How Much of the Evidence for a Closed Universe Comes From the Vacuum and the Prior?

Some analyses of the cosmic microwave background (CMB), the relic radiation from the early universe, report strong Bayesian odds that space is slightly curved and closed rather than flat. One such analysis finds odds above 100:1 under a model of "just-enough" inflation. Inflation is the early burst of expansion that seeded cosmic structure. The preference comes mostly from the largest angular scales, where the measured sky has less power than the standard model predicts. A model of the very early universe can produce exactly this kind of large-scale deficit and ringing pattern in the primordial power spectrum, the initial distribution of fluctuation strength across length scales. It does so through two ingredients: a short pre-inflationary phase in which the inflaton field's kinetic energy dominates, and spatial curvature. Two choices sit underneath that spectrum, and the data cannot decide either of them. The first is the quantum "vacuum", the initial state assigned to each fluctuation mode. The second is the set of prior distributions placed on curvature, on the duration of inflation, and on a leftover free parameter in the most principled curved-space vacuum.

This note proposes to measure how much of the closed-universe evidence survives when those two choices are varied in a controlled way. The main tool is a cheap, validated estimator that reweights the output of nested sampling, the standard Monte Carlo method for computing Bayesian evidence, so that each prior change does not require a new multi-core-year run. In two sentences: we take the curved-inflation pipeline that produced the >100:1 odds, swap in the coordinate-independent vacuum, and map how the Bayes factor and the posteriors move as the priors change. Nested sampling's discarded "dead points" run from the prior bulk to the posterior peak, which makes this sensitivity map cheap and reliable to compute.

Core Claim

The vacuum literature has moved from Hamiltonian-based prescriptions to minimizing the renormalized stress-energy tensor (RSET). Hamiltonian prescriptions give different physical states depending on how the perturbation variable is written (Quantum initial conditions for inflation and canonical invariance). Minimizing the RSET, in practice the local energy density, does not have this problem (Novel quantum initial conditions for inflation). In curved space this requires a new perturbation variable, $\zeta$, whose equation keeps simple-harmonic-oscillator form. That variable is unique only up to one free scalar parameter, which the authors suggest could be fitted to data (Quantum initial conditions for curved inflating universes).

Our claim is that this parameter, once given a prior, reintroduces through the back door the ambiguity that invariance was meant to remove. This becomes testable once the curved RSET vacuum is put inside the full inference pipeline of Finite inflation in curved space. That paper reported the >100:1 odds and $\Omega_{K,0} = -0.031 \pm 0.014$ for curved Starobinsky inflation, with mode initial conditions not set through the curved $\zeta$ construction. Combining this with prior sensitivity analysis lets us state whether the curvature evidence is a property of the data or of the modelling choices.

The methodological contribution makes the study affordable. Each dead point $i$ of a nested sampling run is drawn from the prior restricted to likelihoods above the previous contour. Given its likelihood $L_i$ and estimated prior-volume shell $\Delta X_i$, the evidence under any alternative prior $\pi'$ with the same support is estimated as

$$ \mathcal{Z}' \approx \sum_i L_i \, \Delta X_i \, \frac{\pi'(\theta_i)}{\pi(\theta_i)}, $$

where $\pi$ is the original prior and $\theta_i$ the dead point's parameters. Posterior samples alone cover only the high-likelihood region. Dead points also cover the low-likelihood prior bulk, so reweighting can handle location shifts and changes in prior width. The power-scaling diagnostic of Detecting and diagnosing prior and likelihood sensitivity with power-scaling explicitly cannot do location shifts and was never connected to nested sampling or evidence.

Research Program

The first component is the forward model. We implement the curved RSET vacuum on $\zeta$ inside the primpy/oscode primordial solver used in Finite inflation in curved space. The free scalar of $\zeta$, which we call $c$, is exposed as a sampled parameter. Initial conditions are set at the kinetic-dominance start using the logolinear expansions of Logolinear series expansions with applications to primordial cosmology. The resulting spectra feed a standard Boltzmann code.

The second component is a set of reference runs. We run PolyChord (polychord: next-generation nested sampling) within Cobaya on Planck 2018 TT,TE,EE+lowE plus BK15, matching the earlier finite-inflation analysis. There are four model arms, all crossed with flat versus curved geometry:

  1. Bunch–Davies or Hamiltonian-style vacuum, which reproduces the published setup.
  2. Curved RSET with $c$ fixed to its flat-limit-consistent value.
  3. Curved RSET with $c$ free.
  4. Curved $\Lambda$CDM with a power-law spectrum, as in Curvature tension: Evidence for a closed universe.

The third component is the sensitivity map. From each run's dead points we compute $\log \mathcal{Z}'$ and posterior shifts for a family of alternative priors. The family includes rescaled and shifted bounds on $\Omega_K$, uniform versus logarithmic priors on the pre-pivot duration $N_\dagger$ and on $c$, and power-scaled priors at several exponents. Each reweighting is diagnosed with the Pareto shape $\hat{k}$ of its importance ratios and with the cumulative Jensen–Shannon distance, using the $D_{\text{CJS}} \ge 0.05$ threshold from Detecting and diagnosing prior and likelihood sensitivity with power-scaling. Run-to-run uncertainty is bootstrapped with nestcheck threads (nestcheck: diagnostic tests for nested sampling calculations).

We also compute the Bayesian model dimensionality $d_G$ and the information gain per parameter, which tell us whether the data constrain $c$ at all. If $c$ contributes essentially no constrained dimension, its prior can only add an Occam penalty or nothing. In that case $c$ should be reported as a theoretical choice, not a fitted parameter.

The fourth component addresses discordance. The CMB-only preference for closure conflicts with combined probes (unimpeded: A Public Grid of Nested Sampling Chains for Cosmological Model Comparison and Tension Analysis; Curvature tension: Evidence for a closed universe). We therefore add BOSS DR12 BAO and Planck lensing, and compute suspiciousness $S$ between Planck and BAO under each vacuum arm. This asks whether the physically motivated spectrum changes the 2.5–3σ curvature tension or leaves it intact.

Baselines

The naive reference point is flat $\Lambda$CDM with a power-law spectrum and Bunch–Davies vacuum, run on identical likelihoods and nested-sampling settings. For combined-probe configurations, the public unimpeded grid supplies chains to cross-check our own runs, so every Bayes factor has a common denominator.

The strongest existing competitor on this question is the finite curved inflation analysis of Finite inflation in curved space. It reports Bayesian odds above 100:1 for a closed universe from CMB data, with $\Omega_{K,0} = -0.051 \pm 0.017$ in curved $\Lambda$CDM and $-0.031 \pm 0.014$ in curved Starobinsky inflation. We first reproduce these numbers within nested-sampling error using their vacuum and priors. Only after that reproduction do we change the vacuum, so differences cannot be blamed on pipeline drift. Compute is matched: identical live-point counts and convergence criteria across arms.

"Beating" this baseline here does not mean a larger Bayes factor. It means a robustness statement the baseline cannot make. Concretely, we report the range of log odds over a stated, physically defensible prior family, each value with a reweighting diagnostic, plus the change in odds attributable to the vacuum alone.

A secondary empirical reference is the quantized spectrum of Improved cosmological fits with quantized primordial power spectra, which reports $\Delta\chi^2 = -8.55$ against $\Lambda$CDM on Planck 2018. That figure is a best-fit improvement, not an evidence. We will compare our best-fit $\Delta\chi^2$ to it directly, and our evidences to the $\Lambda$CDM baseline.

Two ablations isolate the key ingredients. Arm 2 against arm 3 isolates the free parameter $c$. Arm 1 against arm 2 isolates the switch to the covariant vacuum. Brute-force PolyChord reruns at a handful of alternative priors serve as the ground-truth ablation for the reweighting estimator itself.

Why not just…?

Why not just rerun nested sampling for a few prior widths? This is the honest gold standard, and we use it as the validator. As the main tool it does not scale. The curvature-tension analysis alone used about 12 single-core years (Curvature tension: Evidence for a closed universe), and a joint prior family over $\Omega_K$, $N_\dagger$ and $c$ needs dozens of variants per vacuum arm. If reweighting fails its diagnostics for the variants we care about, the threat is real. In that case we fall back to reruns over a reduced family and say so.

Why not just use suspiciousness, which is already prior-independent? Suspiciousness removes prior dependence from a tension statistic between two datasets. The >100:1 figure, by contrast, is a model-comparison Bayes factor on one dataset, and that is prior-dependent by construction. Suspiciousness is the right tool for our discordance component, but it cannot say whether closed-versus-flat odds survive a change in the $\Omega_K$ prior.

Why not just use posterior power-scaling via priorsense? It perturbs only posterior shape through samples near the peak. It cannot perform location shifts, as its authors note, and it says nothing about evidence. Our experiment compares the two directly by running power-scaling on resampled posterior points alongside dead-point reweighting, against brute-force reruns. The comparison is decided by which estimator tracks the reruns.

Why not just replace the numerical vacuum with analytic oscillation templates? The templates of Analytical approximations for curved primordial power spectra could speed up likelihood calls. However, their accuracy for the RSET-on-$\zeta$ vacuum with free $c$ is untested. We check template against numerical spectra on a grid and adopt the template only where they agree within cosmic variance at $\ell < 30$. No reported accuracy figure was found on Lacuna for this case.

Proposed Experiments

The first decisive check validates the estimator before it touches real data. On the Gaussian, twin-peaks and Rosenbrock benchmarks used for PolyChord, and on synthetic CMB low-$\ell$ spectra injected with a known curved RSET truth, reweighted $\log \mathcal{Z}'$ must agree with brute-force reruns within nestcheck bootstrap error whenever $\hat{k} < 0.7$. We also expect, and will report, failures when the alternative prior moves mass outside the region the dead points reach.

The second check uses real data. On Planck 2018 plus BK15, we report the log Bayes factor for closed versus flat in each vacuum arm, together with its spread across the prior family. If the odds fall below roughly 10:1 under a defensible prior variant, or move by more than the nested-sampling error when only the vacuum changes, the published preference is shown to be choice-driven. If they stay above 100:1 throughout, the closed-universe signal is shown to be robust in a way it has not been before.

The third check asks whether the data constrain $c$, measured by $d_G$ and $D_{\text{CJS}}$. The answer settles whether the leftover freedom of the covariant vacuum is observationally meaningful or must be fixed by theory.

The fourth check computes suspiciousness between Planck and BAO under each arm. It shows whether a physically derived early-universe spectrum relieves or preserves the curvature discordance.