Hsin-Po's Website


Postdoc @ UC Berkeley

Polar Coding Papers

The following are my works on polar coding, from the oldest to the newest.

In the early days, my research focuses on the moderate deviations principle (MDP) paradigm. MDP addresses the joint relation among block length ($N$), error probability ($P_e$), and code rate ($R$) in a parameter region
$P_e \approx \exp(-N^\pi)$,
$\text{Capacity} - R \approx N^{-\rho}$,
for some positive numbers $\pi$ and $\rho$. The precise goal is to characterize the region of $(\pi, \rho)$-pairs that are achievable for $N \to \infty$.

The MDP paradigm is a combination of the large deviations principle (LDP) paradigm and the central limit theorem (CLT) paradigm. In LDP, one cares about the asymptotic behavior of $P_e$ but not so much about $R$. In the CLT region, it is the opposite, that the asymptotic behavior of $R$ is studied and $P_e$ is more or less ignored.

It all begin with Mondelli, Hassani, and Urbanke’s work Unified Scaling of Polar Codes: Error Exponent, Scaling Exponent, Moderate Deviations, and Error Floors [MHU16]. This work confirms the folklore conjecture that the scaling exponent of polar coding is $\mu \approx 3.627$ over binary erasure channels (BECs) and $\mu < 4.714$ over binary memoryless symmetric (BMS) channels. Here, the scaling exponent $\mu$ is the lowest value $1/\rho$ can take if $P_e$ is a constant. Meanwhile, they give a characterization of the region of $(\pi, \rho)$-pairs. However, their region does not touch the point $(0, 1/3.627)$ for the BEC case or $(0, 1/4.714)$ for the BMS case, which suggests that something nontrivial happens when $P_e$ goes from constant to exponential decay. My work [ModerDevia18] addresses the mismatch and shows that, using a complicated combinatorial counting method, the region of $(\pi, \rho)$-pairs will touch $(0, 1/3.627)$. Hence the slogan moderate deviations recovers the scaling exponent.

While [ModerDevia18] deals with classical polar codes as constructed in Arıkan’s original paper, [LargeDevia18] extends this theory to a wide class of polar codes. Given a kernel $K$, its scaling exponent $\mu$, and its partial distances, we are able to predict how polar codes constructed using $K$ will behave in terms of the region of $(\pi, \rho)$-pairs. Remark: Whereas this result says that it is easy to prove MDP given an estimate of $\mu$. But $\mu$ is usually difficult to estimate.

[LoglogTime18] stands on the result of [ModerDevia18] and shows that, if we would like to tolerate slightly worse $P_e$ and $R$, we can reduce the encoding and decoding complexities from $O(\log N)$ per information bit to $O(\log(\log N))$ per information bit. By worse $P_e$ we mean that $P_e$ scales as $N^{-1/5}$; By worse $R$ we that mean that $\text{Capacity} - R$ scales as $N^{-1/5}$. Note that the constructed code still, achieves capacity, just not as fast as before.

While [LoglogTime18] deals with BECs, [LoglogTime19] handles arbitrary symmetric $p$-ary channels, where $p$ is any prime. The main theorem is a generalization of that in a old paper—by tolerating that $P_e$ converges to $0$ slower and that $R$ converges to the capacity slower, we can reduce the complexity to $\log(\log N)$ per information bit.

Note that, in both [LoglogTime18] and [LoglogTime19], codes are construct with the standard kernel $[^1_1{}^0_1]$; yet the same idea applies if a general kernel $K$ is used. Note also that the log-log behavior generalizes to arbitrary discrete memoryless channels (DMCs). For general channels, however, the standard kernel $[^1_1{}^0_1]$ does not polarize anymore. So the observation that a general kernel $K$ is compatible with the log-log trick is crucial here

Now comes my favorite work.

[Hypotenuse19] shows that it is possible to construct codes whose error probabilities and code rates scale like random codes’ and encoding and decoding complexities scale like polar codes’. On one hand, random codes’ error and rate are considered the optimal. On the other, polar codes’ complexity ($\log N$) is considered low. (Not the lowest possible complexity, as there exist $\log(\log N)$ constructions for general channels and $O(1)$ constructions for BEC.) This result holds for all DMCs, the family of channels Shannon considered in 1948.

[PhDThesis21] is my PhD dissertation. I summarize my earlier works and extend them a little bit.

For a figurative comparison of the region of $(\pi, \rho)$, see Figure 1 on page 3 of [Hypotenuse19] or Figure 5.4 on page 57 of [PhDThesis21]. The pi--rho plot of several milestone works

See also the following table for channels, goals, and references. It is Table 6.1 on page 98 of PhDThesis21. The channels, goals, and the works that achieve them

Here is a table for the error–gap–complexity trade-offs of some well-known capacity-achieving codes and the corresponding channels, obtained from Table 7.1 on page 103 of [PhDThesis21]. Error--gap--complexity trade-offs of some capacity-achieving codes

The following table, from Table 5.1 on page 55 of [PhDThesis21], describes an analog among probability theory, random coding theory, and polar coding theory. Trinitarian analog among probability, random coding, polar coding

In the next two works, [Sub-4.7-mu22] and [TetraErase22], I pivot from MDP to CLT. To be more precise, since any further improvement of MDP will probably come from improvement of the scaling exponent, I try to attack the latter.

[Sub-4.7-mu22] considers the standard setting of polar code: the standard kernel $[^1_1{}^0_1]$ and BMS channels. Before it was proved in [MHU16] that $\mu < 4.71$. We show that $\mu < 4.63$. This improvement is small, but it carries a strong message that considering the iteration one at a time is not enough. In this work, we consider two iterations. We are able to show that, the first iteration will make BMS channel less BSC, and the second iteration will make use of the fact that the channel is never a BSC and polarizes faster.

[TetraErase22] is another attempt to improve scaling exponent. But this time the target is BEC. For BEC, it is known for a long time that $\mu \approx 3.627$ [MHU16]. It is also known for a long time that to improve $\mu$, one can use a larger kernel matrix $K$ [FVY19]. We go for another direction: enlarging the alphabet size. We show that, by enlarging the alphabet size from $2$ to $4$, the scaling exponent improves from $3.627$ to $3.3$. It would have take a $20 \times 20$ (linearly interpolated) matrix to achieve the same improvement.