Strong Functional Representation Lemma

• SFRL is an information-theoretic theorem stating that for any random variables X and Y with finite mutual information, there exists an auxiliary Z (independent of X) and a deterministic mapping g such that Y=g(X,Z) almost surely.
• The lemma precisely bounds the additional dependency I(X;Z|Y) to at most log(I(X;Y)+1)+4 bits, which quantifies the extra coding cost beyond the inherent mutual information.
• Its constructive proof via Poisson functional representation underpins key one-shot coding applications like exact channel simulation, variable-length lossy source coding, and multiple description coding.

The Strong Functional Representation Lemma (SFRL) is an information-theoretic representation theorem for a joint law $P_{XY}$. It states that for any random variables $X$ and $Y$ with finite mutual information, one can construct an auxiliary random variable $Z$, independent of $X$, and a deterministic mapping $g$ such that $Y=g(X,Z)$ almost surely, while the excess dependence carried by $Z$ beyond $Y$ is bounded by only a logarithmic overhead in $I(X;Y)$. In the formulation of Cheuk Ting Li and Abbas El Gamal, this overhead satisfies $X$0 bits, and the lemma serves as a unifying device for one-shot exact channel simulation, variable-length lossy source coding, multiple description coding, Gray–Wyner coding, and a simple achievability proof of the Gelfand–Pinsker theorem (Li et al., 2017).

1. Formal statement and equivalent forms

Let $X$1 be any pair of random variables, possibly discrete, continuous, or mixed, with finite mutual information $X$2. The SFRL asserts the existence of an auxiliary random variable $X$3 and a deterministic function $X$4 such that $X$5, $X$6 almost surely, and

$X$7

Equivalently,

$X$8

For discrete $X$9 and $Y$0, one may choose $Y$1 finite with

$Y$2

These are the central quantitative statements of the lemma (Li et al., 2017).

The identity

$Y$3

is structurally important. It shows that the entropy of the representation residual $Y$4 exceeds the mutual information $Y$5 by exactly the conditional excess $Y$6. The SFRL therefore strengthens the classical Functional Representation Lemma by controlling not only the existence of an independent seed but also the additional information that the seed can reveal about $Y$7 once $Y$8 is known. A later overview formulates this distinction explicitly: the classical FRL guarantees $Y$9 and $Z$0, whereas the SFRL also bounds $Z$1 (Zamani et al., 2024).

The theorem is existential but constructive. Its content is not that $Z$2 becomes independent of $Z$3, but that the stochastic dependence of $Z$4 on $Z$5 can be “functionalized” through an external source of randomness independent of $Z$6, with only a small additive penalty in the residual entropy term. This suggests why the lemma is especially useful in one-shot achievability arguments, where a stochastic encoder or channel can be replaced by a deterministic map driven by shared randomness.

2. Poisson functional representation and proof architecture

The proof in (Li et al., 2017) is based on a Poisson functional representation. One generates a marked Poisson point process over $Z$7: concretely, i.i.d. atoms $Z$8, where the marks $Z$9 are i.i.d. $X$0 and the $X$1 are ordered exponential arrival times, are collected into

$X$2

For a given $X$3, one defines

$X$4

Using the marking and displacement theorems for Poisson processes, one checks that $X$5 (Li et al., 2017).

The entropy bound is obtained by reducing $X$6 to the entropy of the index $X$7. Since $X$8 is a function of $X$9, one has $g$0. The analysis then proves

$g$1

so that averaging over $g$2 yields

$g$3

A standard maximum-entropy bound for integer-valued random variables,

$g$4

then gives

$g$5

In the discrete case, Carathéodory’s theorem is used to reduce the auxiliary alphabet to a finite one with the stated cardinality bound (Li et al., 2017).

This proof strategy is significant because it yields a representation that is uniform across general alphabets and directly compatible with coding constructions. Rather than relying on rejection-sampling arguments specialized to discrete settings, the Poisson construction produces a common randomness object whose structure can be reused across several one-shot coding theorems.

3. Coding-theoretic consequences in the original paper

The original paper develops the SFRL as a general achievability tool rather than as an isolated representation result (Li et al., 2017). Its first major application is one-shot exact channel simulation. In that problem, Alice observes $g$6, Bob must simulate $g$7, unlimited shared randomness $g$8 is available, and Alice sends a prefix-free message $g$9. Using the SFRL, one takes $Y=g(X,Z)$0 and Huffman-encodes $Y=g(X,Z)$1 conditionally on $Y=g(X,Z)$2. This gives

$Y=g(X,Z)$3

The paper states that this strengthens the results by Harsha et al. and Braverman and Garg, gives the same leading terms, extends to general alphabets, and pins down the additive constant $Y=g(X,Z)$4 (Li et al., 2017).

The same template is then applied to one-shot variable-length lossy source coding. For the rate–distortion function

$Y=g(X,Z)$5

Theorem IV.1 shows that for any distortion level $Y=g(X,Z)$6, there exists a one-shot variable-length code with

$Y=g(X,Z)$7

and average distortion at most $Y=g(X,Z)$8. The construction uses SFRL together with Carathéodory mixing between two reconstructions (Li et al., 2017).

The paper also derives achievability results for multiple description coding and lossy Gray–Wyner coding. In the multiple description setting, SFRL is applied four times, and a $Y=g(X,Z)$9-dimensional Carathéodory argument yields a time-sharing variable $Z$0 of size at most $Z$1. The resulting rates satisfy the El Gamal–Cover / Zhang–Berger inner bound up to an additive $Z$2, described as negligible in the asymptotic regime. For lossy Gray–Wyner coding, SFRL is applied to $Z$3, $Z$4, and $Z$5; a $Z$6-dimensional Carathéodory argument gives $Z$7, and the one-shot rates achieve the Gray–Wyner region plus small $Z$8 terms that vanish per dimension as block-length grows (Li et al., 2017).

A further application is the reduction of the Gelfand–Pinsker channel to a point-to-point channel. For a state $Z$9 known noncausally at the encoder,

$Y$0

Applying the SFRL to $Y$1 produces $Y$2, $Y$3, and

$Y$4

From this the paper deduces

$Y$5

and by standard point-to-point coding with block-length $Y$6, this approaches the full Gelfand–Pinsker rate (Li et al., 2017).

4. Relation to the classical FRL and interpretation of the overhead

A central conceptual point is the distinction between the classical Functional Representation Lemma and the SFRL. The FRL asks for an auxiliary variable independent of $Y$7 that makes $Y$8 a deterministic function of $Y$9. The SFRL imposes the additional constraint

$I(X;Y)$0

thereby quantifying how much “extra” information the representation variable can carry about $I(X;Y)$1 once $I(X;Y)$2 has been observed (Zamani et al., 2024).

This extra term is often the decisive quantity in coding applications. When $I(X;Y)$3 is Huffman-encoded given the shared randomness, the achievable average description length depends on $I(X;Y)$4, not just on the existence of a functional representation. The SFRL converts this entropy into the benchmark $I(X;Y)$5 plus a logarithmic penalty, which is exactly why the lemma yields one-shot rates expressed in mutual information rather than in information-density objects.

The finite-alphabet formulations also show that the representation need not require an unbounded auxiliary in the discrete setting. The original theorem gives

$I(X;Y)$6

(Li et al., 2017), while a later overview states a weaker but simpler discrete bound

$I(X;Y)$7

for FRL, SFRL, and their extended variants (Zamani et al., 2024). The difference is not substantive for existence arguments, but it clarifies that the associated optimization problems remain finite-dimensional.

A common misunderstanding is to view the SFRL as merely a repackaging of conditional simulation. The quantitative novelty is the control of the residual coding cost. In the formulation of (Li et al., 2017), the lemma provides a universal way to functionalize any joint distribution so that the additional cost is only roughly $I(X;Y)$8 bits. This suggests why the same construction repeatedly appears in one-shot achievability proofs across source, channel, and network coding problems.

5. Tighter Shannon-type bounds and discrete layered entropy

Subsequent work has refined the numerical constants in the SFRL. The paper "Discrete Layered Entropy, Conditional Compression and a Tighter Strong Functional Representation Lemma" reports a sequence of bounds of the form

$I(X;Y)$9

with decreasing constants over time, and then derives a sharper bound using discrete layered entropy $X$00 (Li, 23 Jan 2025).

The discrete layered entropy is defined for a pmf $X$01 by

$X$02

and equivalently

$X$03

The paper emphasizes that $X$04 is piecewise-linear, concave, Schur-concave, and satisfies a conditioning property that Shannon entropy lacks: $X$05 This property is what permits a sharper tracking of the auxiliary geometric coupling used in the proof (Li, 23 Jan 2025).

The layered-entropy version of the result is

$X$06

Converting back to Shannon entropy gives, in particular,

$X$07

and an alternative parameter choice yields

$X$08

The same source states that this improves on the previous best $X$09 as soon as $X$10 (Li, 23 Jan 2025).

Bound reported in (Li, 23 Jan 2025)	Form
Gohari–Anantharam (2015)	$X$11
Li–El Gamal (2021)	$X$12
Li (2024)	$X$13
New layered-entropy bound	$X$14
Alternative optimized form	$X$15

The refinement does not alter the conceptual content of the SFRL; it sharpens the entropy accounting. The paper attributes the improvement to the fact that $X$16 can collapse the conditional levels of a geometric auxiliary exactly, avoiding Jensen-type losses that arise with Shannon entropy (Li, 23 Jan 2025).

6. Extensions: privacy, robustness, and Rényi generalizations

Later work has extended the SFRL in several directions. In an overview of privacy mechanism design, semantic communication, caching, and compression, the SFRL is embedded in a broader family of functional representation results (Zamani et al., 2024). The Extended FRL permits a prescribed dependence budget $X$17, while the Extended SFRL states that with $X$18,

$X$19

These lemmas are used there to construct low-complexity privacy mechanisms and to analyze bounded-mutual-information privacy, per-letter privacy constraints, cache-aided delivery with privacy, and semantic communication with privacy (Zamani et al., 2024).

A different extension concerns uncertainty in the source marginal. "On Distributionally Robust Lossy Source Coding" develops two robust SFRLs (Serra et al., 23 Jul 2025). For a finite family $X$20, there exists a single auxiliary $X$21, independent of $X$22 under each $X$23, such that

$X$24

For a KL-sphere

$X$25

the bound becomes

$X$26

The same paper uses these extensions to obtain robust one-shot and asymptotic source coding theorems and a characterization of the robust rate–distortion function under KL uncertainty (Serra et al., 23 Jul 2025).

A further generalization replaces Shannon quantities by Rényi quantities. "On the Rényi Rate-Distortion-Perception Function and Functional Representations" establishes a Rényi-generalized SFRL in which the cost of the representation is measured by the Rényi entropy $X$27 of an index variable $X$28, controlled by Sibson’s mutual information $X$29 and its lifted form $X$30 (Wei et al., 17 Jan 2026). The paper identifies two regimes. For $X$31, the optimizing index pmf has polynomial tail

$X$32

so the codebook is effectively unbounded. For $X$33,

$X$34

which vanishes beyond $X$35, giving finite support. As $X$36, the paper states that the construction recovers the classical Shannon SFRL (Wei et al., 17 Jan 2026).

Taken together, these developments show that the SFRL is not only a lemma about exact functionalization of a channel law. It has become a general method for replacing stochastic mappings by deterministic mappings driven by independent shared randomness, while preserving quantitative control of the coding cost under Shannon, robust, and Rényi criteria.