PhD defense, Université Paris Cité, 12 October 2026

Differentiable Field-Level Cosmology

Cosmic Microwave Background, Weak Lensing, and GPU-Accelerated Inference

Wassim KabalanUniversité Paris Cité, APC/CNRS

supervised by Eric Aubourg and Josquin Errard, co-advised by François Lanusse and Alexandre Boucaud

slides at askabalan.github.io/slides/2026_10_PhD

Stage 2 · SDSS, 2.5 m ground, 1.4″ seeing

Stage 4 · Euclid, 1.2 m space, 0.18″

Abell 2390 · Credit: SDSS DR18 (left); ESA / Euclid / Euclid Consortium / NASA, J.-C. Cuillandre & G. Anselmi (right)

Some of the Stage IV Cosmological Surveys

Microwave
λ ≈ 1–10 mm

wavelength

Optical / near-infrared
λ ≈ 0.3–2 µm

Space based

Ground based

LiteBIRD

full sky, from space
12 bands, 40–402 GHz

Euclid

14 000 deg²
30 gal arcmin⁻²
From Space
~1.5 billion galaxies

Simons Observatory

Atacama, Chile
6 bands, 27–280 GHz

Vera C. Rubin Observatory (LSST)

18 000 deg²
30 gal arcmin⁻²
Cerro Pachón in Chile
~2 billion galaxies

Modern observational cosmology

The Friedmann–Lemaître–Robertson–Walker metric

Credit: NASA (photograph); Hubble (1929), PNAS 15, 168, Fig. 1, redrawn

The Expanding Universe

Georges Lemaître: the Universe had a beginning
It expanded from an extremely dense early state

The Cosmic Microwave Background



1965, Holmdel: Penzias & Wilson find a 3 K signal from every direction · Credit: NASA (antenna); ESA / Planck Collaboration (map)

  • A hot, dense beginning should leave relic radiation: predicted by Gamow, Alpher & Herman in 1948, at a few kelvin
  • Found by accident in 1965 by Penzias & Wilson: a 3 K signal from every direction, interpreted by Dicke & Peebles at Princeton

  • The CMB is the oldest light we can observe

The CMB is one of the essential discoveries that confirm the hot Big Bang

Lambda Cold Dark Matter (\(\Lambda\)CDM)

Lambda Cold Dark Matter (\(\Lambda\)CDM)






The limits of \(\Lambda\)CDM

The early Universe

horizon

Credit: NASA / WMAP Science Team

flatness

+ origin of inhomogeneities
+ magnetic monopole problem

The late Universe

dark energy

Data: Perlmutter et al. (1999)

Nobel
2011

dark matter

Rubin & Ford (1970), Fig. 9 (M31)

Outline

Kabalan et al. (2026) · submitted to RASTI

Kabalan et al. (2026b) · accepted to JOSS
Kabalan et al. (2026c) · in prep. for JCAP

From model to observation: the CMB

Kabalan et al. (2026) · submitted to RASTI

Kabalan et al. (2026b) · accepted to JOSS
Kabalan et al. (2026c) · in prep. for JCAP

Inflation

The early Universe

horizon

flatness

  • A brief accelerated expansion solves the horizon and flatness problems
  • Its quantum fluctuations \(\delta\phi\) seed the large scale structure we see today

Primordial gravitational waves

  • Quantum fluctuations perturb the metric of space-time
  • Scalar: \(\Delta_s^2(k) = A_s\,(k/k_*)^{n_s-1}\)
  • Tensor: \(\Delta_t^2(k) = A_t\,(k/k_*)^{n_t}\), the gravitational waves

\[ r \equiv \frac{A_t}{A_s} \;\propto\; E_\mathrm{inf}^4 \]

The tensor-to-scalar ratio determines the energy scale of inflation

Primordial Gravitational Waves leave a unique imprint in the CMB polarisation

Polarisation of the CMB

The last scattering surface

Adapted from: Yacine Ali-Haïmoud (NYU, 2019)

Adapted from: Hu & White (1997)

  • A quadrupole at last scattering + Thomson scattering, \(\propto |\hat{\epsilon}\cdot\hat{\epsilon}'|^2\) → linear polarisation

The Stokes parameters

Adapted from: Emma Alexander, Wikimedia Commons (Stokes params)

Q and U Stokes parameters

\(Q\)
\(U\)
\(Q\)\(U\)
\(U\)\(-Q\)

\(Q\) and \(U\) depend on the reference frame: turning it by \(45^\circ\) swaps them, \(Q'\) \(=\) \(U\) and \(U'\) \(= -\)\(Q\)

\(E\) and \(B\) modes



Planck 2018 SMICA: temperature (colour), polarisation (sticks)

  • Density (scalar) perturbations make only \(E\) modes
  • At linear order, \(B\) modes are only produced by tensor perturbations

Two-point statistics

\(C(\theta) = \langle \Delta T(\hat n_1)\,\Delta T(\hat n_2)\rangle\) over all pairs at separation \(\theta\): its expansion in Legendre polynomials gives the \(C_\ell\)

Planck 2018 SMICA temperature map · smoothed to 1° FWHM

Predicted CMB power spectra

  • The primordial tensor perturbations are very faint (below than lensing B modes)
  • Measurement of \(r\) requires isolating the primordial B modes from all other contaminants.

Current constraints on \(r\)

Tristram et al. (2022); Simons Observatory (Ade et al. 2019); LiteBIRD Collaboration (2023)

The CMB data-analysis pipeline

My contribution

From model to observation: weak gravitational lensing

Kabalan et al. (2026) · submitted to RASTI

Kabalan et al. (2026b) · accepted to JOSS
Kabalan et al. (2026c) · in prep. for JCAP


The history of the expanding Universe



The expanding Universe and dark energy



redshift \(1 + z = \dfrac{\lambda_\mathrm{obs}}{\lambda_\mathrm{emit}} = \dfrac{1}{a}\)

scale factor \(a(t),\qquad a_0 = 1,\qquad H = \dot a / a\)

comoving distance \(\chi(z) = \displaystyle\int_0^z \frac{c\,dz'}{H(z')}\)

expansion rate \(H(z) = H_0\sqrt{\Omega_r(1+z)^4 + \Omega_m(1+z)^3 + \Omega_\Lambda}\)

Mapping dark matter through weak gravitational lensing


Matter along the line of sight bends light: every background galaxy is slightly distorted

Convergence and shear

galaxy

convergence \(\kappa\) (spin 0)

shear \(\gamma_1, \gamma_2\) (spin 2)

M101: NASA / ESA / Hubble (STScI-PRC2006-10a)

\[ \mathbf{A} = \begin{pmatrix} 1-\kappa-\gamma_1 & -\gamma_2 \\ -\gamma_2 & 1-\kappa+\gamma_1 \end{pmatrix} \]

Matter deforms galaxies through weak lensing, by a spin-0 field, the convergence \(\kappa\), and a spin-2 field, the two components of the shear \(\gamma\)

Distant galaxies lensed by the cosmic web

After Mandelbaum (2018), arXiv:1710.03235, Fig. 1 (image: CFHT) · shell of this work’s 3072³ run, z = 0.36

Dark Energy Survey Y3 redshift distribution

Tomographic projection to convergence maps

\[\kappa_i(\hat{\mathbf n}) = \int_0^{\chi_H} q_i(\chi)\,\delta(\chi\hat{\mathbf n},\chi)\,\mathrm{d}\chi\]

Shell rendering inspired by GLASS (Tessore et al. 2023)

From shear to cosmology

DATA

galaxy
shapes

shear \(\gamma_1, \gamma_2\)

Inverse Kaiser–Squires

convergence \(\kappa\) per bin

measured
\(\hat C_\ell^{\,ij}\)

THEORY

\(\theta = (\Omega_\mathrm{m}, \dots)\)

matter spectrum \(P(k, z)\)

Limber, with the lensing kernels \(q_i(\chi)\)

predicted
\(C_\ell^{\,ij}(\theta)\)

compare
↓
infer \(\theta\)

  • In traditional anaysis we compress the maps into angular power spectra \(C_\ell\).
  • This is called a two-point correlation function (2PCF)

The target of 2PCF analysis






  • \(\Omega_\mathrm{m}\): how much matter
  • \(\sigma_8\): how clumpy it is on \(8\,h^{-1}\mathrm{Mpc}\) scales
  • Weak lensing analysis targets the combination \(S_8 = \sigma_8\sqrt{\Omega_\mathrm{m}/0.3}\)

Current constraints on \(S_8\)

Asgari+21, Secco/Amon+22, Dalal+23, DES & KiDS 2023, Wright+25, DES Y6 2026, Planck 2018 VI

Limitation of the two-point analysis

the cosmic web, non-Gaussian

a Gaussian field

  • The 2PCF captures only the Gaussian features of a field

Limitation of the two-point analysis

DES Y6 cosmic shear (2026)

DES Y6 cosmic shear, DES Collaboration (2026), arXiv:2602.10065, Fig. 3 · CMB: Planck 2018, ACT DR6 and SPT-3G, TT+TE+EE+lowE, no lensing

  • We need to go beyond the 2PCF to extract more physical information: Contribution 2

Statistical inference for cosmology

Kabalan et al. (2026) · submitted to RASTI

Kabalan et al. (2026b) · accepted to JOSS
Kabalan et al. (2026c) · in prep. for JCAP

Inverse problems in cosmology

Solving Inverse Problems : Analytical likelihood


Analytical inversion (Stompor et al. 2009)

\[\mathbf{d} = \mathbf{A}(\boldsymbol\theta)\,\mathbf{s} + \mathbf{n}\]

the data are linear in the signal \(\mathbf{s}\), and the mixing \(\mathbf{A}\) depends on a few parameters \(\boldsymbol\theta\)

\[-2\ln\mathcal{L}(\boldsymbol\theta,\mathbf{s}) = \big(\mathbf{d}-\mathbf{A}\mathbf{s}\big)^{\top}\mathbf{N}^{-1}\big(\mathbf{d}-\mathbf{A}\mathbf{s}\big)\]

for Gaussian noise of covariance \(\mathbf{N}\)

\[\hat{\mathbf{s}}(\boldsymbol\theta) = \left(\mathbf{A}^{\top}\mathbf{N}^{-1}\mathbf{A}\right)^{-1}\mathbf{A}^{\top}\mathbf{N}^{-1}\mathbf{d}\]

setting the derivative with respect to \(\mathbf{s}\) to zero

\[-2\ln\mathcal{L}(\boldsymbol\theta) = -\left(\mathbf{A}^{\top}\mathbf{N}^{-1}\mathbf{d}\right)^{\top}\left(\mathbf{A}^{\top}\mathbf{N}^{-1}\mathbf{A}\right)^{-1}\left(\mathbf{A}^{\top}\mathbf{N}^{-1}\mathbf{d}\right)\]

substituting \(\hat{\mathbf{s}}\) back, up to a constant, leaves a likelihood of \(\boldsymbol\theta\) alone

Analytical inversion requires a model linear in \(\mathbf{s}\) with Gaussian noise

Solving Inverse Problems : Forward Modelling

Forward modelling

\[\hat{\mathbf{d}} = \mathcal{A}(\mathbf{s},\boldsymbol\theta)\]

\[-\ln\mathcal{L}(\boldsymbol\theta,\mathbf{s}) \;\propto\; \frac{\lVert \hat{\mathbf{d}} - \mathbf{d} \rVert_2^2}{\sigma^2}\]

Implicit inference

Explicit inference

  • Explicit inference is more interpretable than its implicit counterpart
  • But it comes at a higher computational cost
  • Implementing a survey-level explicit forward model is subject to : contribution 2

Bayesian inference in cosmology

Bayes’ theorem

\[\underbrace{p(\boldsymbol\theta\mid\mathbf{d})}_{\text{posterior}} \;=\; \frac{\overbrace{p(\mathbf{d}\mid\boldsymbol\theta)}^{\text{likelihood}}\;\overbrace{p(\boldsymbol\theta)}^{\text{prior}}}{\underbrace{p(\mathbf{d})}_{\text{evidence}}}\]

\[p(\boldsymbol\theta\mid\mathbf{d}) \;\propto\; p(\mathbf{d}\mid\boldsymbol\theta)\;p(\boldsymbol\theta)\]

Joint probabilistic modelling

\[p(\boldsymbol\theta,\mathbf{s}\mid\mathbf{d}) \;\propto\; \underbrace{p(\mathbf{d}\mid\mathbf{s},\boldsymbol\theta)}_{\text{likelihood}}\;\underbrace{p(\mathbf{s}\mid\boldsymbol\theta)}_{\text{latent prior}}\;\underbrace{p(\boldsymbol\theta)}_{\text{prior}}\]

Exploring the posterior

Finding the maximum a posteriori (MAP)

Sampling the full posterior

How do we measure how much we trust the result?

Differentiable methods for high-dimensional inference

Both methods become more efficient with a differentiable likelihood

Optimisation: the gradient points to the MAP

\[\boldsymbol\theta_{k+1} = \boldsymbol\theta_k + \eta\,\nabla_{\boldsymbol\theta}\ln p(\boldsymbol\theta\mid\mathbf{d})\]

Sampling: Hamiltonian Monte Carlo

\[\dot{\boldsymbol\theta} = \mathbf{p}, \qquad \dot{\mathbf{p}} = \nabla_{\boldsymbol\theta}\ln p(\boldsymbol\theta\mid\mathbf{d})\]

Adapted from Betancourt (2017), arXiv:1701.02434

For high-dimensional problems, differentiable likelihoods are required

Likelihood evaluation per independent sample: random walk \(\sim\) \(d\), HMC \(\sim\) \(d^{1/4}\)

Neal (2011); Beskos et al. (2013)

Differentiable programming



  • JAX computes exact gradients of numerical code by automatic differentiation
  • The same code runs on CPUs, GPUs and multi-GPU clusters
  • I used Optax for optimisation
  • BlackJAX and NumPyro for sampling

Can better numerical methods improve cosmological pipelines?

Contribution 1: spatially varying map-based component separation

Kabalan et al. (2026) · submitted to RASTI

Kabalan et al. (2026b) · accepted to JOSS
Kabalan et al. (2026c) · in prep. for JCAP

From sky data to clean CMB maps

CMB component separation

  • To estimate \(r\), the CMB must be separated from the foregrounds that contaminate it
  • LiteBIRD observes the sky in 12 frequency bands, from 40 to 402 GHz

Foreground spectral energy distributions

  • Each component has its own SED: synchrotron dominates at low frequencies, dust at high frequencies
  • The primordial \(B\) modes are best seen near the foreground minimum, around 80 GHz

Parametric component separation: the spectral model

Synchrotron and dust (Galactic foregrounds)

Synchrotron: power law

\[A^{\mathrm{sync}}_\nu({\color{#3b6fb6}\beta_s}) \;\propto\; \left(\frac{\nu}{\nu_0}\right)^{\color{#3b6fb6}\beta_s}\]

Planck: synchrotron

Dust: modified black body

\[A^{\mathrm{dust}}_\nu({\color{#d68910}\beta_d}, {\color{#d68910}T_d}) \;\propto\; \left(\frac{\nu}{\nu_0}\right)^{{\color{#d68910}\beta_d}+1}\frac{e^{h\nu_0/k{\color{#d68910}T_d}}-1}{e^{h\nu/k{\color{#d68910}T_d}}-1}\]

Planck: thermal dust

\[\mathbf{d}_\nu = A^{\mathrm{cmb}}_\nu\,\mathbf{s}_{\mathrm{cmb}} + A^{\mathrm{dust}}_\nu({\color{#d68910}\beta_d},{\color{#d68910}T_d})\,\mathbf{s}_{\mathrm{dust}} + A^{\mathrm{sync}}_\nu({\color{#3b6fb6}\beta_s})\,\mathbf{s}_{\mathrm{sync}} + \mathbf{n}_\nu\]

\[\hat{\mathbf{s}}(\boldsymbol\beta) = \left(\mathbf{A}^{\top}\mathbf{N}^{-1}\mathbf{A}\right)^{-1}\mathbf{A}^{\top}\mathbf{N}^{-1}\mathbf{d}\]

Rizzieri et al. (2025), arXiv:2510.08534

The spectral likelihood

\[-2\ln\mathcal{L}(\boldsymbol\beta) = -\left(\mathbf{A}^{\top}\mathbf{N}^{-1}\mathbf{d}\right)^{\top}\left(\mathbf{A}^{\top}\mathbf{N}^{-1}\mathbf{A}\right)^{-1}\left(\mathbf{A}^{\top}\mathbf{N}^{-1}\mathbf{d}\right), \qquad \boldsymbol\beta = ({\color{#d68910}\beta_d}, {\color{#d68910}T_d}, {\color{#3b6fb6}\beta_s})\]

Estimating \(r\) from a clean CMB map

\[C_\ell^{BB,\mathrm{obs}} = C_\ell^{BB,\mathrm{CMB}} + C_\ell^{\mathrm{lens}} + C_\ell^{\mathrm{stat}}\]

\[\text{with}\quad C_\ell^{\mathrm{stat}} = \big\langle C_\ell^{BB,\mathrm{noise}} \big\rangle_n\]

\[C_\ell^{BB,\mathrm{model}}(r) = {\color{#C2560A} r}\,C_\ell^{\mathrm{tensor}} + C_\ell^{\mathrm{lens}} + C_\ell^{\mathrm{stat}}\]

\[-2\ln\mathcal{L}_\mathrm{cosmo}(r) = f_\mathrm{sky}\sum_{\ell}(2\ell+1)\left[\frac{C_\ell^{BB,\mathrm{obs}}}{C_\ell^{BB,\mathrm{model}}(r)} + \ln C_\ell^{BB,\mathrm{model}}(r)\right]\]

Spatially varying foregrounds


c1d0s0 one spectrum for the whole sky

synchrotron index \(\beta_s\)

dust index \(\beta_d\)

dust temperature \(T_d\)

SEDs of 300 pixels

c1d1s1 one spectrum per pixel

Sky models: PySM (Thorne et al. 2017; Zonca et al. 2021; Pan-Experiment Galactic Science Group 2025)

  • In c1d1s1 each pixel has its own spectrum. The likelihood must model this spatial variability
  • Both skies have \(r = 0\) and serve for validation

How do we represent sky data numerically?


Hierarchical Equal Area iso-latitude Pixellation (of the sphere)

HEALPix

\(N_\mathrm{side}\) \(= 64\)

Defining parameter patches on the sphere

Multi-resolution (HEALPix) LiteBIRD Collaboration (2023)

\(N_{\rm side} = 4\) · 192 patches

\(N_{\rm side} = 16\) · 3072 patches

\(N_{\rm side} = 32\) · 12288 patches

\(N_{\rm side} = 64\) · 49152 patches

low · mid · high Galactic latitude, 20 % of the sky each

Planck Galactic masks: Planck Collaboration (2020), A&A 641, A4




\(N_{\rm side}\) per parameter in the LiteBIRD analysis

low mid high
\(\beta_d\) 64 64 64
\(T_d\) 8 4 0
\(\beta_s\) 4 2 2

per Galactic region (0: one patch)

Example: the \(T_d\) partition

\(T_d\) at \(N_{\rm side}\) 8 / 4 / 0 · 344 patches

Defining parameter patches on the sphere

Equal-area patches on the sphere

Spherical \(K\)-means

The bias–variance trade-off

Obstacle for full-sky CMB surveys: LiteBIRD

  • The number of patches must be tuned to meet the requirements of future surveys
  • We want to test many configurations and keep the best one
  • This needs a fast optimisation method and a selection criterion for patch configurations

Fast and robust minimisation

How do we minimise?

The Newton method

\[\boldsymbol\beta_{k+1} = \boldsymbol\beta_k - \mathbf{H}_k^{-1}\,\mathbf{g}_k\] \[\mathbf{g} = \nabla_{\boldsymbol\beta}\left(-\ln\mathcal{L}\right), \qquad \mathbf{H} = \nabla^2_{\boldsymbol\beta}\left(-\ln\mathcal{L}\right)\]

140 GHz, LiteBIRD depth

  • Noisy sky data make the Hessian poorly conditioned and hard to invert
  • Quasi-Newton methods suffer the most since they approximate the inverse Hessian.

Adaptive moment-based methods

Adam is the standard optimiser of machine learning, built for very many parameters and very noisy gradients

\[\boldsymbol\beta_{t+1} = \boldsymbol\beta_t - \eta\,\frac{\hat m_t}{\sqrt{\hat v_t}+\epsilon}\]

\(m_t\), \(v_t\): running means of \(\mathbf{g}\) and \(\mathbf{g}^2\)

AdaBelief scales the step by the spread of the gradient instead

\[v_t \;\longrightarrow\; s_t:\ \text{running mean of }(\mathbf{g} - m_t)^2\]

A noisy direction gets a smaller step

AdaTopK builds on AdaBelief for physical parameters, with a robust handling of parameter bounds

Scaling of AdaTopK

Scaling

  • AdaTopK runs on GPUs with first-order gradients only.
  • It scales well as number of patches grows.

Selection criteria for optimal patches

Chanial, Biquard, Kabalan et al. (2026), FURAX, arXiv:2603.19600, submitted to JOSS
Kabalan et al. (2026), arXiv:2604.08463, submitted to RASTI

Selection criteria for optimal patches

Jean Zay, IDRIS
NVIDIA H100 GPUs

We do this for each region

\(\beta_d\) patches

\(T_d\) patches

\(\beta_s\) patches

Run details

Over 10 000 component separation runs on 64 NVIDIA H100 GPUs

Finishing in 10 hours

Results: comparison with multi-resolution

multi-resolution (LiteBIRD PTEP)

\(\beta_d\)

\(T_d\)

\(\beta_s\)

spherical \(K\)-means (this work)

\(\beta_d\)

\(T_d\)

\(\beta_s\)

30 % lower 68 % upper limit on \(r\), with the same instrument, sky and noise: only the partition changes

Contribution 1: conclusion

  • A data-driven method to select the optimal patch configuration, essential for Stage IV full-sky surveys like LiteBIRD
  • A faster, more robust parametric map-based component-separation code
  • Contributed to the CMB data-analysis package FURAX and its component-separation sub-package furax-cs

Kabalan et al. (2026), JAX-accelerated parametric CMB component separation, submitted to RASTI

Now part of the Megatop pipeline for the Simons Observatory

Perspectives

  • preparing its integration into the LiteBIRD component-separation pipeline
  • adding instrument systematics in FURAX

Contribution 2: field-level inference for weak lensing surveys

Kabalan et al. (2026) · submitted to RASTI

Kabalan et al. (2026b) · accepted to JOSS
Kabalan et al. (2026c) · in prep. for JCAP

A weak-lensing forward model, end to end

From a cosmology and an initial density field to convergence and shear on the curved sky

Building a survey-level differentiable forward model

A fast and differentiable gravity solver that can simulate Stage IV survey volumes

The gravity solver: Lagrangian perturbation theory



Each particle moves from its initial grid position \(\mathbf{q}\)

\[\mathbf{x}(\mathbf{q}, a) = \mathbf{q} + D_1(a)\,\mathbf{\Psi}^{(1)}(\mathbf{q}) + D_2(a)\,\mathbf{\Psi}^{(2)}(\mathbf{q}) + \dots\] \[\nabla \cdot \mathbf{\Psi}^{(1)} = -\delta_\mathrm{ini}\]

First order: the Zel’dovich approximation.

  • Perturbation theory does not capture the non-linear growth of structure
  • We need a fast and differentiable gravity solver that can model the non-linear growth of structure

Matter power spectrum at \(z=0\), linear theory against halofit

The gravity solver: particle mesh

Gravity through FFTs on a mesh: fast, differentiable, and limited by the mesh resolution

Accuracy of the particle mesh

  • The accuracy of a particle-mesh simulation is limited by the resolution of its mesh
  • At survey volume, a mesh with that resolution does not fit on one GPU (For \(4096^3\) requires 1.6 TB of memory).
  • We need a scalable gravity solver

Distributing the simulation

  • We split the box into pencils, one per GPU
  • jaxDecomp implements a differentiable, distributed FFT for the particle-mesh solver

Kabalan et al. (2026b), jaxDecomp: JAX library for 3D domain decomposition and parallel FFTs, accepted to JOSS

Building the observable

Building the observable: requirements

Simulation-based inference on real survey data

Applied on DES Y3, Jeffrey et al. (2024): SBI on real data using Gower Street Similations and validated on CosmoGridV1.

Creating the lightcone

Calibrating the lightcone model

  • The integrator and step size
  • The mesh resolution
  • The mass assignment and the spherical painting
  • The number and spacing of the density shells
  • The interpolation of the particles onto the lightcone
  • The gradient (adjoint) of the particle-mesh ODE
  • The performance of the forward model and its gradient

Building the observable

From shells to convergence: the Born approximation

Born approximation: straight lines of sight

\[\kappa(\hat{\mathbf n};\chi_s) = \frac{3}{2}\,\Omega_m\Big(\frac{H_0}{c}\Big)^2 \sum_j W_j(\chi_s)\,\delta_j(\hat{\mathbf n})\]

\[W_j(\chi_s) = \int_{\mathrm{shell}\ j} \mathrm{d}\chi\;\chi\,(1+z)\Big(1-\frac{\chi}{\chi_s}\Big)\]

\[\kappa_i(\hat{\mathbf n}) = \int \mathrm{d}z\; n_i(z)\,\kappa\big(\hat{\mathbf n};\chi_s(z)\big)\]

The Born approximation computes the convergence along undeflected light rays

Post-Born ray tracing computes deflection angle, but its correction to the convergence power spectrum is below the arcminute scale.

Valid scales for real survey analysis

Kabalan et al. (2026c), in prep. for JCAP

Using the forward model for field-level inference

Field-level likelihood

Gaussian likelihood on the convergence pixels

\[-\ln\mathcal{L}(\boldsymbol\theta,\mathbf{s}) = \sum_i \frac{\lVert \color{#c2560a}{\hat{\mathbf{d}}_i} - \mathbf{d}_i \rVert_2^2}{2\,\color{#3b6fb6}{\sigma_i^2}}\]

\(\color{#c2560a}{\hat{\mathbf{d}}_i} = \kappa_i(\mathbf{s},\boldsymbol\theta)\) the predicted and \(\mathbf{d}_i\) the observed convergence map of bin \(i\)

Noise per pixel

\[\color{#3b6fb6}{\sigma_i} = \frac{\sigma_e}{\sqrt{n_{\mathrm{gal},i}\;\Omega_\mathrm{pix}}}\]

DES Y3 bin 1 bin 2 bin 3 bin 4
\(n_\mathrm{gal}\) [arcmin\(^{-2}\)] 1.476 1.479 1.484 1.461
\(\sigma_e\) 0.26 0.26 0.26 0.26

\(\sigma_e\) shape noise per component, \(n_{\mathrm{gal},i}\) galaxies per arcmin² in bin \(i\),
\(\Omega_\mathrm{pix}\) pixel area in arcmin\(^2\)

Reconstructing the initial conditions

Preliminary
DES Y3 noise, \(1200^3\) mesh, 2LPT
\(N_\mathrm{side} = 1024\), painted at \(512\), \(\ell \le 700\)

Contribution 2: conclusion and perspectives

  • The first fully differentiable forward model for Stage IV weak-lensing surveys
  • A massively distributed \(N\)-body solver, lightcone and lensing model
  • One step closer to field-level inference on real survey data

Kabalan et al. (2026b), jaxDecomp, accepted to JOSS · Kabalan et al. (2026c), in prep. for JCAP

Perspectives

  • Jointly infer cosmology and initial conditions from real survey data
  • A better gravity solver: learned force corrections or a differentiable moving mesh
  • Intrinsic alignments and photo-z errors as latent variables

Moving-mesh particle mesh

Moving mesh: B. Horowitz (2026), arXiv:2607.10983

Conclusion and perspectives

Kabalan et al. (2026) · submitted to RASTI

Kabalan et al. (2026b) · accepted to JOSS
Kabalan et al. (2026c) · in prep. for JCAP

The software behind this thesis

Reproducible software for research

Where it is used

furax-cs → the Megatop pipeline for the Simons Observatory; LiteBIRD component separation next

jax-healpy → used within the DESI collaboration

jaxDecomp → beyond cosmology: fluid dynamics, ocean models, imaging

Scientific goal: applications to CMB × LSS with CMB lensing

Conclusion

Removing computational bottlenecks for Stage IV surveys

Early Universe · inflation
Spatial variability for the CMB

Late Universe · dark matter
Explicit field-level inference for large-scale structure

Thank you

Backup

Backup: General

Why not simply the maximum a posteriori?

Backup: CMB

Spherical Harmonic Transforms

\[ f(\theta,\varphi) = \sum_{\ell=0}^{\infty} \sum_{m=-\ell}^{\ell} a_{\ell m}\, Y_\ell^m(\theta,\varphi) \]

\[ C_\ell = \frac{1}{2\ell+1} \sum_{m=-\ell}^{\ell} |a_{\ell m}|^2 \]

I

spin-0

\(|a^T_{\ell m}|^2\)

Q

spin-2

\(|a^E_{\ell m}|^2\)

U

\(|a^B_{\ell m}|^2\)

Credit: ESA / Planck Collaboration · Planck 2018 SMICA maps (inpainted), \(I\), \(Q\), \(U\)

AdaTopK: handling parameter bounds

Patch selection on skies with \(r \neq 0\)

Backup: Weak lensing

From galaxy shapes to shear

M101: NASA / ESA / Hubble (STScI-PRC2006-10a)

ellipticity \[e = e_1 + i\,e_2 = \frac{a-b}{a+b}\,e^{2i\varphi}\]

shear in a pixel \[\hat\gamma_1 = \frac{\sum_k w_k\,e_{1,k}}{\sum_k w_k}, \qquad \hat\gamma_2 = \frac{\sum_k w_k\,e_{2,k}}{\sum_k w_k}\]

Kaiser–Squires on the sphere

From \(\kappa\) to \(\gamma\)

\[ \gamma^E_{\ell m} = -\sqrt{\frac{(\ell+2)(\ell-1)}{\ell(\ell+1)}}\;\kappa_{\ell m}, \qquad \gamma^B_{\ell m} = 0 \] \[ \gamma_1 + i\gamma_2 = \sum_{\ell m} \gamma^E_{\ell m}\;{}_2Y_{\ell m} \]

From \(\gamma\) back to \(\kappa\)

\[ \kappa_{\ell m} = -\sqrt{\frac{\ell(\ell+1)}{(\ell+2)(\ell-1)}}\;\gamma^E_{\ell m} \]

convergence κ

shear γ1

shear γ2

  • Kaiser–Squires is a simple per-\(\ell\) rescaling between \(\kappa\) and the \(E\)-mode of the shear

Linear theory against halofit: lensing \(C_\ell\)

A survey mask leaks

Backup: Field level

Integrators and time steps

Observer positioning

Shell spacing and the lightcone

Shell spacing: convergence against CosmoGrid

Uniform scale factor against equal volume

Drift on the lightcone

10 shells, one redshift per shell

10 shells, drifted

40 shells, one redshift per shell

\[a_{\mathrm{c}} = a\big(\chi = \lVert \mathbf{x} - \mathbf{x}_{\mathrm{obs}} \rVert\big), \qquad \mathbf{x}(a_{\mathrm{c}}) = \mathbf{x}(a_i) + \boldsymbol{\pi}\,\big[D(a_{\mathrm{c}}) - D(a_i)\big]\]

\(a_i\): scale factor of the shell, \(a_{\mathrm{c}}\): scale factor at which the particle crosses the lightcone, \(\boldsymbol{\pi} = \mathrm{d}\mathbf{x}/\mathrm{d}D\): BullFrog momentum

Interpolating the particles onto the observer’s lightcone

Each particle is drifted to the scale factor at which it crosses the lightcone

Impact of the drift on the first shell

Removes the single-redshift bias of wide shells: 17 % on the first equal-volume shell

Painting on spherical shells

evolved particles

⟶

one HEALPix shell

Painting on spherical shells

Bilinear interpolation

Four pixels, weighted by the position between their centres: get_interp_weights, implemented in jax-healpy

Radial basis function interpolation

The pixel and its eight neighbours (get_all_neighbours, implemented in jax-healpy), weighted by a Gaussian kernel of adjustable width

Painting: RBF against nearest grid point

Valid scales: finer bandpowers

Mesh plateau: density shells

Mesh plateau: convergence

The ghost zone

Scale cuts and baryons

map2alm

× \(T_\ell\)

alm2map

\[T_\ell = \begin{cases} 1, & \ell \le \ell_\mathrm{cut} - \ell_\mathrm{width} \\ \tfrac12\left(1 + \cos \pi x\right), & \ell_\mathrm{cut} - \ell_\mathrm{width} < \ell < \ell_\mathrm{cut} \\ 0, & \ell \ge \ell_\mathrm{cut} \end{cases}\]

\[x = \frac{\ell - (\ell_\mathrm{cut} - \ell_\mathrm{width})}{\ell_\mathrm{width}}\]

Tersenov, Guerrini, Starck & Kilbinger (2026), arXiv:2609.09131

Higher-order statistics: the starlet \(\ell_1\) norm

Ajani et al. (2021), arXiv:2101.01542; Tersenov et al. (2026), in prep.

Starlet \(\ell_1\) norm against CosmoGrid

MAP: power spectra of the reconstruction

MAP: DES Y3 against Euclid, coherence

MAP: DES Y3 against Euclid, starlet \(\ell_1\) norm

MAP: starlet \(\ell_1\) norm of the reconstruction

Field-level posterior: convergence

truth

posterior draws

standard deviation

truth − draw

bin 2

bin 3

MCLMC on one A100: \(128^3\) initial conditions and \((\Omega_c, \sigma_8)\) Wall time
Compiling the tuning step 12 min
Tuning the step size and momentum decoherence: 300 steps 1 h 32 min
Compiling the sampler 1.5 min
Sampling: 10 batches of 10 draws, one draw every 5 steps 15.5 min per batch
Total ≈ 4 h 20 min

Field-level posterior: starlet \(\ell_1\) norm

bin 2
ℓ₁ norm

bin 2
draw − truth

bin 3
ℓ₁ norm

bin 3
draw − truth

Cosmology-only inference

NUTS, initial conditions at the truth
Warmup 100 iterations
Sampling 10 batches of 10 draws
\(\Omega_c\) \(0.2603 \pm 0.0013\) (truth 0.2607)
\(\sigma_8\) \(0.8102 \pm 0.0025\) (truth 0.8102)
Wall time 4 h 56 min

Photometric redshift uncertainty: hyperrank

Jeffrey et al. (2024), arXiv:2403.02314; Cordero et al. (2022)

Intrinsic alignment in the forward model

Tsaprazi et al. (2022), arXiv:2112.04484

Backup: Numerical and technical details

The two adjoints

Distributed Fourier transform

Distributed transpositions

Halo exchange

Scaling on supercomputers

Jean Zay, IDRIS
NVIDIA H100 GPUs

Performance of the forward model

Performance of the gradient

Particle mesh against moving mesh

Jean Zay, IDRIS
NVIDIA H100 GPUs

Lightcone from a replicated box

Sgier et al. (2021), arXiv:2007.05735