Data processing inequality: Difference between revisions

Revision as of 13:44, 27 September 2019

Latest revision as of 16:21, 22 August 2024

(12 intermediate revisions by 10 users not shown)

Line 1:

{{Short description|Concept in information processing}}

The '''Datadata processing inequality''' is an [[information theory|information theoretic]] concept whichthat states that the information content of a signal cannot be increased via a local physical operation. This can be expressed concisely as 'post-processing cannot increase information'.<ref name= BeaudryArxiv>{{citation |journal=Quantum Information & Computation |volume=12 |issue=5–6 |pages=432–441 |last1=Beaudry |first1=Normand |title=An intuitive proof of the data processing inequality |date=2012 |arxiv=1107.0740|bibcode=2011arXiv1107.0740B }}</ref> As explained by ''Kinney and Atwal'', the DPI means that information is generally lost (never gained) when transmitted through a noisy channel.<ref>{{cite journal|pmid=24550517 | doi=10.107326421/pnasQIC12.13099331115-6-4 | last1arxiv=Kinney | last2=Atwal | volume=111 | issue=9 | title=Equitability, mutual information, and the maximal information coefficient1107. 0740| datebibcode=Mar 2014 | journal=Proc Natl Acad Sci U S A | pages=3354–9 | pmc=3948249| arxiv=13012011arXiv1107.77450740B | bibcodes2cid=2014PNAS..111.3354K9531510 }}</ref>

==DefinitionStatement==

Let three random variables form the [[Markov chain]] <math>X \rightarrow Y \rightarrow Z</math>, implying that the conditional distribution of <math>Z</math> depends only on <math>Y</math> and is [[Conditional independence|conditionally independent]] of <math>X</math>. Specifically, we have such a Markov chain if the joint probability mass function can be written as

:<math>p(x,y,z) = p(x)p(y|x)p(z|y)=p(y)p(x|y)p(z|y)</math>

In this setting, no processing of <math>Y </math>, deterministic or random, can increase the information that <math>Y</math> contains about <math>X</math>. Using the [[mutual information]], this can be written as :

:<math> I(xX;yY) \geqslant I(xX;zZ),</math>

Withwith the equality <math>I(xX;yY) = I(xX;zZ) </math> if and only if <math> I(X;Y\mid Z)=0 </math>. That is, <math>Z</math> and <math>Y</math> contain the same information about <math>X</math>, and <math>X \rightarrow Z \rightarrow Y</math> also forms a Markov chain.<ref>{{cite book| title=Elements of information theory | last1=Cover | last2=Thomas | date=2012 | publisher=John Wiley & Sons}}</ref>

==Proof==

One can apply the [[Conditional_mutual_information#Chain_rule_for_mutual_information chain rule for mutual information|chain rule for mutual information]] to obtain two different decompositions of <math>I(X;Y,Z)</math>:

:<math>

I(X;Z) + I(X;Y\mid Z) = I(X;Y,Z) = I(X;Y) + I(X;Z\mid Y)

</math>

By the relationship <math>X \rightarrow Y \rightarrow Z</math>, we know that <math>X</math> and <math>Z</math> are conditionally independent, given <math>Y</math>, which means the [[conditional mutual information]], <math>I(X;Z\mid Y)=0</math>. The data processing inequality then follows from the non-negativity of <math>I(X;Y\mid Z)\ge0</math>.

==See also==