LDPC Bit-Flipping Decoding With Inertial Local Energy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing LDPC code decoding methods, such as GDBF, suffer from suboptimality due to local minima issues, resulting in inferior performance compared to message-passing decoding methods, and their randomized variants like PGDBF and NGDBF do not match the error correction performance of BP methods without increased complexity.
Innovation Solution
A decoding method for LDPC codes that incorporates an inertial component into the local energy function to reduce the likelihood of flipping bits that have recently been flipped, thereby avoiding oscillations between local extrema and improving directional coherence, with options for flipping all unreliable bits, flipping them with a probability, or randomizing the set of unreliable bits.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If GDBF decoding method is used, then decoding complexity is reduced compared to BP methods, but error correction performance deteriorates due to local minima issues
Solution Approach 1:
The patent applies preliminary action by initializing a counter for each bit that tracks the number of consecutive iterations without flipping. Before performing gradient descent updates, this counter is reset or incremented based on flipping history. This preliminary tracking action enables the inertial component to be computed and applied during the decoding process, preventing premature convergence to local minima while maintaining the simplicity of gradient descent methods.
Solution Approach 2:
The patent implements dynamics by making the bit-flipping decision dynamic through the inertial component. Instead of deterministically flipping bits with the smallest gradient values, the method introduces probability-based flipping where the probability is modulated by the inertial component (counter value). This dynamic approach allows the decoder to adaptively explore the solution space, escaping local minima while maintaining computational efficiency comparable to standard GDBF methods.
2Reliability
If randomized variants like PGDBF or NGDBF are used to escape local minima, then error correction performance improves, but decoding complexity increases
Solution Approach 1:
The patent applies self-service by having the decoding process generate its own randomization mechanism through the inertial component derived from flipping history counters. Instead of requiring external random noise injection or complex probabilistic models, the system uses its own operational history (which bits have been flipped and how frequently) to create the necessary variability. This self-generated randomness achieves performance comparable to PGDBF/NGDBF methods while maintaining simpler implementation and lower computational overhead.
3Speed
If standard gradient descent is applied without inertial component, then decoding speed is fast, but performance is suboptimal due to oscillations between local extrema
Solution Approach 1:
The patent implements continuity of useful action by maintaining the gradient descent update process as the core continuous operation while layering the inertial component on top. The counter-based inertia mechanism operates continuously across iterations, accumulating information about flipping patterns and smoothly modulating the probability of future flips. This continuous inertial influence prevents oscillations between local extrema without interrupting or significantly slowing down the gradient descent process, preserving fast decoding speed while improving convergence to global minima.
Data Source
AI summary
A method for decoding a Low Density Parity Check code. At each decoding iteration, as long as the syndrome of the estimated word indicates an error, a set ({tilde over (F)}) of the least reliable bits of the word is determined as those where the value of a local energy function ({tilde over (E)}n) is less than a threshold value. The local energy value of a bit includes a first component proportional to the correlation between this bit and a sample corresponding to the observed signal, a second component representing the number of non-satisfied constraints wherein the bit acts, and a third component decreasing with the number (n) of iterations made since the last flipping of this bit. The bits of the estimated word belonging to this set are flipped, where applicable with a predetermined probability, to provide a new estimated word at the following iteration.