Layered dust / B-modes

Preserved original model output. Check claims against the independent audit →

optimist · 20.2 KB
Fable 5.1critic Grok 4.6optimist Fable 5.1

The steelman below is grounded in the checks I just ran. One citation error in the draft is worth flagging up front: the MNRAS 519, 4370 link the draft attaches to Sponseller & Kogut is actually McBride, Bull & Hensley 2023. Sponseller & Kogut is ApJ 936, 8 from 2022. Both belong in the essay.

1. Strongest physical case

Decorrelation is a rank problem, not a spectral-index problem. Single-template and single-column likelihoods assume the polarized dust sky is a rank-one object: one complex map times one real frequency law. The physical sky is a sum over layers, each contributing a complex number whose modulus is column times polarized emissivity times geometric depolarization, and whose phase is twice the plane-of-sky field angle. The three external datasets are exactly the three factors of that complex number. Three-dimensional dust gives the modulus's column and environment. Grain physics gives the modulus's frequency dependence. Magnetic tomography gives the phase, and where geometry is known, the inclination. That is the tight version of the draft's argument, and it explains why no single ingredient suffices.

The effect is second order in layer contrast, which is good news. Decorrelation needs comparable layer columns, different polarized SEDs, and misaligned fields, all at once. Sightlines meeting all three are a minority, and they can be identified before touching CMB data. Pelgrims et al. 2021 selected sightlines using HI cloud counts and HI-derived field orientations, and found the predicted 217 versus 353 GHz angle rotation, with the effect growing with misalignment. A cloud-to-cloud spread of about 15 percent in the polarized 353/217 ratio reproduced the signal. That is an existence proof that external data predict where decorrelation lives.

All three ingredients are now observed to vary layer to layer. Halal, Clark & Tahani 2024 show the polarization fraction drops when the 3D dust column is split among components at different distances within 1.25 kpc. Mehandiratta et al. 2026 find the ratio of 353 GHz polarized emission to starlight polarization differs between two overlapping clouds in one 4 square degree field. Pelgrims et al. 2024 find most sightlines in their tomography field cross more than one cloud. Panopoulou & Lenz 2020 count on average about three HI clouds per pixel in the north Galactic cap. Multi-layer sightlines are the norm, so the layered model is the correct default, not a special case.

High-latitude dust sits inside the best-mapped volume. The dust scale height is roughly 100 to 150 pc, so at the latitudes of B-mode fields the entire column lies within a few hundred parsecs. Skalidis & Pelgrims 2019 find most 353 GHz polarized emission above 60 degrees latitude arises within about 250 pc. Edenhofer et al. 2024 map that volume at 14 arcminute angular and parsec-scale distance resolution out to 1.25 kpc. The Local Bubble wall dominates many high-latitude sightlines, and if the field is tangent to the shell, its full 3D direction including inclination follows from the mapped geometry. O'Neill et al. 2024 built exactly that model. This turns "projected field" into "3D field" for the layer that matters most.

Grain physics removes freedom rather than adding it. Only large aligned grains polarize, a narrower size range than the one emitting in intensity. The polarized SED per layer is therefore closer to a single-temperature modified blackbody than the intensity SED. Planck finds the polarization fraction roughly flat from 250 micron to 3 mm, and a polarized index of 1.53 close to the intensity index. So the per-layer polarized SED can be tied to the intensity-side SED, which is measurable with high signal to noise, with a small and physically bounded offset. The nuisance per layer is a tightly correlated family, not three free numbers.

The spin-moment expansion is the bridge to the likelihood. The first-order complex moment map is what produces decorrelation and angle rotation. Its modulus is set by column-weighted SED differences, and its phase by field misalignment. Planck already sees these residual maps at low multipoles, as Ritacco et al. 2023 report, but the 2026 spin-moment variance analysis finds the SRoll2 and PR4 maps disagree on whether temperature or index variance is responsible. Planck alone cannot pin the origin. External 3D priors therefore add genuinely new information rather than restating what the microwave data already know.

2. Additional mechanisms and observational handles

  • 3D grain-size state from extinction curves. Zhang & Green 2025 mapped R(V) in three dimensions from 130 million stars. R(V) tracks grain growth by accretion and coagulation, so this is a real 3D proxy for the size distribution, which the draft treats as hypothetical knowledge.
  • Aligned-grain size per cloud. Multi-band starlight polarimetry gives the Serkowski peak wavelength, which scales with the size of aligned grains. Combined with distances this constrains the temperature and alignment efficiency of the polarizing population per layer.
  • Intensity-side temperature tomography. Zelko & Finkbeiner 2022 fitted temperature and index per 3D voxel using Bayestar19 plus Planck and IRAS bands, producing 3D temperature maps at up to 27 arcminutes. This is the intensity prototype of the polarization pipeline.
  • Velocity-resolved 3D HI. McCallum et al. 2026 jointly reconstructed HI density and radial velocity within 1.25 kpc using HI4PI and the Edenhofer dust map. This directly attacks the HI-velocity-to-distance ambiguity the draft lists as a limitation.
  • Cloud counts and per-channel field orientation. Panopoulou & Lenz 2020 provide cloud-count maps. Clark & Hensley 2019 provide position-position-velocity Stokes maps from HI filament orientation. Hu & Lazarian 2023 add velocity-gradient estimates of field orientation and strength.
  • Starlight tomography surveys and inversion. PASIPHAE, RoboPol, and the Interstellar Polarization Survey supply the stars. The BISP-1 inversion of Pelgrims et al. 2023 turns them into per-layer angles. The Planck-versus-starlight angle difference isolates the far layers directly.
  • The emission-to-extinction polarization ratio per layer. Planck Int. XXI measured 5.4 MJy/sr on average. Per-layer values, now seen to vary, convert starlight polarization into predicted polarized emissivity for each layer.
  • Free validation observables. The E/B ratio and EB correlation must change with frequency in a specific way if decorrelation is present, as Vacher et al. 2023 derived. Filament misalignment predicts TB and EB, as Clark et al. 2021 and Cukierman et al. 2023 showed. A layered model that reproduces these without tuning is strongly supported.
  • Component-wise SED fits. BICEP/Keck XVI fitted the HI-filament-correlated component separately and found an index of 1.54 with a 10 to 65 percent template correlation at multipoles 20 to 200. This is the first step of a layer-resolved SED measurement in a real B-mode field.
  • Higher frequencies. LiteBIRD's 402 GHz band and the extended configurations studied by Fuskeland et al. 2023 pin per-layer polarized temperatures. Spectral imaging of the QUBIC type can localize decorrelated residuals within a band.
  • Simulation suites for end-to-end testing. PySM d12 uses the six-layer Martínez-Solaeche model. DustFilaments builds millions of 3D filaments with tunable decorrelation. TIGRESS-based maps and Local Bubble cavity synthetic maps provide MHD realism.
  • High-latitude extinction inference. O'Callaghan et al. 2025 use normalizing flows to avoid artificial structure in low-column reddening, which is exactly the regime of B-mode fields.

3. Why this could matter for r

The relevant scale is set by the cleanest deep fields. Using the BK18 dust amplitude and a standard modified blackbody scaling gives the following picture for multipole 80.

QuantityValue
Dust BB power at 353 GHz, BICEP/Keck field4.4 μK²
Same dust scaled to 150 GHzabout 0.01 μK²
Primordial BB power for r = 1about 0.06 to 0.07 μK²
Dust at 150 GHz in r-equivalent unitsabout r = 0.15
r = 0.001 as a fraction of that dust powerabout 0.6 percent

So a next-generation limit requires the dust power at 150 GHz to be modeled to better than one percent at the recombination bump. In two-band template language the residual fraction is one minus the squared cross-frequency correlation, so the 150 by 353 correlation must exceed about 0.997.

The empirical route cannot deliver that. Under the standard log-squared frequency scaling, the decorrelation measured at 217 by 353 is amplified by a factor of about 3 at 150 by 353 and about 7 at 95 by 353. Planck 2018 and BK18 find no decorrelation, but BK18 does not free the parameter in its baseline because BK15 showed a free decorrelation parameter biases r in simulations. The community is currently choosing between fixing the parameter and inflating the error budget. A physically predicted decorrelation spectrum with a central value and a width is the third option, and it is the only one that improves with external data.

A useful conversion is that for a Gaussian spread of effective index, one minus the correlation equals half the index variance times the squared log frequency ratio. The 150 by 353 requirement then translates into an effective map-level index dispersion below about 0.09. Sponseller & Kogut 2022 find line-of-sight index spreads above 0.045 already bias per-pixel modified blackbody fits. The Pelgrims 15 percent cloud-to-cloud ratio spread corresponds to an index contrast near 0.3 between clouds. Those numbers bracket the regime where prediction, not marginalization, is needed.

What the approach can plausibly buy, stated conditionally:

  • Field selection is a pure gain. A decorrelation-risk map built from cloud counts, per-layer angles, and 3D column splitting lets an experiment place or weight its deep field where a single layer dominates. No modeling risk is incurred.
  • Residual scales with unexplained variance. If the layered model explains a fraction f of the variance of the effective complex index at the relevant scales, the residual bias falls by one minus f. Explaining half the variance halves the bias. Explaining 80 to 90 percent brings an initial dispersion of 0.1 down to 0.03 to 0.045, below the threshold above. No published polarization pipeline has demonstrated such fractions yet, and the BK XVI template correlations show the current state.
  • The angle-rotation term is the priority. A scalar index per pixel absorbs the real part of the moment but never the imaginary part. Starlight tomography gives per-layer angles directly, so it targets exactly the term that parametric fits cannot absorb.
  • LiteBIRD's reionization bump is the best case for this method. At multipoles below 30 the decorrelating structures are the nearest ones, the Local Bubble and Loop I, which are the best mapped. Fuskeland et al. 2023 note that assumed perfect correlation with a decorrelating 3D sky leaks straight into the CMB map at those scales.

4. Pipeline sketch

  1. Choose fields and build a risk map. Combine cloud counts, HI channel orientations, 3D column splitting, and starlight-versus-Planck angle differences into a per-pixel decorrelation-risk score. Use it for field selection and as a weight in later fits.
  2. Define layers per sightline. Segment the Edenhofer column into distance intervals using the 3D map, the velocity-resolved 3D HI atlas, and BISP-1 starlight inversions. Cap the layer count at two or three per sightline, with a regularizing prior tied to the data's frequency lever.
  3. Assign per-layer columns and environments. Take differential extinction per layer, R(V) per layer from the 3D extinction-curve map, gas phase from HI line widths and CO, and a Gaia-based 3D radiation field. Calibrate the extinction-to-emissivity conversion on intensity data using the 3D temperature tomography as a template.
  4. Build per-layer polarized SED priors. Use Astrodust plus PAH and THEMIS 2.0 to translate environment and R(V) into priors on temperature, index, curvature, and polarization fraction. Impose the flat polarization-fraction constraint and the small polarized-minus-intensity index offset. Use multi-band Serkowski peaks where available to set the aligned-grain temperature.
  5. Assign per-layer angles and inclinations. Take angles from starlight tomography and HI channel orientation. For the Local Bubble layer, take inclination from shell tangency. For other layers, use the polarization-fraction-versus-dispersion relation as a weak inclination prior.
  6. Forward model and fit hierarchically. Predict Stokes Q and U per band as the complex layer sum, convolved with beams and integrated over bandpasses. Fit jointly with Planck 217 to 857 GHz and the experiment's bands, with synchrotron included. External data enter as priors, not as fixed templates. Retain the full posterior covariance.
  7. Compress to moments and to a decorrelation spectrum. Project the layered posterior onto spin-moment maps and onto a predicted cross-frequency correlation spectrum per region. Report both the central prediction and its width.
  8. Feed the r likelihood in one of three ways. Replace the fixed decorrelation parameter in a BK-style spectral likelihood with the predicted spectrum and its prior width. Or run spin-moment fits with Gaussian priors on the moment maps centered on the prediction. Or subtract the posterior-mean dust map and propagate its covariance into a map-based likelihood.
  9. Validate and try to falsify. Hold out frequencies. Compare predicted and observed 217 versus 353 angle rotation in high-risk versus low-risk pixels. Check the predicted E/B and EB frequency dependence and the TB, EB signs from filament misalignment. Inject the model into PySM d12, DustFilaments, and MHD suites with the experiment's beams, bandpasses, masks, and systematics.
  10. Deliverables. A risk map, per-layer templates, moment priors with covariances, and a simulation suite that shares the same layers.

5. What to keep from the draft

  • The two-condition mechanism, with its clean statement that SED differences alone give curvature and misalignment alone factorizes.
  • The complex moment formulation and the explicit form of the first-order moment as an angle-weighted sum.
  • The honest observational status of decorrelation: an established possibility, detected in selected environments, not a universal measured coefficient.
  • The correct warning that 3D dust maps are extinction maps and that thermal continuum carries no distance label.
  • The point that a projected 2D field is the net line-of-sight quantity that decorrelation changes, so distance resolution is what is needed.
  • The layer forward model equation and the insistence on hierarchical fitting with external data as priors.
  • The full limitations list, especially identifiability and instrumental effects that mimic angle rotation.

6. Upgrade list

  1. Fix the citation: MNRAS 519, 4370 is McBride, Bull & Hensley 2023. Cite Sponseller & Kogut 2022 as ApJ 936, 8, and quote their thresholds of 1.6 K in temperature spread and 0.045 in index spread.
  2. Add the quantitative requirement in Section 3: dust at 150 GHz in the cleanest fields equals about r = 0.15, so r = 0.001 demands sub-percent dust power control and a 150 by 353 correlation above about 0.997.
  3. Add the log-squared amplification argument and BK18's actual practice, which is to hold decorrelation fixed because freeing it biases r. This is the strongest motivation for physical priors.
  4. Replace "knowledge of the grain size distribution" with concrete observables: 3D R(V) from Zhang & Green 2025, Serkowski peak wavelengths, and the coagulation-emissivity link in THEMIS 2.0.
  5. Reframe grain physics as a freedom-reducing constraint: flat polarization fraction, small polarized index offset, and the narrower aligned-grain size range that makes per-layer polarized SEDs simpler than intensity SEDs.
  6. Add the Local Bubble as the dominant first layer with a geometrically determined inclination, citing Alves 2018, Pelgrims 2020, and O'Neill 2024, and flag Halal et al. 2024's null result on the bubble's own imprint as an honest counterweight.
  7. Add Halal et al. 2024 and Mehandiratta et al. 2026 as direct evidence that column splitting depolarizes and that polarized emissivity per unit extinction polarization varies cloud to cloud.
  8. Add the velocity-resolved 3D HI atlas of McCallum et al. 2026 as the answer to the velocity-to-distance objection.
  9. Add the emission-to-extinction ratio as the physical bridge from starlight polarization to per-layer polarized emissivity.
  10. Add the E/B ratio, EB correlation, and TB predictions as free validation tests, citing Vacher et al. 2023 and Clark et al. 2021.
  11. Add BICEP/Keck XVI as the first component-wise SED fit in a real deep field, with its template correlation range as the honest current baseline.
  12. Add the decorrelation-risk map as a low-risk deliverable that helps even if the full forward model is never trusted for subtraction.
  13. State that the 2026 spin-moment analysis finds SRoll2 and PR4 disagree, so Planck alone cannot identify the moment origin, which is why external priors are informative.
  14. Replace the phrase "spectral-index decorrelation" throughout with "frequency decorrelation," and describe the target as the imaginary part of the first-order moment.
  15. Add a short LiteBIRD paragraph: the reionization-bump decorrelators are the nearest, best-mapped structures, and the high-frequency bands pin per-layer polarized temperatures.

Sources: