MDP Convolutional Codes Overview

• MDP convolutional codes are trellis codes that achieve the fastest possible growth in column distances, ensuring optimal error and erasure correction in sliding-window frameworks.
• They are constructed over finite fields and rings like ℤₚʳ using p-encoder techniques and p-linear combinations, which generalize classical algebraic methods.
• Their design supports robust sequential decoding in real-time streaming applications while meeting stringent distance bounds such as generalized Singleton limits.

A Maximum Distance Profile (MDP) convolutional code is a class of convolutional or trellis code that achieves the fastest possible growth in its column distances, providing optimal error and erasure correction in sliding-window settings. The notion, originally formulated over finite fields, extends the classical optimum distance profile criteria and is now well-developed over both fields and finite rings, such as $\Z_{p^r}$. MDP convolutional codes are characterized by their capability to maximize the number of correctable errors or erasures in any given window, making them integral in delay-constrained, streaming, and real-time communication systems.

1. Algebraic and Module-Theoretic Foundations

An $(n,k,\delta)$ convolutional code over a ring $R$ (typically $\Z_{p^r}$ or a finite field $\F_q$) is an $R[D]$-submodule of $R^n[D]$ of $R$-rank $k$ and degree $\delta$ equal to the sum of the row-degrees of any minimal encoder. Over $\Z_{p^r}$, the construction uses $p$-encoders---matrices $G(D)\in\Z_{p^r}^{k\times n}[D]$ whose rows form a $p$-basis, defined using $p$-linear combinations (coefficients in $\A_p[D]$, where $\A_p = \{0,1,\ldots,p-1\}$). Any codeword has a unique $p$-adic expansion, and the structure is closely tied to the $p$-adic and Teichmüller theoretic properties of the ring.

This $p$-basis framework underpins convolutional codes over rings, generalizing the classical field-based minimality and dimension concepts, and facilitates a natural extension of the distance profile theory (Napp et al., 2017). The $p$-dimension replaces the field dimension, and the $p$-degree generalizes the Forney index sum.

2. Column Distances and the MDP Criterion

For a convolutional code, the $j$th column distance $d^c_j$ is defined as

$$d^c_j = \min\left\{\text{wt}(v_{[0,j]}(D))\,:\, v(D)=u(D)G(D),\, u_0\ne0\right\},$$

where $v_{[0,j]}(D)$ is the truncation of the codeword up to time $j$. For codes over rings, $d^c_j$ is computed over all inputs $u(D)\in \A_p^k((D))$ for $p$-encoders, with $u_0\neq0$.

The fundamental upper bound on column distances is (Napp et al., 2017, Lieb et al., 28 Jan 2026): $$d^c_j \le (j+1)\Bigl(n-\left\lceil k/r \right\rceil\Bigr) + 1$$ over $\Z_{p^r}$, where $\left\lceil k/r \right\rceil$ arises from the minimum sum over $p$-basis parameters; for fields, this specializes to the classical

$d^c_j \le (n-k)(j+1)+1.$

MDP convolutional codes attain this bound with equality for $j$ up to a parameter $L$, where

$$L = \left\lfloor \frac{\delta}{k} \right\rfloor + \left\lfloor \frac{\delta}{n-k} \right\rfloor.$$

Thus, an $(n,k,\delta)$ code is MDP over $\Z_{p^r}$ (or field) if

$$d^c_j = (j+1)\left(n-\left\lceil k/r \right\rceil\right)+1, \quad 0 \le j \le L,$$

with $r=1$ recovering the field case.

3. Structural and Matrix Characterizations

The MDP property is characterized by minor nonvanishing in specific truncated sliding generator or parity-check matrices (Dang et al., 14 Jul 2025, Napp et al., 2017). For a $p$-encoder $G(D)$, the block matrix

$$G^c_j = \begin{bmatrix} G_0 & G_1 & \cdots & G_j \ 0 & G_0 & \cdots & G_{j-1} \ \vdots & & \ddots & \vdots \ 0 & 0 & \cdots & G_0 \end{bmatrix}$$

of size $(j+1)k \times (j+1)n$, must have all "nontrivial" full-size minors nonzero for $j\le L$. The same applies to parity-check matrices built from $H(D)$.

Over rings, matrix decompositions such as

$G(D) = \begin{bmatrix} \G_0(D) \ p\,\G_1(D) \ \vdots \ p^{r-1}\G_{r-1}(D) \end{bmatrix}$

enable a direct-sum structure $$\mathcal{C} = \mathcal{C}_0 \oplus p\,\mathcal{C}_1 \oplus \dots \oplus p^{r-1} \mathcal{C}_{r-1}$$, each component being free. The $p$-encoder criterion for the sliding matrix holds as long as the required $p$-generator sequences and minor conditions are met.

4. Explicit Construction Methods

MDP convolutional codes over $\Z_{p^r}$ are constructed by "lifting" field codes (Napp et al., 2017):

1. Take an MDP $(n,\widetilde{k},\widetilde{\delta})$ code over $\Z_p$ with a minimal basic encoder.
2. Partition the encoder to blocks corresponding to the $p$-dimension decomposition.
3. Form a $p$-encoder by stacking $p$-multiples of these blocks in a precise pattern, ensuring row-degree sum $\delta$ and that, up to $j=L$, the truncated sliding matrices have maximal column distances.

This method generalizes the construction of MDP codes via superregular matrices or superregular Toeplitz patterns for field codes, producing (possibly nonfree) ring codes.

5. Theoretical Bounds and Comparative Aspects

The free distance of an MDP convolutional code over $\Z_{p^r}$ is bounded by

$$d_{\rm free} \le n\Bigl(\left\lfloor \frac{\delta}{k} \right\rfloor + 1 \Bigr) - \left\lceil \frac{k}{r} \Bigl( \left\lfloor \frac{\delta}{k} \right\rfloor + 1 \Bigr) - \frac{\delta}{r} \right\rceil + 1,$$

with column distances strictly maximal until this threshold is met. For $r=1$, this recovers the generalized Singleton bound for finite fields.

Rank is generally "lost" over $\Z_{p^r}$ compared to the field case due to the $p$-adic structure, but the MDP property continues to guarantee optimal sliding-window distance growth.

6. Applications, Algorithmic Implications, and Limitations

MDP convolutional codes over $\Z_{p^r}$ are particularly well-suited for sequential decoding (such as Viterbi-type algorithms), and are essential in scenarios involving nonbinary alphabets or channels naturally modeled using $\Z_{p^r}$ (e.g., phase-modulation or network coding over rings). The $p$-basis and $p$-encoder machinery allow coding-theoretic concepts to accommodate non-free codes and to encompass classical ring constructions (e.g., cyclic or Hensel lifts) within a unifying algebraic lifting framework.

However, the encoder structure for ring codes is more complex and requires explicit handling of $p$-adic cancellation and $p$-linear combinations. The classification of codes admitting noncatastrophic $p$-encoders is also an open area, and the structural complexity is higher compared to free module codes over fields.

7. Summary Table: MDP Convolutional Codes over $\Z_{p^r}$ vs. Fields

Feature	Over Fields ($r=1$)	Over $\Z_{p^r}$
Max column distance	$d^c_j = (n-k)(j+1)+1$	$d^c_j = (j+1)\big(n-\lceil k/r\rceil\big)+1$
MDP construction	Superregular matrices, AG, cyclic, lifting	Stacked $p$-multiple lift of field MDP code
Module structure	Free $k$-dimensional $\F_q[D]$-module	$p$-basis; may be nonfree; $p$-span
Distance upper bounds	Classical Singleton, field-based expressions	Singleton-type with $p$-basis correction
Structural decomposition	Field: direct sum free, Forney indices	Ring: direct sum by $p$-multiples of free codes
Encoding complexity	Standard polynomial arithmetic	Complex $p$-adic operations, $p$-span handling

MDP convolutional codes over $\Z_{p^r}$ represent the canonical extension of the maximum distance profile concept to the ring setting, optimally exploiting the column distance growth for as long as allowed by the (generalized) Singleton bound, and providing a uniform algebraic and algorithmic framework for robust streaming and storage applications over both fields and finite rings (Napp et al., 2017).