; g(1)=0.$$
Here $\Omega_m$ is today's matter fraction and $k_\theta$ a kernel with hyperparameters $\theta$. The equation of state follows as $w(a) = -1 - \tfrac{1}{3}\, dg/d\ln a$. Because differentiation is linear, $w$ is jointly Gaussian with $g$. This is the same closure property that Learning Energy Conserving Dynamics Efficiently with Hamiltonian Gaussian Processes uses to obtain a conservative vector field from a prior on energy. We borrow the mechanism, not the Hamiltonian structure, since an expanding universe is not a closed conservative system.
The linear growth of the matter contrast $\delta$ then obeys
$$\delta'' + \left(2 + \frac{d\ln H}{d\ln a}\right)\delta' = \tfrac{3}{2}\,\mu(a)\,\Omega_m(a)\,\delta, \qquad \mu = 1 + r(a), \; r \sim \mathcal{GP}(0, k_\phi).$$
Primes denote derivatives with respect to $\ln a$, $\Omega_m(a)$ is the matter fraction at scale factor $a$, and $\mu$ is the effective strength of gravity relative to Newton's constant.
This "nominal physics plus Gaussian-process residual" split is the structure that Safe Model-based Reinforcement Learning with Stability Guarantees uses for robot dynamics. Keeping a separate latent term for each part of a coupled system follows Learning interacting dynamical systems with latent Gaussian process ODEs, which separated autonomous motion from interaction forces. Here the two parts are background expansion ($g$) and the gravitational coupling ($r$).
Two modelling assumptions fix the overall growth amplitude. First, the growth amplitude is anchored at $z \gtrsim 3$ by the early-universe fluctuation amplitude. Second, the dark-energy prior mean reverts to a cosmological constant at high redshift. Both are stated explicitly and stress-tested.
Research Program
Both processes are represented with a reduced-rank basis expansion of about 20 to 40 basis functions, in the style of Svensson et al.. This keeps each likelihood evaluation to one deterministic ODE solve. Uncertainty is then propagated by full posterior sampling rather than by linearization or moment matching. The problem is one-dimensional in time, so sampling is cheap and exact, and the motivation for analytic moment matching in Gaussian Processes for Data-Efficient Learning in Robotics and Control largely disappears.
Following the transdimensional method of the kernel-marginalization paper, we run PolyChord nested sampling with a categorical kernel index. The index switches among squared-exponential, Matérn, and the non-stationary linear kernel that paper found favored on sparse chronometer data. The sampler marginalizes over the kernel index, the hyperparameters, the basis coefficients, and $\Omega_m$, $H_0$, and the sound horizon $r_d$. It also marginalizes over the noise model, with homoscedastic and heteroscedastic options and an optional per-dataset error inflation, so that apparent tensions are not artifacts of fixed error assumptions. Bayesian evidence for the variants with $r \equiv 0$ versus $r$ free quantifies whether gravity must change.
The data match those already used in the file. The supernova data are Pantheon, Pantheon+, and DES-SN5YR. The galaxy-survey distance data are 6dF, BOSS DR12, eBOSS DR14, and DESI DR1 and DR2. We also use cosmic chronometers, CMB shift parameters and the $r_d$ priors from Planck and WMAP, and the compressed DES Y3 $S_8$–$\Omega_m$ posterior. Growth data enter only in the final predictive check, not in the background fit.
Baselines
The naive reference point is flat ΛCDM with ordinary gravity, fit to the same background data, with $S_8$ computed from the same growth equation. The file gives two of its outcomes. The DES Y3 analysis reports that the DES-only $S_8$ of $0.776 \pm 0.017$ moves to $0.812 \pm 0.008$ once Planck, galaxy-survey distances, and supernovae are added, with tension significance $p = 0.13$ to $0.48$. Through the inverse distance ladder, Model independent H(z) reconstruction using the cosmic inverse distance ladder reports $H_0 = 68.42 \pm 0.88$ km/s/Mpc. Our pipeline must reproduce both numbers when $g$ and $r$ are switched off. This is a correctness gate rather than a target.
The strongest parametric competitor is CPL, $w = w_0 + w_a(1-a)$. It is the model most favored on DESI DR2 plus DES5Y in the DESI DR2 comparison paper, with a Bayes factor of about 2.3 over ΛCDM. We run it through the identical growth equation, priors on shared parameters, nested-sampling settings, and live-point budget. No reported result was found on Lacuna for the $S_8$ that CPL implies on these combinations, so we compute it.
The strongest flexible competitor is the flexknot reconstruction of the flexknots paper, with 1 to 10 knots, integrated through the same growth equation. It is the method that actually exposed the localized clash near $a \approx 0.75$. Our model beats it if it gives a tighter or better-calibrated $S_8$ prediction on mock data at equal or better evidence. A smaller but still meaningful win is that its conclusions stay stable when the prior is changed, whereas knot-count priors are known to sway flexknot evidence.
The closest Gaussian-process baseline is the kernel-marginalized curve fit of $H(z)$. It reports $H_0 = 69 \pm 6$ km/s/Mpc on chronometers plus galaxy-survey distances and a consistency statistic of $\ln R = 12.17 \pm 0.02$. We rerun it on our data and differentiate its $H(z)$ to drive the growth equation. This isolates the value of putting the prior on a physical quantity rather than on the observable curve.
Two ablations isolate our ingredients. The first fixes $r \equiv 0$, which tests whether the gravity residual is needed. The second fixes a single kernel and a single noise model, which tests whether marginalization changes the tension verdicts.
Why not just…?
Why not just use CPL and compute $S_8$ from it? It is cheap, standard, and may give the same answer. The flexknots paper argues that CPL and wCDM are too restrictive and that they hide the localized DESI–DES5Y clash. A history with a dip near $a \approx 0.75$ alters the friction term in the growth equation exactly when growth matters most for $S_8$. If the posterior predictive $S_8$ from the Gaussian process matches CPL's to within the error bars on real data, we will report that the $S_8$ conclusion does not depend on the parametrization. That is a useful result in itself, and it would demote the method to a robustness check.
Why not just push flexknots through the growth equation? This is the real threat. Nothing about piecewise-linear $w$ prevents integrating growth. What the Gaussian process adds is a differentiable $w$ that is exactly tied to the density through the closure property, marginalization over kernel and noise class, and a separate gravity term. Whether these matter is an empirical question. The mock-data coverage test settles it: we inject smooth histories, kinked histories, and modified-gravity histories, and compare how often each method's 68% intervals on $S_8$ and $w(a)$ contain the truth. If flexknots match the Gaussian process in coverage and in the real-data verdicts, we will say so.
Why not just apply existing tension statistics to ΛCDM posteriors? The suspiciousness and $R$ statistics, made easy by the public chains in unimpeded and calibrated in Calibrating Bayesian tension statistics using neural ratio estimators, measure whether datasets disagree inside one model. They cannot say whether some other expansion history reconciles distances with lensing, or whether the fix must sit in gravity. We use these statistics as outputs, not substitutes.
Why not just use a neural ODE? The latent function is one-dimensional, so scaling is not the bottleneck and calibrated uncertainty is the whole point. The Hamiltonian Gaussian process paper reports that neural ODEs underperform on probabilistic error measures. We include a deep-ensemble neural ODE only as a mock-data calibration check, and we do not treat it as a serious contender.
Proposed Experiments
The first experiment is synthetic. We generate mock catalogs with the Pantheon+, DESI DR2, and DES Y3 error budgets from four truths: ΛCDM, CPL, a W-shaped history, and a constant-$w$ history with $\mu = 0.9$ at low redshift. For each method we measure interval coverage, posterior predictive log-likelihood of held-out $S_8$, and whether the evidence correctly flags the need for a nonzero $r$.
The decisive real-data check is a posterior predictive test. We fit only background data in two combinations, DESI DR2 plus Pantheon+ and DESI DR2 plus DES5Y, with $r \equiv 0$. We then compare the predicted $S_8$ distribution with the DES Y3 measurement of $0.776 \pm 0.017$.
The outcomes are interpretable either way. If the prediction overlaps DES Y3 for some kernel-marginalized dark-energy history, the $S_8$ tension can be relieved by late-time expansion alone under ordinary gravity. If the prediction stays near the combined value of $0.812$ while evidence rises sharply once $r$ is freed, the data point to a growth-sector change or to lensing systematics. The baryonic suppression studied by Yuan et al. 2021 is the leading systematic candidate, and we test it by restricting lensing to large scales.
The second real-data check asks whether the clash near $a \approx 0.75$ survives noise-class marginalization. We report the $R$ statistic per dataset pair under single-noise and marginalized-noise fits.
Finally, the inverse distance ladder result gives a sanity bound on $H_0$. If $H_0$ remains near 68 km/s/Mpc under every late-time Gaussian-process history, this independently confirms that the $H_0$ problem cannot be fixed this way. In that case the $S_8$ question is the one late-time dynamics can still answer.
References
- Non-parametric reconstructions of dynamical dark energy via flexknots
- Comparison of dynamical dark energy with ΛCDM in light of DESI DR2
- Kernel-, mean-, and noise-marginalized Gaussian processes for exoplanet transits andH0 inference
- Dark Energy Survey Year 3 results: Cosmological constraints from galaxy clustering and weak lensing
- Learning Energy Conserving Dynamics Efficiently with Hamiltonian Gaussian Processes
- Safe Model-based Reinforcement Learning with Stability Guarantees
- Learning interacting dynamical systems with latent Gaussian process ODEs
- Computationally Efficient Bayesian Learning of Gaussian Process State Space Models.
- Gaussian Processes for Data-Efficient Learning in Robotics and Control
- Model independentH(z) reconstruction using the cosmic inverse distance ladder
- $\texttt{unimpeded}$: A Public Grid of Nested Sampling Chains for Cosmological Model Comparison and Tension Analysis
- Calibrating Bayesian tension statistics using neural ratio estimators
- AbacusHOD: A highly efficient extended multi-tracer HOD framework and its application to BOSS and eBOSS data