Source

Massey (1990), “Causality, Feedback and Directed Information”, Proc. ISITA-90, LIT-1166, read in full in NOTE-939 from the author’s preprint. The proceedings version was not compared.

What was actually shown

Definitions. A discrete channel is memoryless if P(y_n | x^n y^{n−1}) = P(y_n | x_n), and used without feedback if P(x_n | x^{n−1} y^{n−1}) = P(x_n | x^{n−1}). Together they give the textbook P(y^N | x^N) = Π P(y_n | x_n), which noiseless feedback violates. Directed information is I(X^N → Y^N) = Σ_n I(X^n; Y_n | Y

Theorem 1. I(X^N → Y^N) ≤ I(X^N; Y^N), with equality without feedback. With noiseless feedback over a binary symmetric channel (crossover 0.1), N = 3, the reading found I(X;Y) = 1.469 bits against I(X → Y) = 0.531: mutual information counts the fed-back copies.

Theorem 2. For a DMC, I(X^N → Y^N) ≤ Σ I(X_n; Y_n), with equality iff the outputs are independent.

Theorem 3. In a causal system, where a channel sees a source only through its own inputs and outputs, I(U^K; Y^N) ≤ I(X^N → Y^N). With Theorem 2, feedback cannot raise a DMC’s capacity.

What could have come out otherwise: directed information could have coincided with mutual information under feedback. The numerical case shows it does not.

What this does not say

  • That directed information is causal in the interventional sense. Massey’s causal system is a factorization condition about how information may enter a channel. Nothing is set or perturbed.
  • That directed information is transfer entropy. It conditions on the current input X_n. By the chain rule it is the sum of full-history transfer entropies plus the sum of instantaneous terms I(X_n; Y_n | X^{n−1}, Y^{n−1}); the record checked this on a random binary process (NOTE-939). The decomposition is the record’s.
  • Anything for continuous alphabets or time.
  • That it suffices for a theory of feedback systems. Massey leaves that open.

Connections

  • TERM-057 and TERM-058: directed information contains full-history transfer entropy, and so Granger causality in the Gaussian case (THEORY-700), as its lagged part. Its instantaneous part is Geweke’s instantaneous feedback (LIT-1169).
  • CLAIM-333’s counterexample has zero lagged part and positive instantaneous part.
  • THEORY-710 (Ay and Polani) is an interventional measure; this one is not.