Permutation Channels in Communication

• Permutation channels are models where the order of symbols is disregarded, focusing on permutation-invariant multisets rather than ordered sequences.
• The analysis leverages finite-blocklength geometry and intrinsic dimension to characterize capacity via the rank of the channel’s stochastic matrix.
• These channels extend to quantum coding, multiset error correction, and permutation-equivariant architectures for multiuser communication and time-series forecasting.

Permutation channels are models and symmetry classes in which ordering along one axis is unavailable, irrelevant, or deliberately quotiented out, so the operational object is a multiset, a type, or a permutation-invariant representation rather than an ordered sequence. In classical information theory, the canonical noisy permutation channel is a discrete memoryless channel followed by a uniformly random permutation of the output block, which makes the empirical output distribution the sufficient statistic and changes the natural rate normalization from $n$ to $\log n$ (Makur, 2020). In contemporary literature, closely related notions also appear in multiset coding with impairments, identification over permutation channels, permutation-symmetric quantum channels and codes, and channel-axis permutation symmetry in multivariate time-series forecasting (Kovačević et al., 2016, Mančinska et al., 9 Oct 2025, Xu et al., 28 Jan 2026).

1. Classical noisy permutation channels and their capacity

The standard noisy permutation channel has finite input alphabet $\mathcal X$, finite output alphabet $\mathcal Y$, a DMC $W$, and a uniform random permutation of the length-$n$ output sequence. A code consists of an encoder $f_n:\mathcal M_n\to \mathcal X^n$ and a decoder $g_n:\mathcal Y^n\to \mathcal M_n\cup\{\mathtt e\}$, but because the output order is erased, the receiver effectively observes only the multiset, equivalently the empirical distribution, of output symbols. For this reason, the appropriate normalization is $\frac{\log M^\star(n,\varepsilon)}{\log n}$, not $\frac{\log M^\star(n,\varepsilon)}{n}$, and the corresponding capacity is polynomial in blocklength rather than exponential (Makur, 2020).

Makur’s coding theorems established a general achievability lower bound $\log n$0, together with converse bounds for strictly positive channels and an exact characterization for strictly positive, full-rank DMCs (Makur, 2020). The capacity of strictly positive noisy permutation channels was then identified exactly as

$\log n$1

showing that the decisive parameter is the rank of the stochastic matrix, equivalently the affine dimension of the reachable set of output distributions, rather than the detailed numerical values of the transition probabilities (Tang et al., 2021).

This dependence on rank rather than conventional Shannon-theoretic noise parameters is one of the defining peculiarities of the model. For example, the noisy permutation capacity of a binary symmetric channel is $\log n$2 for every nontrivial crossover probability and $\log n$3 only at the completely random point, while the identity channel, which is not strictly positive, has capacity $\log n$4 over a $\log n$5-symbol alphabet because the output type is preserved exactly (Makur, 2020, Makur, 2020).

2. Finite-blocklength geometry and intrinsic dimension

Finite-blocklength analysis makes the geometry explicit. For a DMC $\log n$6 with input alphabet size $\log n$7 and output alphabet size $\log n$8, the reachable output polytope is

$\log n$9

and its affine dimension is

$\mathcal X$0

The analysis is carried out intrinsically on the affine hull $\mathcal X$1, not on the full output simplex, and introduces a minimum-volume $\mathcal X$2-simplex $\mathcal X$3, its coordinate preimage $\mathcal X$4, and the relative volume ratio

$\mathcal X$5

Achievability is obtained by placing messages on a simplex lattice in affine coordinates with resolution $\mathcal X$6, decoding by projecting the empirical output distribution onto $\mathcal X$7, and then applying Euclidean nearest-neighbor decoding. A Voronoi-transfer reduction converts any decoding error into one of $\mathcal X$8 one-dimensional transfer events, which leads to a refined Gaussian achievability bound expressed through averaged local coordinate variances $\mathcal X$9 and $\mathcal Y$0. On the converse side, a modified meta-converse, KL-divergence covering of $\mathcal Y$1, and a local binary-testing lemma yield an upper bound whose blocklength-dependent term is $\mathcal Y$2, up to bounded additive terms (Feng et al., 25 May 2026).

The resulting asymptotic statement is

$\mathcal Y$3

and hence

$\mathcal Y$4

for every fixed $\mathcal Y$5, with a strong converse: any rate strictly larger than $\mathcal Y$6 in $\mathcal Y$7 units drives the average error probability to $\mathcal Y$8 (Feng et al., 25 May 2026). This sharpens earlier weak-converse capacity results associated in the paper with Tang and Polyanskiy by showing that the relevant finite-blocklength term is governed by the intrinsic dimension $\mathcal Y$9, not the ambient simplex dimension $W$0 (Feng et al., 25 May 2026).

The lower-dimensional viewpoint is especially important when $W$1 has redundant outputs. A binary symmetric channel gives $W$2, so $W$3, while a lower-dimensional $W$4 example with fixed fourth output coordinate gives $W$5 and a two-dimensional Gaussian approximation that closely matches simulated lattice-code performance (Feng et al., 25 May 2026).

3. Multiset coding and permutation channels with impairments

A broader coding-theoretic formulation replaces ordered sequences by multisets. Over an alphabet $W$6, a multiset $W$7 of size $W$8 is represented by its multiplicity vector

$W$9

so the code space is the discrete simplex

$n$0

Insertions, deletions, substitutions, and erasures act directly on multiplicity vectors, and for fixed-length codes these error types are equivalent in the worst-case sense: correcting $n$1 insertions, $n$2 deletions, and $n$3 substitutions is equivalent to correcting $n$4 insertions or $n$5 deletions. The natural metric is

$n$6

and a multiset code corrects $n$7 deletions iff its minimum $n$8-distance exceeds $n$9 (Kovačević et al., 2016).

The central construction uses Sidon sets. Given a finite Abelian group $f_n:\mathcal M_n\to \mathcal X^n$0 and a $f_n:\mathcal M_n\to \mathcal X^n$1 set $f_n:\mathcal M_n\to \mathcal X^n$2, one defines

$f_n:\mathcal M_n\to \mathcal X^n$3

If $f_n:\mathcal M_n\to \mathcal X^n$4 is a Sidon set of order $f_n:\mathcal M_n\to \mathcal X^n$5, then $f_n:\mathcal M_n\to \mathcal X^n$6 corrects $f_n:\mathcal M_n\to \mathcal X^n$7 deletions, and the resulting family is asymptotically optimal in redundancy scaling for any error radius and any alphabet size. The same work also gives indexing-based constructions, in which packets are tagged by sequence numbers and then protected by conventional Hamming-space coding, and polynomial-root constructions over finite fields, but shows that the Sidon-set construction is the asymptotically optimal one for linear multiset codes and, in several parameter regimes, optimal in the stronger sense of maximal code cardinality (Kovačević et al., 2016).

This multiset viewpoint clarifies a common ambiguity. A noisy permutation channel in the DMC sense destroys order after symbol corruption; a multiset channel treats order as absent from the outset and admits insertions, deletions, substitutions, and erasures at the multiset level. The two viewpoints coincide when the receiver’s invariant statistic is the multiplicity vector (Kovačević et al., 2016).

4. Identification and multiuser extensions

Permutation symmetry produces markedly different behavior for identification. For the $f_n:\mathcal M_n\to \mathcal X^n$8-ary uniform permutation channel, where the output is a uniformly random permutation of the input vector and thus uniform on the type class of the transmitted sequence, the number of identifiable messages can grow as

$f_n:\mathcal M_n\to \mathcal X^n$9

for any $g_n:\mathcal Y^n\to \mathcal M_n\cup\{\mathtt e\}$0. At the same time, there is no identification code with message size $g_n:\mathcal Y^n\to \mathcal M_n\cup\{\mathtt e\}$1 and both type-I and type-II error probabilities decaying as $g_n:\mathcal Y^n\to \mathcal M_n\cup\{\mathtt e\}$2 for any fixed $g_n:\mathcal Y^n\to \mathcal M_n\cup\{\mathtt e\}$3 and $g_n:\mathcal Y^n\to \mathcal M_n\cup\{\mathtt e\}$4, and if the exponent $g_n:\mathcal Y^n\to \mathcal M_n\cup\{\mathtt e\}$5 tends to infinity then the sum of type-I and type-II errors approaches at least $g_n:\mathcal Y^n\to \mathcal M_n\cup\{\mathtt e\}$6. The no-feedback identification capacity is therefore zero in the natural $g_n:\mathcal Y^n\to \mathcal M_n\cup\{\mathtt e\}$7 normalization, even though identification remains strictly richer than ordinary communication. With causal block-wise feedback, however, the maximum number of identifiable messages becomes doubly exponential and the identification capacity equals $g_n:\mathcal Y^n\to \mathcal M_n\cup\{\mathtt e\}$8 in units of $g_n:\mathcal Y^n\to \mathcal M_n\cup\{\mathtt e\}$9 (Sarkar et al., 2024).

Permutation symmetry also admits a nontrivial multiple-access theory. In the permutation adder multiple-access channel (PAMAC), $\frac{\log M^\star(n,\varepsilon)}{\log n}$0 users send $\frac{\log M^\star(n,\varepsilon)}{\log n}$1-ary codewords through an integer adder MAC, the sum passes through a strictly positive invertible DMC, and the output block is then uniformly permuted. Rates are normalized by $\frac{\log M^\star(n,\varepsilon)}{\log n}$2, and the permutation capacity region is exactly the simplex

$\frac{\log M^\star(n,\varepsilon)}{\log n}$3

Achievability is based on i.i.d. categorical coding together with a permutation-adapted time-sharing construction whose combinatorial core is a mixed-radix representation of users’ message digits; converse bounds match achievability (Lu et al., 2023).

These results indicate that permutation channels are not merely point-to-point curiosities. They support distinct identification asymptotics, admit a genuine multiuser capacity region, and require coding schemes built from types, histograms, and permutation-compatible algebra rather than ordered-symbol typicality (Sarkar et al., 2024, Lu et al., 2023).

5. Quantum permutation-symmetric channels and codes

In quantum information, permutation symmetry arises because i.i.d. channels $\frac{\log M^\star(n,\varepsilon)}{\log n}$4 commute with the action of the symmetric group on tensor factors. If the input state is permutation-invariant, then the output state is permutation-invariant as well, and Schur–Weyl duality yields a block decomposition

$\frac{\log M^\star(n,\varepsilon)}{\log n}$5

This permits coherent-information calculations for permutation-invariant codes of the form

$\frac{\log M^\star(n,\varepsilon)}{\log n}$6

including non-orthogonal repetition-like codes

$\frac{\log M^\star(n,\varepsilon)}{\log n}$7

Using the representation theory of the symmetric and general linear groups, coherent information can be computed blockwise for at least $\frac{\log M^\star(n,\varepsilon)}{\log n}$8 channel copies in the qubit case, and this yields improved lower bounds on quantum capacities for general Pauli channels, the dephrasure channel, the generalized amplitude damping channel, and the damping-dephasing channel. In particular, the paper reports significant threshold improvements for the 2-Pauli and BB84 channel families over the earlier bounds of Fern and Whaley, and attributes those gains to non-orthogonal repetition-like permutation-invariant codes (Bhalerao et al., 13 Aug 2025).

A complementary line of work classifies the channels themselves. Unitary-equivariant and permutation-invariant quantum channels

$\frac{\log M^\star(n,\varepsilon)}{\log n}$9

are classified by extremal points, and every extremal channel factors operationally as

$\frac{\log M^\star(n,\varepsilon)}{n}$0

This classification leads to a streaming implementation ansatz with polynomial-time algorithms and exponential memory improvements for state symmetrization, symmetric cloning, and purity amplification; for symmetric cloning, it yields what the paper describes as the first efficient polynomial-time algorithm with explicit memory and gate bounds (Mančinska et al., 9 Oct 2025).

Permutation can also appear as an admissible encoding resource. For classical communication over a noisy quantum channel with local operations and a shared bipartite state $\frac{\log M^\star(n,\varepsilon)}{n}$1, if Alice is allowed local encodings followed by a global permutation of the channel inputs, then for pure assistance $\frac{\log M^\star(n,\varepsilon)}{n}$2 the permutation-assisted capacity satisfies

$\frac{\log M^\star(n,\varepsilon)}{n}$3

interpolating between the Holevo capacity and the entanglement-assisted capacity depending on the preshared state. The same work proves a strong converse and shows that the increase above the Holevo capacity is upper bounded by the discord of formation of the preshared state (Wang et al., 2020).

6. Permutation symmetry as a computational and learning principle

Permutation symmetry is also used as a computational reduction principle. In optimization problems over $\frac{\log M^\star(n,\varepsilon)}{n}$4 parallel uses of a quantum channel, Schur–Weyl duality compresses semidefinite programs from exponential size in $\frac{\log M^\star(n,\varepsilon)}{n}$5 to polynomial size by working entirely inside the permutation-invariant subspace. The resulting framework supports partial channel application, partial traces, partial transpose, serial composition, and relative-entropy computations in the reduced representation, and yields a symmetric seesaw method for lower-bounding channel fidelity over $\frac{\log M^\star(n,\varepsilon)}{n}$6 uses. The method improves lower bounds for the depolarizing and amplitude-damping channels in the regime of tens of channel uses and was used to demonstrate non-asymptotic superactivation of quantum capacity for $\frac{\log M^\star(n,\varepsilon)}{n}$7 (Bergh et al., 29 Apr 2026).

A distinct but related use of the term appears in multivariate time-series forecasting, where the “channels” are variables or sensors rather than communication links. CPiRi treats the channel dimension as a permutation-symmetric object: with $\frac{\log M^\star(n,\varepsilon)}{n}$8 and $\frac{\log M^\star(n,\varepsilon)}{n}$9, the forecasting task is recast as permutation-equivariant in the channel index, and the architecture combines a frozen channel-wise temporal encoder, a one-layer Transformer spatial module that is permutation-equivariant over channel embeddings, and a frozen channel-wise decoder. Training samples a random channel permutation $\log n$00 at each iteration and minimizes

$\log n$01

thereby enforcing content-driven rather than index-driven inter-channel reasoning. The paper reports that many conventional channel-dependent forecasting models collapse under test-time channel shuffling, whereas CPiRi remains essentially unchanged under full or partial permutations, generalizes to unseen channels, and scales to datasets with up to $\log n$02 channels with $\log n$03 complexity (Xu et al., 28 Jan 2026).

Taken together, these developments show that permutation channels are not a single model but a family of mathematically related constructions. In communication and storage they formalize loss of ordering and force coding over types or multisets; in quantum information they identify the symmetry class naturally induced by tensor-power channels and multi-copy codes; and in learning systems they motivate permutation-equivariant architectures when channel indices are exchangeable rather than semantically ordered (Tang et al., 2021, Mančinska et al., 9 Oct 2025, Xu et al., 28 Jan 2026).