TCSLib

92 Communication Complexity — TV Distance

92.1 Overview

This module develops the theory of total variation (TV) distance between probability measures on measurable spaces. It gives two equivalent definitions—one via the total variation norm of a signed measure and one as a supremum over measurable events—proves their equivalence, establishes the basic bound on event probability gaps, and derives a concrete half-\(\ell ^1\) formula on finite spaces, specialised further to Boolean and product distributions.

92.2 Declarations

Definition 92.1 Signed measure difference
#

Given two probability measures \(\mu \) and \(\nu \) on a measurable space \(\Omega \), the signed measure difference \(\mu - \nu \) is the signed measure on \(\Omega \) obtained by taking the difference of \(\mu \) and \(\nu \) as measures.

Definition 92.2 Total variation distance
#

The total variation distance between two probability measures \(\mu \) and \(\nu \) on \(\Omega \) is defined as \(\tfrac {1}{2}\) times the total variation norm of the signed measure \(\mu - \nu \), i.e. \(\mathrm{TV}(\mu ,\nu ) = \tfrac {1}{2}\| \mu - \nu \| _{\mathrm{TV}}\).

Definition 92.3 Total variation distance via event supremum
#

The supremum definition of total variation distance between \(\mu \) and \(\nu \) is

\[ \mathrm{TV}_{\sup }(\mu ,\nu ) \; =\; \sup \bigl\{ |\mu (S) - \nu (S)| : S \subseteq \Omega \text{ measurable}\bigr\} . \]
Lemma 92.4 Signed measure via Jordan decomposition

For any signed measure \(s\) on \(\Omega \) and any measurable set \(S\), the value \(s(S)\) equals the difference of its Jordan positive and negative parts: \(s(S) = s^+(S) - s^-(S)\).

Lemma 92.5 Signed difference vanishes on the whole space

For any two probability measures \(\mu \) and \(\nu \), the signed difference satisfies \((\mu - \nu )(\Omega ) = 0\).

Lemma 92.6 Jordan parts agree on the whole space

For probability measures \(\mu \) and \(\nu \), the Jordan positive part and the Jordan negative part of \(\mu - \nu \) assign equal real mass to \(\Omega \): \(((\mu -\nu )^+)(\Omega ) = ((\mu -\nu )^-)(\Omega )\).

Lemma 92.7 Total variation mass of the whole space

For any signed measure \(s\), the real-valued total variation mass of \(\Omega \) equals the sum of the Jordan positive and negative part masses: \(\| s\| _{\mathrm{TV}}(\Omega ) = s^+(\Omega ) + s^-(\Omega )\).

Lemma 92.8 Half total variation equals positive Jordan part

For probability measures \(\mu \) and \(\nu \), \(\tfrac {1}{2}\| \mu -\nu \| _{\mathrm{TV}}(\Omega ) = ((\mu -\nu )^+)(\Omega )\).

Lemma 92.9 Event gap bounded by half total variation

For every measurable set \(S \subseteq \Omega \), \(|(\mu -\nu )(S)| \le \tfrac {1}{2}\| \mu -\nu \| _{\mathrm{TV}}(\Omega )\).

Lemma 92.10 Probability gap bounded by half total variation

For every measurable set \(S \subseteq \Omega \), \(|\mu (S) - \nu (S)| \le \tfrac {1}{2}\| \mu -\nu \| _{\mathrm{TV}}(\Omega )\).

Lemma 92.11 Attainment of the signed measure bound

There exists a measurable set \(S \subseteq \Omega \) for which the bound is tight: \(|(\mu -\nu )(S)| = \tfrac {1}{2}\| \mu -\nu \| _{\mathrm{TV}}(\Omega )\).

Lemma 92.12 Attainment of the probability gap bound

There exists a measurable set \(S \subseteq \Omega \) such that \(|\mu (S) - \nu (S)| = \tfrac {1}{2}\| \mu -\nu \| _{\mathrm{TV}}(\Omega )\).

Theorem 92.13 Equivalence of the two TV distance definitions

The total-variation-mass definition and the supremum-over-events definition coincide: \(\mathrm{TV}(\mu ,\nu ) = \mathrm{TV}_{\sup }(\mu ,\nu )\).

Theorem 92.14 TV distance bounds event probability gaps

For every measurable set \(S \subseteq \Omega \), \(|\mu (S) - \nu (S)| \le \mathrm{TV}(\mu ,\nu )\).

Theorem 92.15 TV distance is nonneg

For any two probability measures \(\mu \) and \(\nu \), \(\mathrm{TV}(\mu ,\nu ) \ge 0\).

Lemma 92.16 Indicator sum bounded by positive part sum

For any function \(a : \alpha \to \mathbb {R}\) on a finite type \(\alpha \) and any set \(S \subseteq \alpha \), \(\sum _{x \in S} a(x) \le \sum _{x : \alpha } \max (a(x), 0)\).

Lemma 92.17 Positive part sum equals half the \(\ell ^1\) sum

Let \(a : \alpha \to \mathbb {R}\) be a function on a finite type with \(\sum _x a(x) = 0\). Then \(\sum _x \max (a(x), 0) = \tfrac {1}{2}\sum _x |a(x)|\).

Lemma 92.18 Indicator sum absolutely bounded by half \(\ell ^1\) sum

If \(a : \alpha \to \mathbb {R}\) satisfies \(\sum _x a(x) = 0\), then for every \(S \subseteq \alpha \), \(\bigl|\sum _{x \in S} a(x)\bigr| \le \tfrac {1}{2}\sum _x |a(x)|\).

Lemma 92.19 Positive-part indicator sum equals half \(\ell ^1\) sum

If \(a : \alpha \to \mathbb {R}\) satisfies \(\sum _x a(x) = 0\), then \(\bigl|\sum _x \max (a(x),0)\bigr| = \tfrac {1}{2}\sum _x |a(x)|\).

Definition 92.20 Singleton mass difference

For probability measures \(\mu \) and \(\nu \) on \(\Omega \) and a point \(\omega \in \Omega \), the singleton mass difference is \(\mu (\{ \omega \} ) - \nu (\{ \omega \} ) \in \mathbb {R}\).

Lemma 92.21 Probability gap as sum of singleton mass differences

On a finite measurable space, for any measurable set \(S\), \(\mu (S) - \nu (S) = \sum _{\omega \in S} (\mu (\{ \omega \} ) - \nu (\{ \omega \} ))\).

Lemma 92.22 Singleton mass differences sum to zero

On a finite measurable space, the singleton mass differences sum to zero: \(\sum _{\omega : \Omega } (\mu (\{ \omega \} ) - \nu (\{ \omega \} )) = 0\).

Theorem 92.23 TV distance as half \(\ell ^1\) distance on finite spaces (supremum form)

On a finite measurable space \(\Omega \), the supremum form of total variation distance satisfies

\[ \mathrm{TV}_{\sup }(\mu ,\nu ) \; =\; \frac{1}{2}\sum _{\omega :\Omega } |\mu (\{ \omega \} ) - \nu (\{ \omega \} )|. \]
Theorem 92.24 TV distance as half \(\ell ^1\) distance on finite spaces

On a finite measurable space \(\Omega \),

\[ \mathrm{TV}(\mu ,\nu ) \; =\; \frac{1}{2}\sum _{\omega :\Omega } |\mu (\{ \omega \} ) - \nu (\{ \omega \} )|. \]
Theorem 92.25 TV distance on Bool equals gap at true

For probability measures \(\mu \) and \(\nu \) on \(\mathrm{Bool}\), \(\mathrm{TV}(\mu ,\nu ) = |\mu (\{ \mathrm{true}\} ) - \nu (\{ \mathrm{true}\} )|\).

Theorem 92.26 TV distance sub-additivity for product measures

For probability measures \(\mu _1, \nu _1\) on \(\alpha \) and \(\mu _2, \nu _2\) on \(\beta \) (both finite measurable spaces),

\[ \mathrm{TV}(\mu _1 \otimes \mu _2,\, \nu _1 \otimes \nu _2) \; \le \; \mathrm{TV}(\mu _1,\nu _1) + \mathrm{TV}(\mu _2,\nu _2). \]