Subspace Decoding Task in Communication

• Subspace decoding task is the process of inferring a transmitted subspace from corrupted or mixed observations using linear algebraic constraints.
• It encompasses diverse models such as noncoherent network coding, quantum state discrimination, and analog subspace codes, each leveraging distinct metrics and decoding techniques.
• Advanced strategies like list decoding and layered decoding enable recovery beyond traditional error limits, with significant implications for robust statistics and sensor network applications.

Subspace decoding task denotes a family of decoding problems in which the latent object is a subspace, or a codeword or state constrained by a linear subspace, and the decoder must infer it from a corrupted, mixed, or indirect observation. In noncoherent network coding, this means recovering a transmitted subspace from a received subspace under insertions and deletions. In structured classical-quantum communication, it means minimum-error discrimination among product states indexed by binary strings constrained to lie in a linear subspace. Closely related formulations also appear in one-dimensional analog subspace codes, blind code identification, sparse-array sensing, and robust statistics (Mahdavifar et al., 2012, Piveteau et al., 23 Sep 2025, Riasat et al., 2024, Singh et al., 22 Jan 2026, Gooty et al., 4 Dec 2025, Raghavendra et al., 2020).

1. Core formulations and observation models

The classical formulation arises in the Koetter–Kschischang model of noncoherent network coding. A message is encoded as a subspace $V$ of an ambient vector space, and the receiver observes a subspace $U$ that differs from $V$ because transmitted dimensions may be lost and extraneous dimensions may be injected. This formulation is standard across subspace codes, folded subspace codes, interleaved subspace codes, and layered subspace codes (Mahdavifar et al., 2012, Bartz et al., 2015, Bartz et al., 2014, Chen et al., 2012).

The quantum formulation replaces a received vector space by a received quantum state. For a binary linear code $C \subseteq \mathbb{F}_2^n$ and a binary-input classical-quantum pure-state channel $x \mapsto \ket{\phi_x}$, each codeword $c=(c_1,\dots,c_n)$ is mapped to the product state

$\ket{\phi_c}=\bigotimes_{i=1}^n \ket{\phi_{c_i}}.$

The decoding objective is minimum-error discrimination among the ensemble of codeword states indexed by the subspace $C$, typically with uniform prior (Piveteau et al., 23 Sep 2025).

Other formulations preserve the same structural pattern. In blind code identification on the BSC, the observation is a noisy matrix $Y=AX+E$, and the task is to identify which code in a known family generated the rows (Singh et al., 22 Jan 2026). In sensing subspace codes for direction-of-arrival estimation, a received vector

$\mathbf y=\mathbf h(\theta)x+\mathbf w$

must be matched to a codeword $U$0 determined by the sparse-array geometry (Gooty et al., 4 Dec 2025). In analog subspace coding, the transmitted object is a one-dimensional subspace codeword over complex numbers rather than over a finite field (Riasat et al., 2024).

Setting	Observation model	Decoding objective
Noncoherent network coding	$U$1	Recover transmitted subspace $U$2
Structured cq decoding	$U$3, $U$4	Discriminate among $U$5
Blind code identification	$U$6	Identify the unknown code $U$7
Sensing subspace codes	$U$8	Estimate the DoA or codeword index

A common feature across these settings is that the relevant uncertainty is not attached to an individual symbol alone. What matters is the global linear constraint, the row space, or the codeword/state family induced by a subspace. This suggests a unifying viewpoint in which subspace decoding is a structured inference problem over linear-algebraic constraint sets.

2. Metrics, channels, and unique-decoding criteria

The basic metric in classical subspace coding is the subspace distance

$U$9

Under the operator-channel model $V$0, erasures are

$V$1

insertions are

$V$2

and

$V$3

A subspace code $V$4 with minimum distance $V$5 can correct any pattern satisfying

$V$6

using minimum-distance decoding (Chen et al., 2012, Bartz et al., 2014).

This same distance governs many algebraic decoders. Folded subspace codes and interleaved subspace codes are both decoded under the operator channel, with $V$7 deletions and $V$8 insertions, and both aim to go beyond the traditional half-distance unique-decoding regime (Bartz et al., 2015, Bartz et al., 2014). In the analog setting, the relevant metric is again subspace-based: $V$9 and minimum-distance decoding succeeds if

$C \subseteq \mathbb{F}_2^n$0

This places analog subspace codes within the same formal template even though the ambient field is $C \subseteq \mathbb{F}_2^n$1 rather than $C \subseteq \mathbb{F}_2^n$2 (Riasat et al., 2024).

A substantial refinement is provided by the atomic framework. In the lattice $C \subseteq \mathbb{F}_2^n$3 of subspaces, the atoms are the one-dimensional subspaces. For a subspace $C \subseteq \mathbb{F}_2^n$4, a minimal atomic decomposition is an irredundant set of atoms whose sum is $C \subseteq \mathbb{F}_2^n$5, and the counting function

$C \subseteq \mathbb{F}_2^n$6

depends only on $C \subseteq \mathbb{F}_2^n$7, is isotone, and satisfies supermodularity and log-supermodularity. This leads to the $C \subseteq \mathbb{F}_2^n$8-induced distance

$C \subseteq \mathbb{F}_2^n$9

which is a genuine metric on $x \mapsto \ket{\phi_x}$0. The associated Atomic Operator Channel models corruption directly as atomic insertions and erasures, and minimum-distance decoding is guaranteed whenever

$x \mapsto \ket{\phi_x}$1

In the constant-dimension case, the classical unique-decodability condition under the subspace distance remains sufficient for unique decoding under the atomic metric (Ramirez et al., 22 Mar 2026).

3. Algebraic list decoding and decoding beyond the unique radius

A central theme in the literature is the replacement of unique decoding by list decoding or probabilistic unique decoding. In “Algebraic List-decoding of Subspace Codes” (Mahdavifar et al., 2012), the Koetter–Kschischang construction is modified so that message polynomials lie in a commutative subring of linearized polynomials, and the transmitted tuple is enlarged to

$x \mapsto \ket{\phi_x}$2

The resulting list-$x \mapsto \ket{\phi_x}$3 decoder guarantees recovery when the normalized dimension of the error is at most

$x \mapsto \ket{\phi_x}$4

which exceeds the previously best known error-correction radius $x \mapsto \ket{\phi_x}$5 for low rates.

“List decoding subspace codes from insertions and deletions” (Guruswami et al., 2012) gives the first list-decoding algorithm for subspace codes that tolerates deletions. Its folded linearized Reed-Solomon subspace code can be list decoded in polynomial time whenever

$x \mapsto \ket{\phi_x}$6

equivalently

$x \mapsto \ket{\phi_x}$7

and the output is contained in an affine subspace of dimension at most $x \mapsto \ket{\phi_x}$8, so the list size is at most $x \mapsto \ket{\phi_x}$9.

A related linear-algebraic construction appears in “List-decoding of Subspace Codes and Rank-Metric Codes up to Singleton Bound” (Mahdavifar et al., 2012). There the decoder interpolates a nonzero multivariate linearized polynomial

$c=(c_1,\dots,c_n)$0

on a basis of the received subspace and then solves a linearized factorization problem. It succeeds whenever

$c=(c_1,\dots,c_n)$1

and the output list has size at most $c=(c_1,\dots,c_n)$2. In the paper’s summary, the achievable normalized radius is expressed as $c=(c_1,\dots,c_n)$3.

Folded and interleaved variants sharpen this picture. “List and Probabilistic Unique Decoding of Folded Subspace Codes” (Bartz et al., 2015) introduces an interpolation-based decoder for $c=(c_1,\dots,c_n)$4-folded subspace codes that corrects insertions and deletions whenever

$c=(c_1,\dots,c_n)$5

with normalized radius

$c=(c_1,\dots,c_n)$6

under the stated design approximation. The same machinery can act as a list decoder or as a probabilistic unique decoder. “Efficient Interpolation-Based Decoding of Interleaved Subspace and Gabidulin Codes” (Bartz et al., 2014) decodes interleaved subspace codes whenever

$c=(c_1,\dots,c_n)$7

supports both list decoding and probabilistic unique decoding beyond half the minimum subspace distance, and gives efficient interpolation and root finding with complexities

$c=(c_1,\dots,c_n)$8

respectively.

More recent work extends the algebraic framework to sum-rank analogues. “Explicit List-Decodable Linearized Reed-Solomon Subspace Codes via Subspace Designs” (Shang et al., 5 Feb 2026) restricts message coefficients to a product of subspaces from an explicit subspace design. For the resulting LRS subcodes, list decoding is possible up to

$c=(c_1,\dots,c_n)$9

with rate at least $\ket{\phi_c}=\bigotimes_{i=1}^n \ket{\phi_{c_i}}.$0 in the paper’s parameterization and list size bounded by $\ket{\phi_c}=\bigotimes_{i=1}^n \ket{\phi_{c_i}}.$1. The folded LRS extension produces explicit positive-rate FLRS subcodes that are efficiently list decodable beyond the unique-decoding radius.

4. Structural redundancy, layering, and low-complexity decoding

One way to simplify subspace decoding is to impose explicit structure on the codeword. In layered subspace codes, the overall codeword is a direct sum

$\ket{\phi_c}=\bigotimes_{i=1}^n \ket{\phi_{c_i}}.$2

of component subspace codewords. This yields two decoders. Algorithm I extracts the layer-restricted received spaces $\ket{\phi_c}=\bigotimes_{i=1}^n \ket{\phi_{c_i}}.$3 and decodes each component independently. Algorithm II is SIC-like: it decodes one layer, adds the recovered layer back to the received space, and proceeds iteratively. Both are guaranteed whenever

$\ket{\phi_c}=\bigotimes_{i=1}^n \ket{\phi_{c_i}}.$4

and both can sometimes decode beyond the nominal capability of the overall code, depending on the error geometry across layers (Chen et al., 2012).

The atomic framework adds a different kind of structure. A minimal atomic decomposition of a subspace is characterized by the property that every choice of nonzero generators from the involved atoms yields a basis of the subspace. This identifies minimal atomic decompositions with bases up to scalar equivalence, turns the number of such decompositions into a combinatorial invariant $\ket{\phi_c}=\bigotimes_{i=1}^n \ket{\phi_{c_i}}.$5, and leads to a decoding rule that measures atomic erasure cost, atomic insertion cost, and total atomic distortion directly in the $\ket{\phi_c}=\bigotimes_{i=1}^n \ket{\phi_{c_i}}.$6 metric. A plausible implication is that two subspaces with the same dimensions may exhibit different robustness once their atomic-decomposition structure is taken into account (Ramirez et al., 22 Mar 2026).

Subspace designs provide another structured route to low-complexity decoding. “Majority-logic Decoding with Subspace Designs” (Cruz et al., 2019) shows that codes derived from subspace designs have the same majority-logic decoding capability as codes from geometric designs, while often requiring far fewer parity-check equations. For one-step majority-logic decoding, the number of correctable errors is

$\ket{\phi_c}=\bigotimes_{i=1}^n \ket{\phi_{c_i}}.$7

For the two-step decoder, the correction radius becomes

$\ket{\phi_c}=\bigotimes_{i=1}^n \ket{\phi_{c_i}}.$8

matching Reed’s multistep majority-logic decoder and the Peterson–Weldon two-step decoder. Here the main gain is not a larger radius but a smaller repetition number $\ket{\phi_c}=\bigotimes_{i=1}^n \ket{\phi_{c_i}}.$9, and hence lower decoding complexity.

5. Quantum and analog formulations

In the quantum setting, the subspace decoding task is a structured state-discrimination problem. The receiver is given an ensemble

$C$0

of product states indexed by a linear subspace $C$1, and must perform minimum-error decoding. “Efficient and optimal quantum state discrimination via quantum belief propagation” (Piveteau et al., 23 Sep 2025) shows that when the code has an exact tree or trellis representation, quantum belief propagation computes the required quantum marginals efficiently and achieves optimal minimum-error discrimination: $C$2 The same framework yields efficient optimal decoders for all classical codes with efficient trellis representations, and for turbo codes the reported thresholds surpass the Shannon bound and closely approach the Holevo bound.

Analog subspace codes provide a different generalization. Character-polynomial codes are one-dimensional analog subspace codes over complex numbers. If the character map is ignored, the code becomes a subcode of a Reed–Solomon or generalized Reed–Solomon code in which certain message coefficients are forced to zero. The minimum-distance decoder first maps each received coordinate to the nearest point on the complex unit-circle alphabet, then inverts the character and applies an RS/GRS decoder. The Guruswami–Sudan list decoder can also be reused, and the paper shows that the CP list decoder is

$C$3

times less likely to return more than one codeword than the GS decoder for the ambient RS/GRS code of the same length and degree (Riasat et al., 2024).

A further quantum variant is post-processing by subspace expansions. “Decoding quantum errors with subspace expansions” (McClean et al., 2019) constructs exact or approximate projectors

$C$4

from stabilizers or related symmetry operators and uses classical post-processing rather than online syndrome extraction. On the perfect $C$5 code, the reported pseudo-threshold is

$C$6

under a single qubit depolarizing channel applied to all qubits.

6. Applications beyond classical transmission

Subspace decoding ideas have been adapted to problems in which the transmitted object is not itself a subspace codeword in the classical operator-channel sense. In blind channel-code identification on the BSC, the minimum denoised subspace discrepancy decoder first performs candidate-dependent bounded-distance Hamming denoising of each received row and then measures

$C$7

Under bounded-weight row errors and sufficient rank or uniqueness conditions, the denoised subspace discrepancy of the true code is strictly smaller than that of every wrong code. Simulations for random linear codes show improved performance beyond existing general-purpose techniques across most channel conditions and even with a limited number of received vectors (Singh et al., 22 Jan 2026).

In sparse-array sensing, the codebook itself is induced by array geometry. The Bose–Chowla sensing subspace code is

$C$8

with $C$9. The baseline MAP decoder has complexity $Y=AX+E$0. The proposed window decoder and geometry-based decoders reduce effective complexity to quadratic, and the geometric-reduced MAP decoder uses a pruned candidate set followed by MAP refinement. Monte Carlo simulations show performance that smoothly approaches the MAP performance as the complexity grows from quadratic to cubic in the number of antennas (Gooty et al., 4 Dec 2025).

In routing-like interference channels with cooperative destinations, packets can be treated as generators of a subspace codeword rather than as isolated packets. In the reported simulations, RLNC and RLNC with subspace coding both recover the intended source data at all destinations in the error-free case for all tested values of the interference probability $Y=AX+E$1. With errors fixed to $Y=AX+E$2, subspace coding performs best, and errors are less likely to affect decoding as long as they are under about $Y=AX+E$3 random errors in the data matrix (Brahimi et al., 2022).

Robust statistics supplies an inference-theoretic analogue. In list-decodable subspace recovery, an $Y=AX+E$4 fraction of the data lies near an unknown $Y=AX+E$5-dimensional subspace and the remaining points are arbitrary. A polynomial-time sum-of-squares algorithm outputs a list of

$Y=AX+E$6

candidate projection matrices, one of which is nontrivially correlated with the planted subspace (Raghavendra et al., 2020). This suggests that the phrase “subspace decoding task” can also describe robust recovery of low-dimensional structure from heavily corrupted observations, even when the ambient problem is not a communication channel in the usual sense.