Let n≥2 and let f be a smooth dynamical system on Rn (a
diffeomorphism, or the time-one map of a smooth flow) possessing a compact
attractor A that carries an ergodic natural measure μ, i.e. a probability
measure supported on A such that for Lebesgue-almost every initial point x
in the basin of attraction of
A the orbit time averages converge to the μ-averages,
N→∞limN1k=0∑N−1φ(fk(x))=∫φdμfor every continuous φ.
Let λ1≥λ2≥⋯≥λn be the Lyapunov
exponents of f with respect to μ, listed with multiplicity, i.e. the
exponential rates of growth
λi=N→∞limN1logσi(DfN(x)),
where σ1≥⋯≥σn denote singular values and the limit
exists for μ-almost every x.
Assume λ1≥0 and ∑i=1nλi<0, and let
j∈{1,…,n−1} be the largest index with
i=1∑jλi≥0,i=1∑j+1λi<0,
so that λj+1<0. Define the Kaplan–Yorke (Lyapunov) dimension
DKY=j+∣λj+1∣∑i=1jλi.
Define the information dimension of μ as follows: for ε>0 let {Ci(ε)}i be the cubes of side ε of a
grid partition of Rn, put μi(ε)=μ(Ci(ε)) and
H(ε)=−i:μi(ε)>0∑μi(ε)logμi(ε),DI=ε→0+limlog(1/ε)H(ε).
Then for a typical such system the limit DI exists and
DI=DKY=j+∣λj+1∣∑i=1jλi;
that is, the dimension of the attractor A equals DKY.