Coset Bombe Codes Overview

• Coset Bombe codes are multilevel coset codes built on lattice partition chains that integrate non-binary polar coding and Voronoi shaping for enhanced error performance.
• They employ a layered coset structure that unifies decoding, batch retrieval, and secrecy coding under a common algebraic framework.
• Practical implementations on D4 lattices demonstrate up to 0.8 dB performance gains and reduced latency on short-to-medium blocklength channels.

Coset Bombe codes denote, in the narrow sense introduced by the lattice-coding literature, a class of multilevel coset codes built on a lattice partition chain and designed to implement polar-like coding directly on dense lattices rather than on bit-labels of QAM constellations (Bertholet et al., 7 Apr 2026). The name “Bombe” is pronounced “Bahm-buh” and was chosen in reference to Turing’s WW2 cryptanalytic machine (Bertholet et al., 7 Apr 2026). A broader interpretive usage, reflected in several technically aligned constructions, treats Coset Bombe codes as a family of coset-structured schemes in which messages, retrieval objects, or codewords are indexed by cosets, quotient levels, or subgroup orbits, while decoding exploits algebraic closure under addition, quotienting, or subgroup action. This suggests a unifying viewpoint spanning lattice modulation, batch retrieval, secrecy coding, flash/WOM rewriting, subspace coding, and multi-terminal information theory (Bertholet et al., 7 Apr 2026, Baumbaugh et al., 2020, Lu et al., 2016, Yamawaki et al., 2017, Heinlein et al., 2015).

1. Definition, scope, and unifying structure

In the formal lattice-modulation sense, a Coset Bombe code is a multilevel coset code built on a lattice partition chain

$$\Lambda_0 \supset \Lambda_1 \supset \cdots \supset \Lambda_L,$$

where each quotient $Q_i=\Lambda_i/\Lambda_{i+1}$ is one level of a multilevel coding scheme, each level is protected by a non-binary polar code over $\mathrm{GF}(2^d)$, and decoding is performed by multistage decoding with CRC-aided successive cancellation list decoding at each level (Bertholet et al., 7 Apr 2026). The construction is instantiated on dense lattices such as $D_4$, uses Voronoi shaping, and targets short and medium blocklengths (Bertholet et al., 7 Apr 2026).

A broader interpretive reading places several older coset-based constructions under the same architectural pattern. In that reading, the essential ingredients are a structured ambient object, a partition into cosets or quotient classes, and a decoder or retrieval mechanism that acts on sums, subgroup representatives, or quotient labels rather than on arbitrary codewords. This suggests that the narrow lattice definition is one member of a larger coset-structured design family rather than an isolated construction (Baumbaugh et al., 2020, Heinlein et al., 2015, Kim et al., 2013).

Domain	Representative construction	Coset role
Lattice modulation	Multilevel coset codes on $D_4$	Quotient levels $\Lambda_i/\Lambda_{i+1}$
Batch retrieval	Affine Cartesian codes with quotient-space buckets	Buckets as cosets $p+V$
Wiretap secrecy	Lattice coset codes	Message selects coset, randomness selects point
Flash/WOM	P-RIO via Hamming coset coding	Syndromes and disjoint coset leaders
Subspace coding	Coset construction in $G_q(n,k)$	Prefix–suffix coupling via coset offset
Group/orbit coding	Group coding with subgroup decoding	Minimal coset representatives

Within this unifying picture, “Bombe” emphasizes layered search or layered decoding through a structured coset space. That reading is explicit in the lattice work and is strongly suggested by related decoding frameworks based on subgroup chains, coset leaders, quotient buckets, and nested cosets (Bertholet et al., 7 Apr 2026, Kim et al., 2013, Yamawaki et al., 2017).

2. Lattice multilevel construction

The core lattice construction begins from a full-rank lattice $\Lambda \subset \mathbb{R}^d$ with generator matrix $M$, together with a partition chain $Q_i=\Lambda_i/\Lambda_{i+1}$0 (Bertholet et al., 7 Apr 2026). In the special case emphasized in the paper, $Q_i=\Lambda_i/\Lambda_{i+1}$1 and the chain is

$Q_i=\Lambda_i/\Lambda_{i+1}$2

Then each quotient

$Q_i=\Lambda_i/\Lambda_{i+1}$3

so every level has the same cardinality $Q_i=\Lambda_i/\Lambda_{i+1}$4, and the minimum distance doubles from one level to the next (Bertholet et al., 7 Apr 2026). This is the multilevel analogue of Ungerboeck set partitioning for higher-dimensional lattices (Bertholet et al., 7 Apr 2026).

For coefficient vectors $Q_i=\Lambda_i/\Lambda_{i+1}$5, the decomposition

$Q_i=\Lambda_i/\Lambda_{i+1}$6

induces a corresponding decomposition of the lattice point $Q_i=\Lambda_i/\Lambda_{i+1}$7 into quotient-level contributions (Bertholet et al., 7 Apr 2026). Each quotient level is then protected by a polar code over $Q_i=\Lambda_i/\Lambda_{i+1}$8. The construction uses the same binary polar kernel matrix as in Arıkan’s original construction, so only the additive structure is used, and the transmitter can be implemented as $Q_i=\Lambda_i/\Lambda_{i+1}$9 parallel binary polar encoders per level, while the receiver must perform joint decoding across the $\mathrm{GF}(2^d)$0 components because the lattice couples them (Bertholet et al., 7 Apr 2026).

The reported instantiation fixes $\mathrm{GF}(2^d)$1, $\mathrm{GF}(2^d)$2, and $\mathrm{GF}(2^d)$3, so there are $\mathrm{GF}(2^d)$4 levels and a constellation

$\mathrm{GF}(2^d)$5

points in $\mathrm{GF}(2^d)$6, which is throughput-equivalent to 16-QAM at 4 bits per complex channel use (Bertholet et al., 7 Apr 2026). Per level, the encoder uses a systematic polar code concatenated with a CRC, and decoding uses CA-SCL with list size 8 (Bertholet et al., 7 Apr 2026). Reliability construction is not performed by assigning rates solely from asymptotic bit-channel capacities; instead, the paper performs unified reliability ranking across all bit positions in all levels using Monte Carlo simulation of rate-0 successive cancellation over a range of SNR values (Bertholet et al., 7 Apr 2026).

This construction is explicitly intended as a practical finite-blocklength scheme. The target blocklengths are $\mathrm{GF}(2^d)$7 and $\mathrm{GF}(2^d)$8 bits, both described as relevant for 5G control channels, and the stated design goal is low-latency communication at small and medium block sizes where standard binary-coded QAM is suboptimal (Bertholet et al., 7 Apr 2026).

3. Shaping, decoding, performance, and complexity

The geometric side of the lattice construction is governed by Voronoi shaping. For a lattice $\mathrm{GF}(2^d)$9, the Voronoi region

$D_4$0

is the fundamental cell, and shaping by $D_4$1 provides average power close to that of a Gaussian of the same dimension, with up to $D_4$2 dB shaping gain in the limit of large dimension (Bertholet et al., 7 Apr 2026). In the $D_4$3 construction, the finite constellation $D_4$4 is realized by taking coset representatives of $D_4$5 in $D_4$6 and shifting $D_4$7 so that each representative lies in the Voronoi cell of $D_4$8, which is effectively Voronoi shaping with respect to $D_4$9 (Bertholet et al., 7 Apr 2026).

Over AWGN, the demapper forms a likelihood

$D_4$0

for each lattice point $D_4$1, then marginalizes this PMF to quotient-level PMFs conditioned on previously decoded levels (Bertholet et al., 7 Apr 2026). For level $D_4$2, with previously decoded partial sum $D_4$3, the decoder uses

$D_4$4

where $D_4$5 collects constellation points congruent to $D_4$6 (Bertholet et al., 7 Apr 2026). Multistage decoding proceeds serially from the least significant level upward, and path metrics are inherited across levels (Bertholet et al., 7 Apr 2026).

The reported experimental results compare the $D_4$7-based Coset Bombe code against BICM polar and MLC polar on 16-QAM. For all tested rates $D_4$8 and both blocklengths $D_4$9 and $\Lambda_i/\Lambda_{i+1}$0, the Coset Bombe code outperforms both baselines in BER and BLER (Bertholet et al., 7 Apr 2026). At $\Lambda_i/\Lambda_{i+1}$1 and $\Lambda_i/\Lambda_{i+1}$2, the paper reports about $\Lambda_i/\Lambda_{i+1}$3 dB BER gain and $\Lambda_i/\Lambda_{i+1}$4 dB BLER gain over MLC polar in the plotted region around $\Lambda_i/\Lambda_{i+1}$5–$\Lambda_i/\Lambda_{i+1}$6 BLER, with gains over BICM polar that are similar or slightly larger (Bertholet et al., 7 Apr 2026). The abstract summarizes the same trend as “up to 0.8 dB of gain” and also states that the scheme “reduces block size latency by half while maintaining superior bit and block error rate (BER/BLER) performance on codewords of 256 and 1024 bits” (Bertholet et al., 7 Apr 2026).

The complexity profile is mixed. Demodulation over $\Lambda_i/\Lambda_{i+1}$7 points costs $\Lambda_i/\Lambda_{i+1}$8 per received lattice point, and non-binary polar operations scale roughly with $\Lambda_i/\Lambda_{i+1}$9 because they manipulate $p+V$0-ary PMFs (Bertholet et al., 7 Apr 2026). The total lattice-decoding complexity is written as

$p+V$1

where $p+V$2 and $p+V$3 scale with $p+V$4, while a comparable QAM multilevel polar decoder has

$p+V$5

(Bertholet et al., 7 Apr 2026). The paper therefore emphasizes that the scheme is practical for $p+V$6 but that extension to higher-dimensional lattices such as $p+V$7 requires improved demapping and lower-complexity non-binary polar algorithms (Bertholet et al., 7 Apr 2026).

The authors explicitly note that the work is primarily experimental and does not prove full capacity-achieving properties for Coset Bombe codes (Bertholet et al., 7 Apr 2026). This suggests that the current status of the lattice construction is that of a strong finite-blocklength architecture rather than a completed asymptotic theory.

4. Coset buckets, quotient spaces, and rewriting codes

A broader storage-oriented interpretation of Coset Bombe codes appears in batch retrieval and write-once memory constructions. In the affine Cartesian setting, the code is an evaluation code

$p+V$8

over a Cartesian point set $p+V$9, and the coset structure enters through a subspace $G_q(n,k)$0 whose cosets define the buckets (Baumbaugh et al., 2020). For full Reed–Muller evaluation sets, buckets are $G_q(n,k)$1; for general Cartesian $G_q(n,k)$2, they are $G_q(n,k)$3 (Baumbaugh et al., 2020). Recovery sets are geometric fibers or lines in coordinate directions $G_q(n,k)$4, and line interpolation gives local recovery whenever $G_q(n,k)$5 (Baumbaugh et al., 2020).

The central batch-code claim is that, under a condition

$G_q(n,k)$6

the code has $G_q(n,k)$7 and can satisfy a query of size $G_q(n,k)$8 in the Reed–Muller case, or $G_q(n,k)$9 in the general affine Cartesian case, where $\Lambda \subset \mathbb{R}^d$0 counts the coordinate sets with $\Lambda \subset \mathbb{R}^d$1 (Baumbaugh et al., 2020). The diagonal subspace $\Lambda \subset \mathbb{R}^d$2 yields a particularly simple example: for $\Lambda \subset \mathbb{R}^d$3, $\Lambda \subset \mathbb{R}^d$4, and $\Lambda \subset \mathbb{R}^d$5, the Reed–Muller code on $\Lambda \subset \mathbb{R}^d$6 becomes a batch code with $\Lambda \subset \mathbb{R}^d$7, $\Lambda \subset \mathbb{R}^d$8, and $\Lambda \subset \mathbb{R}^d$9 (Baumbaugh et al., 2020). This is not the narrow lattice definition of Coset Bombe coding, but it explicitly instantiates the coset-bucket design rule that a broader Bombe-style taxonomy emphasizes (Baumbaugh et al., 2020).

In multilevel flash memory, a related coset mechanism appears in parallel random input/output codes. Using Hamming-code coset coding, a sufficient condition for a P-RIO code is the existence, for any syndrome tuple $M$0, of vectors $M$1 with

$M$2

(Yamawaki et al., 2017). The paper applies this to the $M$3 and $M$4 Hamming codes and constructs a $M$5 P-RIO code and a $M$6 P-RIO code, improving on the sequential coset-coded RIO constructions $M$7 and $M$8 from the same Hamming families (Yamawaki et al., 2017). Here the “bombe” aspect is a combinatorial one: all page data are known in advance, and encoding jointly selects disjoint low-weight coset leaders across all pages (Yamawaki et al., 2017).

These retrieval and rewriting constructions reinforce a common pattern. The code is organized around a fixed quotient structure, and decoding or access is simplified because multiple coordinates, writes, or requests can be represented by a small number of coset-level objects rather than by unrestricted tuples. This suggests that the quotient-space viewpoint is central to the broader meaning of Coset Bombe codes, even outside the lattice-modulation setting (Baumbaugh et al., 2020, Yamawaki et al., 2017).

5. Secrecy, wiretap coding, and multi-terminal structured communication

In physical-layer secrecy, coset coding is used to inject controlled randomness. For nested lattices $M$9, the transmitted point is

$Q_i=\Lambda_i/\Lambda_{i+1}$00

where the secret $Q_i=\Lambda_i/\Lambda_{i+1}$01 indexes a coset of $Q_i=\Lambda_i/\Lambda_{i+1}$02 in $Q_i=\Lambda_i/\Lambda_{i+1}$03, and $Q_i=\Lambda_i/\Lambda_{i+1}$04 is random (Lu et al., 2016). A USRP implementation with three radios—Alice, Bob, and Eve—demonstrates that secrecy depends on Eve’s position and on the amount of randomness. At Eve’s “Placement 4”, conventional encoding still yields a recognizable image, coset encoding with 1 bit of confusion increases distortion, and 2 bits of confusion drive Eve’s BER close to $Q_i=\Lambda_i/\Lambda_{i+1}$05, while Bob still reconstructs the image correctly (Lu et al., 2016). The experiments were repeated for $Q_i=\Lambda_i/\Lambda_{i+1}$06, $Q_i=\Lambda_i/\Lambda_{i+1}$07, and $Q_i=\Lambda_i/\Lambda_{i+1}$08-based lattice coset codes, with the same qualitative conclusion that coset encoding significantly raises Eve’s BER at the same SNR (Lu et al., 2016).

A complementary finite-blocklength secrecy analysis on the binary erasure wiretap channel represents a coset code by the column-type vector $Q_i=\Lambda_i/\Lambda_{i+1}$09 of the base generator matrix and then decomposes performance by subspaces of $Q_i=\Lambda_i/\Lambda_{i+1}$10 (Hunn et al., 2024). The resulting equivocation-loss and $Q_i=\Lambda_i/\Lambda_{i+1}$11-divergence formulas depend on the subspace masses

$Q_i=\Lambda_i/\Lambda_{i+1}$12

and the paper proves that codes with the all-zero column are never locally optimal, that the uniform vector fraction code is locally optimal for equivocation and globally optimal for $Q_i=\Lambda_i/\Lambda_{i+1}$13, and that first subspace exclusion codes are locally optimal for equivocation under a radius constraint and globally optimal for $Q_i=\Lambda_i/\Lambda_{i+1}$14 (Hunn et al., 2024). This suggests a subspace-aware design rule for secrecy-oriented Coset Bombe codes: populate column types so that the induced hyperplane profile is favorable to equivocation and ML ambiguity (Hunn et al., 2024).

In multi-terminal information theory, nested coset codes are used to decode sums or functions of interfering codewords instead of the individual codewords themselves. For general non-additive multi-user channels, the authors show examples in which coset codes strictly enlarge regions achievable by iid random coding in 3-user interference channels, 3-user broadcast channels, and MACs with distributed state (Padakandla et al., 2015). A parallel line of work on MACs with distributed state develops nested coset and Abelian group code frameworks in which the decoder recovers $Q_i=\Lambda_i/\Lambda_{i+1}$15 or $Q_i=\Lambda_i/\Lambda_{i+1}$16, leading to regions $Q_i=\Lambda_i/\Lambda_{i+1}$17, $Q_i=\Lambda_i/\Lambda_{i+1}$18, and $Q_i=\Lambda_i/\Lambda_{i+1}$19 that strictly contain the best known unstructured inner bounds in explicit non-additive examples (Padakandla et al., 2013). For 3-user discrete broadcast channels and 3-user interference channels, partitioned coset codes are used to decode bivariate interference components, and the derived PCC-based regions are strictly larger than Marton-style or iid-codebook regions on explicit additive and non-additive examples (Padakandla et al., 2012, Padakandla et al., 2014).

Across these secrecy and network settings, the common principle is that algebraic closure permits decoding of a compressed object—such as a sum, a subgroup label, or a coset index—whose entropy is lower than that of the full tuple of interfering or hidden variables. This is exactly the sort of reduction that the term “Coset Bombe” evokes in its broader technical usage (Padakandla et al., 2015, Padakandla et al., 2013, Padakandla et al., 2012, Padakandla et al., 2014).

6. Algebraic generalizations, related realizations, and open directions

Subspace coding provides one of the clearest algebraic generalizations. In the coset construction for constant-dimension codes, a codeword in $Q_i=\Lambda_i/\Lambda_{i+1}$20 is built from a prefix codeword $Q_i=\Lambda_i/\Lambda_{i+1}$21, a suffix codeword $Q_i=\Lambda_i/\Lambda_{i+1}$22, and a matrix $Q_i=\Lambda_i/\Lambda_{i+1}$23 via a reduced-row-echelon matrix of the form

$Q_i=\Lambda_i/\Lambda_{i+1}$24

and the resulting constant-dimension code has cardinality

$Q_i=\Lambda_i/\Lambda_{i+1}$25

(Heinlein et al., 2015). Distance is guaranteed either by subspace distances in the $Q_i=\Lambda_i/\Lambda_{i+1}$26- and $Q_i=\Lambda_i/\Lambda_{i+1}$27-parts or by the rank distance of the $Q_i=\Lambda_i/\Lambda_{i+1}$28-matrices (Heinlein et al., 2015). The construction attains the MRD bound for infinite parameter families such as

$Q_i=\Lambda_i/\Lambda_{i+1}$29

(Heinlein et al., 2015). In a broad Bombe-style reading, this is a three-level coset design: prefix, suffix, and rank-metric offset (Heinlein et al., 2015).

Group-based geometric coding offers another perspective. For a finite group $Q_i=\Lambda_i/\Lambda_{i+1}$30 of complex isometries acting on $Q_i=\Lambda_i/\Lambda_{i+1}$31, the code is the orbit $Q_i=\Lambda_i/\Lambda_{i+1}$32, and subgroup decoding uses a chain

$Q_i=\Lambda_i/\Lambda_{i+1}$33

with coset leaders chosen so that greedy decoding over the chain is correct or robust (Kim et al., 2013). For the complex reflection groups $Q_i=\Lambda_i/\Lambda_{i+1}$34, the paper gives explicit subgroup chains and greed-compatible coset leaders, yielding codes of size $Q_i=\Lambda_i/\Lambda_{i+1}$35 that can be decoded in relatively few steps (Kim et al., 2013). The same paper proves that greed compatibility is equivalent to region minimality when the initial vector has full orbit, and that subgroup decoding complexity drops from $Q_i=\Lambda_i/\Lambda_{i+1}$36 to $Q_i=\Lambda_i/\Lambda_{i+1}$37 comparisons (Kim et al., 2013).

An even more explicit algebraic embodiment appears in the representation of Best’s $Q_i=\Lambda_i/\Lambda_{i+1}$38 code as the Gray image of a $Q_i=\Lambda_i/\Lambda_{i+1}$39-code that is itself a coset $Q_i=\Lambda_i/\Lambda_{i+1}$40 of a subgroup $Q_i=\Lambda_i/\Lambda_{i+1}$41 of the unit group of the group ring $Q_i=\Lambda_i/\Lambda_{i+1}$42 (Greferath et al., 2011). One representation is

$Q_i=\Lambda_i/\Lambda_{i+1}$43

and this leads to a simple decoding algorithm based on normalization by $Q_i=\Lambda_i/\Lambda_{i+1}$44, a subgroup membership test, and a small pattern table (Greferath et al., 2011). The same methodology is applied to Julin’s $Q_i=\Lambda_i/\Lambda_{i+1}$45 code, a $Q_i=\Lambda_i/\Lambda_{i+1}$46 Hadamard code, and a $Q_i=\Lambda_i/\Lambda_{i+1}$47 code (Greferath et al., 2011). This suggests that the unit-group-coset viewpoint is a particularly concrete algebraic realization of the broader Coset Bombe concept (Greferath et al., 2011).

Related analytic tools also arise in nonlinear source-channel constructions. For overlapped arithmetic codes, the coset cardinality spectrum $Q_i=\Lambda_i/\Lambda_{i+1}$48 describes how the source space is partitioned into cosets, while the Hamming distance spectrum $Q_i=\Lambda_i/\Lambda_{i+1}$49 describes pairwise distances inside cosets (Fang, 2023). The paper proves that, in some regimes, HDS can be calculated from CCS, with a fast approximation

$Q_i=\Lambda_i/\Lambda_{i+1}$50

for $Q_i=\Lambda_i/\Lambda_{i+1}$51, where $Q_i=\Lambda_i/\Lambda_{i+1}$52 (Fang, 2023). Although this is not a coding family called Coset Bombe codes, it provides a quantitative way to analyze how coset partitioning influences internal distance structure (Fang, 2023).

Several open directions recur across the literature. In the lattice setting, the main limitations are exponential complexity in $Q_i=\Lambda_i/\Lambda_{i+1}$53, current restriction to $Q_i=\Lambda_i/\Lambda_{i+1}$54, $Q_i=\Lambda_i/\Lambda_{i+1}$55, and two levels, and the lack of full capacity-achieving theorems (Bertholet et al., 7 Apr 2026). In batch codes, batch size is limited by ambient dimension or by $Q_i=\Lambda_i/\Lambda_{i+1}$56, and the bucket subspace $Q_i=\Lambda_i/\Lambda_{i+1}$57 must satisfy strong intersection constraints (Baumbaugh et al., 2020). In subspace coding, optimal decompositions of $Q_i=\Lambda_i/\Lambda_{i+1}$58 and $Q_i=\Lambda_i/\Lambda_{i+1}$59 and the general tightness of MRD bounds remain difficult (Heinlein et al., 2015). In multi-terminal communication, the general problem is to identify which algebraic combination of interfering codewords is the right one to decode and which algebraic object—field, ring, group, lattice, or orbit—is best matched to that combination (Padakandla et al., 2015, Padakandla et al., 2013, Padakandla et al., 2012, Padakandla et al., 2014).

Taken together, these constructions support a coherent encyclopedia-level characterization. In the narrow sense, Coset Bombe codes are multilevel lattice coset codes with non-binary polar protection, Voronoi shaping, and multistage decoding on quotient lattices (Bertholet et al., 7 Apr 2026). In the broader technical sense suggested by closely related work, they are coset-organized coding schemes in which quotient structure is not merely a convenient representation but the primary mechanism for shaping, alignment, secrecy, rewriting, retrieval, or decoding (Baumbaugh et al., 2020, Lu et al., 2016, Yamawaki et al., 2017, Heinlein et al., 2015, Kim et al., 2013, Greferath et al., 2011).