---
title: One-Shot Random Coding Bound
url: https://www.emergentmind.com/topics/one-shot-random-coding-bound
type: topic
---

# One-Shot Random Coding Bound

The one-shot random coding bound is a nonasymptotic achievability statement for communication, compression, or simulation when a source or channel is used once, or when a finite-blocklength instance is analyzed directly rather than through asymptotic typicality. In the literature, the term covers several closely related constructions: stochastic encoders and decoders weighted by information density, Poisson functional representations and Poisson matching arguments, hypothesis-testing bounds, and Rényi-entropy or decoupling formulations. A particularly influential formulation is the one developed from the strong functional representation lemma, which yields explicit universal constants for exact channel simulation, variable-length lossy source coding, multiple description coding, Gray–Wyner coding, and a reduction of the Gelfand–Pinsker problem to point-to-point coding [1701.02827].

## 1. Nonasymptotic meaning and canonical form

One-shot information theory studies source coding and channel coding when the signal blocklength is \(1\); in that regime, each source and channel can be used only once, and the sources and channels are arbitrary and not required to be memoryless or ergodic [2511.19556]. A one-shot random coding bound is then an achievability inequality obtained by averaging performance over a randomized code ensemble and then, typically, derandomizing or extracting a deterministic code.

A common single-user form is an information-density bound. For a channel \(P_{Y|X}\), an input law \(P_X\), and message set size \(M\), one Poisson-matching-based formulation gives
\[
P_e \le \mathbb{E}\!\left[\min\{M\cdot 2^{-\iota_{X;Y}(X;Y)},1\}\right],
\]
where \(\iota_{X;Y}(x;y)=\log \frac{dP_{Y|X}(y|x)}{dP_Y(y)}\) [2511.19556]. An earlier stochastic-coder technique yields the closely related point-to-point lower bound on correct decoding
\[
\mathbb{E}P[C]\ge \mathbb{E}_{q_{XY}}\!\left[\frac{1}{1+(M-1)2^{-(X;Y)}}\right],
\]
with stochastic likelihood or stochastic mutual-information decoding and Jensen’s inequality as the key analytic step [1303.0696].

These formulations already display the characteristic structure of one-shot random coding: a finite-blocklength trade-off expressed directly through information quantities rather than asymptotic rates. In later work, the same pattern reappears in multiterminal settings, but with more elaborate denominators, auxiliary variables, or logarithmic penalties.

## 2. Strong functional representation as a coding mechanism

The strong functional representation lemma (SFRL) states that for random variables \(X\) and \(Y\) on a Polish space with Borel probability measures and \(I(X;Y)<\infty\), there exists a random variable \(Z\) such that \(Z \perp X\), \(Y=g(X,Z)\), and
\[
I(X;Z|Y)\le \log(I(X;Y)+1)+4.
\]
Equivalently, since \(Z\perp X\) and \(H(Y|X,Z)=0\),
\[
H(Y|Z)\le I(X;Y)+\log(I(X;Y)+1)+4.
\]
If \(X\) and \(Y\) are discrete with finite alphabets, \(Z\) can be chosen with cardinality \(|Z|\le |X|(|Y|-1)+2\) [1701.02827].

The proof uses a marked Poisson point process representation. One draws a rate-\(1\) Poisson point process \(\{T_i\}\) with i.i.d. marks \(\{\tilde Y_i\}\sim P_Y\), defines \(Z=\{(\tilde Y_i,T_i)\}\), sets
\[
K(x,z)=\arg\min_i T_i\cdot \frac{dP_Y}{dP_{Y|X=x}}(\tilde Y_i),
\]
and outputs \(g(x,z)=\tilde Y_{K(x,z)}\). Bounding \(\mathbb{E}[\log K]\) and then using a maximum-entropy argument yields the additive constant \(+4\) in the SFRL bound [1701.02827].

This construction matters because it turns a probabilistic channel \(P_{Y|X}\) into a deterministic map fed by independent randomness. In the discrete case, the same idea admits a simpler representation with independent exponentials \(Z_y\sim \mathrm{Exp}(1)\) and
\[
g(x,z)=\arg\min_y \frac{z_y}{p_{Y|X}(y|x)}
\]
after appropriate scaling by \(p_Y(y)\) [1701.02827]. The paper interprets the Poisson representation as a form of “soft random coding”: instead of nearest-codeword selection in a conventional random codebook, the encoder selects an index by a likelihood-weighted competition among random candidates.

## 3. Exact channel simulation and variable-length lossy coding

In one-shot exact channel simulation, Alice observes \(X\sim P_X\), Alice and Bob share unlimited common randomness \(W\), Alice sends a prefix-free message \(M\), and Bob must generate \(Y\) so that the joint law \((X,Y)\) matches a target channel \(P_{Y|X}\). Choosing \(W=Z\) from the SFRL and Huffman-coding \(Y|Z\) gives
\[
\mathbb{E}[L(M)]\le H(Y|Z)+1\le I(X;Y)+\log(I(X;Y)+1)+5.
\]
A converse lower bound is immediate from the Markov chain \(X-M-Y\) given \(W\):
\[
\mathbb{E}[L(M)]\ge H(M|W)\ge I(X;Y).
\]
For arbitrary inputs \(x\in\mathcal X\) and channel capacity \(C\), the same approach yields
\[
\mathbb{E}[L(M)]\le C+\log(C+1)+5,
\]
and in the discrete case the common randomness can be chosen with \(|W|=|Z|\le |X|(|Y|-1)+2\) in the distribution-dependent setting or \(|Z|\le |X||Y|+1\) in the capacity-based worst-case-input setting [1701.02827].

The same lemma gives a one-shot variable-length lossy source-coding theorem. For source \(X\sim P_X\), distortion \(d\), and rate-distortion function
\[
R(D)=\inf_{P_{Y|X}:\,\mathbb E[d(X,Y)]\le D} I(X;Y),
\]
the achievable expected description length satisfies
\[
\bar R \le R(D)+\log(R(D)+1)+6.
\]
The proof first selects \(P_{Y|X}\) with \(I(X;Y)\le R(D)+\epsilon\), applies SFRL to obtain \(Y=g(X,Z)\) with \(H(Y|Z)\le I(X;Y)+\log(I(X;Y)+1)+4\), and then uses a two-point Carathéodory mixture so that the entropy and distortion constraints hold simultaneously before applying Huffman coding [1701.02827].

The Poisson representation yields an explicit operational picture. The “soft codebook”
\[
Z=\{(\tilde Y_i,T_i)\}
\]
is shared conceptually, the encoder chooses
\[
K=\arg\min_i T_i\cdot \left(\frac{dP_Y}{dP_{Y|X}}\right)(\tilde Y_i|X),
\]
and \(K\) can be encoded with a Zipf-based prefix code. In blocklength \(n\), the excess logarithmic term scales like \(\log n\), so the per-symbol overhead vanishes [1701.02827].

## 4. Multiterminal and state-dependent regimes

The same representation method extends to several network source-coding problems. For one-shot multiple description coding, if there exists \(P_{U,Y_0,Y_1,Y_2|X}\) such that
\[
\bar R_1 \ge I(X;Y_1,U)+2\eta,\qquad
\bar R_2 \ge I(X;Y_2,U)+2\eta,
\]
and
\[
\bar R_1+\bar R_2 \ge I(X;Y_0,Y_1,Y_2|U)+2I(X;U)+I(Y_1;Y_2|U)+5\eta,
\]
with
\[
\eta=\log\!\big(I(X;Y_0,Y_1,Y_2,U)+I(Y_1;Y_2|U)+1\big)+7,
\]
then the tuple \((\bar R_1,\bar R_2,D_0,D_1,D_2)\) is achievable under the stated distortion constraints [1701.02827].

For the Gray–Wyner system, if \(P_{U|X_1,X_2}\), \(P_{Y_1|X_1,U}\), and \(P_{Y_2|X_2,U}\) satisfy the distortion requirements, then one-shot achievability holds under
\[
\bar R_0 \ge I(X_1,X_2;U)+\log(I(X_1,X_2;U)+1)+8,
\]
\[
\bar R_1 \ge I(X_1;Y_1|U)+\log(I(X_1;Y_1|U)+1)+5,
\]
\[
\bar R_2 \ge I(X_2;Y_2|U)+\log(I(X_2;Y_2|U)+1)+5.
\]
In both multiple description and Gray–Wyner settings, the added logarithmic terms scale like \(\log n\) at blocklength \(n\), so they disappear asymptotically relative to \(n\) [1701.02827].

The Gelfand–Pinsker problem admits a particularly transparent reduction. For fixed \(p_{U|S}\) and \(x(u,s)\), SFRL applied to \(S\to U\) yields \(V\perp S\) with
\[
H(U|V)\le I(U;S)+\log(I(U;S)+1)+4.
\]
Therefore
\[
I(V;Y)\ge I(U;Y)-I(U;S)-\log(I(U;S)+1)-4.
\]
At blocklength \(n\), one obtains
\[
I(V;Y^n)\ge n[I(U;Y)-I(U;S)]-\log(nI(U;S)+1)-4,
\]
and standard channel coding over \(p_{Y^n|V}\) recovers the classical Gelfand–Pinsker rate \(I(U;Y)-I(U;S)\) as \(n\to\infty\) [1701.02827]. The interpretation given in the paper is that SFRL “orthogonalizes” \(U\) from \(S\) by introducing \(V\) independent of \(S\).

## 5. Poisson matching, general networks, and codebook diversity

A later unifying step is the Poisson matching lemma (PML), which replaces both packing and covering lemmas. For measures \(P,Q\ll \mu\) and Poisson functional representatives \(\tilde U_P,\tilde U_Q\),
\[
\mathbf P\!\left\{\left.\tilde U_Q\neq \tilde U_P\,\right|\,\tilde U_P\right\}
\le 1-\left(1+\frac{dP}{dQ}(\tilde U_P)\right)^{-1}.
\]
This single inequality yields one-shot achievability results for DMCs, channels with state known at the encoder, lossy source coding with decoder side information, joint source-channel coding, broadcast channels, Berger–Tung coding, multiple access channels, channel resolvability, and wiretap channels [1812.03616].

For general acyclic noisy networks, the exponential process refinement lemma extends the same philosophy. The deterministic one-shot error theorem takes the generic form
\[
P\big((\tilde X^N,\tilde Y^N)\in \mathcal E\big)
\le \mathbb E\!\left[\min\left\{\mathbf 1\{(X^N,Y^N)\in\mathcal E\}+\sum_{i=1}^N\sum_{j=1}^{d_i'} B_{i,j},\,1\right\}\right],
\]
with explicit multiplicative terms \(B_{i,j}\) built from products of information-density differences and logarithmic factors \(\gamma_{i,j}=\prod_{k=j+1}^{d_i}(\ln|\mathcal U_{a_{i,k}}|+1)\) [2402.06021]. Point-to-point channel coding is recovered as the specialization
\[
P_e \le \mathbb E\!\left[\min\{\mathsf L\cdot 2^{-\iota(X;Y)},1\}\right].
\]

A distinct but related development appears in one-shot broadcast joint source-channel coding with multiple decoders. With disjoint codebooks at \(K\) decoders, the error probability satisfies
\[
P_e \le \mathbb{E}\!\left[\Big(1+K\,P_Z(\mathcal B_D(W))\,2^{\iota_{X;Y}(X;Y)}\Big)^{-1}\right].
\]
With a shared codebook, the baseline bound becomes
\[
P_e \le \mathbb{E}\!\left[\Big(1+P_Z(\mathcal B_D(W))\,\max_{1\le k\le K}2^{\iota_{X;Y}(X;Y_k)}\Big)^{-1}\right],
\]
and a hybrid grouped scheme interpolates between them:
\[
P_e \le \mathbb{E}\!\left[\Big(1+J\,P_Z(\mathcal B_D(W))\,\max_{1\le k\le L}2^{\iota_{X;Y}(X;Y_k)}\Big)^{-1}\right].
\]
The paper interprets the multiplicative \(K\) or \(J\) as a codebook-diversity gain distinct from channel diversity, and the second-order expansion exhibits a \(-\log K\) or \(-\log J\) improvement in the backoff term [2601.10648].

## 6. Tightness, structured-code variants, and quantum generalizations

One recurrent issue is whether the logarithmic overhead in one-shot random coding is an artifact. In the SFRL framework the relevant quantity is the excess functional information
\[
\Psi(X\to Y)=\inf_{Z\perp X,\;H(Y|X,Z)=0} I(X;Z|Y)
= \inf_{Z\perp X}\big[H(Y|Z)-I(X;Y)\big].
\]
For discrete \(Y\), there is a general lower bound on \(\Psi(X\to Y)\); for \(|Y|=2\), equality holds and the Poisson functional representation attains the infimum; and there exist examples with \(I(X;Y)\) arbitrarily large for which
\[
\Psi(X\to Y)\ge \log(I(X;Y)+1)-1.
\]
Thus the \(\log(I+1)\) term is essentially tight, within approximately \(5\) bits of the SFRL upper bound [1701.02827]. A common misconception is therefore that one-shot overhead should collapse to a universal constant independent of mutual information; the examples in the paper rule that out.

Another misconception is that one-shot random coding is synonymous with unstructured i.i.d. coding. For Abelian group codes, one-shot achievability is expressed through group-constrained hypothesis-testing mutual information, with classical and classical-quantum bounds of the form
\[
\epsilon' \le \epsilon + \sum_{\hat\theta\neq \mathbf s}
2^{(1-\omega_{\hat\theta})R-I_{\mathrm H}^{\epsilon_{\hat\theta}}(\overline X_{\eta,b};[X_{\eta,b}],Y)}
\]
in the classical case, and an analogous quantum bound with an additional prefactor in the cq case [2406.13243]. The paper uses a random homomorphism ensemble and a group-averaged auxiliary distribution, showing that algebraic structure alters both achievability and converse terms through subgroup/coset penalties.

Quantum one-shot random coding bounds extend the same nonasymptotic philosophy. For a classical-quantum channel \(W:x\mapsto \rho_x\), one recent theorem proves
\[
\epsilon(\{M\},p_X)\le c_\alpha\cdot 2^{-[(1-\alpha)/\alpha]\,[I_\alpha(X;B)_\rho-\log_2(M-1)]},
\]
for \(\alpha\in[1/2,1]\), where \(c_\alpha<1.102\), \(I_\alpha(X;B)_\rho\) is the Petz–Rényi mutual information, and the decoder is an integral \(\alpha\)-PGM equivalent to a randomized Holevo–Helstrom measurement [2507.06232]. The same work recovers the optimal error exponent of classical-quantum channels for rates above the critical rate and extends to fully quantum channels, constant-composition codes, and compression with quantum side information.

Taken together, these developments suggest a broad classification of one-shot random coding bounds. One branch uses explicit random-code ensembles and stochastic or threshold decoders; a second uses representation lemmas and Poisson processes; a third uses hypothesis testing or Rényi quantities, especially in structured or quantum settings. Across these branches, the central object remains the same: a finite-blocklength achievability estimate that can be evaluated directly in the one-shot regime and that reproduces first-order and, in many cases, second-order asymptotics when specialized to memoryless models.

Source: https://www.emergentmind.com/topics/one-shot-random-coding-bound