Why do cosmological simulations need strict conservation laws when game engines tolerate small approximation errors?

Great question—it cuts to the heart of why cosmology is fundamentally different from video games.

Game engines can tolerate approximation errors because they're *interactive systems* where the user perceives only the present frame. A tiny energy leak or momentum fudge in frame *n* doesn't compound visibly into frame *n*+1 because rendering is decoupled from physics; what matters is that objects look plausible *right now*. Cosmological simulations, by contrast, are *predictive systems* where errors accumulate over billions of simulated years.

The accumulation problem:

In cosmology, you're trying to predict how small density fluctuations grow into galaxies and clusters across cosmic time. A 0.1% energy leak per timestep sounds harmless, but when you run millions of timesteps, that leak compounds into a systematic bias [AG-2024.03-936]. Energy non-conservation isn't just sloppy—it changes the temperature, pressure, and dynamics of your simulated matter. Since you're testing models against real observations of structure in today's universe, even small biases in total energy propagate into wrong predictions of where galaxies should be.

Why conservation laws are the frontier challenge:

The archive shows this is an *active research problem*. In dark matter self-interactions, [energy leaks emerge from "asymmetric gravity calculations, time-asymmetric particle stepping, and collision kicks" [AG-2024.03-936]. For relativistic hydrodynamics, maintaining positive density and pressure while conserving energy in curved spacetime requires specially designed numerical schemes [AG-2024.10-129, AG-2025.05-504]. Even standard Newtonian codes show significant field-level differences when you compare cosmological N-body codes like Gadget and Gizmo—some of which trace to how they handle energy and momentum [AG-2025.05-214].

The practical reason: *cosmology is falsifiable*. We measure the power spectrum of matter distribution in the real universe to parts per million. Your simulation's answer has to match that, which means systematic biases from energy slop are fatal.