Fisher Information

Information gradient, statistical curvature, parameter sensitivity
distributionaldim statistical manifold3 metrics

What It Measures

How sharply the signal's distribution changes when you perturb the histogram.

Treats the 16-bin histogram as a point on a statistical manifold — a curved space where each point represents a probability distribution. The Fisher information matrix measures the curvature at that point: high curvature means a small change in the data would radically shift the distribution (informationally sensitive). Low curvature means the distribution is robust to perturbation.

Metrics

effective_dimension

How many independent directions matter in the Fisher matrix? Computed as the participation ratio of eigenvalues. De Bruijn, phyllotaxis, and circle map quasiperiodic all score 16.0 (the maximum — all 16 bins are equally important, so the statistical manifold is fully 16-dimensional). Seismic b-value scores 2.27 (its distribution has only 2-3 effective degrees of freedom, despite occupying many bins). This measures the intrinsic dimensionality of the signal's distributional footprint.

log_det_fisher

Logarithm of the determinant of the Fisher matrix. This is the log-volume element of the statistical manifold at the data's location. Collatz parity (137.4), Symbolic Henon (137.4), and Fibonacci word (137.3) score highest — their sparse, peaked distributions create enormous Fisher curvature (tiny probabilities in many bins produce large 1/p terms). De Bruijn scores 44.4 (the minimum for a 16-bin uniform — all probabilities equal 1/16). The 93-unit range across the atlas spans 40 orders of magnitude in actual determinant value.

trace_fisher

Sum of diagonal Fisher matrix entries (sum of 1/p_i for each bin). Collatz parity (229,605), Symbolic Henon (229,604), and Fibonacci word (229,604) score highest for the same reason as log_det: near-empty bins dominate the trace. De Bruijn scores 256 (16 bins, each with probability 1/16, so 16 * 16 = 256). Trace and log_det are correlated but not identical: log_det captures the product of per-bin information (sensitive to the emptiest bin), while trace captures the sum (sensitive to the total information budget).

Atlas Rankings

effective_dimension
SourceDomainValue
De Bruijn Sequencenumber_theory16.0000
Phyllotaxisbio15.9959
Circle Map Quasiperiodicchaos15.9959
···
Constant 0xFFnoise0.0000
Constant 0x00noise0.0000
Seismic b-value (SoCal)geophysics2.2681
log_det_fisher
SourceDomainValue
Collatz Paritynumber_theory137.3795
Symbolic Henonexotic137.3695
Fibonacci Wordexotic137.3158
···
Constant 0xFFnoise0.0000
Constant 0x00noise0.0000
De Bruijn Sequencenumber_theory44.3614
trace_fisher
SourceDomainValue
Collatz Paritynumber_theory229604.5187
Symbolic Henonexotic229604.4734
Fibonacci Wordexotic229604.2396
···
Constant 0xFFnoise0.0000
Constant 0x00noise0.0000
De Bruijn Sequencenumber_theory256.0000

When It Lights Up

Fisher Information geometry detects a specific distributional property: how many bins are nearly empty. Signals that use only a few of 16 bins — binary sequences, periodic orbits, sparse event streams — create Fisher matrices with explosive curvature because the nearly-empty bins contribute 1/p terms near infinity (Laplace smoothing prevents actual infinity but preserves the relative ranking). In the atlas, effective_dimension separates "genuinely multi-valued" signals (dimension near 16) from "effectively binary" signals (dimension 2-4), providing a distributional complexity measure independent of entropy.

Open in Atlas
← WassersteinZipf–Mandelbrot (8-bit) →