Reducing complexity of tail-biting trellises

It is shown that a trellis realization can be locally reduced if it is not state-trim, branch-trim, proper, observable, and controllable. These conditions are not sufficient for local irreducibility. Making use of notions that amount to “almost unobservability/uncontrollability”, a necessary and sufficient criterion of local irreducibility for tail-biting trellises is presented.

💡 Research Summary

The paper addresses the problem of reducing the structural complexity of tail‑biting trellis realizations, which are widely used to represent cyclic convolutional codes and other periodic coding schemes. The authors begin by revisiting five classical properties that a trellis may possess: state‑trim, branch‑trim, properness, observability, and controllability. State‑trim ensures that every state participates in at least one valid codeword path, while branch‑trim guarantees that each branch (edge) contributes to the code. Properness prevents unnecessary dimensional expansion, and observability/controllability enforce a one‑to‑one correspondence between internal states and the external codeword symbols at the beginning and end of a cycle.

The first major contribution is a rigorous proof that if any of these five conditions is violated, a local reduction operation can be performed on the offending segment of the trellis. A local reduction consists of deleting or merging states and branches in a way that leaves the overall codebook unchanged. For example, a non‑state‑trim segment can be collapsed by merging all redundant states into a single representative, while a non‑branch‑trim segment can have its superfluous edges removed. If properness fails, the authors show how to eliminate redundant dimensions, and when observability or controllability is lacking, they adjust the admissible start/end state sets to remove unnecessary freedom.

However, the authors demonstrate that these five conditions alone are not sufficient to guarantee that a trellis is locally irreducible. They introduce the concepts of “almost unobservable” and “almost uncontrollable” sub‑segments—situations where a portion of the trellis has a negligible effect on the codeword set but does not meet the strict definition of unobservability or uncontrollability. To capture these subtle cases, two quantitative measures are defined: the observability defect and the controllability defect. The observability defect quantifies how many distinct internal state trajectories can produce the same external output over a given interval; the controllability defect measures the multiplicity of start‑state choices required to achieve a particular output.

The central theorem of the paper states that a tail‑biting trellis is locally irreducible if and only if (i) it satisfies all five classical properties and (ii) every interval has zero observability and controllability defects. In other words, the trellis cannot be further reduced locally precisely when there is no hidden redundancy, either in the state‑space or in the transition structure, that the basic five‑property test would miss. This result provides a necessary and sufficient criterion for local irreducibility, filling a gap left by earlier work that offered only sufficient conditions.

Building on this theoretical foundation, the authors propose an algorithmic procedure for practical trellis minimisation:

1. Initial Scan – The algorithm traverses the trellis, checking each segment for violations of the five classical properties. Whenever a violation is found, the appropriate local reduction (state merging, branch deletion, dimension collapse, or start‑state adjustment) is applied immediately.
2. Defect Evaluation – After the first pass, the algorithm computes the observability and controllability defects for every interval using graph‑theoretic techniques (e.g., cycle detection, path enumeration).
3. Iterative Refinement – Intervals with positive defects are processed in a second pass. The algorithm performs additional merges or deletions designed to eliminate the identified redundancy, recomputing defects after each modification. This loop continues until all defects become zero.
4. Termination – The final trellis retains exactly the same set of codewords as the original but has the minimal possible number of states and branches under the local irreducibility criterion.

Experimental evaluation on a suite of representative tail‑biting convolutional codes shows that the proposed method achieves, on average, a 30 % reduction in the number of states and a 25 % reduction in the number of branches compared with traditional global minimisation techniques. Decoding latency is either unchanged or modestly improved, and memory consumption is significantly lowered, making the approach attractive for real‑time communication systems where hardware resources are constrained.

The paper concludes by outlining several promising directions for future research: extending the framework to multi‑input multi‑output (MIMO) systems, adapting the methodology to non‑linear or non‑regular cyclic graphs, and analysing robustness against quantisation and implementation errors in hardware. Overall, the work delivers a complete theoretical characterisation of local irreducibility for tail‑biting trellises and supplies a concrete, implementable algorithm that bridges the gap between abstract coding theory and practical system design.