---
title: Monotone Chain Polar Codes
url: https://www.emergentmind.com/topics/monotone-chain-polar-codes
type: topic
---

# Monotone Chain Polar Codes

Monotone chain polar codes are a class of polar-code constructions in which the polarized symbols are decoded according to a total ordering that is monotone within each component sequence, and they appear in two closely related research lines: as multivariate polar constructions for distributed lossless coding, and as algebraic instances of polar decreasing monomial-Cartesian codes over finite fields [2509.03128] [2002.00325]. In the multivariate viewpoint, a monotone chain is a decoding order on the joint collection of polarized variables from multiple terminals, chosen so that each terminal’s own copy-indices remain in natural order; by varying this chain, one can realize the entire admissible rate region on the dominant face of the Slepian–Wolf problem without time-sharing [2509.03128]. In the algebraic viewpoint, polar-code information sets defined by thresholding monomials under a preorder refining divisibility yield decreasing monomial-Cartesian codes, thereby placing classical binary polar codes, multikernel constructions, and finite-field Reed–Solomon-kernel variants in a common framework [2002.00325].

## 1. Algebraic formulation via decreasing monomial-Cartesian codes

Let $\mathbb{F}=\mathbb{F}_q$ be a finite field, and let
$$
R=\mathbb{F}[x_1,\dots,x_m]
$$
be the polynomial ring in $m$ variables. Fix nonempty subsets $S_i\subseteq\mathbb{F}$ of sizes $n_i:=|S_i|$, and form the Cartesian set
$$
S=S_1\times\cdots\times S_m=\{s^{(1)},\dots,s^{(n)}\}\subseteq \mathbb{F}^m,\qquad n=\prod_{i=1}^m n_i,
$$
with an arbitrary but fixed linear ordering on $S$. A set $\Delta$ of monomials in $R$ is called closed under divisibility, or decreasing, if whenever $M\in\Delta$ and $M' \mid M$, then $M'\in\Delta$ [2002.00325].

Writing $L(\Delta)$ for the $\mathbb{F}$-span of $\Delta$, the evaluation map
$$
\operatorname{ev}_S:L(\Delta)\to\mathbb{F}^n,\qquad
f\mapsto \bigl(f(s^{(1)}),\dots,f(s^{(n)})\bigr)
$$
is $\mathbb{F}$-linear. Under the assumption that every monomial in $\Delta$ has $x_i$-degree $<n_i$, the map is injective. The associated decreasing monomial-Cartesian code is
$$
C(S,\Delta):=\operatorname{ev}_S(L(\Delta))\subseteq\mathbb{F}^n
$$
[2002.00325].

This formulation is the algebraic backbone of the monotone-chain connection. The key point is that a polar-code information set can be indexed by monomials, and when that set is selected by a threshold in an order refining divisibility, the resulting code is a decreasing monomial-Cartesian code. A plausible implication is that many design questions for monotone-chain polar codes can be rephrased as structural questions about monomial ideals on Cartesian point sets.

## 2. Partial orders, threshold monomials, and the monotone-chain viewpoint

To isolate the polar information sets among decreasing monomial sets, Camps, López, Matthews, and Sarmiento introduce a partial order on monomials inspired by Bardet, Dragoi, Otmani, and Tillich. For monomials
$$
M=x_1^{a_1}\cdots x_m^{a_m},\qquad
M'=x_1^{b_1}\cdots x_m^{b_m},
$$
one declares $M'\prec M$ if one of three conditions holds: divisibility $M'\mid M$; a tie-break in one variable block when $S_1=\cdots=S_k$ but $S_k\neq S_{k+1}$, comparing exponent vectors $(a_1,\dots,a_k)$ and $(b_1,\dots,b_k)$ in the product order; or a recursive extension across two variable blocks, where the first-block projections satisfy the same order and the second block is arbitrary [2002.00325].

In practice, this preorder refines divisibility and is compatible with the usual Kronecker-product construction of polar kernels. Whenever the information set of the polar-code construction from a kernel sequence $\{T_i\}_i$ is taken to be those row-monomials $M$ satisfying $M\prec M'$ for some threshold monomial $M'$, the information set forms a divisibility-closed chain and recovers exactly a decreasing monomial-Cartesian code [2002.00325].

This is the point at which monotone-chain polar codes enter the algebraic picture. In the distributed-coding literature, a monotone chain is a total ordering of polarized variables from multiple sources; in the monomial-Cartesian literature, the corresponding information set is specified by a preorder threshold on monomials. The shared feature is a structured monotonicity constraint: order is unrestricted globally, but restricted locally so that divisibility or per-source index order is preserved.

## 3. Polarization over finite fields and multikernel constructions

For a discrete memoryless channel $W:\mathbb{F}\to\mathcal{Y}$, the finite-field framework distinguishes additive-symmetric and multiplicative-symmetric channels. The channel is additive-symmetric if for every $a\in\mathbb{F}$ there is a permutation $\pi_a$ of $\mathcal{Y}$ such that
$$
W(y\mid x)=W(\pi_a(y)\mid x+a),
$$
and multiplicative-symmetric if for every $a\in\mathbb{F}^\times$ there is a permutation $\nu_a$ of $\mathcal{Y}$ such that
$$
W(y\mid x)=W(\nu_a(y)\mid a\cdot x).
$$
If both conditions hold, the channel is symmetric over the field, or SOF [2002.00325].

Given invertible matrices $T_i$ of sizes $\ell_i\times\ell_i$, the multikernel construction uses
$$
G_m=T_1\otimes T_2\otimes\cdots\otimes T_m
$$
up to bit-reversal. The sequence $\{T_i\}_i$ polarizes $W$ if, for all $\varepsilon>0$, the split channels $W^{(i)}$ satisfy
$$
\frac{\#\{i:I(W^{(i)})>1-\varepsilon\}}{n}\to I(W),\qquad
\frac{\#\{i:I(W^{(i)})<\varepsilon\}}{n}\to 1-I(W).
$$
The stated sufficient condition is that, if $p$ is the characteristic of $\mathbb{F}_q$ and each $T_i$ has a non-identity standard form $T_i'$ whose entries generate the full field,
$$
\mathbb{F}_p(T_i')=\mathbb{F}_q,
$$
then the sequence polarizes every SOF channel over $\mathbb{F}_q$ [2002.00325].

A particularly important instance is the generalized Reed–Solomon evaluation matrix on any $n_i\le q$ points of $\mathbb{F}_q$,
$$
(T_i)_{jk}=(s_{i,j})^{k-1},\qquad 1\le j\le n_i,\ 1\le k\le n_i.
$$
The original Arıkan $2\times2$ kernel $G_A$ yields the binary decreasing-monomial codes of Bardet et al., while Reed–Solomon kernels appear by taking the $S_i$ to be $\ell$ distinct points in $\mathbb{F}_q$; the summary states that they achieve the best possible exponent
$$
E(T_i)=\frac{1}{\ell}\log_\ell(\ell!)
$$
known in the literature and polarize SOF channels under the finite-field theorem [2002.00325].

The algebraic unification is therefore not limited to a single kernel family. It subsumes classical binary polar coding and extends it to arbitrary field sizes and arbitrary sequences of kernels satisfying the field-generation condition.

## 4. Code parameters, minimal generators, and duality

A subset $B\subseteq\Delta$ is a generating set if every monomial $M\in\Delta$ divides some $B\in B$, and it is minimal if no two elements of $B$ are divisible by one another. Every decreasing $\Delta$ has a unique minimal generator $B(\Delta)$ [2002.00325]. This object controls the basic parameters of the code.

The length is
$$
\operatorname{length}(C(S,\Delta))=n=\prod_{i=1}^m n_i.
$$
If each $B\in B(\Delta)$ is written as $x_1^{b_1}\cdots x_m^{b_m}$, then the dimension is given by the inclusion–exclusion formula
$$
k=\dim C(S,\Delta)
=\sum_{\emptyset\neq P\subseteq B(\Delta)}(-1)^{|P|-1}\prod_{i=1}^m\bigl(t_i(P)+1\bigr),
$$
where for each nonempty $P\subseteq B(\Delta)$,
$$
t_i(P):=\min_{B\in P}\{\text{exponent of }x_i\text{ in }B\}.
$$
The minimum distance is
$$
d(C(S,\Delta))
=\min_{B=x_1^{b_1}\cdots x_m^{b_m}\in B(\Delta)}\prod_{i=1}^m (n_i-b_i)
$$
[2002.00325].

The same work proves a duality theorem: the Euclidean dual $C(S,\Delta)^\perp$ is monomially equivalent to another decreasing monomial-Cartesian code $C(S,\Delta^\perp)$, where
$$
\Delta^\perp=
\left\{
x_1^{\,n_1-1-a_1}\cdots x_m^{\,n_m-1-a_m}:
x_1^{a_1}\cdots x_m^{a_m}\notin\Delta
\right\}.
$$
Moreover, a basis of the dual can be written in terms of residue-polynomials, or Lagrange residues, attached to the monomials in $\Delta^\perp$ [2002.00325].

For monotone-chain polar codes, these formulas are significant because they make length, dimension, and minimum distance explicit in terms of the thresholded monomial set. The summary further states that the minimal-generator formula for the minimum distance shows that the minimum distance of the $m$-fold Kronecker product code grows super-polynomially in $m$ as soon as one picks information monomials all of high total degree, yielding explicit lower bounds on $d_N$ versus $N$ that in many cases exceed those known for random binary kernels of the same rate [2002.00325]. It also states that the dual-code equivalence yields explicit LCD polar codes by exchanging $\Delta\leftrightarrow\Delta^\perp$.

## 5. Multiterminal monotone chains and the Slepian–Wolf rate region

In the distributed lossless coding formulation, let $M$ be the number of correlated sources, with alphabet sizes $q_1,\dots,q_M$, and let $N=2^n$. For $\gamma=1,\dots,M$, define
$$
X^\gamma_{1:N}=(X^\gamma_1,\dots,X^\gamma_N),
$$
with each sequence an i.i.d. copy of source $X^\gamma$. The polar transform
$$
G_N=G_2^{\otimes n},\qquad
G_2=\begin{bmatrix}1&0\\1&1\end{bmatrix},
$$
is applied separately to each source:
$$
U^\gamma_{1:N}=X^\gamma_{1:N}\cdot G_N.
$$
The joint system contains the $MN$ polarized symbols $\{U^{\gamma_t}_{i_t}:t=1,\dots,MN\}$, where $(\gamma_t,i_t)$ is a permutation of $(1,1),(1,2),\dots,(M,N)$ [2509.03128].

A chain is a total ordering $t\mapsto (\gamma_t,i_t)$, and it is monotone if, whenever $t<t'$ with $\gamma_t=\gamma_{t'}$, one has $i_t<i_{t'}$. Equivalently, each source’s own copy-indices appear in natural order, so the monotone chain is fully specified by the length-$MN$ sequence $\gamma_{1:MN}$ [2509.03128].

The entropy-chain-rule formulation is
$$
N\cdot H(X^{1:M})
=
\sum_{t=1}^{MN}
H\bigl(U^{\gamma_t}_{i_t}\mid U^{\gamma_1}_{i_1},\dots,U^{\gamma_{t-1}}_{i_{t-1}}\bigr).
$$
For each position define
$$
h_t=
H\bigl(U^{\gamma_t}_{i_t}\mid U^{\gamma_1}_{i_1},\dots,U^{\gamma_{t-1}}_{i_{t-1}}\bigr).
$$
As in classical polar coding, these conditional entropies polarize to $0$ or full $\log(q_\gamma)$. Fixing $\delta\in(0,1)$ and taking $N$ large, one chooses a high-entropy set $F\subseteq\{1,\dots,MN\}$ of size approximately $\sum_\gamma \lceil N\cdot R_\gamma\rceil$, where $t\in F$ corresponds to large $h_t$ and thus frozen symbols, while $t\notin F$ corresponds to small $h_t$ and thus information symbols [2509.03128].

The achievable-rate statement is that, in the Slepian–Wolf problem, one seeks $(R_1,\dots,R_M)$ on the dominant face of
$$
\left\{
R:\sum_{\gamma\in S}R_\gamma \ge H(X^S\mid X^{S^c})
\text{ for all }S\subset [M]
\right\},
$$
and Bilkent–Arıkan showed that by choosing different monotone chains $\gamma_{1:MN}$, the induced rates
$$
R_\gamma=\frac{1}{N}\,|\{t:\gamma_t=\gamma,\ t\notin F\}|
$$
cover every point on the dominant face, without time-sharing [2509.03128]. The same chain rule and monotonicity argument extend to arbitrary $M\ge 2$ and $q_\gamma\ge 2$, with quasigroups or random permutations used to ensure polarization in the non-binary setting.

## 6. Successive cancellation decoding, complexity, and list decoding

General monotone-chain polar codes introduce decoding issues that do not arise in the classical one-dimensional chain. The successive cancellation task is to estimate $U^{\gamma_t}_{i_t}$ sequentially for $t=1,\dots,MN$, given the previously decoded symbols, by computing
$$
\Pr\bigl(U^{\gamma_t}_{i_t}=u\mid
U^{\gamma_1}_{i_1}=\hat u^{\gamma_1}_{i_1},\dots,
U^{\gamma_{t-1}}_{i_{t-1}}=\hat u^{\gamma_{t-1}}_{i_{t-1}}
\bigr).
$$
Once $\hat u^{\gamma_t}_{i_t}$ is chosen, whether frozen or by an ML rule, a partial decision is imposed at the corresponding leaf by making the probability tensor deterministic in the known components and uniform over the unknown ones [2509.03128].

The decoding framework in [2509.03128] represents any multivariate distribution $P^{X^{1:M}}$ by an $M$-dimensional tensor $\mathcal{P}$ and uses three tensor operations: circular convolution $\circledast$, dual convolution $\circledast'$, and normalized product $\odot$. On a size-$2$ butterfly with relations $P_i=L_i+R_i$ and $P_{i+\ell}=R_i$, the inference updates are
$$
\text{calcLeft:}\quad \mathcal{P}^L=\mathcal{P}^P\circledast(\mathcal{P}^{P_{i+\ell}}\odot \mathcal{P}^R),
$$
$$
\text{calcRight:}\quad \mathcal{P}^R=\mathcal{P}^{P_{i+\ell}}\odot(\mathcal{P}^L\circledast' \mathcal{P}^P),
$$
$$
\text{calcParent:}\quad (\mathcal{P}^P,\mathcal{P}^{P_{i+\ell}})=
(\mathcal{P}^L\circledast' \mathcal{P}^R,\mathcal{P}^R).
$$
The full polar transform is viewed as a directed computation graph with $2N-1$ edges and $N-1$ vertices, each edge carrying a probability tensor and each vertex implementing one of these local update rules [2509.03128].

Two complexity facts are central. First, any SC decoder must spend $\Omega(N\log N)$ operations, since each of the $N$ leaves must receive the root prior at least once along a path of length $O(\log N)$. Second, a general monotone chain requires at most $O(N^2)$ work, because each of the $MN$ decoding steps may force a recomputation along the path between successive leaves, which costs $O(N)$ in the worst case [2509.03128]. Corner-point chains, in which one decodes each source in a single block, reduce to classical SC in each block and therefore achieve total $O(N\log N)$ complexity; alternating chains can force leaf jumps across the root at every step and therefore incur $\Theta(N^2)$ complexity.

A common misconception is that the classical $O(N)$-space optimization for SC decoding extends unchanged to general monotone chains. The 2025 decoding paper states explicitly that this is not universally applicable: in a general monotone chain, frequent back-and-forth between leaves invalidates one-time deletion of graph memory, so one must keep $O(N\log N)$ tensors or accept $O(N^2)$ time through recomputation [2509.03128].

For list decoding, the same work introduces a pointer-based computational-graph implementation with two core data structures, `Edge` and `Vertex`, and a head pointer `decHead`. The `initGraph()` procedure allocates all $2N-1$ edges, `decodeAt(t)` moves the head along a path of vertices in $O(\log N)$, and constant-time decoder forking is achieved by copying only the `decHead` pointer. This enables time-efficient list decoding without relying on $O(N)$-space tricks [2509.03128].

The reported runtime comparison for list size $L=2$, corner chain, and binary sources shows the following timings:

| Blocklength | Lazy-copy time (s) | Proposed (s) |
|---|---:|---:|
| $N=64$ | $5.87\times10^{-3}$ | $5.79\times10^{-3}$ |
| $N=256$ | $2.46\times10^{-2}$ | $2.41\times10^{-2}$ |
| $N=1\,024$ | $1.03\times10^{-1}$ | $9.49\times10^{-2}$ |
| $N=4\,096$ | $4.29\times10^{-1}$ | $3.75\times10^{-1}$ |
| $N=16\,384$ | $2.51$ | $1.44$ |
| $N=65\,536$ | $20.2$ | $5.43$ |
| $N=262\,144$ | $255$ | $21.8$ |

The corresponding improvements reported in the source range from $-1.4\%$ at $N=64$ to $-91.5\%$ at $N=262\,144$, averaged over $100$ runs with no fast-pruning or min-sum optimizations [2509.03128]. For a non-binary example with $M=2$, ternary $X^1$, quinary $X^2$, and a joint pmf of size $15$, the same source reports that BLER improves dramatically when moving from SC ($L=1$) to list decoding ($L=32)$, and that some non-corner chains slightly outperform corner chains at the same sum-rate.

Taken together, these results present monotone-chain polar codes as both an algebraic and an algorithmic generalization of classical polar coding: algebraically, they are captured by decreasing monomial-Cartesian codes and their duals; operationally, they achieve the full dominant face of the multiterminal Slepian–Wolf region and require decoding methods whose complexity depends sensitively on chain structure [2002.00325] [2509.03128].

Source: https://www.emergentmind.com/topics/monotone-chain-polar-codes