How Much Evidence Do the CMB's Largest Scales Really Carry? Calibrating Low-Multipole Model Comparison with Forensic Likelihood-Ratio Methods

Several proposed explanations for the missing power on the largest angular scales of the cosmic microwave background (CMB) now claim statistically meaningful support. The CMB is the relic radiation map whose lowest "multipoles", or angular frequencies, are anomalously weak. One explanation lets only a discrete comb of Fourier wavenumbers exist, rather than a continuum. Another adds a brief kinetically dominated phase before inflation, when the field driving inflation is moving too fast for inflation to have started yet. That phase stamps a cutoff and ringing onto the primordial spectrum. The first idea improves the fit to Planck data by $\Delta\chi^2=-8.55$. Variants of the second report Bayesian odds above 100:1. These numbers come from a few hundred sky modes, and the feature models were partly built after the anomaly was seen. Forensic statistics has studied this exact failure for years. When a likelihood ratio is computed from a small reference sample, the "plugin" number is often overconfident by orders of magnitude, and the field has developed tools to measure and repair this.

This note proposes treating each low-multipole model comparison as a forensic evidence-reporting system and testing whether it is calibrated. We would generate many simulated skies under standard cosmology and under each feature model. We would then run the same fitting or evidence pipeline on every sky and score the resulting log-likelihood ratios with the forensic log-likelihood-ratio cost. Mock skies are expensive to analyze, so the number we can afford plays the role of the forensic "reference population size". This lets us ask directly how many mocks are needed before a reported weight of evidence can be trusted.

Core Claim

The headline numbers for discrete-wavevector and pre-inflationary models are discriminating scores, not calibrated weights of evidence. Once they are calibrated against simulations under both hypotheses, the conclusions can change in both size and sign. The comparison between the two feature families, a discrete comb versus kinetic-dominance ringing, may be undecidable with temperature data alone. We claim three things will be new. The first is a calibration map that turns a reported $\Delta\chi^2$ or Bayes factor into an honest likelihood ratio. The second is the stability of that map as a function of mock count, measured the way forensic studies measure population-size effects. The third is a discriminability measurement between the two physical feature families, separate from any calibration question.

The proposal assigns the four directions different roles. Discrete wavevector quantization and kinetic-dominance oscillations supply the competing hypotheses. Large-scale power suppression is the data regime. The population-size literature supplies the method and the evaluation metric.

Why this is not already done, and what the forensic analogy does and does not buy

Improved cosmological fits with quantized primordial power spectra replaces the integral over wavenumbers in the CLASS Boltzmann code (the standard program that turns a primordial spectrum into CMB spectra) with a sum over $k_n=k_0+n\Delta k$. It finds the best fit by Nelder-Mead optimization against Planck 2018 and reports $\Delta\chi^2=-8.55$ at $k_0=3.225\times10^{-4}$ and $\Delta k=2.257\times10^{-4}\,\mathrm{Mpc^{-1}}$. That paper motivates our work: the score comes from maximizing over a likelihood surface that oscillates in $\Delta k$ and therefore has many local optima. That situation has no simple reference distribution, and the paper does not compute evidence.

Finite inflation in curved space does compute nested-sampling evidence, using PolyChord within the Cobaya inference framework, for kinetic-dominance models on closed backgrounds. It integrates over continuous wavenumbers. Constraining the kinetically dominated universe finds that in flat space the kinetic-dominance model with a Starobinsky potential merely ties standard slow roll in evidence. That result constrains the proposal: the Occam penalty already eats much of the likelihood gain, so any calibration effect must be compared against an evidence baseline, not only against raw $\Delta\chi^2$.

On the forensic side, Bayesian calibration for forensic evidence reporting integrates calibration parameters over their posterior so that tiny reference sets yield tempered likelihood ratios. How many do we need? exploration of the population size effect on the performance of forensic speaker classification shows that log10 likelihood ratios swing by up to 13 across 10-speaker reference sets. Reducing uncertainty at the score-to-LR stage in likelihood ratio-based forensic voice comparison using automatic speaker recognition systems shows that skewed score distributions make plugin calibration erratic. These papers supply our method.

We should be honest about where the analogy breaks. The research file frames Gap 3 as cosmology "ignoring cosmic variance". That framing is too strong. Standard low-multipole likelihoods already model the variance of a Gaussian sky with finitely many modes. The real exposure lies elsewhere. Search over oscillatory templates inflates the maximized $\Delta\chi^2$. Evidence values depend on priors that were chosen after the anomaly was seen. Neither the score nor the evidence has been checked against skies where the truth is known. The small-sample problem that transfers cleanly from forensics is the number of mock skies used to calibrate, not the number of sky modes. That is where the population-size results apply.

Research Program

We first build a fast low-multipole forward model for multipoles $\ell\le30$ in temperature and E-mode polarization. The background cosmology is held at the Planck best fit, so the transfer functions $\Delta_\ell(k)$ are computed once with CLASS. The angular spectrum for any primordial spectrum $P(k)$, continuous or discrete, then becomes a weighted sum:

$$C_\ell=\sum_n w_n\,P(k_n)\,\Delta_\ell(k_n)^2,$$

where $k_n$ are the allowed wavenumbers and $w_n$ their quadrature or comb weights. For the continuum case, $w_n$ is a fine integration grid. With this sum, one model evaluation takes milliseconds, so thousands of mock fits become affordable. We will validate it against full CLASS runs for the published best-fit comb.

Three feature families serve as the alternatives to standard flat cosmology with a power-law spectrum. The first is the linear comb $k_n=k_0+n\Delta k$. The second is the kinetic-dominance cutoff-plus-ringing spectrum. We generate it with the analytic matching templates of Analytical approximations for curved primordial power spectra and spot-check it against numerical Mukhanov-Sasaki solutions, i.e. the equation governing each perturbation mode, as in Finite inflation in curved space. The third is a smooth exponential cutoff with no oscillation, included to isolate what discreteness and ringing add.

Mock skies are Gaussian $a_{\ell m}$ realizations with a realistic sky mask and noise, drawn under each hypothesis. Feature-model mocks use parameters drawn from the same priors the evidence computation uses. Three scoring pipelines then run on every mock. The first maximizes $\Delta\chi^2$ with Nelder-Mead, replicating the published procedure, including its multistart settings. The second computes a PolyChord log-Bayes factor. The third gives the Wilks-style naive conversion of $\Delta\chi^2$, a $\chi^2$ distribution with two degrees of freedom.

Each pipeline's raw output is a score $s$. We calibrate scores to log-likelihood ratios with three methods: plugin logistic regression, regularized logistic regression, and the Bayesian calibration of Bayesian calibration for forensic evidence reporting. We then evaluate each on held-out mocks with the log-likelihood-ratio cost:

$$C_{\mathrm{llr}}=\tfrac12\Big[\tfrac{1}{N_1}\textstyle\sum_{i\in H_1}\log_2(1+\mathrm{LR}_i^{-1})+\tfrac{1}{N_0}\textstyle\sum_{j\in H_0}\log_2(1+\mathrm{LR}_j)\Big],$$

where $H_1$ is the feature hypothesis, $H_0$ standard cosmology, and $N_1,N_0$ the mock counts under each. $C_{\mathrm{llr}}$ splits into a discrimination floor $C_{\mathrm{llr}}^{\min}$ and a calibration loss. Calibration stability is measured as in System Performance as a Function of Calibration Methods, Sample Size and Sampling Variability in Likelihood Ratio-Based Forensic Voice Comparison: we report the range of $C_{\mathrm{llr}}$ across bootstrap draws of the calibration mocks, with calibration-set sizes swept from 10 to several thousand. Finally, we apply the fitted calibration map to the real Planck score. The output is a calibrated likelihood ratio with an uncertainty band, in place of 8.55 or "100:1".

Baselines

The do-nothing reference is the published score itself. That means the $\Delta\chi^2=-8.55$ of Improved cosmological fits with quantized primordial power spectra, read either as a raw likelihood ratio, $\exp(8.55/2)$, or through the Wilks two-degree-of-freedom approximation. We rerun that exact procedure on every mock with the same optimizer and start budget. On mocks this baseline is beaten if its implied log-likelihood ratios have a calibration loss well above zero, meaning they are systematically overconfident on $H_0$ skies. On the real data it is beaten if the calibrated ratio differs from the naive one by more than the bootstrap spread.

The strongest existing method is nested-sampling Bayesian evidence with PolyChord, as used by Finite inflation in curved space and described in PolyChord: nested sampling for cosmology. It is the strongest competitor because it already marginalizes over parameter uncertainty, which is the cosmological counterpart of Bayesian score calibration. If the priors are right, its output should already be calibrated. We run it on the same mocks with identical priors and live-point settings, and the comparison is fair because both see the same data and the same prior volume. Evidence is beaten only if its $C_{\mathrm{llr}}$ on mocks exceeds that of the calibrated score by more than the bootstrap range. If evidence turns out to be calibrated, that is a publishable negative result in its own right. The reported figure for the baseline is "Bayesian odds of over 100:1" for closed kinetic-dominance models from Finite inflation in curved space. For any cosmological $C_{\mathrm{llr}}$, no reported result was found on Lacuna.

Among the calibration methods, the reference numbers come from forensics. Plugin logistic regression reached a mean $C_{\mathrm{llr}}$ of 0.37–0.38 against 0.54–0.56 for the Bayesian model in System Performance as a Function of Calibration Methods, Sample Size and Sampling Variability in Likelihood Ratio-Based Forensic Voice Comparison. With 20 speakers and skewed scores, logistic regression had a $C_{\mathrm{llr}}$ range of 1.3 and the Bayesian model about 0.31 in Reducing uncertainty at the score-to-LR stage in likelihood ratio-based forensic voice comparison using automatic speaker recognition systems. We expect the same accuracy–stability trade-off. A method counts as winning at a given mock budget if it lies on the Pareto front of mean $C_{\mathrm{llr}}$ against $C_{\mathrm{llr}}$ range.

Two ablations isolate the key ingredients. The first replaces the comb with the smooth cutoff, so the gain due to discreteness is measured with the cutoff held fixed. The second swaps the comb for kinetic-dominance ringing, which measures whether the data can tell the physical mechanisms apart at all.

Why not just…?

Why not just compute a frequentist p-value from ΛCDM mocks and stop? This is the most serious threat, and we partly adopt it. The $H_0$ mocks yield exactly that look-elsewhere-corrected p-value as a by-product, and we will report it. A p-value only says how surprising the data are under standard cosmology, not how much more probable they are under a specific alternative. That second quantity needs mocks drawn under $H_1$ and a calibration map, which is what the forensic machinery provides. If the p-value alone already settles the question, for example a corrected p well above 0.1, the comparison between feature families becomes moot. We would then report that outcome plainly.

Why not just trust PolyChord evidence, since it already integrates out parameters? Evidence is calibrated only if the priors and likelihood match how the data were generated. Priors on $k_0$ and $\Delta k$ were chosen with knowledge of where the anomaly sits. Constraining the kinetically dominated universe shows that the Occam factor can erase a likelihood gain entirely, so evidence values are sensitive to exactly these choices. Our mock study tests evidence calibration directly, and evidence is the baseline we must beat, not a straw man.

Why not just fit a smooth exponential cutoff? If the smooth cutoff matches the comb in calibrated likelihood ratio, the discreteness hypothesis has no empirical support beyond "less power at low multipoles". The ablation is designed to show this. The published claim that the comb also reproduces the dip at $\ell\approx20$–30 is a likelihood statement that calibration may or may not preserve.

Why not just clip extreme likelihood ratios with empirical bounds? Empirical lower and upper bounds were the least stable method in System Performance as a Function of Calibration Methods, Sample Size and Sampling Variability in Likelihood Ratio-Based Forensic Voice Comparison, with mean $C_{\mathrm{llr}}$ of 0.65–0.66. We include clipping cheaply but do not expect it to compete.

Proposed Experiments and Decisive Checks

The first check concerns the published comb. We run the Nelder-Mead comb fit on several thousand $H_0$ mocks. If more than roughly 5% reach $\Delta\chi^2\le-8.55$, the published improvement is mostly a search effect. If almost none do, it survives.

The second check is the calibrated likelihood ratio for the real Planck low-multipole data under each pipeline. It is reported with its bootstrap band at each calibration-set size. The result we would find most informative is the mock count at which the $C_{\mathrm{llr}}$ range falls below 0.1. That is the cosmological "how many do we need". It can be compared directly with the forensic threshold of about 30 reference speakers reported in How many do we need? exploration of the population size effect on the performance of forensic speaker classification.

The third check is discriminability between the comb and kinetic-dominance ringing, measured as $C_{\mathrm{llr}}^{\min}$ on mocks. It is computed first with temperature only and then with E-mode polarization added. Polarization contributes independent modes, which is the cosmological equivalent of enlarging the reference population. A $C_{\mathrm{llr}}^{\min}$ near 1 would show that current data cannot separate boundary quantization from pre-inflationary dynamics, however each fares against standard cosmology. A clear drop with polarization would identify the measurement that could separate them.

Two extensions are out of scope for the core study. One is quantization on closed backgrounds, using the integer-labeled eigenvalues of Primordial power spectra for curved inflating universes. The other is the curvature tension with lensing and baryon acoustic oscillations. Both need high-multipole mocks and are deferred until the low-multipole calibration machinery is validated.