Coarse-grained multifractal analysis: core definitions #
This file lays the abstract, measure-free foundation for the coarse-grained multifractal analysis
of a finite weight family p : ι → ℝ (think p i = μ(cell i) for the cells of a partition at
scale ε). All quantities are defined on a bare family p; the hypotheses one needs for the
theory (0 ≤ p i, ∑ i, p i = 1, 0 < ε < 1) are carried on the lemmas, never baked
into the definitions, so the algebra stays clean and the same definitions specialize to any
measure.
The exponent ^ throughout is Real.rpow (real-base, real-exponent power), which is the right
notion for the continuous parameter q : ℝ.
Main definitions #
Oseledets.Multifractal.partitionFunction: the generalized partition functionZ_q = ∑_{i : p i > 0} (p i) ^ q. The guard0 < p iis load-bearing atq = 0: it forces empty cells to contribute0(soZ_0counts occupied cells) instead ofReal.rpow 0 0 = 1.Oseledets.Multifractal.massExponent: the mass exponentτ(q) = log Z_q / log ε.Oseledets.Multifractal.renyiDim: the Rényi / generalized dimensionD_q = τ(q) / (q - 1), with theq = 1information-dimension branch(∑ i, p i log p i) / log εsupplied directly (theL'Hôpitallimit), since the general formula is0 / 0there.Oseledets.Multifractal.singularitySpectrum: the singularity spectrumf(α) = ⨅ q, q α - τ(q), the Legendre transform ofτ. This is an infimum: becauseτis concave, the supremum would be+∞.
Main results #
Oseledets.Multifractal.partitionFunction_nonneg:Z_q ≥ 0when0 ≤ p i.Oseledets.Multifractal.partitionFunction_eq_sum_rpow: forq ≠ 0the guard is removable,Z_q = ∑ i, (p i) ^ q, sinceReal.zero_rpowmakes empty cells vanish anyway.Oseledets.Multifractal.partitionFunction_one_eq_one:Z_1 = 1for a probability weight family.Oseledets.Multifractal.massExponent_one_eq_zero:τ(1) = 0for a probability weight family.
The generalized partition function Z_q = ∑_{i : p i > 0} (p i) ^ q of a finite weight
family p : ι → ℝ, with exponent ^ interpreted as Real.rpow. The guard 0 < p i is
load-bearing at q = 0: it ensures empty cells (p i = 0) contribute 0 rather than
Real.rpow 0 0 = 1, so that Z_0 counts the number of occupied cells.
Instances For
The mass exponent τ(q) = log Z_q / log ε of a finite weight family p at scale ε.
This is defined for every q with no case split.
Equations
Instances For
The Rényi (generalized) dimension D_q of a finite weight family p at scale ε. For
q ≠ 1 it is τ(q) / (q - 1). At q = 1 the general formula is the indeterminate 0 / 0,
so the L'Hôpital value — the information dimension (∑ i, p i log p i) / log ε — is
supplied directly. (By the Mathlib convention Real.log 0 = 0, the q = 1 numerator needs
no guard: empty cells contribute 0 automatically.)
Equations
Instances For
The singularity spectrum f(α) = ⨅ q, q α - τ(q) of a finite weight family p at
scale ε, the Legendre transform of the mass exponent τ. This is an infimum: since τ
is concave, the supremum of q α - τ(q) would be +∞.
Equations
- Oseledets.Multifractal.singularitySpectrum p ε α = ⨅ (q : ℝ), q * α - Oseledets.Multifractal.massExponent p ε q
Instances For
For q ≠ 0 the positivity guard in the partition function is removable: empty cells
(p i = 0) contribute Real.rpow 0 q = 0 anyway, so Z_q = ∑ i, (p i) ^ q.
For a probability weight family (0 ≤ p i and ∑ i, p i = 1), the partition function at
q = 1 is 1. Occupied cells contribute (p i) ^ 1 = p i, empty cells contribute 0 either
way.
If at least one weight is strictly positive, the partition function is strictly positive: the
contributing cell i adds (p i) ^ q > 0 to a sum of nonnegative terms (each guarded summand is
either a positive rpow of a positive base, or 0).