Downlink Massive Random Access (DMRA)

• Downlink Massive Random Access (DMRA) is a communication paradigm where a base station sends a common message that allows each of k active users (out of n) to independently recover its unique data.
• Techniques like random codebook construction and covering array methods reduce the identity overhead from O(k log n) to O(log k) or even O(1), especially for i.i.d. or uniform source distributions.
• The integration of DMRA with massive MIMO beamforming and ARQ protocols enables scalable, low-latency, and reliable downlink transmissions, with overhead bounds independent of the total user pool size.

Downlink Massive Random Access (DMRA) refers to the setting in which a base station communicates individual messages or data symbols to a randomly selected, typically small, subset of active users among an extremely large pool of total users, by transmitting a single common downlink message. The DMRA paradigm arises in large-scale wireless networks and IoT scenarios requiring efficient downlink one-to-many communication, where each active user must recover its unique intended message given only the common broadcast and its own identity, without knowledge of the other active users or their messages.

1. Problem Formulation and Mathematical Structure

Let $n$ be the total user pool, from which $k \ll n$ users are randomly activated in each communication epoch. The set of active user indices is denoted

$$\mathbf{A} = (A_1, \ldots, A_k) \in \mathcal{A}^{(n,k)}, \qquad \mathcal{A}^{(n,k)} = \{\mathbf{a} \in [n]^k: a_i \ne a_j \,\, \forall i \ne j\}.$$

The base station (BS) has $k$ independent information symbols $\mathbf{X} = (X_1, \ldots, X_k)$, intended respectively for $A_1, \ldots, A_k$. The task is to construct a single common message $M \in \{0,1\}^*$ such that each active user $u$ can recover its unique message $X_i$ solely from its own index $u$ and the common $M$.

Formally, an encoder $$f: \mathcal{X}^k \times \mathcal{A}^{(n,k)} \to \{1,2,\dots\}$$ is defined, with user-specific decoder $d_u$ guaranteeing

$$d_{A_i}\bigl(f(\mathbf{x}, \mathbf{a})\bigr) = x_i \quad \forall (\mathbf{x}, \mathbf{a}),\, i=1,\ldots,k.$$

The minimal achievable rate is $R^* = \min H(f(\mathbf{X}, \mathbf{A}))$, where entropy coding is applied to $f$.

This model encapsulates the key DMRA statistical challenge: efficiently mapping $(\mathbf{X}, \mathbf{A})$ into $M$ so per-user retrieval is guaranteed without revealing the entire active set.

2. Identity-Encoding Overhead and Its Elimination

The naive DMRA coding strategy explicitly labels each source symbol with the recipient's index, incurring an overhead of $k \log_2 n$ bits for active user identities, plus $H(X_1,\dots,X_k)$ bits for the joint symbols. For $n \gg 1$, this identity-overhead dominates and rapidly becomes prohibitive (Song et al., 2024).

However, DMRA exploits statistical symmetries to eliminate this dependence on $n$. When the sources are i.i.d., one can design coding schemes where identity-overhead depends only on $k$ or even disappears entirely:

• i.i.d. sources: Overhead is at most $O(\log k)$ bits.
• Uniform i.i.d. sources: Overhead is $O(1)$ bits for large $k$.
• Exchangeable distributions: Overhead is $O(k)$ bits; for finite-alphabet exchangeable sources, $O(\log k)$ bits.

The elimination of $n$-dependence relies critically on symmetry properties of the message assignment and codebook construction.

3. Coding Techniques: Probabilistic and Deterministic Methods

Random Codebook Construction

For i.i.d. sources $X_1, \dots, X_k \sim p(x)$, a random codebook $\{\mathbf{c}^{(t)}\}$ is constructed by sampling ($n$-dimensional) codewords i.i.d. from $p(x)$. The encoder selects the smallest index $t$ such that for all $i$, $c^{(t)}_{A_i} = X_i$. The number of required bits for $t$ is shown to satisfy

$R^* \leq k H(p) + \log(k H(p) + 1) + 1,$

demonstrating $O(\log k)$ identity-overhead (Song et al., 2024).

Covering Array Constructions

Recent work recognizes DMRA coding as a covering array problem: for alphabet-size $q$, a covering array $CA(M; k, n, q)$ is constructed such that for every choice of $k$ distinct columns and every $q$-ary $k$-tuple, there is at least one row corresponding to that pattern. The mapping is deterministic: the encoder picks the first covering row matching the message pattern, and the index $m$ is encoded.

A strict upper bound on the expected codeword length is shown to be $k \log_2 q + 1 + \log_2 e$ bits, independent of $n$ (Liao et al., 22 Jan 2026). The covering array method thus provides explicit, deterministic codes with tight overhead bounds.

Key Comparison Table

Coding Scheme	Overhead	Dependencies
Naive (explicit ID)	$k\log_2 n$	$n$
Random codebook	$O(\log k)$ / $O(1)$	$k$
Covering array	$k\log_2 q + 1 + \log_2 e$	$k$, $q$

The explicit covering array approach is comparable to probabilistic random coding but yields explicit codebooks and a deterministic worst-case overhead (Liao et al., 22 Jan 2026).

4. Exchangeability, Source Distributions, and Connection to de Finetti

If the message assignment exhibits exchangeable structure (the joint law of $(X_1, \ldots, X_k)$ is invariant under permutations), the rate can be bounded by

$$R^* \leq H(\mathbf{X}) + D\bigl(p\|q\bigr) + \log(H(\mathbf{X}) + D(p\|q) + 1) + 1,$$

where $q$ is an i.i.d. mixture law and $D(p\|q)$ denotes the KL divergence. For urn-based codebooks, $$D(p\|q) \leq \min\{k\log e, |\mathcal{X}| \log(k+1)\}$$, giving overhead $O(k)$ or $O(\log k)$ for finite alphabets (Song et al., 2024).

A major conceptual contribution is the connection to the finite de Finetti theorem, providing KL divergence bounds between finite exchangeable and i.i.d. mixture distributions. For $d$-extendable exchangeable sources,

$$D(p\|q) \leq \min\left\{\log\left(\frac{d^k}{d^{\underline{k}}}\right), (|\mathcal{X}|-1)\log\left(\frac{d-1}{d-k}\right)\right\}$$

vanishes as $d \to \infty$ for fixed $k$—recovering classical results in the limit.

This connection underpins the scaling-optimality of the coding schemes in DMRA, substantiating their fundamental rate bounds.

5. Downlink Massive ARQ and ACK Protocols

In two-way random access (IoT, wireless access), acknowledgment (ACK) messages must confirm successful packet delivery to active users. Listing identities in the ACK would require $B \geq \log_2|\Omega|$ bits where $|\Omega|$ is the number of possible decoded user subsets, incurring large overhead. Efficient joint encoding allows a negligible increase in $B$ by permitting a small false-positive rate $\epsilon_{fp}$:

$B \geq \log_2 |\Omega| - \log_2(1 - \epsilon)$

(Kalør et al., 2022). For practical settings (e.g., $K=10^4$, $A=100$), allowing $\epsilon_{fp}=1\%$ increases $B$ by only $\approx 0.014$ bits, while reducing failure probability by over $8\times$.

Multiround ARQ, in which users retransmit unless positively ACKed, achieves per-user failure rates as low as $10^{-3}$–$10^{-4}$ and mean retransmission count $E[L] \approx 1.05$ (Kalør et al., 2022). The approach is robust under massive user populations and stringent reliability needs.

6. Physical-Layer DMRA and Massive MIMO Beamforming

In crowded TDD massive-MIMO systems, DMRA is operationalized through uplink random access (RA), timing advance (TA) estimation, user grouping, and downlink random access response (RAR) beamforming (Mukherjee et al., 2018). The base station uses Zadoff–Chu sequence correlators to detect RA preambles, estimates group-common TA, and assigns users into spatial groups based on delay spread overlap.

For each group, a short RAR packet (24 bits) is BPSK-modulated and beamformed onto designated subcarriers. Maximum ratio transmission (MRT) is applied using per-group channel estimates. Per-user transmit and RAR beamforming power scales down as $1/\sqrt{M}$ for $M$ base station antennas (yielding $\approx 1.5$ dB gain per doubling):

$$p_u \propto 1/\sqrt{M}, \quad P_T \propto 1/\sqrt{M}.$$

Group RAR transmission resolves preamble collisions, reduces access latency, and increases reliability compared to LTE-like protocols. All processing steps scale linearly in $M$, making the approach practical for hundreds of antennas.

7. Practical Implications, Limitations, and Open Problems

The theory and explicit constructions show DMRA overhead can be bounded independently of $n$, provided suitable codebooks—random or covering-array-based—are used. The per-user overhead can be held below $2.443$ bits ($1 + \log_2 e$) (Liao et al., 22 Jan 2026). In practice, for large $n$ one can use $\lceil \log_2 M \rceil = O(\log \log n)$-bit indexes.

Limitations include computational cost and storage for covering arrays with large $k$ or $q$, and lack of structure complicating rapid addressing. Efficient algebraic constructions and joint source–channel coding extensions remain open research directions. A plausible implication is that integrating these combinatorial codes with channel coding could further reduce overall overhead in noisy environments.

8. Summary of Fundamental DMRA Rate Limits

The key DMRA rate overheads beyond $H(\mathbf{X})$ are:

• i.i.d. sources: $O(\log k)$, $O(1)$ if uniform.
• Arbitrary exchangeable: $O(k)$.
• Finite-alphabet exchangeable: $O(\log k)$.
• $d$-extendable exchangeable: $\log(d^k / d^{\underline{k}}) = O(k^2/d)$, vanishes as $k^2 \ll d$.

These results establish the DMRA regime as supporting scalable, practical, and rate-optimal one-shot downlink schemes, independent of user pool size and fundamentally governed by the statistical properties of the source assignment and codebook construction (Song et al., 2024, Liao et al., 22 Jan 2026).