The Optimization of Random Tree Codes for Limited Computational Resources

The probability of error that \mathbf{\hat{m}}\neq \mathbf{m} can be decomposed into two terms as follows:

Now, we will give individual upper bound for the RHS terms in (Proof).

First, observe that any node at branching time h+1 is not checked if its parent node \mathbf{m'}_{1:s(b_{h})} satisfies d_{\mid \mathbf{y}}(\mathbf{m'}_{1:s(b_{h})})> d_{\mid \mathbf{y}}(\mathbf{m}). The reason is that the stack of the decoding algorithm always contains an ancestor of \mathbf{m} which has a decoding error cost at most as d_{\mid \mathbf{y}}(\mathbf{m}) (due to the definition of AEC measures) hence a node with a decoding error cost lower than of m'_{1:s(b_{h})} always exists. Therefore, the number of nodes that the decoding algorithm checks N_{c} satisfies:

with probability 1. Accordingly, we have:

Secondly, observe that if \mathbf{\hat{m}} \neq \varepsilon, a wrong message is decoded only if there exists at least two terminal nodes \mathbf{m'} that satisfy d_{\mid \mathbf{y}}(\mathbf{m'}) \leq d_{\mid \mathbf{y}}(\mathbf{m}) as \mathbf{m'}=\mathbf{m} always satisfies the condition. Accordingly, we have:

Combining (Proof) with (?) and (?) proves (1).

First observe that:

where the expectation is taken over only on \mathbf{\bar{m}}_{1:s(b_{h})}. Then, consider:

where the first inequality is due to that \Pr(X \geq 1)\leq \mathbb{E}[[X]_{\leq 1}] for a non-negative random variable X, the third equality is due to (Proof), the fourth equality is due to c_{h}=2^{s(b_{h+1})-s(b_{h})}, and the second inequality is due to Jensen’s inequality for the commutation of [\cdot]_{\leq 1} and \mathbb{E}[\cdot \mid \mathbf{x}, \mathbf{y}, \mathbf{\bar{m}}] operations.

Similar to (Proof), observe that:

where the expectation is taken over only on \mathbf{\bar{m}}. Then, consider:

where the third equality is due to that \Pr(X \geq 1)=\mathbb{E}[[X]_{\leq 1}] for X \in \lbrace 0, 1, \cdots, 2^{k}\rbrace, the fourth equality is due to (?), and the inequality is due to Jensen’s inequality for the commutation of [\cdot]_{\leq 1} and \mathbb{E}[\cdot \mid \mathbf{x}, \mathbf{y}, \mathbf{\bar{m}}] operations.

Combining (Proof) and (?) proves (2).

Applying expectation for both sides on (?) gives (3).

If s(1)=k, then t=1 is the only branching time and hence \mathbf{G}\mathbf{\bar{m}} and \mathbf{G}\mathbf{m} are conditionally indepedent for the conditioning on \mathbf{\bar{m}} \neq \mathbf{m}. Accordingly,

where the second equality is due that \mathbf{G}\mathbf{\bar{m}} and \mathbf{G}\mathbf{m} are conditionally indepedent for the conditioning on \mathbf{\bar{m}} \neq \mathbf{m}.

Combining (Proof) and (4) with (?) gives (4).

Consider a node \mathbf{m'}_{1:s(t_{1})} and its descendant \mathbf{m'}_{1:s(t_{2})} where 1 \leq s(t_{1}) < s(t_{2}) \leq k and 1 \leq t_{1} < t_{2} \leq n where t_{1}=r_{s(1)} and t_{2}=r_{s(2)}. Observe that:

where the third and fifth inequalities follow from the definition of channels that are causal, and the inequality follows from the fact that -\log_{2}(P) \geq 0 for P \in [0,1].

Observe that:

where the first inequality follows from D_{\text{CLE}}^{\text{(U)}}[s,d]\leq D_{\text{CLE}}^{\text{(M)}}[s,d] and (?), the second inequality follows from that a \leq a^{\varrho} for a, \varrho \in [0,1], and the third inequality follows from Chernoff bound.

Then, observe that:

where the first inequality follows from applying Jensen’s inequality on (?) for the commutation of [\cdot]_{\leq 1} and \mathbb{E}[\cdot \mid \mathbf{y}] operations, the second inequlity follows from that [a + b]_{\leq 1}\leq [a]_{\leq 1}+[b]_{\leq 1}, the third inequlity follows from that [a]_{\leq 1} \leq a^{\rho} for \rho \in [0,1] and the fourth inequlity follows from Chernoff bound.