Epsilon-entropy

of a set in a metric space

The logarithm to the base 2 of the smallest number of points in an $ \epsilon $- net for this set. In other words, the $ \epsilon $- entropy $ {\mathcal H} _ \epsilon ( C , X ) $ of a set $ C $ lying in a metric space $ ( X , \rho ) $ is

$$ {\mathcal H} _ \epsilon ( C , X ) = \ \mathop{\rm log} _ {2} N _ \epsilon ( C , X ) , $$

where

$$ N _ \epsilon ( C , X ) = \inf \left \{ {n } : { \exists x _ {1} \dots x _ {n} , x _ {i} \in X :\ C \subset \cup _ { i= } 1 ^ { n } B ( x _ {i} , \epsilon ) } \right \} $$

and $ B ( \zeta , \alpha ) = \{ {x \in X } : {\rho ( x , \zeta ) \leq \alpha } \} $ is a ball of radius $ \alpha $ with centre at $ \zeta $. The definition of $ {\mathcal H} _ \epsilon ( C , X ) $ was given by A.N. Kolmogorov in [1], motivated by ideas and definitions of information theory (cf. Information theory). $ {\mathcal H} _ \epsilon ( C , X ) $ is also called the relative $ \epsilon $- entropy; it depends on the space $ X $ in which the set $ C $ is situated, that is, on the metric extension of $ C $. The quantity

$$ {\mathcal H} _ \epsilon ( C) = \inf {\mathcal H} _ \epsilon ( C , X ) , $$

where the infimum is taken over all metric extensions $ X $ of $ C $, is called the absolute $ \epsilon $- entropy of $ C $. It may also be defined directly (Kolmogorov, [1]): for a metric space $ C $, $ {\mathcal H} _ \epsilon ( C) $ is the logarithm to the base 2 of the cardinality $ N _ \epsilon ( C) $ of the most economic (by the number of sets) $ 2 \epsilon $- covering of $ C $. Here a system of sets $ \{ C _ {i} \} $ is called a $ 2 \epsilon $- covering of $ C $ if $ C _ {i} \subset C $, $ \cup _ {i=} 1 ^ {n} C _ {i} = C $ and the diameter of each set does not exceed $ 2 \epsilon $. It has been shown [2] that the absolute $ \epsilon $- entropy is the minimum relative $ \epsilon $- entropy. The quantities $ N _ \epsilon ( C) $ and $ N _ \epsilon ( C , X ) $ are inverse to the widths (cf. Width) $ \epsilon ^ {N} ( C) $ and $ \epsilon _ {N} ( C , X ) $. This characterizes the fact that it is possible to recover the elements of $ C $ from tables of $ N $ elements and to approximate by $ N $- point sets.

The investigation of the asymptotic behaviour of the $ \epsilon $- entropy of various function classes is a special branch of approximation theory.

References

[1]	A.N. Kolmogorov, "On certain asymptotic characteristics of completely bounded metric spaces" Dokl. Akad. Nauk SSSR , 108 : 3 (1956) pp. 385–388 (In Russian)
[2]	A.G. Vitushkin, "Theory of transmission and processing of information" , Pergamon (1961) (Translated from Russian)
[3]	A.N. Kolmogorov, V.M. Tikhomirov, "-entropy and -capacity of sets in functional spaces" Amer. Math. Soc. Transl. Ser. 2 , 17 (1961) pp. 277–364 Uspekhi Mat. Nauk , 14 : 2 (1959) pp. 3–86

Comments

Kolmogorov's original definition was formulated to investigate the question of whether functions of $ n \geq 3 $ variables are representable as compositions of functions of $ r < n $ variables. This has a great deal to do with Hilbert's 13th problem, cf. [2] and [a1].

The notion of $ \epsilon $- entropy has also proved useful in probability theory, cf. [a2]. It is also called metric entropy.

The $ \epsilon $- entropy $ {\mathcal H} _ \epsilon ( C , X) $ is defined by considering coverings of $ C $ by $ \epsilon $- balls. An $ \epsilon $- packing $ {\mathcal P} = \{ B ( y _ {i} , \epsilon ) \} $ of $ C $ is a collection of $ \epsilon $- balls such that $ B( y _ {i} , \epsilon ) \cap B ( y _ {j} , \epsilon ) = \emptyset $ if $ i \neq j $ and $ B ( x _ {i} , \epsilon ) \subset C $. Let $ M ( C , \epsilon ) $ be the maximum cardinality of a packing in $ C $. Then $ \mathop{\rm log} M( C , \epsilon ) $ is called the $ \epsilon $- capacity of $ C $ in $ X $.

There are a number of additional (but not unrelated) notions of $ \epsilon $- entropy for random variables and random processes. They are defined as follows. Let $ \xi $ be a random variable with values in a metric space $ ( X , \rho ) $. Let $ W _ \epsilon $ be the set of all random variables $ \xi ^ \prime $ with values in $ X $ such that $ {\mathsf E} ( \rho ^ {2} ( \xi , \xi ^ \prime ) ) \leq \epsilon ^ {2} $, where $ {\mathsf E} $ denotes mathematical expectation. Then the $ \epsilon $- entropy of $ \xi $ is ([a3])

$$ {\mathcal H} _ \epsilon ^ \prime ( \xi ) = \ \inf \{ {I ( \xi , \xi ^ \prime ) } : { \xi ^ \prime \in W _ \epsilon } \} , $$

where $ I ( \xi , \xi ^ \prime ) $ is the usual information function for a pair of random variables (cf. Information). This idea can be easily extended to random processes $ \{ \xi ( t) \} _ {t \in [ a, b ] } $ by regarding such a process as a random variable with values in $ L _ {2} ([ a , b ]) $, say. [a4] contains some asymptotic estimates of $ {\mathcal H} _ \epsilon ^ \prime ( \xi ) $.

Now let $ ( X , \rho , \mu ) $ be a probabilistic metric space, i.e. $ ( X, \rho ) $ is a metric space and $ ( X , \mu ) $ is a probability space. For $ \epsilon > 0 $, let $ {\mathcal A} _ \epsilon $ be the set of all partitions of $ X $( cf. Decomposition) into measurable sets of diameter $ \leq \epsilon $. Then ([a5]) one defines

$$ {\mathcal H} _ \epsilon ^ {\prime\prime} ( X) = \ \inf \{ {H ( {\mathcal U} ) } : { {\mathcal U} \in {\mathcal A} _ \epsilon } \} , $$

where $ H ( {\mathcal U} ) $ denotes the ordinary entropy of the partition $ {\mathcal U} $. Define

$$ I _ \epsilon ( X) = \lim\limits _ {n \rightarrow \infty } \ \frac{1}{n} {\mathcal H} _ \epsilon ^ {\prime\prime} ( X ^ {n} ) , $$

where $ X ^ {n} = X \times \dots \times X $( $ n $ factors) with the product measure and distance. (This limit exists.) For a measure $ \pi $ on $ X \times X $ such that $ \pi ( A \times X ) = \pi ( X \times A) = \mu ( A) $ for all measurable $ A \subset X $, let

$$ I ( \rho ) = \frac{d \rho }{d ( \mu \times \mu ) } $$

be the Radon–Nikodým derivative. Let $ R _ \epsilon ( X) $ be the set of all measures $ \pi $ on the measurable space $ X \times X $ such that $ \pi ( A \times X ) = \pi ( X \times A) = \mu ( A) $ for $ A $ measurable and

$$ \pi \{ {( x, y) } : {\rho ( x, y) \leq \epsilon /2 } \} = 1. $$

Then

$$ I _ \epsilon = \inf _ {\rho \in R _ \epsilon ( X) } \ I ( \rho ) , $$

which is a noisy channel coding theorem; cf. [a6]. There are also estimates both ways relating $ {\mathcal H} _ \epsilon ^ {\prime\prime} $ and $ I _ \epsilon ( X) $; cf. [a6]. Cf. also [a7] for (more) relations between $ {\mathcal H} _ \epsilon ^ {\prime\prime} $ and (extended notions of) Shannon entropy.

Define the absolute $ \epsilon $- entropy $ {\mathcal H} _ \epsilon ( C) $( as in the main article above) by means of $ \epsilon $- partitions of $ ( X , \rho ) $ by means of Borel sets. Then

$$ {\mathcal H} _ \epsilon ^ {\prime\prime} \leq {\mathcal H} _ {\epsilon /2 } , $$

but it need not be true that the supremum of the $ {\mathcal H} _ \epsilon ^ {\prime\prime} $ for varying $ \mu $ on $ ( X, \rho ) $ is equal to $ {\mathcal H} _ {\epsilon /2 } $, cf. [a5].

Now let again $ \xi ( t) $, $ t \in [ 0, 1] $, be a stochastic Gaussian process (real-valued), let $ K ( \cdot , \cdot ) $ be its covariance function and let $ T $ be the integral operator with kernel $ K( \cdot , \cdot ) $, i.e.

$$ ( Tf ) ( t) = \int\limits _ { 0 } ^ { 1 } K( s, t) f( s) ds . $$

Let $ \lambda _ {1} \geq \lambda _ {2} \geq \dots \geq 0 $ be the eigen values of $ T $. For $ \epsilon \in [ 0, ( \sum _ {i=} 1 ^ \infty \lambda _ {i} ) ^ {1/2} ] $ define a function $ f ( \epsilon ) $ by the equality

$$ \epsilon ^ {2} = \ \sum _ { i= } 1 ^ \infty \min \{ \lambda _ {i} , f ( \epsilon ) \} . $$

Then Pinsker's theorem says that

$$ {\mathcal H} _ \epsilon ^ \prime ( \xi ) = \frac{1}{2} \sum _ { i= } 1 ^ \infty \mathop{\rm log} \max \ \left \{ \frac{\lambda _ {i} }{f ( \epsilon ) } , 1 \right \} $$

(where log stands for the logarithm to the base 2). One also has ([a8]): For all $ a > 0 $,

$$ \lim\limits _ {k \rightarrow \infty } \frac{S( T ^ {k} , a f ( \epsilon ) ^ {k} ) }{k} = 2 {\mathcal H} _ \epsilon ^ \prime ( \xi ) , $$

where for a suitable operator on a Hilbert space $ T : H \rightarrow H $ the entropy $ S( T, \alpha ) $ is defined as the metric entropy of the set $ \{ {Tx } : {\| x \| \leq 1 } \} \subset H $:

$$ S ( T , \alpha ) = {\mathcal H} _ \alpha ( T( B( 0, 1)) , H ). $$

References

[a1]	G.G. Lorentz, "The 13-th problem of Hilbert" F.E. Browder (ed.) , Mathematical developments arising from Hilbert problems , Proc. Symp. Pure Math. , 28 , Amer. Math. Soc. (1976) pp. 419–430
[a2]	R.M. Dudley, "Metric entropy and the central limit theorem in $C ( S )$" Ann. Inst. Fourier (Grenoble) , 24 (1974) pp. 49–60
[a3]	A.N. Kolmogorov, "Theory of transmission of information" Amer. Math. Soc. Transl. Ser. 2 , 33 (1963) pp. 291–321 Acad. R. P. Romine An. Romino-Sov. , 13 : 1 (1959) pp. 5–33
[a4]	J. Biria, M. Zakai, J. Ziv, "On the -entropy and the rate distortion function of certain non-Gaussian processes" IEEE Trans. Inform. Theory , 20 (1974) pp. 517–524
[a5]	E.C. Posner, E.R. Rodenich, H. Rumsey, "-Entropy of stochastic processes" Ann. Math. Stat. , 38 (1967) pp. 1000–1020
[a6]	E.C. Posner, E.R. Rodenich, "Epsilon entropy and data compression" Ann. Math. Stat. , 42 (1971) pp. 2079–2125
[a7]	M. Katětov, "On extended Shannon entropies and the epsilon entropy" Comm. Math. Univ. Carolinae , 27 (1986) pp. 519–534
[a8]	S. Akashi, "On operator theoretical characterization of epsilon-entropy in Gaussian processes" Kodai Math. J. , 9 (1986) pp. 58–67