How Much of the Early-Universe Evidence Is the Prior? Vacuum Ambiguity and Prior Sensitivity at Low Multipoles

Several recent analyses of the cosmic microwave background (CMB), the relic radiation from the early universe, draw strong conclusions from its largest angular scales. They report odds above 100:1 that space is closed rather than flat, and they find that a pre-inflationary phase dominated by the inflaton's kinetic energy fits as well as standard inflation. These conclusions rest on a few dozen low multipoles, where each measured number carries an irreducible random scatter (cosmic variance) because there is only one sky. In that regime Bayesian answers can be decided by the prior, the probability assigned to parameters before seeing data. The theory of the initial quantum state adds a second, hidden source of choice. Even the canonically invariant "vacuum" prescription leaves some freedom: when to impose it, and one free constant in the curved-space perturbation variable. Each choice reshapes the predicted low-multipole spectrum.

This note proposes to treat all of these choices together and uniformly, as prior choices, and to measure how far the headline evidence ratios move under reasonable changes to them. The method runs nested sampling (an algorithm that computes both the posterior and the Bayesian evidence) once, under a deliberately broad reference prior that includes the vacuum's residual freedoms as sampled parameters. The evidence under any narrower or reshaped prior, including "fix the vacuum this way," is then recovered by importance reweighting or by the Savage–Dickey density ratio. A small number of brute-force refits checks the reweighted results. The output is a sensitivity map for each published conclusion: closed versus flat, kinetic dominance versus slow roll, and one vacuum versus another. The map says whether the data or the analyst's choices carry the conclusion.

Core Claim

The claim has two parts. The first is testable: we expect that at least one of the low-multipole conclusions in this literature changes by more than about one unit of log-evidence under priors a reasonable physicist could defend. Candidates are the parity between kinetic-dominance and slow-roll models in Constraining the kinetically dominated universe and the >100:1 closed-universe odds in Finite inflation in curved space. We also expect that the gap between competing vacuum prescriptions is smaller than this prior-induced spread. The second part is methodological. The residual freedoms of the covariant vacuum can be sampled as ordinary parameters. Doing so turns the open theoretical question of when to impose the vacuum into a measurable quantity: the Savage–Dickey Bayes factor for each imposition time, plus a statement of whether the data can discriminate imposition times at all.

What would be new is the combination. The vacuum-theory papers evaluate spectra at fixed best-fit points. The sampling papers fix the vacuum. The sensitivity-diagnostic literature has never been applied to evidence ratios for inflationary initial conditions.

Background and What It Constrains

The vacuum work supplies the physics we reuse. Quantum initial conditions for inflation and canonical invariance shows that two common ways of picking the initial state are not invariant under canonical transformations. These are Hamiltonian diagonalization, which picks the state that minimizes the instantaneous energy operator, and the related Danielsson prescription. Under a canonical transformation, meaning a rewriting of the same physics with rescaled fields or added boundary terms, both prescriptions give a whole family of different states. The paper also shows that minimizing the renormalized stress-energy tensor (the local energy density with ultraviolet infinities subtracted covariantly) gives a unique state. That analysis was spatially flat. It already used Planck data to narrow the non-invariant family, which motivates our choice to treat vacuum freedom as a parameter. Novel quantum initial conditions for inflation introduced the stress-energy prescription and states openly that the epoch at which to impose it is unresolved. Quantum initial conditions for curved inflating universes extends the prescription to curved space. It builds a new perturbation variable that behaves like a simple oscillator and is unique only up to one scalar constant. It computes spectra only at Planck best-fit points. We sample both residual freedoms, the imposition time and that constant, rather than fixing them.

The kinetic-dominance work supplies the signal and the pipeline. When the inflaton starts with most of its energy in motion, a comoving horizon maximum produces a large-scale power cutoff and ringing oscillations. Constraining the kinetically dominated universe found the duration parameter $N_\dagger$ (e-folds of inflation before the reference scale exits the horizon) peaking near 6. It reported evidence parity with slow roll but ran no sensitivity checks. Finite inflation in curved space shows that curvature bounds the amount of inflation, which pushes these features into the observed window. That paper built the exact-spectrum nested sampling pipeline (primpy, oscode, PolyChord inside Cobaya) that we reuse, and it acknowledges that its odds are debated on prior grounds. Case for kinetically dominated initial conditions for inflation shows that for plateau potentials, kinetic-dominated starts are prior-insensitive at the level of initial field values. This constrains our proposal. We should not expect sensitivity in whether kinetic dominance occurs. We should look for it in the late-time parameters, $N_\dagger$ and $\Omega_K$, that actually enter the likelihood.

The statistics work supplies tools and warns about their limits. Detecting and diagnosing prior and likelihood sensitivity with power-scaling perturbs the prior by raising it to a power and reweights existing samples. It reports a heuristic warning threshold of cumulative Jensen–Shannon distance $\ge 0.05$. Power-scaling leaves a uniform prior unchanged, and the priors on $N_\dagger$ and $\Omega_K$ are uniform. Applied naively here, the method would therefore report no sensitivity at all. That is why we replace it with width, location and shape perturbations. Curvature tension: Evidence for a closed universe uses the suspiciousness statistic to remove prior-volume dependence from measures of dataset disagreement. That removes volume effects for dataset agreement, but it does not address model comparison or prior shape. Nested sampling with any prior you like trains normalizing flows so that nested sampling can use arbitrary sample-defined priors, with evidence errors well below one unit of log-evidence. We use it to impose physically derived priors that have no analytic form.

Research Program

The central construction is a single broad reference run per model. The reference prior $p_0(\theta)$ is wider than any target prior we intend to study. It includes $N_\dagger$ and $\Omega_K$ over generous ranges, the vacuum imposition time over the whole pre-inflationary interval, and the free constant of the curved perturbation variable. Any target prior $p'$ whose support lies inside that of $p_0$ then has evidence

$$\log \mathcal{Z}' - \log \mathcal{Z}_0 = \log \mathbb{E}_{\text{post},0}\!\left[\frac{p'(\theta)}{p_0(\theta)}\right],$$

where $\mathcal{Z}_0$ is the reference evidence and the expectation is over reference-posterior samples. The identity follows directly from the definition of the evidence as the likelihood averaged over the prior. We stabilize the weights with Pareto smoothing and use the Pareto shape diagnostic from the power-scaling work to flag targets that are too far from the reference to trust. To fix a residual freedom at a single value $\lambda^*$, such as imposing the vacuum at one specific time, we use the Savage–Dickey ratio instead. This is the ratio of marginal posterior density to prior density at $\lambda^*$, and it gives the Bayes factor for the restricted model without a new run.

We study three families of target priors. The first is width and location changes to $N_\dagger$ and $\Omega_K$ across ranges used in the cited papers and plausible alternatives. Location shifts use exponential tilting, which the power-scaling paper names as underdeveloped. The second is shape changes: uniform versus log-uniform on $N_\dagger$, and a physically derived prior obtained by pushing Planck-surface initial conditions forward to $N_\dagger$ and imposing it through a trained flow bijector. The third is vacuum choices: bare stress-energy minimization at each imposition time, a Bunch–Davies-like state, and members of the Hamiltonian-diagonalization family. The last are separate models that each get their own reference run.

For each published conclusion we report the range of log Bayes factors over the defensible target priors. A conclusion counts as data-driven if its sign and approximate size survive that range, and as prior-driven otherwise.

Baselines

The naive reference point is to take the published numbers at face value with no sensitivity analysis. For the closed universe, Finite inflation in curved space reports odds "over 100:1" and a Starobinsky-model posterior $\Omega_{K,0} = -0.031 \pm 0.014$. Curvature tension reports Planck-alone odds over 50:1 against flat. For kinetic-dominance parity, Constraining the kinetically dominated universe states parity qualitatively, and the file reports no numerical log-evidence difference for it. We first reproduce these values with our pipeline on the same data (Planck 2018 temperature and polarization with low-multipole polarization, plus BK15 where the original used it) and the same priors. If we cannot reproduce them within nested sampling error, the sensitivity map is not meaningful.

The strongest competitor on the scientific question is brute-force refitting: rerunning the full PolyChord pipeline for each target prior. This is the ground truth and not a strawman. We run it for a fixed subset of about six target priors, chosen to span easy and hard reweighting cases, with the same live-point settings and likelihoods as the reference run. Our method "beats" it only in cost at equal accuracy. The criterion is that reweighted log-evidence agrees with the refit within the combined nested sampling uncertainty, at a small fraction of the compute. Curvature tension needed 12 single-core years for one curvature analysis, which is why refitting cannot cover a full sensitivity map.

The second competitor is the suspiciousness statistic from the same paper. It already removes prior volume, so we compute it alongside our map. It beats our approach wherever the question is dataset agreement. Our contribution must show changes it cannot capture: shape and vacuum effects on model comparison.

The ablation that isolates the key ingredient keeps the vacuum's residual freedoms fixed at the values of Quantum initial conditions for curved inflating universes instead of sampling them. The difference between the two maps measures how much of the total sensitivity comes from the unresolved vacuum theory rather than from ordinary cosmological priors.

Why not just…?

Why not just rerun nested sampling with a few different prior widths? For width changes on uniform priors this is nearly analytic: the evidence shifts by the log ratio of widths when the posterior sits well inside both ranges. Rerunning therefore adds little there. It becomes expensive exactly where the interesting effects lie, in shape changes, correlated physical priors and a continuum of vacuum imposition times. Refitting is still our validation standard. If reweighting fails the refit comparison for the hard cases, those entries of the map fall back to refits and the method claim narrows to the easy cases.

Why not just use power-scaling from priorsense directly? Raising a uniform prior to a power returns the same uniform prior. The central priors here would therefore show zero sensitivity by construction, and the tool itself only perturbs spread, not location (Detecting and diagnosing prior and likelihood sensitivity with power-scaling). We keep its smoothing and reliability diagnostics and its posterior-distance measure, but replace the perturbation family.

Why not just rely on the suspiciousness statistic? It was designed to cancel prior volume when asking whether two datasets disagree (Curvature tension). Closed versus flat on Planck alone, and one vacuum versus another, are model-comparison questions where prior shape still matters. Suspiciousness would also give no answer about imposition time. This alternative is a real threat only for the curvature-tension question itself. There we report both, and if they agree we say so.

Why not just fix the vacuum to stress-energy minimization and move on? That is the ablation above. If sampling the imposition time and the free constant barely changes the evidence or posteriors, the vacuum ambiguity is observationally irrelevant at current precision. That is a publishable negative result, and it would close the question the vacuum papers leave open.

Proposed Experiments

The first experiment reproduces the baselines under their original priors. The second builds the broad reference runs for flat and curved Starobinsky inflation with kinetic dominance. It uses the stress-energy vacuum with free imposition time and free perturbation constant, plus separate runs for the Bunch–Davies-like and Hamiltonian-diagonalization vacua, all on Planck 2018 data, with CMB lensing and BOSS DR12 BAO added in a second tier. The third is the validation grid of refits described above. The fourth produces the sensitivity maps.

The decisive checks are four. Reweighted and refit evidences must agree within error on the validation subset; otherwise the method fails. We ask whether the closed-versus-flat preference or the kinetic-dominance parity changes sign, or moves by more than one unit of log-evidence, across defensible priors. We ask whether the Savage–Dickey ratios over imposition time are flat, meaning the data are silent on the theoretical open question, or peaked. Finally, we run a mock-data test. We simulate low-multipole skies from a stress-energy-vacuum model with known parameters and cosmic variance, and ask how often the pipeline recovers the true vacuum over the Hamiltonian-diagonalization alternative. If recovery is near chance, then at current precision no observational claim about the vacuum can be made at these scales, whatever the real-data evidence says.