Implicit vs Explicity field level inference

Carolina Cuesta-Lazaro (Harvard U, Cambridge)

Sep 23. 2025, 09:20 — 09:40

Field-level inference (FLI) directly models the full density field rather than compressed summary statistics. Two distinct approaches emerge: explicit FLI, which jointly infers cosmological parameters and initial conditions p(δ_ICs, θ|δ_Obs) to create true "digital twins" of our Universe, and implicit FLI, which directly constrains cosmology p(θ|δ_Obs) without reconstructing initial conditions and thus cannot provide a complete digital twin.

Explicit FLI enables reconstruction of dark matter distributions and formation histories but requires sampling in O(10^9) dimensional spaces. Implicit FLI extracts optimal summaries from the data, reducing the problem to O(10-100) dimensions while potentially achieving similar constraining power.

Machine learning advances in generative modeling now make high-dimensional cosmological inference tractable. The choice between approaches involves fundamental trade-offs: explicit FLI provides complete physical reconstruction at a high computational cost, while implicit FLI offers efficiency but sacrifices reconstruction capabilities. Key challenges include developing robust neural validation frameworks and scalable sampling methods for realistic forward models.

Further Information
Venue:
ESI Boltzmann Lecture Hall
Associated Event:
Putting the Cosmic Large-scale Structure on the Map: Theory Meets Numerics (Workshop)
Organizer(s):
Oliver Hahn (U of Vienna)
Cornelius Rampf (RBI, Zagreb)
Cora Uhlemann (Bielefeld U)