Telescopes that map the polarization of the cosmic microwave background (the relic radiation from the early universe) are hunting a signal three to four orders of magnitude fainter than the unpolarized brightness it sits on (LiteBIRD scanning strategy). Small instrument flaws make a tiny fraction of that brightness show up as fake polarization. Examples are an elliptical beam, two paired detectors with slightly different frequency responses, or a slowly drifting pointing. This contamination is called temperature-to-polarization leakage. The size of the flaw is not constant. It changes with the spin of the spacecraft (20 minutes for LiteBIRD), with its precession (about 3.2 hours), and with thermal and orbital cycles over days to months. Corrections today use one of two extremes. Some assume the leakage coefficients are fixed for a whole season or day, as in static leakage matrices and the per-day interferometric leakage solutions used by the Event Horizon Telescope. Others correct on one hand-tuned short timescale, as in the 167 ms destriping of Planck's NPIPE pipeline.
This note proposes estimating the leakage coefficients as a sum of components at several timescales at once, and choosing how much freedom each timescale gets by how well the corrected data agree with an independent experiment, not by how well they fit the data being corrected. The reason is a specific failure we expect from the obvious within-dataset choice. A leakage template allowed to vary quickly can soak up real sky polarization during stretches where the scan crosses the sky from too few angles. The fit then looks better while the science signal is quietly removed. Agreement with a second instrument, whose flaws are unrelated, cannot be improved by that kind of absorption.
Core Claim
The leakage coefficients have a natural multi-scale temporal structure. The right resolution at each scale cannot be identified from one dataset alone, because fast leakage freedom and true sky polarization are degenerate wherever scan cross-linking is poor. The cross-linking factor measures how uniformly a sky pixel is crossed from many angles. We claim that cross-dataset agreement is a usable selection signal for this resolution. It is scored scale by scale, in the spirit of the QMAP/Saskatoon "null-buster" differencing test (QMASK). We claim it recovers the injected leakage timescales in simulation, and that it yields less biased estimates of the tensor-to-scalar ratio $r$ (the amplitude of the primordial gravitational-wave signal) than either fixed-scale correction or within-dataset model selection.
The model for one detector's time stream is
$$d_t = \sum_{k} \gamma_k(t)\, T_k(p_t) + s^{\rm pol}_t + n_t,$$
Here $d_t$ is the sample at time $t$, $p_t$ is the sky pixel being observed, and $T_k$ are leakage templates. The templates are the temperature map and its first and second spatial derivatives along the beam axes, which are the shapes produced by gain, pointing, and ellipticity mismatch. They are built from an external temperature map such as Planck's. The term $s^{\rm pol}_t$ is the true polarization signal and $n_t$ is noise. The coefficients are expanded on a multi-scale basis,
$$\gamma_k(t) = \sum_{s \in \mathcal{S}} \sum_j c_{k,s,j}\, \phi_{s,j}(t), \qquad \text{penalty} = \sum_{s} \lambda_s \sum_{k,j} c_{k,s,j}^2 .$$
Here $\phi_{s,j}$ are window functions of length $s$ placed at offset $j$, and $\mathcal{S}$ spans sub-second, spin, precession, daily, and season scales. The weight $\lambda_s$ controls how freely leakage may vary at scale $s$. The new step is choosing $\{\lambda_s\}$ to minimize the scale-binned power of the difference between our corrected map and an independent experiment's map on the overlapping sky. The usual choice would minimize the within-dataset residual.
Relation to Prior Work
Two papers motivate the timescale question. The LiteBIRD study optimizes spin and precession parameters for cross-linking, but its authors state that their map-based approach cannot capture time-dependent effects such as detector drifts or signals locked to the scan (LiteBIRD scanning strategy). We work directly in the time domain where those effects live. NPIPE shows that aggressive time-domain processing pays off, with 10–20% lower polarization noise variance than the 2018 release (Planck intermediate results). However, it fixes its baseline length by hand for one mission. We make the scale a quantity to be selected, and we make the selection transferable.
Two papers motivate the degeneracy. MeqSilhouette v2 shows that recovered interferometric leakage terms depend on the source's own polarization and on baseline geometry (MeqSilhouette v2). The Sgr A* polarization analysis keeps leakage solutions static per day precisely because the source varies within minutes, and the imaging pipelines still disagree at the 5–10% level on circular polarization (EHT Sgr A* VII). Both are cases where leakage and real signal trade off. We treat this tradeoff as the central obstacle, not as a nuisance.
The multi-scale structure borrows from video temporal modeling. There, start-time × duration maps at several scales let one model handle events of very different lengths (Multi-scale 2D Representation Learning). The same work names the two failure modes our basis must avoid: short events averaged away in long windows, and long context lost in short ones. We take that representational idea, not the neural architecture. Instrument drift is continuous, so boundary-pairing approaches like BSN do not fit it.
The validation criterion comes from the QMASK null-buster. That work's scale-by-scale cross-experiment differencing exposed a roughly 37% power suppression at $\ell\in[201,300]$ from unmodeled beam ellipticity (QMASK). QMASK used the test as a diagnosis, and then handled the problem by injecting noise to discard unreliable scales. We use the same statistic as an objective for model selection.
The domain-generalization methods constrain the proposal more than they supply it. Orthogonal feature subspaces in Compositional Feature Alignment and causal masking in Adaptive-CaRe both separate invariant signal from environment-specific features. Adaptive-CaRe also warns that hard penalties discard useful predictive features. In our setting the physics already gives the separation: leakage lives on temperature-derived templates and the sky lives on polarization. So we test CFA-style orthogonality only as a competitor, not as the main method.
Research Program
We will build a time-domain simulation suite with known, time-varying leakage. LiteBIRD-like pointing will be generated with Falcons.jl. A ground-based scan will follow the AliCPT-1 setup at 90/150 GHz on 17% of the northern sky (AliCPT-1 forecasts). We will inject beam ellipticity, gain mismatch, and bandpass mismatch whose amplitudes drift through four regimes: constant, spin-synchronous, precession-scale thermal, and random walks with $1/f$ character. Sky polarization will include a lensing component and a tensor component of known $r$. Each scan gets an independently simulated partner experiment, satellite against ground, with its own unrelated leakage. The partner provides the cross-dataset score.
The estimator solves the penalized regression above for many settings of $\{\lambda_s\}$. It is a large sparse least-squares problem, of similar size to destriping. It then scores each setting by cross-experiment difference power in multipole bins and keeps the best. We will also analyze the degeneracy directly. For each window, we will relate the fraction of injected polarization absorbed by the leakage terms to the window's local cross-linking factor. Our prediction is that within-dataset selection fails in proportion to poor cross-linking, which would give a cheap way to forecast where it fails. A secondary arm repeats the logic for interferometry. It uses MeqSilhouette v2 with time-varying leakage terms and scores against an independent array configuration (EHT2017 against ALMA extended).
Baselines
The naive reference applies no leakage correction beyond standard map-making. It sets the scale of the problem and confirms that the injected leakage actually biases $r$.
The classical competitor is static template regression. It uses the same templates $T_k$, but fits one coefficient set per detector per season, or per day for the interferometric arm. This matches common practice and the per-day static solutions in EHT Sgr A* VII. It gets identical data and templates, so any gain from our method comes from the temporal structure alone.
The fixed-short-scale competitor is NPIPE-style. It uses one short window length chosen by a grid search with the same compute budget as our selection. It represents the "just destripe harder" position (Planck intermediate results).
The key ablation is our multi-scale estimator with $\{\lambda_s\}$ chosen by within-dataset criteria. We use two such criteria: held-out time-segment residuals and a Bayesian information criterion. Both get the same search grid. This isolates the claim that the cross-dataset signal is what matters.
A learned competitor is a multi-scale temporal network, in the pyramid style of Multi-scale 2D Representation Learning. It maps time stream plus templates to cleaned time stream, and is trained by pooled empirical risk minimization on both experiments. A second version adds CFA orthogonal heads separating experiment identity from signal (Compositional Feature Alignment). Both versions get equal training compute and tuning trials.
For real-data context, the strongest reported number is from the AliCPT-1 Data Challenge 1. Five pipelines achieved $\sigma(r)\sim 0.02$ and recovered $r$ between 0.012 and 0.029 against an embedded truth of 0.023, with a null-test limit of $r<0.041$ at 95% confidence (AliCPT-1 forecasts). Those simulations did not, as far as the file reports, include time-varying leakage. Beating the baselines therefore means three things. First, on our drifting-leakage simulations, our method must keep the bias in $r$ below $0.25\,\sigma(r)$ while static and fixed-scale corrections do not. Second, it must do no worse than DC1-level spread when leakage is constant. Third, within-dataset selection must show measurable signal loss that cross-dataset selection avoids.
Why not just…?
Why not just destripe on a short fixed baseline, as NPIPE does? Destriping removes slow offsets, not leakage proportional to the temperature-sky gradient. A single scale also cannot follow spin-locked and month-long drifts at the same time. Even after Planck's processing, residual systematics at $\ell<30$ remain comparable to the cosmological signal (Planck 2018 results). The fixed-scale baseline, given its best window by grid search, tests this directly. If it matches us on all four drift regimes, the multi-scale basis is unnecessary and we will say so.
Why not just rely on a rotating half-wave plate? Modulating polarization with a spinning half-wave plate moves the signal away from slow detector noise. However, imperfections in the plate couple with pointing offsets and produce scan-synchronous leakage on spin and precession timescales (LiteBIRD scanning strategy). That is exactly the multi-scale drift we model. We will include a simulated plate with differential transmission so the residual it leaves becomes one of the test regimes.
Why not just invert a calibrated response matrix? If leakage were constant, this would be the right answer. The static-regression baseline is that method run at its best. Its known weakness is that the recovered coefficients depend on the source's intrinsic polarization (MeqSilhouette v2), and coefficients that drift break it. In the constant-leakage regime we expect it to tie with our method. The comparison is decisive only in the drifting regimes.
Why not just pick the timescale by within-dataset cross-validation? This is the most serious threat, because it is cheaper and needs no second experiment. Our argument is that held-out time segments share the same scan geometry, and so share the same degeneracy between fast leakage and sky polarization. If the within-dataset ablation selects the same $\{\lambda_s\}$ and loses no signal, the cross-dataset ingredient is not needed. The absorption-versus-cross-linking analysis would then tell us which scan strategies make it safe.
Why not just train a multi-scale network on pooled data? Pooled training learns experiment-specific shortcuts, as the 99% train versus ~76% test collapse in Adaptive-CaRe illustrates. A flexible network is also the model most able to absorb real polarization. The network baselines, with and without orthogonality, will show whether learned capacity helps or simply moves the absorption into a black box.
Decisive Checks
Three results would settle the proposal. First, on drifting-leakage simulations, the cross-dataset-selected weights recover the injected timescales, while within-dataset selection picks scales that are too fine. That over-fine choice should show up as removed tensor power that grows as the local cross-linking factor worsens. Second, the bias in $r$ under our method stays within a quarter of $\sigma(r)\sim0.02$ across all drift regimes, while static and fixed-scale baselines exceed it in at least the spin- and precession-scale regimes. Third, swapping which experiment plays the reference changes the selected weights only within noise. That would show the criterion is responding to a real property of the leakage and not to the partner's own flaws. If the first check fails, so that within-dataset selection is just as good, the proposal reduces to a multi-scale template fit, and we will report it as that.
References
- Multi-dimensional optimisation of the scanning strategy for the LiteBIRD space mission
- CMB power spectrum at<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>l</mml:mi><mml:mo>=</mml:mo><mml:mn>30</mml:mn><mml:mn/><mml:mo>–</mml:mo><mml:mn>200</mml:mn></mml:math>from QMASK
- Planck intermediate results
- MeqSilhouette v2: spectrally resolved polarimetric synthetic data generation for the event horizon telescope
- First Sagittarius A* Event Horizon Telescope Results. VII. Polarization of the Ring
- Multi-scale 2D Representation Learning for weakly-supervised moment retrieval
- BSN: Boundary Sensitive Network for Temporal Action Proposal Generation
- Enhancing Compositional Generalization via Compositional Feature Alignment
- Adaptive-CaRe: Adaptive Causal Regularization for Robust Outcome Prediction
- Performance forecasts for the primordial gravitational wave detection pipelines for AliCPT-1
- Planck2018 results