WorksPhysics and Astronomy
Millimeter-Wavelength Constraints on Circular Polarization of the Cosmic Microwave Background at Arcminute Scales with the Simons Array
Abstract
We present constraints on the cosmic microwave background (CMB) circular polarization anisotropy at arcminute angular scales, using dedicated measurements from the Simons Array taken with a quarter-wave plate and rotating half-wave plate. A detection of circular polarization in the CMB would probe a range of proposed generation mechanisms, from Standard Model astrophysics to physics beyond it. At an observing frequency of 90 GHz, we place constraints over the multipole range , corresponding to angular scales of 5.4-36 arcmin, in bands of width , roughly tripling the highest multipole previously reached by a -mode power spectrum measurement. Within this domain, we constrain the 95% confidence upper bound of to range from at the bandpower to at the bandpower. These limits are derived from 2.5 months of observations of each of two small regions of the sky located away from the Galactic plane, totaling 12 deg of effective sky area. Because the methodology of this study marginalizes known foreground contamination, these bounds serve as robust, conservative limits on the true CMB circular polarization signal down to the sensitivity limits of the instrument. Additionally, we provide an experimental and analytical framework for implementing circular polarization observations on other telescopes by utilizing a quarter-wave plate and half-wave plate design.
Introduction
The cosmic microwave background (CMB) is a mostly unpolarized blackbody radiation signal which is the most distant detectable direct evidence of the Big Bang and its immediate aftermath. Understanding how the post-Big Bang events affect the CMB signal is one of the foremost ongoing scientific goals in the field of cosmology, and experiments have constrained the temperature anisotropies at both small and large angular scales [11, 61, 6, 47]. Most contemporary experiments measuring the CMB focus on characterizing the linear polarization ( and -modes), which can shed light on inflationary physics and other early-Universe phenomena [28, 68] (W. Hu & M. White 1997). The standard cosmological model, Lambda cold dark matter (ΛCDM), predicts a measurable linear polarization signal, but no appreciable circular polarization in the CMB. However, there are several potential secondary mechanisms that could produce it.
Before decoupling, a primordial magnetic field acting on the cold pre-decoupling plasma sources modes in the presence of adiabatic curvature perturbations [21]. After decoupling, the most studied channel is Faraday conversion, whereby a magnetic field threading a relativistic plasma converts pre-existing linear polarization, considered both for magnetized galaxy clusters [10] and for the supernova remnants of Population III (Pop. III) stars (S. De & H. Tashiro 2015). Circular polarization may also be generated after decoupling through scattering of CMB photons off the cosmic neutrino background (CνB) [39] and through the nonlinear photon–photon interactions of the Euler–Heisenberg effective Lagrangian [56, 40]. A smaller set of proposals instead invokes physics beyond the Standard Model, coupling photons to an external vector field through a Chern–Simons term [1] or to a cosmological pseudoscalar field acting as dark matter [17]. A comprehensive summary of these origins is provided by S. King & P. Lubin [30]. Distinct from all of these are foregrounds, which emit circular polarization of their own rather than acting on CMB photons. Galactic components, such as synchrotron and thermal dust emission, are not expected to contribute appreciably at 90 GHz [30]. The far brighter atmospheric signal, produced by Zeeman splitting of molecular oxygen, is dominant at larger angular scales than those within this study’s scope [42, 19], and is in any case removed by the filtering described in § 4.2. A diffuse signal observed here would therefore be dominated by converted CMB photons, and we refer to it as CMB circular polarization throughout.
Observational upper limits on the circular polarization anisotropy of the CMB have historically spanned disparate angular scales and employed varied instrumental techniques. The earliest came from Very Large Array (VLA) interferometry at 5 GHz, whose highly localized measurements constrained the root-mean-square fractional circular polarization to , or at the widest synthesized beam, on scales [44]. On much larger angular scales, the Milano Polarimeter (MIPOL) utilized a correlation receiver at 33 GHz to set constraints at scales of to , bounding the -mode power spectrum at the level of [35]. The SPIDER balloon-borne experiment significantly improved these large-scale constraints at 95 GHz and 150 GHz; by leveraging nonidealities in their half-wave plate polarization modulators, SPIDER placed upper limits on scales at the level of [41]. Most recently, the Cosmology Large Angular Scale Surveyor (CLASS) achieved a direct measurement of circular polarization at 40 GHz using a Variable-delay Polarization Modulator. After characterizing and filtering a bright atmospheric oxygen dipole at the level of , CLASS established the tightest limits at the time on the large-scale CMB circular polarization power spectrum [42]. That analysis has since been extended to the CLASS 90 GHz, 150 GHz and 220 GHz bands, improving the large-scale limits by a factor of and reaching [14].
In this paper, we present new millimeter-wavelength constraints on the circular polarization of the CMB using data from the PolarBear-2b (PB-2b) receiver of the Simons Array, the successor telescope to PolarBEAR. This study presents a novel approach to efficiently measuring circular polarization, using a stationary quarter-wave plate (QWP) positioned in front of a constantly spinning half-wave plate (HWP). The technique is available to any experiment already operating linear polarization sensitive detectors and a rotating HWP, along with established CMB mapmaking and power spectrum estimation analysis tools. The remainder of this paper is organized as follows: § 2 describes the instrument, optical configuration, and observations; § 3 describes the calibration techniques; § 4 details the data analysis pipeline and our systematic error characterization; § 5 presents our unblinded limits; and we conclude in § 6.
Instrument and Observations
The measurements presented in this work utilize data collected by the PB-2b receiver, the second deployed receiver of the Simons Array [4, 62, 60], located in the Atacama Desert in Chile. It features a continuously rotating cryogenic HWP [23] to modulate incoming polarization and systematically suppress low-frequency atmospheric noise. In this section, we detail the optical and cryogenic architecture of the receiver, the properties of the transition-edge sensor (TES) bolometer array, and the specific observational campaigns that form the basis of our circular polarization constraints.
The POLARBEAR-2b Receiver
The Simons Array is located at an altitude of 5200 m in the Atacama Desert in Chile. The PB-2b receiver was mounted on an off-axis Gregorian telescope featuring a 2.5-meter primary mirror aperture [62, 38], shown in Figure 1. The PB-2b focal plane consists of 7,588 TESs coupled to dichroic, dual-polarization pixels utilizing lenslet-coupled sinuous antennas, providing sensitivity to radiation at 90 GHz and 150 GHz [67] (however, only 90 GHz is utilized for this study as explained in § 2.2), so chosen to avoid foregrounds, specifically atmospheric absorption bands and Galactic synchrotron radiation at lower frequencies. The detectors are read out using a 40 digital frequency-domain multiplexing (DfMUX) system [5]. The focal plane detectors are split into -pixels and -pixels, while the latter are rotated by 45° from the former.

Figure 1. Schematics describing the setup of PB-2b. Left: Telescope receiver mounted, with the waved orange arrow indicating path of light rays as they travel across the two mirrors, through the QWP, and into the receiver. Right: Interior of the receiver, with the path traced through the QWP at ambient air temperature, through the cold HWP (CHWP), lenses and filters, and into the detectors on the focal plane. The QWP is placed 15 cm from the vacuum window.
Following the initial deployment of the PB-2b receiver to the Simons Array in 2022, the instrument’s baseline optical performance and readout electronics were successfully validated [34, 54]. However, in-field evaluations determined that the on-sky sensitivity did not reach the initial design targets. The median detector noise-equivalent temperature (NET) for PB-2b at 90 GHz is , against a design value of , a loss driven by low optical throughput and elevated readout noise [15]. Furthermore, overall detector performance was hampered by an excess noise component that is strongly suspected to originate from environmental radio frequency (RF) interference at the observing site [15]. This characterization predates the installation of the QWP of § 2.2, whose transmission loss [55] further degrades the effective sensitivity of the observations presented here. Regardless, the elevated instrument noise primarily propagates as statistical uncertainty rather than as a systematic bias.
Polarization Modulators
To modulate the incoming celestial polarization above unpolarized atmospheric fluctuations, the receiver utilizes a cryogenic HWP. The HWP is situated inside the optics tube at 50 K and rotates continuously at approximately 2.025 Hz, placing the polarization modulation at Hz. While a standard continuously rotating HWP provides modulation of linear polarization (Stokes and ) and was the primary focus of the initial PB-2b commissioning [54], explicitly isolating the circular polarization (Stokes ) signal at high efficiency requires additional optical configuration. For the observations presented in this work, our optical system utilizes both a static QWP and the rotating HWP. The addition of the QWP introduces a known phase retardance that, when coupled with the rotating HWP, allows the incoming Stokes signal to be modulated and effectively separated from the linear polarization terms in the timestream, as shown in K. Sakaguri et al. [55].
The physical characteristics and optical performance of this QWP-HWP system have been extensively evaluated and optimized for the 90 GHz band [55]. The QWP is installed on the sky-side of the receiver and is kept at ambient air temperature. It is constructed from Rexolite 1422 (cross-linked polystyrene)1, which is not itself a birefringent material; instead, grooves machined into both faces produce form birefringence, giving different effective indices parallel and perpendicular to the grooves. These grooves are machined with depth and separations to optimize performance at 90 GHz. The 150 GHz band is disfavored in turn because the atmosphere is both brighter and less stable there, sitting between the 118 GHz oxygen line and the wing of the 183 GHz water line [43]. Laboratory measurements demonstrate that the QWP achieves a band-averaged circular-to-linear conversion efficiency () of 99.0%, with a band-averaged unconverted fraction () of ; because only , , and are physically reconstructed downstream of the QWP, with sky recovered from the channel, this leakage represents sky that fails to convert and remains as spurious , directly contaminating the recovered measurement. Spatial variations across the plate differ by less than 1% in amplitude and 3° in phase [55]. The average transmission is approximately 87%. When coupled with the cryogenic HWP, a three-stacked sapphire design structure with anti-reflective coatings achieving modulation efficiency, the combined system provides highly efficient conversion and subsequent temporal modulation of the incoming Stokes signal. The QWP has a Mueller matrix,
where for an ideal QWP, and ; the antisymmetric block encodes the 90° phase retardance, so and both have the same magnitude. Here describes the QWP alone. The quantities measured on the sky instead belong to the effective Mueller matrix of the full optical chain in the QWP basis, schematically , where represents the telescope mirrors and the remaining receiver optics (window, filters, and lenses). Its elements , written without superscript, follow the same input-to-output index convention as Equation (1). The HWP modulation is treated separately in § 4.2, and, following K. Sakaguri et al. [55], the polarization efficiency of everything other than the QWP is factored out of as a single scalar that multiplies the polarized signal in that modulation model. The remaining Stokes mixing is dominated by the QWP, so that , since the cross-polarization of the telescope optics is small and the receiver optics are approximately rotationally symmetric (K. Sakaguri et al. [55]).
Based on astrophysical point-source calibrations in K. Sakaguri et al. [55], we express the measured efficiencies and leakages in terms of the elements of , with both and normalized to unit transmission so that . The linear polarization efficiency of the system independent of the QWP, , was measured on-sky from Tau A by fitting the observed fraction as a function of parallactic angle. Both and are measured from the same detector data and so carry the overall transmission as a common factor; their ratio is therefore insensitive to it, and returns in the same normalization as Equation (1). K. Sakaguri et al. [55] obtain from timestreams that were not deconvolved for the detector time constants, so that value also contains the single-pole attenuation of the signal at , for . Because the science timestreams here are deconvolved (§ 3.3), we repeat the Tau A fit on the same observations after the identical deconvolution with stimulator time constants, obtaining . The net circular polarization efficiency and the leakages that enter the converted- channel are defined from the elements of , together with their approximate forms in terms of the QWP alone, as
The QWP conversion efficiency is near unity, , so is limited primarily by . The retained sky is constrained on-sky from Tau A to , consistent with the laboratory value of quoted above. Because the Tau A fit sees the full optical chain, this value is the effective ratio , whereas the laboratory value is its QWP-only approximation. Leakage from sky is not a property of the plate itself, so its QWP-only form above does not apply. It enters through residual miscalibration of the detector polarization angles relative to the QWP axis, which rotates into the channel. Calibrating the angles from the HWPSS (§ 3.4) intrinsically suppresses this effect. The residual is set by the alignment of the QWP to the telescope, measured at the site to within , which bounds (K. Sakaguri et al. Unlike , this bound comes from the alignment measurement rather than from the sky. The remaining term, -to- leakage , is retained rather than set to zero. An ideal birefringent plate couples intensity only to , but we do not assume that this structure holds exactly. An additional intensity leakage is generated downstream of the optics, in the detector response itself, and therefore lies outside this Mueller-matrix description; it is discussed in § 4.2.4. Both are absorbed into the effective coefficient measured in Section 4.8.1.
In summary, the static QWP maps the Stokes on-sky into the instrument-frame Stokes , which the HWP then modulates at four times its rotation frequency. The circular-polarization measurement channel of this analysis is therefore the demodulated timestream; all subsequent steps construct per-run, per-detector maps, estimate their cross spectra and convert these to , disentangling instrument systematics. The companion channel carries the linear polarization of the measured sky, and is not used in this analysis.
Scan Strategy and Observations
Two CMB patches, RA23 and RA4.5, were chosen for this analysis, both legacy patches from Polarbear (Polarbear Collaboration et al. [49]). These patches are so named due to the right ascension (RA) of their approximate centers in units of hours, which are for RA23 and RA4.5, , (J2000) and , . These locations were chosen to lie away from the Galactic plane, avoiding significant contamination in the sky signal from synchrotron radiation. The dimensions of these patches are , small enough that we work in the flat-sky approximation for later analysis. The small patch sizes are utilized in order to target sub-degree size angular scales (). The geometric distortion introduced by the flat-sky approximation is and strictly negligible compared to the statistical noise.
An observation of a given patch is taken with a constant elevation scan (CES), which tracks back and forth a minimum and maximum azimuth at a constant elevation. The telescope scans at in azimuth with a sampling rate of ; the duration of an observation is typically around 860 s. The CMB patches are observed both rising and setting, at elevations , high enough to avoid ground pickup, and , to maintain elevations roughly in the same range as where detector tuning occurs () (K. Sakaguri et al. [55]). Throughout, a detector is called “tuned” when it has been successfully biased onto its superconducting transition, and “optical” when it is coupled to the sky through its antenna rather than being a dark channel; only detectors that are both can contribute usable data. The scan strategy of observing both when rising and setting provides multiple parallactic angles, which decreases the non-radial systematics of the beam maps (discussed in more detail in § 3.2) and the amount of leakage (see § 4.8).
We observed RA23 and RA4.5 for about 2.5 months each, in staggered periods. During that time, corruption with the telescope timestamp data resulted in a large amount of data becoming unusable. For RA23 (RA4.5), 114 (99) of the 798 (576) runs with nonzero tuned and optical detectors have technical issues resulting in a total loss of usable detector-observations (the data from a single detector within a single observation), arising from corrupted timestream files, improperly stored identifiers in the database tracking each observation, failed tuning, or faults within the timestream data despite it being saved. In addition to these observation-wide technical failures, a subset of detectors in each observation are not usable because they do not have time constant or absolute gain calibration factors; this is explained in § 3.3. For these reasons, 356,385/566,694 (232,866/414,612), about 62% (56%) of detector-observations across all observations for RA23 (RA4.5), have usable timestreams for the analysis detailed in § 4. A typically behaved observing run yields between 600–700 tuned and optical detectors in the 90 GHz band. These instrument-level losses, together with the data quality cuts imposed after map-making, are tabulated in § 4.4.
Calibrations
This section describes the calibrations that fix the instrumental quantities entering the forward model of § 4.9. The pointing model (§ 3.1) sets the suppression of small-scale power by residual pointing error, and Jupiter observations set the beam window function (§ 3.2). A thermal stimulator provides the detector time constants and the absolute gain (§ 3.3), while the detector polarization angles are calibrated from the half-wave plate synchronous signal (HWPSS) fourth harmonic so that every detector can be rotated into the common QWP basis (§ 3.4). The circular polarization efficiency and the leakages and follow from the optical chain described in § 2.2, while the pipeline transfer function and the and leakage spectra are determined from the processed data and are deferred to § 4.7 and § 4.8. Table 3 collects all of these, with the role each plays in the likelihood and the section in which it is determined.
| Quantity | Value | Uncertainty | Role | Dim. | Section |
| — | Data | — | 4.5.2 | ||
| — | — | Sampled (likelihood) | — | ||
| 0.528 | 0.032 | Sampled (prior) | 1 | 2.2 | |
| 1.000 | 0.046 | Sampled (prior) | 1 | 3.3 | |
| – | per bandpower | Sampled (prior) | 3.2 | ||
| – | 0.1 | Sampled (prior) | 4.7 | ||
| – | 0.007 | Sampled (prior) | 1 | 3.1 | |
| Fixed nuisance | — | 4.8.1 | |||
| 1.000 | 0.0037 / 0.0062 | Fixed nuisance | — | 3.3 | |
| — | Fixed nuisance | — | 2.2 | ||
| Fixed nuisance | — | 2.2 | |||
| 0.001 | Fixed nuisance | — | 4.8.2 | ||
| 0.001 | Fixed nuisance | — | 4.8.2 | ||
| 100 | Fixed nuisance | — | 4.8.2 |
Table 3. Inputs of the likelihood estimator, Equation (43), and the model it evaluates, Equation (42). The spectral values are roughly estimated to demonstrate relative significance, and are approximately the same in both patches, except , quoted as RA23 / RA4.5; no central value is quoted for the measured bandpowers because they scatter about zero, so only their scale is meaningful. Each quantity is classified by the role it plays: the measured bandpowers are the data; sampled parameters carry their own chain dimension, with the stated distribution acting as a prior except for , which is constrained by the likelihood itself; fixed nuisances are held at their central values during sampling and marginalized afterwards. Uncertainties are and carry the units of the adjacent value; that of is set per bandpower by the scatter of the Jupiter observations rather than by a single number. The “Dim.” column gives the number of chain dimensions each sampled quantity contributes, with bandpowers and counted per patch, so that the per-patch chains are dimensional and the joint chain of Section 4.10, which carries a separate for each patch, is dimensional. enters through the single amplitude of Equation (44) rather than one dimension per bandpower, and the uncertainty quoted for it is that at the highest bandpower. The last column gives the section in which each quantity is determined.
Pointing
The PB-2b pointing model is a 14-parameter TPoint-style correction (P. T. Wallace [66]) that maps the encoder azimuth/elevation to the true boresight pointing before the az/el RA/Dec conversion, which is a more expansive model than the five-parameter model used in PB14. The corrections (in arcminutes) decompose into structural, thermal-flexure, solar-flexure, and second-harmonic “nod” terms. The 14 coefficients follow F. T. Matsuda [38] and are fit by minimizing a against a catalog of point-source centroid offsets. The fit leaves a per-axis residual scatter of and (≈ total). A few corrections and physical improvements to the model were found to be necessary, detailed in J. Ito et al. (2026, in preparation). These adopt parameters from the pointing models of the Atacama Large Millimeter/submillimeter Array (ALMA) (J. G. Mangum [36]) and the Very Long Baseline Interferometry (VLBI) Field System (W. E. Himwich [24]) as adopted by Square Kilometre Array Observatory (SKAO) (SKA Observatory [58]), reducing the RMS further. However, as the existing model already has reduced pointing suppression significantly from the no-model case, as shown in Figure 2, and its residual uncertainty is carried through the likelihood as a sampled parameter, as discussed in § 4.10, the adjustments were not implemented due to computational restrictions. A comparison of the two models is described in that work, showing comparable performance. The updated model lowers the residual pointing RMS from to , which changes the pointing transfer function by at most 0.13% for , no more than of its sampled uncertainty, a negligible difference.
Figure 2. Suppression factor in -space with (solid line) and without (dashed line) pointing model correction. The dotted black line shows a Gaussian suppression from the determined RMS for reference; both patches are Gaussian at low- but have less suppression at higher .
Because the relevant question for this analysis is how the pointing affects the recovered spectrum, we validate the model with geometry-only simulations, isolating the az/el RA/Dec geometry to quantify the impact as an effective amplitude transfer function. Concretely, we synthesize a temperature-only sky with the Boltzmann code CAMB (A. Lewis et al. [33]) at the Planck best-fit cosmology (Planck Collaboration et al. [48]), sample it along the azimuth and elevation actually recorded during the real observations, both with and without the pointing model applied, and coadd the resulting maps following the same procedure used for the data. A pointing error is a purely geometric remapping of where each sample lands on the sky, so its effect on the recovered spectrum does not depend on which Stokes parameter carries the signal; a -only input sky is therefore sufficient, and this is what “geometry-only” denotes here. The model transfer function is in amplitude out to . The uncorrected maps yield pointing scatter up to 5% at , compared to for those with the pointing correction, as shown in Figure 2. In addition, by propagating the parameter covariance and cosmic variance through Monte Carlo simulations into -space, we find a pointing suppression uncertainty at the highest bin. Because this term arises from smearing it grows as , so it is negligible below but becomes comparable to other spectral sensitivity systematics such as the beam suppression and mapmaking pipeline transfer function (visible in Figure 5) in the highest bandpowers. It is therefore sampled in the likelihood rather than propagated afterwards (§ 4.10).
Figure 5. Per-mode amplitude transmission of the analysis versus multipole. Dashed steps show the simulation transfer function (Equation (33)) for RA23 (blue) and RA4.5 (orange), the dotted lines represent the pointing transfer function for each patch (blue and orange), and the solid teal line represents the independently measured beam window function . The solid blue and orange curves represent the product per patch, with bands from the combined , , and uncertainties. The two patches are nearly identical.
Beam Characterization
To estimate the beam window function at 90 GHz, we analyze 68 Jupiter observations taken over our observing season, extending the beam analysis in N. Farias et al. [15] (see Figure 8 of that work for the inverse-variance average beam map). Jupiter is chosen because it is a bright point-like source that is negligibly polarized at 90 GHz. Each map is apodized with a 2′ two-dimensional cosine taper so it smoothly tapers to zero at the edge, then zero-padded for finer -sampling. A 2D fast Fourier transform (FFT) is performed on the resulting padded map, and the result is normalized by the DC mode (effectively the monopole term in a flat-sky approximation). This is then divided by the Fourier transform of the Airy disk to deconvolve the planet’s angular size, which is determined from ephemeris data for each observation date. The resulting function in -space is squared, radially binned in rings to match the science binning scheme (with denoting the bin center hereafter), and the square root is taken. Finally, we divide by the pixel-window function , where is the map resolution.
The resulting beam window function is shown in Figure 5. As noted in Figure 8 from N. Farias et al. [15], the beams are not perfectly radially symmetric, despite the radial-symmetry assumption used in the binning above. This, in conjunction with the statistical uncertainty, is reflected in the error bars in the plot, computed as standard errors of the points in each -bin. Observing over a range of azimuths and elevations provides multiple parallactic angles that mitigate the azimuthal artifacts in the beam, as discussed in § 2.3. The ellipticity of the 90 GHz beam is quoted at 3% (N. Farias et al. [15]), sufficiently low to contribute negligibly to the overall systematic uncertainty, following the PoLARbear analyses ((Polarbear Collaboration et al. 2017, 2020)).
Time Constants and Absolute Gain Calibration
PB-2b detectors are about four times slower than Polarbear’s, with a median effective time constant of 12.1 ms at 90 GHz ([15]) compared to ms ([49] (Polarbear Collaboration et al. 2017)). This imposes non-negligible attenuation and phase retardance on the polarization signal at the modulation frequency , and the timestreams therefore require deconvolution before demodulation. As discussed in N. Farias et al. [15], effective time constants are determined by fitting the first-harmonic amplitude of a chopped thermal signal (known hereafter as a stimulator) to a single-pole model at a series of seven frequencies; 5 Hz, 9 Hz, 13 Hz, 19 Hz, 27 Hz, 37 Hz, and 44 Hz; so chosen to avoid the significant HWP harmonics. Fit parameters from this methodology are saved for later use as long as they produce plausibly physical values (single-pole model amplitude analog-to-digital converter (ADC) counts and ms) and have at least three surviving frequency points within the sweep. These calibration runs are taken immediately before and after each science observation, and the effective time constant used is their mean, . We deconvolve per timestream before demodulation by Fourier-transforming and dividing by the single-pole model evaluated at this time constant.
Absolute gain calibration factors convert the instrument’s measured current units (ADC counts) into K. The stimulator, operating simultaneously with a Jupiter observation, is used to determine these factors. Jupiter’s known thermodynamic temperature, measured by Planck Collaboration et al. [46], is first converted into a Rayleigh–Jeans brightness temperature and then into an equivalent K temperature fluctuation. This conversion incorporates the Simons Array instrument characterization ([67, 62]), including the effective detector bandpasses and Lyot efficiencies ( and 0.849 for the 90 GHz and 150 GHz bands, respectively). Furthermore, to account for sources with different spectral indices, color-correction scaling factors are calculated internally by integrating the source spectra over the instrument’s simulated diplexer bandpass. The 90 GHz bandpass is flat, spanning 76.1 GHz to 94.3 GHz, utilizing the Fourier-transform spectrometer measurement adopted by K. Sakaguri et al. [55], which gives a band center of 85.18 GHz and a bandwidth of 18.25 GHz (a fractional bandwidth of 0.214). The color correction is normalized to a Rayleigh–Jeans source, for a flux density , which is appropriate for Jupiter. The spectrum relevant to CMB anisotropies is instead , whose effective index is at 90 GHz, where ; the resulting correction relative to the Rayleigh–Jeans normalization is .
The QWP conversion efficiency does not enter this absolute gain calibration, which is an intensity measurement. Its frequency dependence across the band instead enters through the band-averaged value used in (Equation (2a)). As in the typical planet scans, we fit a 2D Gaussian to the source in ADC counts. By comparing the amplitude of this Gaussian fit and the measured beam size to the expected temperature and solid angle of Jupiter, a conversion factor in K/ADC is determined. Ten such Jupiter scans with the stimulator were used to build the absolute gain calibration template database.
As a secondary calibration source, we use the HWPSS, a spurious non-sky signal generated due to imperfections in the HWP, following a procedure similar to the relative gain calibration in K. Sakaguri et al. [55]. We build a template of calibration factors from observations with reliable stimulator calibration, determining, per detector, a conversion between the amplitude of the HWPSS fourth harmonic (at ) and the equivalent stimulator amplitude. Likewise, we estimate the effective time constant from the phase. The phase is primarily set by the physical angle between each detector and the QWP axis, but the time constant introduces an additional phase offset. We build a template for the phase corresponding to the physical angle; the difference between each observation’s phase and this template gives the effective time constant. This method agrees well with the stimulator-based approach, as validated on observations for which both stimulator calibration and the HWPSS signal are available. This alternative method is utilized when there are technical issues with the stimulator that prevent it from being used as a reliable calibration source.
The time constant during an observation may differ from the mean used for deconvolution to the extent that it drifts between the two calibrations, which are h apart. Taking the true value to lie uniformly between them gives for each detector-observation, which the single-pole model converts into a fractional error on the deconvolution gain of , with . The drift is largely common to the array within one calibration interval, so we treat each interval’s error as coherent across its detectors and independent between intervals, and average over each patch with the map weights. Time constants from the HWPSS phase are measured within the observation and carry no such term. The result is a unit-mean gain factor , with uncertainty 0.37% for RA23 and 0.62% for RA4.5, which multiplies the data in the same way as . An overall bias in the stimulator time constants does not enter , as it cancels against the calibration of (§ 2.2).
Because the gain enters the measured spectrum only as an overall factor of , a fractional calibration error rescales by twice that fraction at every multipole—its effect is therefore analytic rather than something that must be propagated by simulation, and we account for it explicitly through the residual gain coefficient defined in Equation (8). All science observations analyzed here were taken in the QWP configuration, and the applied calibration is therefore the mean over only the four planet observations available in that configuration; their 8.9 % common-mode scatter averages down to 4.45 %, which dominates the budget. The remaining terms are the Jupiter brightness-temperature model of Planck Collaboration et al. [46] ( %), the per-detector calibration scatter (7.6 % across 706 detectors, or 0.29 % after averaging), and the HWPSS template fit error ( %). Summing in quadrature gives a total absolute gain uncertainty of %, corresponding to a 9.1 % multiplicative scale uncertainty on the quoted limits. Rather than propagate this through the pipeline, we carry it into the model of § 4.9 as a prior on .
Detector Polarization Angles
The QWP converts sky into the Stokes of its own frame, for which the efficiency and leakage terms of Equations (2a)–(2d) are defined. Circular polarization is invariant under rotation about the line of sight, so were the only signal present no orientation reference would be needed. It is the linear polarization, spin-2 and hence basis-dependent, that forces one—a detector whose polarization axis lies at an angle to the QWP axis measures and mixed by , so without correcting for it each detector would fold an uncontrolled fraction of sky linear polarization into its circular polarization channel. We therefore rotate every detector into the common QWP basis, both to keep sky linear polarization out of the measured and because , , and are themselves ratios of Mueller elements about the QWP axis, quantifying the leakage only once the data are expressed in that frame.
Because the HWPSS fourth-harmonic signal is physically related to the QWP transmission (K. Sakaguri et al. [55]; N. Farias et al. [15]) and is dominant over the other harmonics by several orders of magnitude, we can calibrate these angles directly from the timestream data. This permits us to rotate from the individual detector orientations into the common QWP basis. For each detector, the mean-subtracted timestream prior to HWPSS removal is binned into 128 uniform bins in the encoder-measured HWP angle over . Because the samples are sorted by measured angle rather than by time, no constant rotation rate is assumed and departures from the nominal rotation rate are absorbed by the binning. The resulting binned HWPSS profile is then fit by least squares to a harmonic series in ,
The physical alignment angle per detector is determined from the fourth-harmonic amplitudes as
which is defined on .
Data Analysis
This section describes the analysis pipeline in the order in which the data flow through it. Timestreams are calibrated, cleaned, and demodulated into per-detector Stokes components (§ 4.2); these are binned into per-observation maps (§ 4.3) and screened by the data quality cuts of § 4.4. The surviving maps are combined into cross-spectra, yielding the measured bandpowers and their covariance (§ 4.5), whose internal consistency is then established with the null test suite of § 4.6. The remaining sections quantify the terms needed to invert those bandpowers into a circular polarization spectrum: the pipeline transfer function (§ 4.7) and the instrumental leakages (§ 4.8). These are assembled into the forward model described in § 4.9, which is inverted in the likelihood of § 4.10 to give . Figure 3 traces the Stokes parameters through this chain and defines the subscripts used throughout.
Figure 3. Schematic describing the Stokes parameters at each part of the observing and analysis pipeline, with indicated leakages and efficiencies and associated bases. This is the path traced by Figure 1. The subscripts label successive stages along this path: “sky” denotes the true celestial signal; “in” the same signal transformed into the telescope basis; “conv” the Stokes parameters after the QWP has applied its phase delay, converting into and vice versa; “mod” the , , and of the converted signal modulated by the rotating HWP into a detector timestream; and “obs” the signal actually recorded by the detectors, either in the time domain as a function of or in the map domain as a function of . This figure is based on Figure 7 from K. Sakaguri et al. [55].
Blinding Scheme
To prevent analysis choices from being biased, consciously or otherwise, toward the final circular polarization result, we developed the analysis pipeline on a benchmark dataset with full access to its power spectra, then applied it to the science data under a blinding scheme that hides the bandpowers. The benchmark set comprises % of the usable RA23 data, on which we had unblinded access to its power spectra. These observations were selected on observing conditions and detector yield alone, requiring more than 500 tuned and optical detectors and a precipitable water vapor (PWV; a measure of water vapor in the line of sight to the zenith) below 3 mm, so that pipeline development was carried out on data of reliably good quality rather than on observations chosen for any property of their recovered spectra. All analysis choices were initially set using this benchmark set alone, including the time-domain filtering parameters, glitch-flagging and HWPSS-removal procedures, data-quality cut thresholds (§ 4.4), the power-spectrum estimator configuration (§ 4.5), and the leakage and transfer-function treatments (§ 4.7 and § 4.8). Because the RA4.5 observations were acquired chronologically later, and to maintain a uniform analysis pipeline across both sky patches, we did not define a separate benchmark dataset for RA4.5. Instead, the analysis pipeline developed on the RA23 benchmark was applied directly to the RA4.5 dataset, validated by the success of the RA4.5 null tests (§ 4.6). A separate set of RA4.5 observations was set aside for the transfer-function simulations of § 4.7. It was not used to develop or validate the pipeline, so it was not excluded from the RA4.5 science data.
We then applied this pipeline to the remaining of the data. Throughout this stage the bandpowers themselves were blinded; while the intermediate pipeline products (filtered timestreams, detector and coadded maps, and the various calibration and leakage estimates) remained accessible, the recovered science spectra were withheld from inspection. The null tests of § 4.6 were run in this blinded configuration for both patches, with only the summary probabilities-to-exceed (PTEs) exposed and the underlying bandpowers hidden, so that internal-consistency failures could be diagnosed without sight of the science spectra. Failures were addressed with additional data-quality cuts and adjustments to the pipeline, made in parallel and based only on these PTEs and on secondary data products such as the timestreams and maps; the bandpowers, and hence the upper limits derived from them, remained blinded throughout. The unblinding sequence was fixed in advance rather than settled as results became available. The pipeline was frozen, with no analysis choice or data cut changed thereafter, once the null-test passage criteria of § 4.6 were met, and only then were the bandpowers revealed. The benchmark observations are themselves excluded from the final science dataset, at a sensitivity cost, so that the limits reported in § 5 derive entirely from data whose spectra were unseen during pipeline construction.
Timestream Processing
In this section we will detail the data processing that occurs from the intake of raw modulated detector data to isolated Stokes timestreams, including the time-domain filters meant to improve the quality of the data. A schematic of how the data passes through each component, both physically in the telescope and later in the analysis pipeline, is shown in Figure 3.
Calibration and Modulation
The raw detector data are first converted from ADC counts into K using the procedures in § 3.3. That conversion is a single scalar responsivity value per detector, which propagates to being a scalar multiplier in -space. We parameterize any residual error in it with a dimensionless coefficient ,
defined as the ratio of the applied calibration factor to the true one, where is what a perfectly calibrated detector would record. The calibration of § 3.3 has already been applied to the data, so is unity by construction and only quantifies how far that calibration may be from the truth, with a width given by the absolute gain uncertainty derived there. As mentioned in § 3.3, we account for its residual uncertainty once, in spectrum space (§ 4.9), to avoid propagating the parameter through hundreds of thousands of individual detector timestreams.
The time constants determined from the same procedures are also deconvolved at first order. We also mask the first and last second of each subscan, a single left- or right-going sweep in azimuth at constant elevation, to exclude transients from the turnaround between sweeps. At this point, we also determine the corresponding RA and declination (Dec) based on each detector’s position in the focal plane and the focal plane’s boresight pointing at a given timestamp, and these are corrected according to the pointing model from § 3.1. The HWP encoder data is also reconstructed per the procedure discussed in B. Bixler et al. [7], which is used in the following subsection to demodulate into individual Stokes components.
As discussed in K. Sakaguri et al. [55], the HWP modulates polarized signal according to the following equation for each detector ,
where “conv” represents the Stokes components after passing through the QWP, is the angle of the HWP as it rotates, is the detector’s polarization angle relative to the QWP axis (§ 3.4), is the HWPSS signal, is white noise, and is low-frequency noise caused by changes in atmospheric loading. In addition, as noted in § 2.2, the detector response is not strictly linear, so a second-order term mixes the HWP-synchronous signal with the unmodulated intensity and produces sidebands about that survive demodulation. This leaks the intensity timestream into the demodulated Stokes components, which we denote with the complex leakage parameter , where the subscript denotes that it originates in the detector response rather than the optics. The same term was identified in POLARBEAR as an excess leakage and derived from the bolometer response model ([64, 63]).
HWPSS Removal
The first step before demodulating is to remove the HWPSS, which we model as a sum of harmonics of the HWP angle with slowly varying complex amplitudes,
Following POLARBEAR Collaboration et al. [53], for each we band-pass the timestream within Hz of (1023-tap finite impulse response filter), demodulate by , and low-pass at 0.1 Hz to obtain , then subtract the reconstructed harmonic before proceeding to the next. Notably, is included, so any polarization sky signal that varies slowly enough to lie within Hz of is subtracted along with the HWPSS, while faster HWP-modulated polarization signal content (at ) is preserved for downstream demodulation. The effect this has on real data is encapsulated when we calculate the pipeline transfer function, which will be discussed in § 4.7.
Filtering and Glitch Removal
The modulated data is put through two polynomial filters, one on the entire timestream and one per subscan. Both utilize Chebyshev polynomials, the former with and the latter with . These are designed to remove the low-frequency noise primarily associated with fluctuations in atmospheric loading and the gain drifts that accompany them. These dominate the low-frequency content of the modulated timestream because the bulk atmospheric emission is unpolarized and therefore not modulated by the HWP.
In addition, during timestream processing, we identified populations of transient glitches and scan-synchronous spikes. Transient glitches are well documented in bolometric detector arrays generally, where energetic particle interactions with the silicon detector wafers produce events spanning a wide range of amplitudes and time constants ([9]). The scan-synchronous component is instead tied to the observing environment, arising from far-sidelobe ground pickup and from radio frequency interference (RFI) coupling into the multiplexed readout ([5]), the latter also being the suspected origin of the excess noise reported for PB-2b ([15]). Because the precise coupling mechanisms of out-of-band RFI and ground emission cannot be easily modeled from first principles, we employ a masking procedure and replacement with Gaussian noise for the purposes of demodulation. First, the autocorrelation function (ACF) is used as a convergence criterion; glitches are repeatedly flagged and the ACF is repeatedly recomputed, cleaning until the maximum ACF in the 5–400 s lag window drops below 0.08 or after four iterations. Each pass flags samples whose median-subtracted absolute deviation exceeds , then extends each flagged region outward from its peak to the nearest local minimum on either side to capture the glitch’s full rise and fall, saving the mask. The masked regions are replaced with a linear baseline interpolated between the five-sample means on either side, plus a realization of white Gaussian noise scaled to that detector’s 6.5 Hz to 7.5 Hz noise floor. The noise is multiplied by a half-sine window that goes to zero at the first and last masked sample, so that no noise discontinuity is spliced against the surrounding real data. This Gaussian-filled timestream is used for demodulation, but those timestamps remain masked for later mapmaking.
Demodulation
This section primarily follows the framework described in S. Takakura et al. [64]. To demodulate, we apply a bandpass filter spanning Hz to the modulated signal, isolating the polarization modulation at . This effectively removes the intensity signal, as with the scan speed of the lower bound of our bandpass is above the beam-imposed diffraction limit of the intensity timestream ( with beam size [15]). This allows us to retain resolving power while not contaminating the modulated polarized signal with low-frequency noise. The remaining signal after HWPSS removal and bandpass filtering is, with for brevity (where the detector polarization angle is calibrated per § 3.4),
where “bp” denotes a bandpass-filtered timestream. This is then shifted to baseband, i.e., the polarization signal is moved from the carrier to zero frequency, by multiplying Equation (10) by , resulting in
where “bb” stands for “baseband.” A low-pass filter is then applied at , which removes the term shifted up to . The remaining modulated timestream is,
The timestream is then rotated into the QWP axis basis by multiplying by ,
with as defined above. This removes the detector-angle dependence of the sky terms exactly, so the real and imaginary components may be saved directly as and without any residual dependence on . The intensity is recovered separately, from the unmodulated low-frequency component of the timestream (denoted “lp”) rather than from the band around , so it carries no factor and retains the full low-frequency atmospheric noise. Together,
where the intensity leakage coefficients are the real and imaginary parts of ,
and the noise terms follow the same demodulation as the sky signal. The intensity inherits the low-pass-filtered noise directly, , while the polarization channels carry the bandpassed noise projected onto the two demodulation quadratures, and .
To express this in terms of the Stokes components on the sky in the QWP basis, we use the effective matrix of § 2.2, . For the observed Stokes ,
with , assuming negligible polarization leakage to intensity [55]. Using Equations (2a)–(2d), the observed Stokes becomes
The intensity term of Equation (19) collects two contributions of distinct origin: , generated downstream of the optics in the nonlinear detector response, and , any static intensity leakage in the optics. We assume a value for neither, and therefore define the total intensity-to--channel leakage, including the transferred -to- as well as the downstream -to-, to be
as the single quantity carried through the remainder of this work. The observed Stokes timestream is then, in final form,
Because is measured as a ratio of beam window functions (Section 4.8.1), it is in general multipole dependent, and is written wherever it appears in -space. As mentioned earlier, Equation (17) will be discussed in more detail in § 4.8. From the bandpass, the two noise terms are between Hz in baseband, which corresponds to between Hz to Hz in the modulated regime. Generally the atmospheric low-frequency noise shows up below Hz, which is primarily decimated by the bandpass, and while the white noise is flat in frequency space, it becomes independent in and and is independent detector-to-detector ([31]), so it averages down with more observing time.
Additional Filtering
In the event that low-frequency noise is not fully mitigated by the initial bandpass filtering, we reapply the polynomial filters described in § 4.2.3. This residual noise is mostly unpolarized and driven by inhomogeneously distributed, turbulent water vapor and line-of-sight temperature gradients. In addition, because this secondary filtering stage acts on the demodulated timestreams, it also suppresses circularly polarized emission from atmospheric oxygen via Zeeman splitting ([22, 59, 45, 19], see also Appendix § 10). Unlike water vapor, oxygen is well-mixed, producing a signal that varies smoothly with the telescope’s pointing relative to Earth’s magnetic field. While this signal is carried at through the initial filtering stages, demodulation downconverts it to low frequencies. Because the telescope sweeps monotonically in azimuth during a subscan, a low-order polynomial filter in the time domain acts effectively as a low-order expansion in azimuth, absorbing the bulk of this smoothly varying atmospheric signal. Any surviving residuals remain tied to telescope azimuth rather than celestial coordinates, and are subsequently removed by the scan-synchronous filter (see Appendix § 10).
To address this remaining azimuth-locked pickup, such as ground spillover and far-sidelobe emission that recur at fixed pointings, we apply a scan-synchronous filter. Because left- and right-going sweeps generally couple differently to the local topography, the timestreams for each detector are divided into individual subscans and processed independently by scan direction. Within each direction, the subscans are pooled to construct an empirical template of the detector’s azimuthal response. The data are binned in fixed azimuth intervals, and the median of each bin is taken to suppress transient noise and outliers; bins with insufficient samples are discarded. We then fit a Chebyshev polynomial of degree to this profile, weighted by the intra-bin variance so that well-determined bins govern the fit. Finally, this smooth, direction-dependent azimuth template is evaluated at each sample’s true pointing and subtracted directly from the timestream.
Mapmaking
The result at the end of this processing is a set of timestream data, consisting of , , and , for which we have associated turnaround buffer and glitch masks. For each detector in each observation we construct flat-sky maps of its associated timestream. A tangent-plane (gnomonic) projection is defined about the field center, spanning on a side at a fixed pixel scale with width , and detectors are mapped independently before being combined. The best-fit planar gradient in sky coordinates is then subtracted from each detector’s retained samples, suppressing residual large-scale scan-synchronous and atmospheric gradients that survive the time-domain filtering. The cleaned samples are accumulated into the map by inverse-variance weighting, scaled according to the number of timestamp hits within a given pixel. Following collection of all detector maps within a single observation, the coadded observation map is computed as the per-pixel inverse-variance-weighted mean of the detector maps, utilizing the weight map saved alongside each map. This produces one map per CES, yielding 448 maps for RA23 and 341 for RA4.5 that survive the data quality cuts of § 4.4.
At /pix () over a field, the map oversamples the science band. The pixel Nyquist multipole, , sits well above the highest analyzed multipole (a factor in margin), so aliasing from pixelization is negligible over the analyzed multipoles. Pixelization does still smooth the map, suppressing power by the azimuthally averaged pixel window function,
where is the relative azimuthal angle of the 2D Fourier mode. The value of this function is at and at . This suppression is folded into the measurement of the transfer function, discussed in § 4.7, as the simulated maps are generated at the same resolution as the observed maps. The pixel window therefore enters the forward model of § 4.9 exactly once, through . The explicit of Equation (22) is used only in Equation (38), to remove a second copy from a quantity that would otherwise carry one.
Data Quality Cuts
We impose additional data quality cuts to ensure that known systematics do not propagate into the final analysis, motivated by the null test suite of § 4.6. We split these cuts into both observation-level and detector-level cuts, as was done in the null tests. The yields in the following table are relative to the previous cut. Instrument loss refers to detector-observations lost to the technical failures described in § 2.3, such as corrupted timestream files, improperly stored database identifiers, or failed detector tuning. It also includes detectors that lack time constants or absolute gain calibration factors. The two observation cuts are whether a given run has associated temperature and PWV from the Atacama Pathfinder EXperiment (APEX, km from the PB-2b site) and focal plane temperatures within 0.30 K to 0.38 K. APEX weather readings were required to run null tests with associated criteria and the focal plane temperature ranges exclude a focal plane with elevated temperatures. With the detector splits, we eliminate detectors with extreme non-physical limits of calibration parameters. Gain calibration factors must be negative (due to TES polarity) and mK/ADC, time constant must be ms, and measured white noise must be between K and mK. These ensure the calibration factors are physical and the saved maps are within expected bounds. Finally, we make selections based on quantities derived from the maps themselves. Detector maps with known issues, such as spatial artifacts resulting from filtering, HWPSS removal, or glitch masking, are intentionally not used. These only appear in the RA23 patch. Runs with pixels that point to the edge of the patch footprint are not used; this is typically reflective of the boresight getting stuck and not tracking properly as the sky rotates. Finally, detector maps with abnormally high RMS, mK, are also dropped, as a final removal of any detector maps that are unphysical yet survive the earlier cuts.
The results of these cuts, as well as the earlier losses due to instrument technical issues, are represented in Table 1.
| Selection Type | RA23 | RA4.5 |
| Instrument Loss Yield | 62.9% | 56.2% |
| Observation selection | 82.0% | 83.2% |
| Weather data exists | 98.1% | 95.6% |
| Focal plane temperature | 83.6% | 87.0% |
| Detector selection | 89.3% | 83.2% |
| Gain calibration factor | 92.3% | 84.6% |
| Time constant | 99.6% | 99.4% |
| White noise | 97.1% | 99.0% |
| Map selection | 94.1% | 91.3% |
| Map artifacts | 96.1% | 100% |
| Boresight pointing drift | 99.9% | 92.5% |
| High variance | 98.1% | 98.8% |
| Overall Yield | 43.3% | 35.5% |
Table 1. Summary of data retention yields from saved detector-observations. The first row reflects the instrument losses described in § 2.3, both observation-wide and from detectors lacking calibration factors; all subsequent yields are relative to the preceding cut.
Power Spectrum Estimation
MASTER Estimator
The angular auto-power spectrum is estimated from a set of cross-power spectra, computed from pairs of independently coadded maps using the flat-sky implementation of the MASTER (Monte-Carlo Apodised Spherical Transform Estimator) pseudo-power-spectrum formalism via NaMASTER and its Python wrapper pymaster ([25, 3]). Because each pair of maps is built from two disjoint sets of observations, their instrumental noise is uncorrelated and approximately averages to zero in the cross-correlation, so the resulting bandpowers are an unbiased estimate of the sky signal free of the noise-bias term that contaminates a single-map auto-spectrum. Each map is treated as a small patch of the celestial sphere projected onto a flat grid in right ascension and declination; the physical angular extent of the grid is set by the patch size, with the right-ascension axis rescaled by the cosine of the central declination to account for the convergence of meridians, so that the two-dimensional Fourier modes correspond to well-defined angular multipoles.
Before the transform, each map is lightly conditioned. Unobserved pixels are set to zero, the pixel-value distribution is iteratively sigma-clipped to obtain a robust per-map noise amplitude used to weight that pair relative to others, and a weighted monopole is subtracted so that the mean within the observed region vanishes. The coverage mask for each field is the region of valid coverage, defined as pixels with hit count dB of the maximum. The threshold is set logarithmically rather than as a linear fraction because the hit count of a constant-elevation scan falls off across more than an order of magnitude between the deeply covered center of the field and its turnaround-dominated perimeter, so a decibel cut places the boundary on the natural scale of the coverage profile. The dB value, retaining pixels above of the peak hit count, balances sky area against edge quality. Pixels below it carry the highest per-pixel noise and the largest residuals from the ground-template and polynomial filtering of § 4.2, so they contribute little statistical weight relative to the systematic risk they introduce; they also lie where the coverage boundary becomes most irregular, which broadens the mode-coupling kernel and degrades the conditioning of the binned coupling matrix. For RA4.5 a -radius circular hole is additionally cut at the position of the bright point source J04404333 (see Appendix § 11) (M. Bonato et al. [8]), the only significant potentially contaminating point source within either footprint (K. Sakaguri et al. [55]), so that its compact emission does not contribute to the bandpowers. The mask is apodized with a -wide cosine taper. At five map pixels, this width is well resolved by the grid, so its smooth roll-off replaces the slowly decaying mode-coupling sidelobes of a sharp edge with a compact kernel that keeps the binned coupling matrix well conditioned. It nonetheless remains small compared to both the effective footprint retained by the hit-count threshold and the angular wavelength of the lowest analyzed multipole (). The apodized footprint is not applied as a binary weight; it is multiplied by the per-pixel inverse-variance weight map from mapmaking (§ 4.3), and it is the power spectrum of this combined weight map that sets the mode-coupling matrix.
The masked, apodized maps are then carried through the MASTER estimator, which deconvolves the mode coupling induced by the finite apodized footprint using the coupling matrix computed from the power spectrum of the mask itself (E. Hivon et al. [25]). Because that matrix is singular at the level of individual multipoles, the spectrum is first averaged into uniform bandpowers over the range where the binned matrix can be inverted. No beam is deconvolved at this stage, so these bandpowers remain beam-smoothed. The beam is removed only in the final inversion of § 4.9.
In the context of this study, this MASTER estimation procedure is primarily used to determine the measured circular polarization spectra as cross-spectra between pairs of independently coadded observation maps. However, we also use this estimator within the context of simulated maps, in order to estimate transfer functions (see § 4.7) and leakages (see § 4.8).
Measured UU Spectra
The quantity accessible directly from the maps is the power spectrum of , not the circular polarization spectrum itself. As shown in Equation (21), the observed carries the sky signal converted by the QWP, but scaled by the circular polarization efficiency and accompanied by leakage from sky , , and , as well as suppression by the instrument beam, residual pointing error, and the filtering applied to the timestreams. Recovering therefore requires measuring the observed bandpowers first and then disentangling each of these contributions from them. This section presents that measurement; the transfer function and leakage terms are quantified in § 4.7 and § 4.8, and the inversion that yields is carried out in § 4.9.
We estimate the average of the cross-spectra from the set of coadded observation maps, which are immune to the noise bias of an auto-spectrum because the noise in two different maps is uncorrelated. Forming cross-spectra between the individual CES maps of § 4.3 would require of order pairs per patch. These are therefore first combined into a smaller number of deeper maps. Each patch is observed in blocks of several consecutive CESes, and the maps within a block are coadded together. Each CES contributes to exactly one such map, so no observation is shared between any two of them and their noise remains uncorrelated. This reduces the 448 CES maps of RA23 and the 341 of RA4.5 to and independent maps, respectively, which we index by . We then form the cross-power spectra of all distinct map pairs, indexed by , where the pair is formed from the two distinct maps and . Each pair spectrum is computed with the flat-sky MASTER estimator on monopole-subtracted, apodized maps and binned into bandpowers over . The pairs are not equally informative, because the maps entering them differ in depth and different pairs overlap over different fractions of the patch. A uniform average would therefore be suboptimal, and the bandpower estimate is instead formed as an inverse-variance weighted mean over pairs,
which is the quantity plotted as the data points. Throughout, a hat denotes a quantity estimated from the data and a bar an average, so that is a single measured cross-spectrum and its mean over pairs; spectra carrying a bar alone, such as the leakage terms of § 4.8, are averages over simulations rather than measurements. The one exception is , which denotes a direction on the sky.
The weight is a single -independent inverse-noise-variance factor per pair,
where and are per-map noise proxies for the two maps forming pair , each obtained by taking the standard deviation of the unmasked pixels, discarding those deviating by more than , and adopting the standard deviation of the pixels that remain. The effective sky fraction is the mean over pixels of the product of those two maps’ apodized coverage masks. Because the cross-spectrum variance scales as , Equation (24) is the corresponding inverse-variance weight; the common sky-signal contribution to the map variance is shared by every pair and cancels in Equation (23), so a scalar weight encoding only per-pair noise and mask overlap well approximates the optimal weight.
Estimating the uncertainty on requires accounting for correlations between the pair spectra, because the pairs are formed from a common set of maps, with each map entering pairs. In principle this couples pairs that share a map, which would make the maps, rather than the pairs, the independent units. Two such pairs are correlated, however, only through a component common to their unshared maps. These maps are dominated by instrument noise, which is independent from map to map, and each is itself a coadd of separate observations, so their atmospheric fluctuations are independent as well. The only component common to every map is the sky signal itself. Theoretical predictions place CMB circular polarization well below our sensitivity [30, 15], so this shared component is expected to be negligible. We confirm this directly by estimating the shared-map covariance between distinct pairs and that have one of their two maps in common, which is consistent with zero in every -bin; the pairs are therefore effectively independent. We accordingly estimate the covariance of with a pair-level bootstrap. In each of bootstrap realizations, indexed by and labeled throughout by the superscript , we draw pairs with replacement and recompute the weighted mean of Equation (23) over the resampled set,
where is the multiset of pair indices drawn in realization , so that is the weighted mean recovered from that realization. The bandpower covariance is the sample covariance of the resampled means,
where averages over realizations; the uncertainty per -bin is . The covariance is therefore measured directly from the resampled scatter, rather than assumed from an analytic form evaluated at a fiducial spectrum; the Gaussian likelihood of § 4.10 uses only this measured second moment. We verified that the alternative map-level bootstrap, in which whole maps rather than pairs are resampled and each pair is weighted by the product of the number of times its two maps were drawn, inflates the bandpower uncertainties by a factor of ; since , this excess reflects the scatter in those draw counts from one realization to the next rather than genuine map-to-map correlation, and would bias the covariance high.
Null Tests
We assess the internal consistency of the power spectrum, measured in § 4.5, with a suite of sign-flip jackknife null tests, following the framework developed for earlier POLARBEAR analyses [52, 53]. Each test is defined by a criterion that partitions the data into two halves according to a particular systematic. The null test criteria are listed in Appendix § 9 and are adapted from PB24, with a few additional criteria specific to PB-2b. Differencing the two half-coadds cancels the common sky signal, so the resulting null bandpower spectrum (estimated via MASTER as discussed in § 4.5.1) is consistent with zero in the absence of systematic contamination or miscalibration. Throughout this section indexes the same bandpowers over used elsewhere in this work; we label by the discrete index rather than their center only to keep the sums and maxima over bins compact. The null tests and the data quality cuts of § 4.4 were developed together rather than in sequence—a criterion that failed motivated a candidate cut or pipeline adjustment, which was applied before the full suite was rerun, iterating until every criterion passed. The cuts quoted in § 4.4 are the endpoint of that loop, and the PTEs reported below are those of the final iteration.
The null distribution for each criterion is estimated directly from the data by Monte Carlo simulation. We generate random sign-flip realizations , , each constructed identically to the data null spectrum except that the observations are assigned to the two halves at random rather than by the criterion . Each realization therefore inherits the noise, filtering, and survey geometry of the data but contains no coherent split-dependent signal. From this ensemble we form the per-criterion mean and standard deviation , the standardized per-bin residuals of the data, , and the bandpower covariance , which defines a covariance-weighted statistic for each criterion,
where boldface denotes vectors over the bandpower index . Evaluated on the data split, Equation (28) yields the observed statistic , while the identical expression applied to each random shuffle (replacing with ) yields its null distribution . Because the same statistics are evaluated on every shuffle, each quantity carries its own empirical null distribution and we make no assumption about its analytic form. A PTE is read off as
where is the number of shuffle realizations whose value of the statistic is at least as extreme as the observed (data) value, so that the fraction of trials (data included) at least as extreme as the observed value is well defined and unbiased at finite .
To avoid reporting many correlated PTEs, we compress the full set of residuals into four summary statistics, each sensitive to a distinct failure mode. These are the grand mean , which detects a coherent additive bias and is evaluated two-sided; the largest single excursion , sensitive to one anomalous bandpower; the worst criterion from Equation (28), sensitive to one bad split; and the total , which captures a diffuse, low-level excess spread across many splits. Each statistic is referred to its own shuffle distribution via Equation (29) to yield a PTE (Table 2).
| Statistic | RA23 Value | RA23 PTE | RA4.5 Value | RA4.5 PTE |
| 0.024 | 0.765 | 0.020 | 0.915 | |
| 11.22 | 0.195 | 12.00 | 0.145 | |
| 34.86 | 0.060 | 33.81 | 0.055 | |
| 543.3 | 0.265 | 533.1 | 0.315 | |
| 0.190 | 0.160 | |||
| 0.575 | 0.300 |
Table 2. Summary of the PB-2b VV null tests for the RA23 and RA4.5 patches.
Because four statistics are examined simultaneously, the most extreme of them is biased toward small values even under a clean null, and the minimum PTE cannot be interpreted at face value. We therefore treat the “worst of four,” , and the “best of four,” , as test statistics in their own right and calibrate them against the null. The identical four-statistic reduction is applied to each shuffle, producing distributions of and from which we compute the look-elsewhere–corrected meta-PTEs,
By construction the shuffles “look elsewhere” exactly as the analysis does, so the multiple-testing penalty is incorporated automatically, without assuming Gaussianity or independence among the four statistics. The condition for passage is that both meta-PTEs fall within (0.05, 0.95). Each is uniformly distributed under a clean null, so either tail is informative, and the two bounds catch complementary failures. A low or a high indicates residual contamination; in the first case the least consistent of the four statistics is more extreme than the shuffle ensemble permits, and in the second even the most consistent one is not consistent enough. A high or a low instead signals the opposite pathology of over-consistency, the hallmark of error bars that are too conservative. Both patches pass, and we find and for RA23 (RA4.5), with the smallest individual PTE in either patch being 0.055 (Table 2, Figure 4).
Figure 4. Look-elsewhere calibration of the null tests for RA23 (blue) and RA4.5 (orange). Left: distribution over the random sign-flip shuffles of the most significant of the four summary-statistic PTEs, ; right: the least significant, . Dashed vertical lines mark the values measured in the data, and the corrected meta-PTEs of Equation (24) are simply the shuffle fractions at or beyond them (to the left in the left panel, to the right in the right), and are tabulated alongside the individual statistics in Table 2. In both patches the data fall well within the null distributions, showing neither an anomalously low minimum PTE (left) nor the clustering near unity that would signal over-consistency (right).
Transfer Functions
The combination of ground-template removal, polynomial and subscan filtering, and demodulation suppresses and mixes power in the recovered maps; this is an effect we characterize with an end-to-end simulation transfer function . For each patch, given that true sky is an unknown signal, we generate realizations of a known input sky signal, in this case white noise smoothed with a Gaussian kernel of 5′ standard deviation so that it carries power across the full range of analyzed multipoles. The simulated sky is then scanned into timestream using the real pointing and detector set, with noise added following the measured PB-2b noise spectrum but scaled down in amplitude. The resulting timestreams are propagated through the identical map-making pipeline applied to the data (time-constant deconvolution, turnaround masking, degree-five polynomial and subscan-polynomial filtering, ground-template and notch HWPSS removal, demodulation, and per-subscan binning and coaddition). Retaining noise at all keeps the stochastic character of the timestreams that the noise-dependent stages of the pipeline respond to, such as glitch flagging, sigma-clipping, and inverse-variance weighting, while scaling it down reduces the scatter on and so the number of realizations needed for it to converge. Aside from this reduced noise level, the only difference from the data run is the substitution of the known input sky for the measured timestreams, so the resulting output map carries the full multiplicative and mode-mixing response of the analysis to a known input .
We estimate the transfer function from the cross-spectrum of the processed output with the input realization. For each input sky realization we first average over a set of observations reserved for these simulations, which is the benchmark set for RA23 and a separate set for RA4.5 (§ 4.1),
and then average over realizations,
binned in the same bandpowers as our science spectra. Here is the input sky of realization and is the map recovered by propagating that realization through observation ; the index runs over that set rather than over the full data set. Using the outputinput cross-spectrum rather than the output auto-spectrum makes the estimator insensitive to any signal uncorrelated with the input (e.g. map noise). Normalizing by the per-realization input auto-spectrum, rather than the ensemble-mean or theory spectrum, cancels the sample variance of each realization in the ratio.
The effective number of modes per bandpower, measured directly from the scatter of the input auto-spectra, is – over the range considered, so this removes a – per-realization fluctuation and reduces the variance of by 20–25%. Because is formed bin by bin as this ratio, it characterizes the response of the filtering pipeline rather than any property of the sky, and is insensitive to the choice of input spectrum provided each bandpower carries sufficient power.
Circular polarization is a spin-0 field with a single auto-spectrum and no decomposition, and the analysis correspondingly operates on the single demodulated map, so the transfer function suffices; the lensed-CDM realizations used to evaluate the leakage are described separately in § 4.8.2.
The uncertainty on is estimated from the scatter among the ,
Because every observation in the set processes the same set of input skies, any realization-dependent bias in is common to all and averages down only with ; the independent samples are therefore the input skies rather than the observations, and the covariance must be taken over after the average over has been performed.
The transfer function is determined independently for each patch. is deconvolved by , together with the beam and the circular polarization efficiency , before the upper limits are derived, and its covariance is propagated into the final error budget; the results are shown together with in Figure 5. Additionally, Figure 5 demonstrates low- suppression, which validates that the polynomial and scan-synchronous filtering stages (described in § 4.2.3 and 4.2.5, respectively) effectively marginalize the low- contamination associated with atmosphere, most notably the circularly polarized signal induced by Zeeman splitting and the loading fluctuations induced by variable water vapor.
Leakage Estimation
In K. Sakaguri et al. [55], the leakages between Stokes parameters due to instrument systematics were discussed. In this section we will derive bandpower intensity leakage and forecast the polarization leakage based on the scan strategy of PB-2b. The contributions of both leakage sources are shown in Figure 6, where
Figure 6. Leakages determined from Sections 4.8.1 and § 4.8.2. The left is for RA23, and the right RA4.5. The joint contribution of both leakage sources is shown as the solid step plot with associated uncertainty ranges, accounting for . The uncertainties in the plots are propagated from the circular polarization uncertainty and variance of the individual leakage contributors.
based on the leakage terms present in Equation (41).
Total Leakage in -Space
As mentioned earlier, K. Sakaguri et al. [55] report a small but non-negligible intensity leakage into the channel. However, there is additional leakage from intensity generated in the detector response (S. Takakura [63]), downstream of the optics and therefore outside the Mueller-matrix description; this is the of Equation (17). The quantity measured here is therefore the total of Equation (20), which additionally absorbs the static optical leakage . We measure it directly from observations of Jupiter, which at 90 GHz is a bright and, to good approximation, unpolarized source (both in linear and circular). Any response in its co-added map is therefore dominated by instrumental leakage of . We co-add all 90 GHz Jupiter maps in and , with Stokes data transformed into the QWP basis, apodize the finite map with a taper, zero-pad, and Fourier transform; both transforms share a common normalization (the zero-mode), which cancels in the leakage ratio below. In essence, this is the same procedure as the estimation of the instrumental beam, detailed in § 3.2, and uncertainty in each -bin is propagated similarly.
As a result, the total leakage is the ratio of the two amplitudes,
where are the beam functions of the demodulated Stokes component maps and , respectively, following the same procedure in § 3.2. The uncertainty on the leakage in each multipole bin is estimated with a leave-one-out jackknife over the individual planet observations, re-coadding Stokes and and re-running the full pipeline with each observation removed in turn. Across , we find . This is the total leakage into the -carrying channel, containing both the optical term of § 2.2 and the detector-response term of § 4.2.4; we do not attempt to separate the two. We carry and its uncertainty in -space as systematic terms in the error budget.
We estimate the total leakage in -space,
where is computed from CAMB (A. Lewis et al. [33]), using best-fit cosmological parameters from Planck (Planck Collaboration et al. [48]). Measured maps from PB-2b were not utilized, due to high low-frequency noise from weather contaminating the demodulated intensity maps. Because and are spin-0 fields, they are invariant under the rotation represented by the detector transform.
Overall P-to-V Leakage
As discussed in § 2.2, residual instrumental coupling converts a fraction of the linear polarization into the circular-polarization channel, so the measured map contains a leakage term . where was measured and bounded in K. Sakaguri et al. [55], as described in § 2.2. Both are defined from the effective Mueller matrix in Equations (2a)–(2d). At the power-spectrum level, with pixel window function from Equation (22), this contributes
for each patch, where the bar indicates this is a measured quantity from simulations, rather than a deterministic, analytical systematic. The explicit factor of is what prevents the pixel window from being counted twice. These leakage spectra are estimated from simulated theory skies synthesized with the same pixel grid and resolution as the data, so they carry the pixel window of that grid. The forward model of Equation (41) then multiplies them by , which already includes the pixel window by construction (§ 4.3). Dividing it out here leaves the pixel window applied exactly once, to the sky signal and the leakage alike. Propagating it into the budget requires the auto- and cross-spectra of the linear polarization as it actually projects onto our maps due to the multitude of QWP orientations relative to the sky per observation. Because the detectors measure polarization in the QWP frame, the relevant and are the sky signal rotated by the parallactic angle and then averaged over the full set of observations, which partially cancels the leakage wherever the patch is observed over a range of parallactic angles.
We estimate these spectra from simulations. For each realization we draw a Gaussian lensed-CDM sky (; flat-sky and synthesized on the survey patch geometry with the Boltzmann code wrapper CAMB (A. Lewis et al. [33])), using best-fit cosmological parameters from Planck (Planck Collaboration et al. [48]). From the real observations we build effective rotation maps, accumulating the weighted parallactic-angle factors over all observations ,
where is the parallactic angle at the patch center for observation . For the sake of this leakage analysis, and are treated as spin-0 fields, assuming independent leakage per Stokes component. The effective instrument-frame polarization is then and , restricted to the well-observed region (pixels within dB of the peak hit count). Their mask-deconvolved (MASTER) auto- and cross-spectra give one realization of , , and .
Repeating over sky realizations yields the mean leakage spectra and their bandpower covariance, which are the products carried forward (per patch). Combined with the calibrated and through Equation (38), they give the leakage contribution and its uncertainty in the final error budget.
Estimation of CVV
With the understanding of the instrument systematics and leakages, we can now extract . Starting with Equation (21), the mapmaking procedure in § 4.3 is utilized to produce one map per observation (). This brings the domain from time () to map pixel space, (). Given that the instrument systematics happen in map space, the effects of the beam, pointing model, and transfer function are represented as operators , , and , respectively.
Averaging Equation (40) across all cross-spectra following the methodology of § 4.5.2, the resulting -space equation relates the noise-free observed-frame spectrum , of which of Equation (23) is the measurement, to the sky spectrum,
with the leakage spectra defined in § 4.8. Solving for the sky spectrum,
The absolute gain, , of Equation (8) enters on the measured spectrum alone. The beam, pointing, transfer function, and leakage coefficients are each ratios of maps built from the same calibrated timestreams, so cancels identically in all of them, while the simulated input spectra are generated in true and carry no instrumental gain. In the joint fit of § 4.10, is common to both patches and so cancels exactly in the transfer-function ratio of Equation (45). It enters the RA4.5 prediction only through the small patch-difference leakage term. Because the limits scale as , we sample it as a free parameter with the Gaussian prior of § 3.3, giving it its own chain dimension rather than fixing it and marginalizing afterwards. The time-constant factor of § 3.3 enters in the same way, but it is specific to each patch and small, so it is not sampled; it is drawn with the fixed nuisances after sampling (§ 4.10). In the joint fit, is drawn independently for each patch, and the two draws are averaged in each bandpower with weights given by each patch’s statistical inverse variance on , computed from the diagonal of its covariance and the nominal and .
Likelihood and Upper Limits
The bandpowers are compared to the model of § 4.9 through a Gaussian likelihood. Writing for the measured bandpowers of patch (Equation (23)), for their pair-level bootstrap covariance (§ 4.5.2), and for the predicted noise-free spectrum,
where is the prior. It is Gaussian on the calibrated parameters , , , , and (Table 3), truncated by the positivity bounds given below, and uniform on . The sampled parameters are the three global scalars, the circular polarization efficiency , the absolute gain , and the pointing amplitude of Equation (44), together with the per-bandpower latent spectrum , beam , and pipeline transfer function ; the model is for the patch in which the signal is parameterized, and Equation (45) for RA4.5. It is essential to distinguish the two Gaussians here: the covariance is the likelihood, constraining the latent spectrum by the data, whereas , , , , and enter through as Gaussian priors set by their external calibration measurements (Table 3).
The coadded map is the true sky convolved with the distribution of pointing displacements across the coadded observations, so in harmonic space the pointing error multiplies the spectrum by the characteristic function of that distribution. A displacement common to every observation is then a rigid translation, which contributes pure phase and drops out exactly, while binning in retains only the trace of the displacement covariance, whose anisotropy is suppressed to of the effect at these scales. The pointing uncertainty therefore carries a single degree of freedom, the dispersion induced across observations by the uncertainty in the fitted pointing parameters (12.6″ for RA23 and 12.1″ for RA4.5), on the fixed shape , and we sample its amplitude,
where is the measured transfer function of § 3.1 and is its uncertainty template. This single degree of freedom is what distinguishes from , whose filtering is mode-selective rather than a convolution, so that no single scalar controls its curve and each bandpower must be sampled separately. Diagonalizing the measured covariances confirms the distinction—one eigenmode carries 99.8% of the variance of , against the 15 of 17 modes needed for . A single is shared by both patches, with per-patch templates, because one pointing-model fit underlies both. We sample Equation (43) with the affine-invariant ensemble sampler EMCEE (D. Foreman-Mackey et al. [18]). The remaining, weakly coupled systematics (, , , , and are held at their central values during sampling and marginalized afterwards by drawing nuisance realizations from their priors for every posterior sample and recomputing Equation (42). Hard physical limits , , , , are imposed throughout. We discard the first 30% of each chain as burn-in and assess convergence with the integrated autocorrelation time , estimated separately for every sampled parameter. We require , where is the largest over all sampled parameters, including the global as well as the per-bandpower ones, and both quantities are expressed in stored steps.
Because individual bandpowers may fluctuate negative, we set upper limits with the unified Feldman–Cousins construction [16] rather than a positivity-truncated posterior, which collapses asymmetrically against the boundary. From the marginalized posterior we take the per-bin mean and standard deviation , treat with the physical bound , and obtain the 95% interval from the likelihood-ratio ordering. We quote as the 95% upper limit; would flag a bandpower inconsistent with (a two-sided detection band).
We analyze both cases:
Per-patch fits.—RA23 and RA4.5 are analyzed independently, each with a -dimensional chain and plus the per-bandpower , , ; , so 54 dimensions). We run 512 walkers for stored steps, storing every third sample; both chains are well converged ().
Joint fit.—Assuming the circular-polarization signal is statistically isotropic, a single is common to both patches, so a joint fit combines their constraining power. We parameterize the signal in the RA23 frame and predict the RA4.5 spectrum from it,
where is the patch-difference of each leakage spectrum, each deconvolved by its own pixel window for the reason given below Equation (38); the two differ because the right-ascension rescaling of § 4.5.1 gives each patch a slightly different effective pixel solid angle. The two patches contribute the two terms of the sum in Equation (43). The pointing transfer functions are evaluated at the shared amplitude of Equation (44), so the coherent pointing error largely cancels in the ratio. Only the difference in fractional susceptibility, , survives. At this moves the ratio by 0.11%, against the 0.72% by which itself moves, a sevenfold cancellation. This is the expected behavior, as a single pointing model underlies both patches and its common mode should not affect their relative calibration; the per-patch fits retain the full 0.72%. Because the RA4.5 prediction depends on the transfer-function ratio, is sampled separately for each patch (raising the chain to dimensions) so that the -ratio uncertainty propagates rather than being held fixed, which would make the RA4.5 prediction artificially sharp. The larger parameter space lengthens the autocorrelation time, so the chain must be run correspondingly longer. We use 256 walkers for stored steps, retaining every sixth sample, which is roughly 4.7 times as many raw steps per walker as the per-patch fits. The heavier thinning leaves unchanged, since it rescales both quantities equally, but keeps the stored chain within memory; like the individual patch likelihood estimation, this ratio is also . The joint fit yields the tightest limits, as both patches inform the common .
To gauge the relative importance of each component in the error budget, we propagate the measured uncertainty of each parameter through the inversion of Equation (42) one at a time. For a given parameter we draw Monte Carlo realizations from its measurement distribution (Gaussian, or uniform for ) while holding all others at their central values, and take the per-bin standard deviation of the resulting as that parameter’s contribution to the uncertainty; varying all parameters simultaneously gives the total , to which the individual terms add in quadrature. Because the standard deviation is sign-symmetric, this budget is unaffected by bandpowers that fluctuate below zero, and as no positivity is imposed it reproduces the unconstrained uncertainty that the Feldman–Cousins construction (§ 4.10) consumes.
Figure 7 shows the relative contributions of each parameter to the overall uncertainty, separating the six parameters that carry their own chain dimension, the latent spectrum , the circular polarization efficiency , the absolute gain , the beam , the transfer function , and the pointing transfer function , from the quantities propagated after sampling: , shown separately, and the and leakages, shown as a single combined curve. The median contribution of is 0.3% (0.7%) of the statistical term for RA23 (RA4.5), below every sampled calibration, which is why it is propagated rather than sampled. In both patches the statistical term traces the total at every multipole, while all calibration, beam, transfer, and leakage contributions lie one to three orders of magnitude below it across . The upper limits are therefore statistics-limited, as no instrumental systematic contributes appreciably to the error, and the result is robust to their calibration.

Figure 7. Relative contribution of overall uncertainty in a Monte Carlo realization distribution of draws. This result shows the uncertainty is significantly dominated by the statistical uncertainty of the measured spectra; the next most significant contributions, from the absolute gain and the circular polarization efficiency , are comparable, with slightly larger, and typically lie one to two orders of magnitude below it. The sampled parameters and their error ranges are detailed in Table 3.
Results
We find no evidence for circular polarization in either field. The coadded maps entering the analysis are shown in Figure 8; both are consistent with noise, with no visible structure apart from the point source J0440−4333 in RA4.5. Figure 9 shows the measured detector power spectra, , for the RA23 and RA4.5 patches across , where the error bars are the bandpower uncertainties of Section 4.5.2 and include none of the aforementioned leakages or systematics. As shown by Figure 7, the uncertainty is dominated by statistical uncertainty of the measured , which we know from K. Sakaguri et al. [55] is primarily sky . From the measured spectra, (PTE = 0.304) for RA23 and 1.00 (PTE = 0.450) for RA4.5, consistent with the null hypothesis. We therefore place 95% confidence upper limits on , constructed bin-by-bin with the Feldman–Cousins unified prescription (G. J. Feldman & R. D. Cousins [16]), which respects the physical boundary . The resulting limits on are given in Table 4 and Figure 10; the combined two-field limit ranges from at the bin centered at to at the bin centered at . Figure 10 also shows representative predicted signals from Pop. III supernova remnants, the CB, and Galactic synchrotron. These lie roughly six orders of magnitude in below the limits reported here, equivalently a factor of in circular polarization amplitude, so detecting any of them remains far out of reach at this sensitivity. Furthermore, the contribution from Galactic synchrotron as represented in Figure 10, is estimated in S. King & P. Lubin [30] from full-sky simulations utilizing HAMMURABI (A. Waelkens et al. [65]). Unlike the Pop. III and CB sources, Galactic synchrotron is not isotropic and homogeneous, and given that RA23 and RA4.5 are specifically chosen to lie away from the Galactic plane, this Galactic synchrotron signal is a maximum estimate. In practice, the emission signal would likely be much smaller relative to the other two sources within RA23 and RA4.5, therefore the actual on-sky signal within this -range and at an observing frequency of 90 GHz can justifiably be claimed to consist primarily of CMB photons. Therefore, it is reasonable to claim that the upper limit constraints placed by this study are primarily on the CMB circular polarization signal with minimal contamination from foregrounds.
Figure 8. Coadded maps of the RA23 (left) and RA4.5 (right) patches, in gnomonic projection with an RA/Dec graticule. is the observed Stokes essentially as it comes out of the detector timestreams. It retains all leakages and has not been deconvolved for the beam, transfer function, or pointing. The maps are coadded over all observations and smoothed with a 5′-FWHM Gaussian, and both panels share a common color scale. The footprint is set by a −5 dB cut on the pixel hit count relative to its maximum, as described in § 4.5.1. These maps are noise-dominated. The point source J0440−4333, circled in the RA4.5 panel, is discussed in Appendix § 11.
Figure 9. Measured for the RA23 (left) and RA4.5 (right) fields, shown as bandpowers over . Points mark the bandpower values; vertical bars are the bandpower uncertainties obtained from the pair-level bootstrap of § 4.5.2, while the horizontal bars denote each bandpower’s -width and are not uncertainties. The dashed line is the null hypothesis of zero signal.
| RA23 | RA4.5 | Combined | |
| 350 | 24.3 | 82.3 | 21.8 |
| 450 | 134 | 113 | 95.7 |
| 550 | 115 | 215 | 133 |
| 650 | 72.8 | 176 | 71.5 |
| 750 | 172 | 290 | 157 |
| 850 | 108 | 271 | 98.7 |
| 950 | 384 | 150 | 147 |
| 1050 | 311 | 426 | 239 |
| 1150 | 482 | 292 | 228 |
| 1250 | 615 | 1380 | 756 |
| 1350 | 952 | 1130 | 763 |
| 1450 | 1200 | 688 | 642 |
| 1550 | 2340 | 1820 | 1750 |
| 1650 | 1430 | 3440 | 1800 |
| 1750 | 1020 | 2710 | 1140 |
| 1850 | 2730 | 2850 | 2180 |
| 1950 | 5520 | 3850 | 3490 |
Table 4. PB-2b 95% C.L. upper limits on (), in bandpowers. RA23 and RA4.5 are the individual fields; “Combined” is the joint two-field limit.

Figure 10. Upper panel: 95% confidence upper limits on the CMB circular-polarization power spectrum, . Downward arrows denote the 95% upper limit in each bin; horizontal bars indicate the multipole width of the bandpower and are not uncertainties. The PB-2b measurements (this work) are shown for combined (bold magenta) upper limit constraints, together with previous limits from MIPOL [35], SPIDER [41], CLASS [42, 13, 14], and VLA [44] (at much higher , note the break in the -axis). The paired SPIDER colors denote 95 GHz and 150 GHz, the paired CLASS 40 GHz colors the 2020 and 2022 analyses, and the paired 2026 CLASS colors the 90 GHz band and the 90 GHz, 150 GHz and 220 GHz co-add, which assumes the circular polarization follows the CMB frequency dependence. Also shown are representative predictions from S. King & P. Lubin [30], evaluated at 90 GHz: Faraday conversion by Pop. III supernova remnants and scattering off the cosmic neutrino background. The simulated intrinsic full-sky circular polarization emission of Galactic synchrotron, a foreground, is shown as well [30]. The two Pop. III curves correspond to and , where is the number of Pop. III stars per halo, a poorly constrained quantity to which is quadratically sensitive. Lower panel: the corresponding bandpower sensitivities of the RA23, RA4.5, and combined PB-2b datasets, i.e. the standard deviation of the deconvolved posterior, shown as rather than so that the comparison between datasets is not dominated by the steep rise.
These constraints extend the measured -mode power spectrum to , or angular scales of 5.4′ to 36′, roughly three times the highest multipole previously reached. Prior bandpower limits probe larger scales and lower multipoles: MIPOL at (R. Mainini et al. [35]), SPIDER over at 95 GHz and 150 GHz (J. M. Nagy et al. [41]), and CLASS at 40 GHz over (I. L. Padilla et al. [42]; J. R. Eimer et al. [13]) and subsequently to at 90 GHz, 150 GHz and 220 GHz (T. Essinger-Hileman et al. [14]). The Very Large Array observations of R. B. Partridge et al. [44] reach smaller angular scales still, but bound the fractional circular polarization within individual fields rather than the angular power spectrum. Our limits are weaker at the maximum in absolute than the large-scale SPIDER and CLASS constraints; rises steeply toward high , and the instrument noise grows after beam and transfer-function deconvolution, but they are complementary, covering multipoles above that have not previously been accessible to a -mode power spectrum measurement.
The bin-to-bin scatter of the limits in Figure 10 is an intrinsic feature of noise-dominated measured spectra of Figure 9 rather than a numerical artifact. The posterior chains are well converged (Gelman–Rubin in every bin, with a Monte Carlo uncertainty on each limit of ), so each bandpower’s limit is fixed by where that bin’s measured power fluctuates relative to zero. The limits vary by dex (a factor ) from bin to bin about a smooth trend, comparable to the dex scatter of the CLASS 40 GHz limits (I. L. Padilla et al. [42]) and in contrast to the much smoother SPIDER limits ( dex). This difference reflects noise structure, not sampling methodology. SPIDER utilizes leakage in its HWP to measure circular polarization, and so uncertainty is dominated by a HWP calibration coupling that is strongly correlated across (J. M. Nagy et al. [41]). This suppresses bin-to-bin scatter, whereas the PB-2b and CLASS limits are dominated by per-bin statistical noise.
The underlying bandpower sensitivities, shown in the lower panel of Figure 10, are better behaved than the limits themselves. The combined dataset is the most sensitive in every bandpower, with each individual field
exceeding it by an average factor of , consistent with the expected from combining two fields of comparable depth. The upper limits do not follow this ordering bin by bin, however, because a Feldman–Cousins limit depends on the central value as well as on its uncertainty.
Conclusions
We present results on the circular polarization of the sky at 90 GHz from a series of targeted observations with PB-2b for two specific patches over , at μKCMB sensitivity. As established in § 1, these limits constrain the total diffuse circular polarization of those fields, which includes the circularly polarized CMB signal and trace foregrounds not removed by location choice, masking, and filtering. Both contributions lie far below the instrument noise that dominates the measured bandpowers, so the limits reported here are set by that noise floor rather than by any detected sky signal. This extends the measured -mode power spectrum into the few-arcminute regime, reaching multipoles roughly three times higher than previous measurements, and improves on the prior upper limit placed by SPIDER (J. M. Nagy et al. [41]) at 95 GHz by nearly two orders of magnitude in the single bandpower where those two analyses overlap. The CLASS constraints of T. Essinger-Hileman et al. [14] overlap this analysis over ; above that range these are the only -mode bandpower limits reported to date.
The current limitations of CMB circular polarization studies are that no current experiment utilizing a QWP has the sensitivity necessary for a statistically significant detection (S. King & P. Lubin [30]). Figure 10 conclusively demonstrates this, as the limits reported here sit roughly six orders of magnitude in above the predicted signals. In addition, because the measurement is noise-dominated, with the bandpower uncertainty scaling inversely with integration time, closing that gap with PB-2b would demand of order times the observing time presented here, or an equivalent growth in detector count. Focusing study on lower observing frequency, however, could reduce the discrepancy between noise sensitivity and signal magnitude without requiring an unprecedented improvement in instrument sensitivity or detector count. The Faraday conversion contribution scales as , so the factor of in amplitude gained between 90 GHz and 10 GHz accounts for nearly the entire gap. S. King & P. Lubin [30] correspondingly forecast a signal-to-noise ratio above unity for a 10 m telescope observing for 20 months at 10 GHz, provided . In addition, the single observing band utilized in this experiment does not permit differentiation between potential sources, meaning more exotic sources of circular polarization cannot be rigorously identified among known contaminants, such as Milky Way synchrotron radiation.
CMB circular polarization provides an interesting window into early-Universe physics not encapsulated by standard linear polarization measurements probing ΛCDM. While very few other experiments with comparable sensitivity to the Simons Array are measuring microwave-band circular polarization directly, similar procedures to those of J. M. Nagy et al. [41] might be able to constrain it via leakage from linear polarization. For example, the Simons Observatory (SO) Small Aperture Telescopes (SATs) utilize rotating cold HWPs with nearly identical design to that of PB-2b, which can be used to characterize leakage to even higher sensitivity than that of SPIDER’s limits (J. M. Nagy et al. [41]). Because such a measurement would rely on the small leakage of the HWP rather than on the direct conversion afforded by a QWP, the effective circular polarization efficiency of the SATs is likely to fall well below the achieved here. That penalty may be offset by raw sensitivity, as each mid-frequency SAT carries 12,292 detectors and the SAT survey is projected to reach a combined map depth of 2 μK-arcmin at 93 GHz and 145 GHz over five years (N. Galitzki et al. [20]), orders of magnitude higher than what was achievable with PB-2b. In addition, the SATs will observe at six distinct frequency bands (30 GHz, 40 GHz, 90 GHz, 150 GHz, 220 GHz, and 280 GHz), which may allow discrimination between the contributions of circular polarization sources as a function of frequency, most importantly foregrounds. This allows simultaneous probing of multiple circular polarization sources and foregrounds, which differ in their dependence on observing frequency. These prospects would allow for research in this field to persist beyond the end of PB-2b’s active observation.
Acknowledgments
This work is supported by the Gordon and Betty Moore Foundation Grant No. GBMF7939, the Simons Foundation Grant No. 034079, the National Science Foundation grants AST-1207892 and AST-1440338, the World Premier International Research Center Initiative (WPI) of MEXT, Japan, and the John Templeton Foundation grant 58724. In Japan, this work was supported by JSPS KAKENHI grant Nos. 18H05539 and 19H00674, and the JSPS Core-to-Core Program JPJSCCA20200003. Work at LBNL is supported in part by the U.S. Department of Energy, Office of Science, Office of High Energy Physics, under Contract No. DE-AC02-05CH11231. Calculations were performed on the Central Computing System, owned and operated by the Computing Research Center at KEK, and the National Energy Research Scientific Computing Center, which is supported by the Department of Energy under Contract No. DE-AC02-05CH11231. The Simons Array operated at the James Ax Observatory in the Parque Astronómico Atacama in Northern Chile under the stewardship of the Agencia Nacional de Investigación y Desarrollo (ANID). We acknowledge support from the ASI-COSMOS Network (cosmosnet.it) and from the INDARK INFN Initiative (web.infn.it/CSN4/IS/Linea5/InDark). Support from the Ax Center for Experimental Cosmology at UC San Diego is gratefully acknowledged. M. Hazumi is partially supported by the Yushan Fellow Program by the Ministry of Education (MOE), Taiwan. YC acknowledges support from JSPS KAKENHI Grant Number JP24K00667.
Appendix
Null Test Criteria
We adopt a similar strategy for null splits as in Polarbear Collaboration et al. [52], both for detector-based and observation-based splits, however, we adopt several additional criteria specific to the performance of PB-2b. As with Polarbear, some of these splits are not necessarily evenly split into halves. For splits which are repeated from the detector-based splits, we average over all detectors within a given observation, then split on the median of all observation averages in the observing season, which probes changes in instrument performance over time. The full set of criteria is listed in Table 5.
| Split | Motivation |
| Detector-based splits | |
| A/B pixel handedness | The sinuous antennae come in two symmetric handednesses; probes the frequency dependence of the polarization angle, which is relevant here through linear-to-circular leakage. |
| type pixel | Detectors are aligned horizontally/vertically or rotated by . Especially crucial given that the QWP modulates into the parameter the detectors measure. |
| Top/bottom detector | The two detectors within each antenna, oriented along the two polarization vectors of the pixel. |
| Focal plane position | Left/right and upper/lower physical location, probing location-dependent optical systematics not captured by the pointing model. |
| Gain calibration factor | Per-observation gain factor, without distinguishing stimulator from HWPSS calibration. |
| Effective time constant | Per-observation detector speed, again without distinguishing the calibration method. |
| Calibration source | Whether a detector draws its calibration mostly from the stimulator or from the HWPSS. |
| amplitude | Dominated by contamination from the QWP; validates optical and polarization systematics as well as the time-domain HWPSS removal. |
| amplitude | Several orders of magnitude weaker than the amplitude, but associated with leakage, which is small but not negligible. |
| White noise | Time-domain white noise level, probing noise modeling and data selection. |
| Input glitch ACF | Glitch autocorrelation prior to cleaning. |
| Observation-based splits | |
| Median detector quantities | Observation-level analogues of the detector splits above: gain calibration factor, amplitude and phase, amplitude and phase, white noise, and effective time constant. |
| Calibration source | Which method calibrates the majority of detectors, since stimulator calibration is wholly or partly missing for some observations. |
| “Glitchy” runs | Median fraction of detectors with glitches exceeding an autocorrelation threshold. |
| 1st/2nd half of season | A time-based split probing overall drift in instrument performance across the observing season. |
| Median PWV | Loading and general weather systematics. |
| PWV standard deviation | Variability of PWV within a single observation, probing rapid changes in loading. |
| Sun/Moon distance | Angular distance to the Sun or Moon, probing far-sidelobe contamination. |
| Elevation | Median scan elevation, i.e. the atmospheric column along the line of sight. |
| Bolometer tuning strategy | Tuning to a fixed fractional resistance of 0.75 versus an algorithm setting a different fractional resistance per detector. |
| Ambient air temperature | Components held at air temperature, notably the QWP, may have emission or transmission properties correlated with temperature. |
| Focal plane temperature | Sets the thermal noise level, which could bias the gain and leave residual sky signal in the noise maps. |
| Pointing correction | Magnitude of the geometric correction required by the pointing model, which depends on the patch azimuth and elevation when observed. |
Table 5. Null test splits used to assess the internal consistency of the measured spectra, following the strategy of Polarbear Collaboration et al. [52] with additional criteria specific to PB-2b.
Atmospheric Polarization
Atmospheric circular polarization emission is the result of Zeeman splitting of oxygen due to Earth’s magnetic field (W. B. Lenoir [32]; B. Keating et al. [29]; S. Hanany & P. Rosenkranz [22]; S. Spinelli et al. [59]; M. A. Petroff et al. [45]; T. Fujino et al. [19]) and presents a major foreground for CMB circular polarization measurements. Through the Zeeman effect, Earth’s magnetic field splits the rotational energy levels of atmospheric oxygen, allowing transitions that emit polarized radiation. The imbalance between the right- and left-handed circularly polarized components produces a net circular polarization signal near the oxygen resonance lines. As such, CMB experiments observe at frequencies away from these resonance lines, which are well characterized (see Figure 11). However, even away from the resonances, atmospheric emission is still a dominant foreground and therefore must be accounted for.

Figure 11. Frequency dependence of the absolute value of the atmospheric circular polarization evaluated at zenith. The dark dashed vertical line marks the measured band center at 89.5 GHz, and the surrounding dark shaded region indicates its uncertainty. The lighter dashed vertical lines mark the nominal lower and upper band edges, while the adjacent light shaded regions indicate the uncertainty associated with the measured bandwidth. The model is made using the software suite of M. A. Petroff et al. [45], adapted to the PB-2b observation date and frequency.
Due to its coupling with Earth’s magnetic poles, the atmospheric circular polarization obtains a dipole-like structure which depends on the projection of the telescope’s line-of-sight onto the local magnetic field vector. This produces a characteristic azimuthal dependence, with the signal largest when the line of sight is most aligned with the magnetic-field direction and smallest when it is perpendicular (see Figure 12). Further, the atmospheric signal amplitude decreases with increasing elevation due to the reduced atmospheric column density along the telescope’s line of sight. At a fixed telescope pointing, this signal remains constant in the sky, but in the detector timestream it will get modulated to such that the corresponding raw detector signal is not constant. Demodulation then removes the HWP-angle dependence, and recovers the constant atmospheric Stokes signal.

Figure 12. Predicted azimuthal dependence of atmospheric circular polarization. Solid lines show the nominal fit, while dashed lines indicate the 16th and 84th percentile bounds of the Monte Carlo predictions. It can be seen that the amplitude decreases with increasing elevation. Large uncertainties can be observed, reflecting the strong sensitivity of the predicted signal to variations in the detector band center. The model is as in Figure 11, additionally accounting for the measured bandwidth.
We are able to mitigate any major contribution from the atmospheric circular polarization by a combination of polynomial and scan-synchronous signal filtering. We apply a polynomial filter both on the full timestream as well as individual subscans, as described in § 4.2.3, and reapply it after demodulation (§ 4.2.5), where the atmospheric signal is no longer modulated. Over the limited azimuth range of an individual subscan, the atmospheric circular polarization signal is expected to vary slowly and can be approximated by a low-order expansion in azimuth. Therefore, the constant and first-order terms of the subscan polynomial filter remove much of this atmospheric contribution and any other slow varying offsets. The scan-synchronous signal filter provides an additional suppression of components that are stable in telescope azimuth. Since atmospheric circular polarization is tied to the telescope line of sight relative to Earth’s magnetic field, rather than to fixed celestial coordinates, it is expected to appear primarily as an azimuth- and elevation-dependent signal and is there-
fore similarly able to be removed by the constant and low order terms of the scan-synchronous filter.
To describe the atmospheric circular polarization at the PB-2b site, we evaluate the empirical model of M. A. Petroff et al. [45] at 90 GHz, taking both azimuthal and zenith angle dependence into account. This theoretical model has previously been confirmed at 40 GHz by the CLASS experiment M. A. Petroff et al. [45] and at 150 GHz by the Polarbear project T. Fujino et al. [19]. We use a band center of GHz and a fractional bandwidth of , as defined in pre-deployment detector measurements B. Westbrook et al. [67]. To determine the uncertainty in the theoretical atmospheric emission, we draw Monte Carlo realizations of the band center and fractional bandwidth from independent Gaussian distributions. We then calculate the band-averaged emission for each realization and take the 16th–84th percentile range as the bandpass-induced uncertainty. The resulting uncertainty is much larger than the nominal signal, mainly due to the strong dependence of the atmospheric signal on observing frequency. Across the frequency ranges sampled by the Monte Carlo bandpasses, the atmospheric spectrum changes sign, therefore broadening the distribution of predicted band-averaged signals. Because the zenith angle dependence is an empirical approximation, we only evaluate the model within its calibrated range from 30 to 60 degrees. Further, as the Earth’s magnetic field is not perfectly aligned with true north, we also must determine the magnetic declination. We calculate the declination using the International Geomagnetic Reference Field model P. Alken et al. [2], evaluated at a site elevation of 5200 m.
Field model P. Alken et al. [2], evaluated at a site elevation of 5200 m. Since the main geomagnetic field varies slowly in time, we evaluate the field at a representative midpoint date of the observing period and approximate it as constant over the data set. The resulting atmospheric circular-polarization curve is therefore slightly offset from true north. As expected, the atmospheric circular polarization varies slowly in azimuth and as such is mitigated by the post-demodulation filtering of § 4.2.5. Due to the limited azimuthal range of our scans, we do not currently detect any atmospheric emission.
Gain Validation with J0440-4333
The RA4.5 patch contains the bright, compact radio source J0440–4333 (redshift , a Q-type blazar) Z.-Q. Shen et al. [57], which provides a unique opportunity to validate our calibration and mapmaking procedures. We measure its flux in our map (Figure 13) and compare it to a model prediction based on its cataloged intensity and expected polarization behavior.

Figure 13. Zoomed-in section of the RA4.5 patch showing J0440–4333 with a −8 dB hit mask. The green circle is the 10′ aperture and the region between the solid and dashed white circles is the background annulus; together they define the aperture estimator used as a cross-check on the matched filter of Equation (46).
The pipeline’s response to a compact source is not simply the beam. The filtering of § 4.2 suppresses the source core and surrounds it with a shallow negative trench from ∼11′ to ∼32′, so a 10′ aperture with a 12′–22′ background annulus measures the suppressed core against the source’s own trench. Because that trench is coherent with the source rather than noise, it is properly treated as signal, and we estimate the flux with a matched filter built on the measured response,
where is the estimated source flux at map position , is the two-dimensional Fourier wavevector, IFFT denotes the inverse fast Fourier transform, and the asterisk denotes complex conjugation. Here is the measured map (which contains the converted Stokes signal), is the point-source template, built from the same observations as the map, and is the map’s own two-dimensional power spectrum, kept anisotropic because the residual noise is striped along the scan direction. Calibrating the response on a synthetic unit-flux source makes the estimator unbiased independently of Fourier conventions; it round-trips on simulations to 1.0098. Evaluating at the J2000 position of J0440−4333 gives the measured integrated flux , and evaluating it at off-source positions in the same filtered map gives the reference set , from which the uncertainty follows as , where MAD denotes the median absolute deviation.
Converting the filtered amplitude to a flux density requires the point-source flux transfer , the ratio of recovered to injected integrated flux. This differs from the transfer function in that it also carries the fraction of the source flux that the estimator collects, and the two effects cannot be separated within a single scalar. Analytically, a mask that is unity inside an aperture of radius and solid angle , and within a background annulus of solid angle , has transform , giving
where is the first-order Bessel function of the first kind and . Since , the annulus subtraction makes this a differential estimator, so the recovered fraction falls below unity even before filtering; the aperture efficiency matters here because the measured 90 GHz beam has broad non-Gaussian wings that leave only 89.5 % of the source flux within . We evaluate Equation (C2) over – in bins, holding constant outside that range, which gives . The kernel places only % of its weight above , since it first vanishes near , so although is uncharacterized there, varying it over its full physical range shifts by only .
We nevertheless quote the value obtained end-to-end by injection, which assumes no model of the beam, filtering, or estimator. A source of exactly known integrated flux, carrying the measured beam profile, is added to the Stokes channel of an input sky map, propagated through the unmodified pipeline of Sections 4.2–4.3, and photometered with the same matched filter, so is the recovered-to-injected ratio by construction; differencing against a baseline over identical observations and CMB realization cancels the sky exactly. Over the RA4.5 observations, , 7 % below the analytic value, a shortfall expected in sign because Equation (C2) treats the map-making operator as diagonal in Fourier space. Because detector weights derive from each timestream’s own noise and the simulations are far quieter per sample than the data, the injection must stay in the linear regime, which we verify through the ratio of total map weight to baseline, 0.9912 here. The measured flux density is then , converted to Jansky.
Since circular polarization data for J0440−4333 is unavailable, we model its emission and propagate it through the same leakage and flux transfer chain,
where denotes the blazar. The total-intensity flux is estimated from ALMA catalog queries ([8]) and the circular polarization fraction assumed to be at 15 GHz ([26]). Since the linear polarization angle is unconstrained, and are marginalized uniformly on , introducing a variance
on the prediction, where and are the uncertainties on the leakage and on the linear polarization fraction, the latter taken as scaled from the 18 GHz measurements of M. Massardi et al. [37]; is neglected as small. The leakage spectrum of Equation (20) is convolved with the same matched-filter weighting to give , a value insensitive to that weighting. It is measured from Jupiter, itself a point source, and is flat enough across the band (0.0166–0.0188) that any reasonable choice alters it by at most 7 %.
The comparison is shown in Figure 14. The matched filter gives mJy, a detection of the source in the circular polarization channel, while the aperture on the same map returns mJy, consistent but 3.2 times less precise. Retaining only the leakage terms of Equation (C3) predicts mJy, differing by mJy , while adding the intrinsic polarization priors predicts mJy, differing by mJy . Both are quoted because J0440−4333 is a variable FSRQ, so neither a single assumed pair of polarization fractions nor their omission is defensible alone; both move the comparison toward agreement, making the leakage-only prediction the conservative statement.
Figure 14. Circular polarization flux density of J0440−4333, measured against prediction. The measurement is the matched filter of Equation (46) applied to the coadded RA4.5 map, divided by the injection-calibrated flux transfer . Two predictions are shown: one retaining only the instrumental leakage terms of Equation (49), and one additionally assuming the intrinsic polarization priors described in the text. Error bars are , propagated at first order.
The statistical power of this check is limited and should be stated plainly. Taking the prediction with intrinsic polarization priors as truth, the measurement constrains the flux transfer to , so it is sensitive only to errors above . The comparison therefore does not itself determine , which rests on Equation (C2) and the injection; what the agreement establishes is that no gross error is present in the calibration and mapmaking chain, complementing the characterization in N. Farias et al. [15] and K. Sakaguri et al. [55]. This appendix does not bear on the science result, since Equation (42) deconvolves and bandpower by bandpower rather than through any scalar collapse.
References
- [1]Alexander, S., Ochoa, J., & Kosowsky, A. 2009, Physical Review D, 79, 063524, doi: 10.1103/PhysRevD.79.063524DOI
- [2]Alken, P., Thébault, E., Beggan, C. D., et al. 2021, Earth, Planets and Space, 73, doi: 10.1186/s40623-020-01288-xDOI
- [3]Alonso, D., Sanchez, J., & Slosar, A. 2019, Monthly Notices of the Royal Astronomical Society, 484, 4127, doi: 10.1093/mnras/stz093DOI
- [4]Arnold, K., Stebor, N., Ade, P. A. R., et al. 2014, in Millimeter, Submillimeter, and Far-Infrared Detectors and Instrumentation for Astronomy VII, Vol. 9153 (SPIE), 91531F, doi: 10.1117/12.2057332DOI
- [5]Barron, D., Mitchell, K., Groh, J., et al. 2021, IEEE Transactions on Applied Superconductivity, 31, 1, doi: 10.1109/TASC.2021.3067190DOI
- [6]Bennett, C. L., Larson, D., Weiland, J. L., et al. 2013, The Astrophysical Journal Supplement Series, 208, 20, doi: 10.1088/0067-0049/208/2/20DOI
- [7]Bixler, B., Arnold, K., Hill, C., & Kusaka, A. 2026, IEEE Transactions on Applied Superconductivity, 36, 1, doi: 10.1109/TASC.2026.3680443DOI
- [8]Bonato, M., Liuzzo, E., Giannetti, A., et al. 2018, Monthly Notices of the Royal Astronomical Society, 478, 1512, doi: 10.1093/mnras/sty1173DOI
- [9]Catalano, A., Ade, P., Atik, Y., et al. 2014, Astronomy & Astrophysics, 569, A88, doi: 10.1051/0004-6361/201423868DOI
- [10]Cooray, A., Melchiorri, A., & Silk, J. 2003, Physics Letters B, 554, 1, doi: 10.1016/S0370-2693(02)03291-4DOI
- [11]Das, S., Louis, T., Nolta, M. R., et al. 2014, Journal of Cosmology and Astroparticle Physics, 2014, 014, doi: 10.1088/1475-7516/2014/04/014DOI
- [13]Eimer, J. R., Li, Y., Brewer, M. K., et al. 2024, CLASS Angular Power Spectra and Map-Component Analysis for 40 GHz Observations through 2022, arXiv, doi: 10.48550/arXiv.2309.00675
- [14]Essinger-Hileman, T., Ryan, C., Li, Y., et al. 2026, Improved Constraints on Cosmic Microwave Background Circular Polarization with CLASS, arXiv; submitted to ApJ, doi: 10.48550/arXiv.2609.40125
- [15]Farias, N., Adkins, T., Chu, M., et al. 2026, Review of Scientific Instruments, 97, 083101, doi: 10.1063/5.0319093DOI
- [16]Feldman, G. J., & Cousins, R. D. 1998, Physical Review D, 57, 3873, doi: 10.1103/PhysRevD.57.3873DOI
- [17]Finelli, F., & Galaverni, M. 2009, Physical Review D, 79, 063002, doi: 10.1103/PhysRevD.79.063002DOI
- [18]Foreman-Mackey, D., Hogg, D. W., Lang, D., & Goodman, J. 2013, Publications of the Astronomical Society of the Pacific, 125, 306, doi: 10.1086/670067
- [19]Fujino, T., Takakura, S., Arani, S. S., et al. 2025, The Astrophysical Journal, 981, 15, doi: 10.3847/1538-4357/ada89b
- [20]Galitzki, N., Tsan, T., Spisak, J., et al. 2024, The Astrophysical Journal Supplement Series, 274, 33, doi: 10.3847/1538-4365/ad64c9
- [21]Giovannini, M. 2009, Physical Review D, 80, 123013, doi: 10.1103/PhysRevD.80.123013
- [22]Hanany, S., & Rosenkranz, P. 2003, New Astronomy Reviews, 47, 1159, doi: 10.1016/j.newar.2003.09.017
- [23]Hill, C. A., Kusaka, A., Ashton, P., et al. 2020, Review of Scientific Instruments, 91, 124503, doi: 10.1063/5.0029006
- [24]Himwich, W. E. 1993, Pointing Model Derivation, Mark IV field system reference manual, version 8.2, NASA/Goddard Space Flight Center
- [25]Hivon, E., Górski, K. M., Netterfield, C. B., et al. 2002, The Astrophysical Journal, 567, 2, doi: 10.1086/338126
- [26]Homan, D. C., & Lister, M. L. 2006, The Astronomical Journal, 131, 1262, doi: 10.1086/500256
- [28]Kamionkowski, M., Kosowsky, A., & Stebbins, A. 1997, Physical Review D, 55, 7368, doi: 10.1103/PhysRevD.55.7368
- [29]Keating, B., Timbie, P., Polnarev, A., & Steinberger, J. 1998, The Astrophysical Journal, 495, 580, doi: 10.1086/305312
- [30]King, S., & Lubin, P. 2016, Circular polarization of the CMB: Foregrounds and detection prospects, arXiv, doi: 10.48550/arXiv.1606.04112DOI
- [31]Kusaka, A., Essinger-Hileman, T., Appel, J. W., et al. 2014, Review of Scientific Instruments, 85, 024501, doi: 10.1063/1.4861834
- [32]Lenoir, W. B. 1968, Journal of Geophysical Research (1896-1977), 73, 361, doi: 10.1029/JA073i001p00361
- [33]Lewis, A., Challinor, A., & Lasenby, A. 2000, The Astrophysical Journal, 538, 473, doi: 10.1086/309179
- [34]Lowry, L. N., Ito, J., & POLARBEAR Collaboration. 2022, in Millimeter, Submillimeter, and Far-Infrared Detectors and Instrumentation for Astronomy XI, ed. J. Zmuidzinas & J.-R. Gao (Montréal, Canada: SPIE), 73, doi: 10.1117/12.2629114
- [35]Mainini, R., Minelli, D., Gervasi, M., et al. 2013, Journal of Cosmology and Astroparticle Physics, 2013, 033, doi: 10.1088/1475-7516/2013/08/033
- [36]Mangum, J. G. 2001, A Telescope Pointing Algorithm for ALMA, ALMA Memo 366, National Radio Astronomy Observatory, Tucson, AZ. http://legacy.nrao.edu/alma/memos/html-memos/alma366/memo366.pdf
- [37]Massardi, M., Burke-Spolaor, S. G., Murphy, T., et al. 2013, Monthly Notices of the Royal Astronomical Society, 436, 2915, doi: 10.1093/mnras/stt1717
- [38]Matsuda, F. T. 2017, PhD thesis, UC San Diego. https://escholarship.org/uc/item/4340f15m
- [39]Mohammadi, R. 2014, The European Physical Journal C, 74, 3102, doi: 10.1140/epjc/s10052-014-3102-1
- [40]Motie, I., & Xue, S.-S. 2012, Euler-Heisenberg Lagrangian and photon circular polarization, arXiv, doi: 10.48550/arXiv.1104.3555DOI
- [41]Nagy, J. M., Ade, P. A. R., Amiri, M., et al. 2017, The Astrophysical Journal, 844, 151, doi: 10.3847/1538-4357/aa7cfd
- [42]Padilla, I. L., Eimer, J. R., Li, Y., et al. 2020, The Astrophysical Journal, 889, 105, doi: 10.3847/1538-4357/ab61f8
- [43]Pardo, J., Cernicharo, J., & Serabyn, E. 2001, IEEE Transactions on Antennas and Propagation, 49, 1683, doi: 10.1109/8.982447
- [44]Partridge, R. B., Nowakowski, J., & Martin, H. M. 1988, Nature, 331, 146, doi: 10.1038/331146a0
- [45]Petroff, M. A., Eimer, J. R., Harrington, K., et al. 2020, The Astrophysical Journal, 889, 120, doi: 10.3847/1538-4357/ab64e2
- [46]Planck Collaboration, Akrami, Y., Ashdown, M., et al. 2017, Astronomy & Astrophysics, 607, A122, doi: 10.1051/0004-6361/201630311
- [47]Planck Collaboration, Akrami, Y., Arroja, F., et al. 2020a, Astronomy & Astrophysics, 641, A1, doi: 10.1051/0004-6361/201833880
- [48]Planck Collaboration, Aghanim, N., Akrami, Y., et al. 2020b, Astronomy & Astrophysics, 641, A6, doi: 10.1051/0004-6361/201833910
- [49]POLARBEAR Collaboration, Ade, P. A. R., Akiba, Y., et al. 2014, The Astrophysical Journal, 794, 171, doi: 10.1088/0004-637X/794/2/171
- [52]POLARBEAR Collaboration, Adachi, S., Adkins, T., et al. 2024, Physical Review D, 110, 063013, doi: 10.1103/PhysRevD.110.063013
- [53]POLARBEAR Collaboration, Adkins, T., Arnold, K., et al. 2026, Physical Review D, 113, 043044, doi: 10.1103/9wqc-tgzq
- [54]Russell, M., Sakaguri, K., Lowry, L. N., et al. 2024, Journal of Low Temperature Physics, 216, 237, doi: 10.1007/s10909-024-03127-9DOI
- [55]Sakaguri, K., Barron, D. R., Chinone, Y., et al. 2026, Development and characterization of a millimeter-wave circular polarization observation method using a quarter-wave plate with Simons Array, arXiv; submitted to JCAP, doi: 10.48550/arXiv.2610.19621DOI
- [56]Sawyer, R. F. 2014, Photon-photon interactions can be a source of CMB circular polarization, arXiv, doi: 10.48550/arXiv.1408.5434DOI
- [57]Shen, Z.-Q., Wan, T.-S., Moran, J. M., et al. 1998, The Astronomical Journal, 115, 1357, doi: 10.1086/300284DOI
- [58]SKA Observatory. 2023, Mathematical Details of Antenna Pointing Correction, katpoint documentation, Square Kilometre Array Observatory. https://developer.skao.int/projects/katpoint/en/latest/math_pointing.html
- [59]Spinelli, S., Fabbian, G., Tartari, A., Zannoni, M., & Gervasi, M. 2011, Monthly Notices of the Royal Astronomical Society, 414, 3272, doi: 10.1111/j.1365-2966.2011.18625.xDOI
- [60]Stebor, N., Ade, P., Akiba, Y., et al. 2016, in Millimeter, Submillimeter, and Far-Infrared Detectors and Instrumentation for Astronomy VIII, Vol. 9914 (SPIE), 99141H, doi: 10.1117/12.2233103
- [61]Story, K. T., Reichardt, C. L., Hou, Z., et al. 2013, The Astrophysical Journal, 779, 86, doi: 10.1088/0004-637X/779/1/86DOI
- [62]Suzuki, A., Ade, P., Akiba, Y., et al. 2016, Journal of Low Temperature Physics, 184, 805, doi: 10.1007/s10909-015-1425-4DOI
- [63]Takakura, S. 2017, PhD thesis, Osaka University, doi: 10.18910/67054DOI
- [64]Takakura, S., Aguilar, M., Akiba, Y., et al. 2017, Journal of Cosmology and Astroparticle Physics, 2017, 008, doi: 10.1088/1475-7516/2017/05/008DOI
- [65]Waelkens, A., Jaffe, T., Reinecke, M., Kitaura, F. S., & Enßlin, T. A. 2009, Astronomy & Astrophysics, 495, 697, doi: 10.1051/0004-6361/200810564
- [66]Wallace, P. T. 2002, in Advanced Telescope and Instrumentation Control Software II, Vol. 4848, 125–136, doi: 10.1117/12.460914
- [67]Westbrook, B., Ade, P. A. R., Aguilar, M., et al. 2018, Journal of Low Temperature Physics, 193, 758, doi: 10.1007/s10909-018-2059-0DOI
- [68]Zaldarriaga, M., & Seljak, U. 1997, Physical Review D, 55, 1830, doi: 10.1103/PhysRevD.55.1830DOI







