Error-Building Decoding (EBD) Overview

• Error-Building Decoding (EBD) is a dual framework that applies a soft-decision ML decoding method for linear block codes and a biologically plausible error broadcast mechanism in neural networks.
• For linear block codes, EBD employs a recursive, syndrome-driven block construction with dynamic penalty minimization and exclusion strategies, ensuring maximum-likelihood performance.
• In neural network learning, EBD replaces traditional backpropagation with error decorrelation and local three-factor updates, achieving competitive accuracy on benchmarks like MNIST and CIFAR-10.

Error-Building Decoding (EBD) encompasses two distinct frameworks within contemporary coding theory and neural network learning research. In the context of linear block codes, EBD refers to a soft-decision maximum-likelihood (ML) decoding algorithm utilizing only the code’s parity-check matrix and a recursive construction to efficiently find optimal error patterns. In neural network optimization, EBD denotes an error broadcast and decorrelation method for credit-assignment in deep architectures, based on the orthogonality principles of MMSE estimation. Both areas are grounded in rigorous algebraic and statistical principles, offering principled alternatives to traditional approaches.

1. EBD for Linear Block Codes: Definition and Core Principles

Error-Building Decoding is a ML soft-decision decoding framework for binary linear block codes defined via their parity-check matrix $\mathbf H \in \mathbb F_2^{Q\times N}$, with $Q = N - K$ for a code of blocklength $N$ and dimension $K$.

Error-Building Block:

A $t$-error-building block is a multiset of $t$ column indices $\{a_0,a_1,\dots,a_{t-1}\}$ where each $a_i \in \{0,\dots,N-1\}$, such that the mod-2 sum $\bigoplus_{i=0}^{t-1} \mathbf h_{a_i} = \mathbf v$ for some $\mathbf v \in \mathbb F_2^Q$. The set of all such $t$-blocks for $\mathbf v$ is denoted $\mathcal B_t(\mathbf v)$; the union over $t=1,\dots,N$ gives $\mathcal B(\mathbf v)$.

Block Penalty:

Given a received LLR vector $\boldsymbol{\lambda}$, the penalty associated with a block $\{a_0,\dots,a_{t-1}\}$ is $$M(\{a_0,\dots,a_{t-1}\}) = \sum_{i=0}^{t-1} |\lambda[a_i]|$$.

Optimality:

• The locally optimal block $\mathcal O_t(\mathbf v)$ minimizes penalty within $\mathcal B_t(\mathbf v)$.
• The globally optimal block $\mathcal O(\mathbf v)$ is the overall minimum-penalty block in $\mathcal B(\mathbf v)$, with block size $t^* \leq Q$ due to linear dependence constraints (Qiu et al., 5 Jan 2026).

This builds the foundation for a recursive, syndrome-driven search for the ML error pattern.

2. Recursive Construction Theorem and Algorithmic Structure

Central to EBD is a recursive block construction:

Theorem 1 (Block Size Bound):

For any $\mathbf v \neq \mathbf 0$, the globally optimal block $\mathcal O(\mathbf v)$ has size $t \le Q$.

Theorem 2 (Recursive Construction):

Larger optimal blocks can be built from the multiset union of smaller, locally optimal blocks. Specifically, $$\mathcal O_{t_1+t_2}(\mathbf v) = \mathcal O_{t_1}(\mathbf v_b) \uplus \mathcal O_{t_2}(\mathbf v_b \oplus \mathbf v)$$, where $\mathbf v_b$ is chosen to minimize the combined penalty over all $\mathbf x \in \mathbb F_2^Q$.

Algorithmic Steps:

1. Compute syndrome from hard-decision on LLRs.
2. Precompute locally optimal 1-blocks for all vectors in syndrome space.
3. Recursively build locally optimal $t$-blocks for $2 \leq t \leq \lceil Q/2\rceil$.
4. Complete construction for $t > \lceil Q/2\rceil$ only for the observed syndrome.
5. Return the closest codeword by correcting flip indices corresponding to the globally optimal block.

Through this approach, syndrome decoding is performed without trellis representations or error-pattern lists, leveraging only the parity-check matrix (Qiu et al., 5 Jan 2026).

3. Customization and Complexity Reduction

The EBD framework exploits code-specific structure through exclusion mechanisms:

Offline Exclusion:

For extended Hamming codes with all-1’s parity row, only block sizes matching syndrome parity can appear. This removes entire subspaces of infeasible blocks from consideration.

Online Exclusion:

During decoding, blocks with penalty exceeding the current best for the target syndrome $s$ are pruned dynamically, as they cannot yield a minimum-penalty global solution.

These strategies yield substantial complexity reduction while provably preserving ML performance for codes where such algebraic manipulations apply. Appendix A in (Qiu et al., 5 Jan 2026) establishes that minimum-penalty blocks found using these exclusions are identical to those found in the unrestricted search.

4. Complexity and Performance Analysis

EBD’s core complexity arises from computation across $2^Q$ syndrome vectors and recursive block construction up to $t=\lceil Q / 2 \rceil$. For general codes, the operation count is upper-bounded by $$\lceil \frac Q2-1\rceil \cdot 2^{2Q+1} + \lfloor \frac Q2 \rfloor \cdot 2^{Q+1}$$.

Optimizations for extended Hamming codes yield the following FLOP counts:

Code (N,K,d)	Trellis Viterbi	EBD-OffOpt (s[0]=0)	EBD-OffOpt (s[0]=1)
(64,57,4)	7,531	7,937	16,065
(128,120,4)	31,435	32,383	64,897
(256,247,4)	128,395	130,303	261,885

For EBD-FullOpt (both offline and online exclusion), the average FLOP at FER $= 10^{-3}$ is 839, 3,231, and 13,213 for the above codes, respectively—representing $\approx$10\% of the Trellis-Viterbi cost (Qiu et al., 5 Jan 2026).

Performance:

EBD achieves ML decoding performance: its FER curves are indistinguishable from trellis-based ML decoders and consistently outperform Chase-II. At FER $=10^{-3}$, the EBD ML threshold is about 0.2 dB better than Chase-II for the tested codes, with no loss of optimality from either exclusion mechanism.

5. EBD in Neural Network Learning: Error Broadcast and Decorrelation

A second, unrelated contemporary use of the EBD acronym arises in neural network learning, denoting Error Broadcast and Decorrelation (Erdogan et al., 15 Apr 2025). Here, EBD refers to a biologically plausible credit-assignment mechanism that eschews weight symmetry, replacing conventional backpropagation with global error broadcasts and a decorrelation penalty.

Key Principle:

EBD is anchored in the stochastic orthogonality property for MMSE estimators: for any measurable transform $g(\mathbf x)$, the estimation error is orthogonal to $g(\mathbf x)$ in expectation. In deep networks, this motivates penalizing the correlation between each layer's activation $g^{(k)}(\mathbf h^{(k)})$ and the output error.

Training Mechanism:

• For each layer, maintain an exponentially averaged error-activation correlation matrix.
• Backpropagate only a decorrelation-driven local error, forming the update

$$\Delta W_{ij}^{(k)} \propto (\text{presynaptic})\times(\text{postsynaptic})\times(\text{modulatory error})$$

(three-factor rule).

• No backward weight symmetry is required.

Experimental Results:

EBD achieves accuracy competitive with backpropagation and Direct Feedback Alignment on MNIST and CIFAR-10 for MLP, CNN, and locally connected architectures. In a recurrent CorInfoMax network, EBD approaches the accuracy of Equilibrium Propagation (97.53% vs. 97.58% on MNIST) (Erdogan et al., 15 Apr 2025).

6. Biological Plausibility, Algorithmic Properties, and Limitations

Biological Plausibility (Neural EBD):

• Updates are inherently three-factor and align with neuromodulated plasticity observed in cortex.
• No explicit weight transport is required; errors are broadcast via learned correlations.
• Updates are parallel across layers.

Algorithmic Properties (Coding EBD):

• Requires only the code’s parity-check matrix.
• Block construction and syndrome-driven recursion are general for all linear block codes.
• Algebraic constraints and dynamic pruning are directly leveraged in codes with suitable matrix properties.

Limitations:

• Neural EBD demands substantial memory for storing per-layer correlation matrices ($O((N^{(k)})^2)$).
• Hyperparameter tuning (e.g., decorrelation rates) introduces complexity for large-scale deployment.
• Coding EBD’s generic version scales exponentially with $Q$ unless structural properties can be exploited.

7. Open Questions and Extensions

For Linear Block Codes:

• Further generalization to codes beyond extended Hamming and exploitation of additional algebraic code properties remain open, especially regarding the scalability of EBD in high-rate or high-dimension codes.
• Analytical investigation of trade-offs between penalty approximation, exclusion strategies, and performance for non-binary or non-systematic codes.

For Neural Network Learning:

• Scalability of EBD to large networks (e.g., transformers or ImageNet-scale CNNs) is yet to be established.
• Extensions to convolutional, locally connected, and recurrent architectures are feasible via error-activation local convolutions.
• Theoretical implications for convergence, implicit regularization, and optimal choice of activation transforms $g^{(k)}$.

Summary Table: EBD Distinctions

Domain	Core Mechanism	Theoretical Foundation
Linear Block Codes	Recursive syndrome-space block search	Parity-check algebra and ML error bounds
Neural Learning	Error decorrelation via broadcast	MMSE orthogonality, three-factor rule

EBD, in both domains, denotes principled, structurally efficient alternatives to prevailing methodologies, grounded in algebraic and statistical optimality (Qiu et al., 5 Jan 2026, Erdogan et al., 15 Apr 2025).