Measuring the cosmic web’s statistical memory—and the limits of the measurement

Comparing the Universe with its earlier self

Gravity transforms small density fluctuations into a web of sheets, filaments and dense clusters. Measuring how strongly matter clusters at one time is not the same as measuring how much its arrangement reveals about another time.

In a revised arXiv preprint, Biswajit Pandey and Anindita Nandi propose using mutual information to address that second question. This statistical quantity measures how much knowing one variable reduces uncertainty about another. It can detect nonlinear relationships, making it a potential complement to conventional clustering statistics rather than a replacement for them.

Their study compares simulated matter densities at matching spatial locations across cosmic epochs. The resulting measure describes shared statistical dependence—not a directional flow of information or a complete inventory of the Universe’s physical information.

Turning simulated matter into probabilities

The researchers use particle-mesh simulations of collisionless dark matter in the standard cosmological framework containing cold dark matter and a cosmological constant. Their cubic simulation volumes have side lengths of 128, 256 and 512 megaparsecs, with ten independent realizations for each size to estimate uncertainties.

Each volume is divided into cubic cells, or voxels. The team counts particles inside each cell at two epochs, groups those counts into density bins, and estimates how frequently different pairs of density bins occur at matching locations. These probabilities provide the mutual information.

Finite samples can produce misleading information estimates. The authors therefore apply a Miller–Madow entropy correction and subtract a reference measurement from independent, structureless random particle distributions with matching particle numbers and box sizes. They also test a shuffled-count reference that preserves each snapshot’s distribution of cell counts while disrupting the correspondence between epochs.

These procedures reduce statistical contamination; they do not make the estimator immune to limited sampling.

A peak that is not a cosmic ruler

For the 512-megaparsec simulation, comparing the present-day density field with redshift one produces little excess mutual information at cell sizes below roughly eight megaparsecs. The signal grows as cells become larger, peaks near 64 megaparsecs, and then falls.

That peak might suggest a preferred scale of cosmic coherence. The robustness tests argue against that interpretation.

In smaller simulations, the maximum shifts to about one-eighth of the box length: approximately 16 megaparsecs for the 128-megaparsec box and 32 megaparsecs for the 256-megaparsec box. Crucially, 128-megaparsec subvolumes extracted from larger simulations show similar peak locations to an independent 128-megaparsec simulation, despite inheriting fluctuations from larger parent volumes.

The authors conclude that the maximum is governed primarily by the number of cells available to estimate probabilities. Enlarging cells reduces that number rapidly. Changing the density binning or reference baseline leaves the broad pattern intact. The peak is therefore an empirical warning about this estimator’s reliability, not evidence for a universal physical boundary.

Filaments retain stronger statistical overlap

The team also separates the simulated cosmic web into sheets, filaments and clusters using a classification based on the gravitational tidal field, with the underlying density field smoothed over two megaparsecs.

Between redshift one and the present, filament-specific density fields show greater shared information than those associated with sheets or clusters. This hierarchy persists when all three environments are restricted to the same particle count: 46,370.

Sheets generally have higher individual density-field entropies at small and intermediate cell sizes, yet filaments have higher mutual information. Their advantage therefore cannot be explained simply by greater marginal information content.

The authors suggest that filaments’ comparatively persistent organization may matter, while sheets continue collapsing and clusters undergo mergers and mixing. The precise dynamical explanation remains unresolved. This analysis compares environment-specific density fields at each epoch; it does not follow an unchanged collection of structures through time.

Accessible memory is not all information

Comparing the present with initial conditions at redshift approximately 124 yields an extremely weak shared signal. Some coarse-cell estimates fall below the reference baseline. Those negative differences indicate estimator uncertainty, not negative mutual information.

Nor does weak overlap demonstrate that gravity destroys information: fixed-location, coarse-grained densities omit particle motions and detailed phase-space relationships.

Models with larger cosmological-constant contributions also show higher shared information in these tests, consistent with slower late-time structural evolution. This is sensitivity in simulations, not an observational cosmological constraint. The preprint’s central lesson is that information-based comparisons can reveal useful differences—but only with explicit sampling, baseline and spatial-resolution checks.