Directed information explained

Directed information is an information theory measure that quantifies the information flow from the random string

X ^{n} = (X _{1,X} _{2,...,X} _{n)} to the random string Y ^{n} = (Y _{1,Y} _{2,...,Y} _{n)} . The term directed information was coined by James Massey and is defined as^[1] I(X ^{n\to} Y ^{n)} \triangleq \sum ^{n} _{i =1} I(X ^{i -1} _{i|Y} ) where I(X ^{i} ;Y ^{i -1} _{i|Y} ) is the conditional mutual information I(X _{1,X} _{2,...,X} _{i} ;Y _{i|Y} _{1,Y} _{2,...,Y} _{i-1} ) .

Directed information has applications to problems where causality plays an important role such as the capacity of channels with feedback,^{[1] [2] [3] [4]} capacity of discrete memoryless networks,^[5] capacity of networks with in-block memory,^[6] gambling with causal side information,^[7] compression with causal side information,^[8] real-time control communication settings,^{[9] [10]} and statistical physics.^[11]

Causal conditioning

The essence of directed information is causal conditioning. The probability of

x ^{n} causally conditioned on y ^{n} is defined as^[5] P(x ^{n||y} ^{n)} \triangleq \prod ^{n} _{i =1} P(x ^{i -1} _{i|x} ,y ^{i} ) .This is similar to the chain rule for conventional conditioning P(x ^{n|y} ^{n)} = \prod ^{n} _{i =1} P(x ^{i -1} _{i|x} ,y ^{n} ) except one conditions on "past" and "present" symbols y ^{i} rather than all symbols y ^{n} . To include "past" symbols only, one can introduce a delay by prepending a constant symbol:P(x ^{n||(0,y} ^{n-1} )) \triangleq \prod ^{n} _{i =1} P(x ^{i -1} _{i|x} ,y ^{i-1} ) .It is common to abuse notation by writing P(x ^{n||y} ^{n-1} ) for this expression, although formally all strings should have the same number of symbols.

One may also condition on multiple strings:

P(x ^{n||y} ^{n,z} ^{n)} \triangleq \prod ^{n} _{i =1} P(x ^{i -1} _{i|x} ,y ^{i} ,z ^{i} ) .

Causally conditioned entropy

The causally conditioned entropy is defined as:^[2]

H(X ^{n} || Y ^{n)=E\left[-log} {P(X ^{n||Y} ^{n)}} \right] = \sum ^{n} _{i =1} H(X _{i} |X ^{i-1} ,Y ^{i} ) Similarly, one may causally condition on multiple strings and writeH(X ^{n} || Y ^{n,Z} ^{n)=E\left[-log} {P(X ^{n||Y} ^{n,Z} ^{n)}} \right] .

Properties

A decomposition rule for causal conditioning^[1] is

P(x ^{n,} y ^{n)} = P(x ^{n||y} ^{n-1} ) P(y ^{n||x} ^{n)} .This rule shows that any product of P(x ^{n||y} ^{n-1} ), P(y ^{n||x} ^{n)} gives a joint distribution P(x ^{n,} y ^{n)} .

The causal conditioning probability

P(y ^{n||x} ^{n)} = \prod ^{n} _{i =1} P(y ^{i -1} _{i|y} ,x ^{i} ) is a probability vector, i.e., P(y ^{n||x} ^{n)\geq} 0 and \sum _{y ^{n}} P(y ^{n||x} ^{n)=1} forall (x ^{n,y} ^{n)} .

Directed Information can be written in terms of causal conditioning:^[2]

I(X ^{N} → Y ^{N)=E\left[log} (P(Y ^{N||X} ^{N)})/(P(Y ^{N)}) \right] = H(Y ^{n)-} H(Y ^{n} || X ^{n)} .

The relation generalizes to three strings: the directed information flowing from

X ^{n} to Y ^{n} causally conditioned on Z ^{n} isI(X ^{n\to} Y ^{n} || Z ^{n)} = H(Y ^{n} || Z ^{n)-} H(Y ^{n} || X ^{n,} Z ^{n)} .

Conservation law of information

This law, established by James Massey and his son Peter Massey,^[12] gives intuition by relating directed information and mutual information. The law states that for any

X ^{n,} Y ^{n} , the following equality holds:I(X ^{n;Y} ^{n)=} I(X ^{n} \to Y ^{n)+I(Y} ^{n-1} \to X ^{n).}

Two alternative forms of this law are^{[2] [13]}

I(X ^{n;Y} ^{n)} = I(X ^{n} \to Y ^{n)} + I(Y ^{n} \to X ^{n)} - I(X ^{n} \leftrightarrow Y ^{n)} I(X ^{n;Y} ^{n)} = I(X ^{n-1} \to Y ^{n)} + I(Y ^{n-1} \to X ^{n)} + I(X ^{n} \leftrightarrow Y ^{n)} where I(X ^{n} \leftrightarrow Y ^{n)} = \sum ^{n} _{i =1} I(X _{i} ; Y _{i} | X ^{i-1} , Y ^{i-1} ) .

Estimation and optimization

Estimating and optimizing the directed information is challenging because it has

n terms where n may be large. In many cases, one is interested in optimizing the limiting average, that is, when n grows to infinity termed as a multi-letter expression.

Estimation

Estimating directed information from samples is a hard problem since the directed information expression does not depend on samples but on the joint distribution

\{P(x _{i,y} ^{i -1} ^{i|x} ,y ^{i-1} ) ^{n\}} _{i =1} which may be unknown. There are several algorithms based on context tree weighting^[14] and empirical parametric distributions^[15] and using long short-term memory.

Optimization

Maximizing directed information is a fundamental problem in information theory. For example, given the channel distributions

\{P(y ^{i} _{i|x} ,y ^{i-1} \} ^{n)} _{i =1} , the objective might be to optimize I(X ^{n\to} Y ^{n)} over the channel input distributions \{P(x ^{i -1} _{i|x} ,y ^{i-1} \} ^{n)} _{i =1} .

There are algorithms to optimize the directed information based on the Blahut-Arimoto,^[16] Markov decision process,^{[17] [18] [19] [20]} Recurrent neural network,^[21] Reinforcement learning.^[22] and Graphical methods (the Q-graphs).^{[23] [24]} For the Blahut-Arimoto algorithm,^[16] the main idea is to start with the last mutual information of the directed information expression and go backward. For the Markov decision process,^{[17] [18] [19] [20]} the main ideas is to transform the optimization into an infinite horizon average reward Markov decision process. For a Recurrent neural network,^[21] the main idea is to model the input distribution using a Recurrent neural network and optimize the parameters using Gradient descent. For Reinforcement learning,^[22] the main idea is to solve the Markov decision process formulation of the capacity using Reinforcement learning tools, which lets one deal with large or even continuous alphabets.

Marko's theory of bidirectional communication

Massey's directed information was motivated by Marko's early work (1966) on developing a theory of bidirectional communication.^{[25] [26]} Marko's definition of directed transinformation differs slightly from Massey's in that, at time

n , one conditions on past symbols X ^{n-1} ,Y ^{n-1} only and one takes limits:T _{12} = \lim _{n} E\left[ -log (P(X _{n} |X ^{n-1} ))/(P(X _{n} |X ^{n-1} ,Y ^{n-1} )) \right] and T _{21} = \lim _{n} E\left[ -log (P(Y _{n} |Y ^{n-1} ))/(P(Y _{n} |Y ^{n-1} ,X ^{n-1} )) \right]. Marko defined several other quantities, including:

• Total information:

H _{1} = \lim _{n} E\left[ -log P(X _{n} |X ^{n-1} ) \right] and H _{2} = \lim _{n} E\left[ -log P(Y _{n} |Y ^{n-1} ) \right]

• Free information:

F _{1} = \lim _{n} E\left[ -log P(X _{n} |X ^{n-1} ,Y ^{n-1} ) \right] and F _{2} = \lim _{n} E\left[ -log P(Y _{n} |Y ^{n-1} ,X ^{n-1} ) \right]

• Coincidence:

K = \lim _{n} E\left[ -log (P(X _{n} |X ^{n-1} ) P(Y _{n} |Y ^{n-1} ))/(P(X _{n} ,Y _{n} |X ^{n-1} ,Y ^{n-1} )) \right]. The total information is usually called an entropy rate. Marko showed the following relations for the problems he was interested in:K = T _{12} + T _{21} H _{1} = T _{12} + F _{1} and H _{2} = T _{21} + F _{2} He also defined quantities he called residual entropies:R _{1} = H _{1} - K = F _{1} - T _{21} R _{2} = H _{2} - K = F _{2} - T _{12} and developed the conservation law F _{1} + F _{2} = R _{1} + R _{2} + K = H _{1} + H _{2} - K and several bounds.

Relation to transfer entropy

Directed information is related to transfer entropy, which is a truncated version of Marko's directed transinformation

T _{21} .

The transfer entropy at time

i and with memory d isT _{X} = I(X _{i-1} , ... ,X _{i-d} ; Y _{i} | Y _{i-1} , ... ,Y _{i-d} ). where one does not include the present symbol X _{i} or the past symbols X ^{i-d-1} ,Y ^{i-d-1} before time i - d .

Transfer entropy usually assumes stationarity, i.e.,

T _{X} does not depend on the time i .

Notes and References

1. Massey . James . Causality, Feedback And Directed Information . Proceedings 1990 International Symposium on Information Theory and its Applications, Waikiki, Hawaii, Nov. 27-30, 1990 . 1990.
2. Doctoral . Kramer . Gerhard . 1998 . Directed information for channels with feedback . 10.3929/ethz-a-001988524 . ETH Zurich . 20.500.11850/143796 . en.
3. Doctoral . Tatikonda . Sekhar Chandra . 2000 . Control under communication constraints . Massachusetts Institute of Technology. 1721.1/16755 .
4. Permuter . Haim Henry. Weissman. Tsachy. Goldsmith . Andrea J.. Finite State Channels With Time-Invariant Deterministic Feedback. IEEE Transactions on Information Theory . February 2009 . 55. 2 . 644–662 . 10.1109/TIT.2008.2009849. cs/0608070. 13178.
5. Kramer . G. . Capacity results for the discrete memoryless network. IEEE Transactions on Information Theory. January 2003. 49. 1 . 4–21. 10.1109/TIT.2002.806135.
6. Kramer . Gerhard . Information Networks With In-Block Memory . IEEE Transactions on Information Theory . April 2014 . 60 . 4 . 2105–2120 . 10.1109/TIT.2014.2303120. 1206.5389 . 16382644 .
7. Permuter. Haim H.. Kim . Young-Han. Weissman . Tsachy. Interpretations of Directed Information in Portfolio Theory, Data Compression, and Hypothesis Testing. IEEE Transactions on Information Theory. June 2011. 57. 6 . 3248–3259. 10.1109/TIT.2011.2136270. 0912.4872 . 11722596.
8. Simeone. Osvaldo . Permuter. Haim Henri. Source Coding When the Side Information May Be Delayed . IEEE Transactions on Information Theory. June 2013 . 59. 6. 3607–3618 . 10.1109/TIT.2013.2248192. 1109.1293. 3211485.
9. Charalambous. Charalambos D. . Stavrou . Photios A.. Directed Information on Abstract Spaces: Properties and Variational Equalities . IEEE Transactions on Information Theory. August 2016. 62. 11. 6019–6052 . 10.1109/TIT.2016.2604846 . 1302.3971. 8107565.
10. Tanaka . Takashi . Esfahani . Peyman Mohajerin . Mitter . Sanjoy K. . LQG Control With Minimum Directed Information: Semidefinite Programming Approach . IEEE Transactions on Automatic Control . January 2018 . 63 . 1 . 37–52 . 10.1109/TAC.2017.2709618 . 1401958 . 1510.04214 .
11. Vinkler . Dror A . Permuter . Haim H . Merhav . Neri . Analogy between gambling and measurement-based work extraction . Journal of Statistical Mechanics: Theory and Experiment . 20 April 2016 . 2016 . 4 . 043403 . 10.1088/1742-5468/2016/04/043403 . 1404.6788 . 2016JSMTE..04.3403V . 124719237.
12. Book: Massey . J.L. . Massey . P.C. . Proceedings. International Symposium on Information Theory, 2005. ISIT 2005 . Conservation of mutual and directed information . September 2005 . 157–158 . 10.1109/ISIT.2005.1523313. 0-7803-9151-9 . 38053218 .
13. Amblard . Pierre-Olivier . Michel . Olivier . The Relation between Granger Causality and Directed Information Theory: A Review . Entropy . 28 December 2012 . 15 . 1 . 113–143 . 10.3390/e15010113 . 1211.3169 . 2012Entrp..15..113A . free .
14. Jiao . Jiantao . Permuter . Haim H. . Zhao . Lei . Kim . Young-Han . Weissman . Tsachy . Universal Estimation of Directed Information . IEEE Transactions on Information Theory. October 2013 . 59 . 10 . 6220–6242 . 10.1109/TIT.2013.2267934 . 1201.2334 . 10855063 .
15. Quinn . Christopher J. . Kiyavash . Negar . Coleman . Todd P. . Directed Information Graphs . IEEE Transactions on Information Theory. December 2015 . 61 . 12 . 6887–6909 . 10.1109/TIT.2015.2478440. 1204.2003 . 3121664.
16. Naiss . Iddo . Permuter . Haim H. . Extension of the Blahut–Arimoto Algorithm for Maximizing Directed Information . IEEE Transactions on Information Theory. January 2013 . 59 . 1 . 204–222 . 10.1109/TIT.2012.2214202 . 3115749 . 1012.5071.
17. Permuter . Haim . Cuff . Paul . Van Roy . Benjamin . Weissman . Tsachy . Capacity of the Trapdoor Channel With Feedback . IEEE Transactions on Information Theory. July 2008 . 54 . 7 . 3150–3165 . 10.1109/TIT.2008.924681. cs/0610047 . 1265.
18. Elishco . Ohad . Permuter . Haim . Capacity and Coding for the Ising Channel With Feedback . IEEE Transactions on Information Theory. September 2014 . 60 . 9 . 5138–5149 . 10.1109/TIT.2014.2331951 . 1205.4674 . 9761759.
19. Sabag . Oron . Permuter . Haim H. . Kashyap . Navin . The Feedback Capacity of the Binary Erasure Channel With a No-Consecutive-Ones Input Constraint . IEEE Transactions on Information Theory. January 2016 . 62 . 1 . 8–22 . 10.1109/TIT.2015.2495239 . 476381.
20. Peled . Ori . Sabag . Oron . Permuter . Haim H. . Feedback Capacity and Coding for the $(0,k)$ -RLL Input-Constrained BEC . IEEE Transactions on Information Theory. July 2019 . 65 . 7 . 4097–4114 . 10.1109/TIT.2019.2903252 . 1712.02690 . 86582654.
21. Book: Aharoni . Ziv . Tsur . Dor . Goldfeld . Ziv . Permuter . Haim Henry . 2020 IEEE International Symposium on Information Theory (ISIT) . Capacity of Continuous Channels with Memory via Directed Information Neural Estimator . 2003.04179 . June 2020 . 2014–2019 . 10.1109/ISIT44484.2020.9174109 . 978-1-7281-6432-8 . 212634151.
22. Aharoni . Ziv . Sabag . Oron . Permuter . Haim Henri . Reinforcement Learning Evaluation and Solution for the Feedback Capacity of the Ising Channel with Large Alphabet . 18 August 2020 . cs.IT . 2008.07983.
23. Sabag . Oron . Permuter . Haim Henry . Pfister . Henry . A Single-Letter Upper Bound on the Feedback Capacity of Unifilar Finite-State Channels . IEEE Transactions on Information Theory. March 2017 . 63 . 3 . 1392–1409. 10.1109/TIT.2016.2636851 . 1604.01878 . 3259603 .
24. Sabag . Oron . Huleihel . Bashar . Permuter . Haim Henry . Graph-Based Encoders and their Performance for Finite-State Channels with Feedback . IEEE Transactions on Communications. 2020 . 68 . 4 . 2106–2117 . 10.1109/TCOMM.2020.2965454 . 1907.08063 . 197544824.
25. Marko . Hans . Die Theorie der bidirektionalen Kommunikation und ihre Anwendung auf die Nachrichtenübermittlung zwischen Menschen (Subjektive Information) . Kybernetik . 1 September 1966 . 3 . 3 . 128–136 . 10.1007/BF00288922 . 5920460 . 33275199 . de . 1432-0770.
26. Marko . H. . The Bidirectional Communication Theory--A Generalization of Information Theory . IEEE Transactions on Communications . December 1973 . 21 . 12 . 1345–1351 . 10.1109/TCOM.1973.1091610. 51664185 .