Turbo Coding with Iterative Decoding
- Introduced by
- Claude Berrou, Alain Glavieux, and Punya Thitimajshima
- Year
- 1993
- Type
- iterative error-correcting code
- Family
- Coding theory
What does it do?
Turbo codes, introduced by Berrou, Glavieux, and Thitimajshima in 1993, are a landmark in channel coding history.
They achieve performance within 0.5 dB of the Shannon limit—the theoretical boundary for reliable communication—a feat previously thought impossible with practical complexity. Turbo codes use concatenated convolutional codes with an interleaver and iterative decoding via belief propagation. They were adopted in 3G (UMTS) and remain important in 4G/5G systems alongside LDPC codes.
The intuition behind the method
Standard error-correcting codes add redundancy to protect data but typically require complex decoding (e.g., maximum likelihood, which grows exponentially with code length). Turbo codes flip the problem: use simple encoders (two or more convolutional codes) in parallel, separated by a pseudo-random interleaver, and let decoders exchange information iteratively. The interleaver ensures that errors that are hard for one code are likely easy for another (due to the permutation), so iterative refinement converges to the true message. This 'turbo effect' (hence the name) produces exponentially better performance than single-pass decoding, approaching the theoretical limit.
History
Turbo codes were announced at ICC 1993, shocking the coding theory community with near-Shannon performance. They were a major driver of 3G (UMTS, cdma2000) adoption in the late 1990s. Understanding turbo codes led to a broader theory of iterative decoding and a renaissance in understanding of classical codes (e.g., Hamming codes were shown to have iterative structure). In the 2000s, LDPC codes (rediscovered by Gallager, 1962) emerged as competitors with comparable performance and simpler architecture. Polar codes (Arikan, 2009) offered theoretical optimality in the limit. Modern standards (5G NR) support all three families; turbo codes are legacy but widely implemented.
When to use it
Turbo codes excel in deep-fading or noisy channels where strong error correction is needed (e.g., satellite, deep space, underwater). Use turbo coding when low bit-error rates (BER < 10^-6) are critical and latency permits iterative decoding (few milliseconds). Turbo codes are less suited to rate-critical applications or very short packets (interleaver and decoder state dominate overhead). With the rise of LDPC and polar codes (slightly better performance, simpler), turbo codes are now legacy in many new standards, though still used in compatibility layers.
Limitations and considerations
- Iterative decoding introduces latency (tens of milliseconds for long interleavers)
- Performance is sensitive to interleaver design; poor choice undermines the turbo effect
- Errors tend to propagate in clusters (error floor) at very low BERs; requires code tailoring to mitigate
- Decoding complexity is proportional to information length (not exponential, but still high)
- Short packets suffer from interleaver overhead and insufficient iteration depth for convergence
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Performance within 0.5 dB of Shannon limit with practical complexity
- Works well over a wide range of signal-to-noise ratios without threshold phenomenon (unlike classical codes)
- Iterative decoding improves performance with each iteration, allowing rate control via iteration count
- Separable design: any convolutional code can be used, enabling flexibility
- Robust to burst errors due to interleaver spreading
Applications
- 3G/4G cellular (UMTS, LTE): uplink and downlink traffic channels with turbo codes providing near-Shannon performance
- Satellite communications: deep space missions (e.g., Mars rovers) use turbo codes for reliable data recovery from extreme noise
- Military secure communications: encrypted data transmission with strong error correction in jamming-prone environments
- Wireless sensor networks: low-power nodes use short turbo codes for reliable multi-hop forwarding
- Medical telemetry: biomedical devices transmitting critical patient data over noisy wireless links
How it works
How it works
Information bits are encoded into two parallel codewords using identical convolutional encoders (RSC: recursive systematic convolutional).
Between the two encoders, a pseudo-random bit interleaver permutes the information sequence. This ensures the two codewords have different weight distributions for the same information.
Encoded bits (systematic and parity from both encoders) are transmitted over the channel, corrupted by noise.
At the receiver, soft-input soft-output (SISO) decoders iterate: decoder 1 uses the received signal and extrinsic information from decoder 2 to refine its estimate. Decoder 2 does likewise. De-interleaver/interleaver relays information between decoders.
After a fixed number of iterations (e.g., 8-20), a final hard decision is made by thresholding the likelihood ratio. Convergence can be tracked via 'turbo cliff': performance drops sharply once SNR crosses a threshold where iteration begins converging.
Common pitfalls
- Using a deterministic or too-short interleaver; pseudo-randomness is essential for the iterative gain
- Assuming fixed iteration count is optimal; performance varies with SNR; adaptive iteration count is better
- Neglecting the error floor; turbo codes can exhibit a threshold (turbo cliff), below which performance degrades rapidly
- Applying turbo codes to very short information blocks (<100 bits); relative overhead becomes prohibitive
- Not accounting for feedback delay in iterative decoding; coherent timing between decoders is required
Frequently asked questions
What is the 'turbo' in turbo codes?
The name refers to the iterative decoding process resembling a turbo engine: each decoder 'boosts' the signal for the next decoder. The exponential performance gain per iteration is sometimes called the turbo effect. It is not literally related to turbochargers.
Why is the interleaver so important?
The interleaver breaks correlations in the error patterns from the first decoder before the second decoder sees them. Without it, both decoders would make the same mistakes repeatedly. A good interleaver ensures that errors from one decoder are spread out randomly, giving the second decoder a fresh perspective and enabling productive iteration.
What is the turbo cliff?
Turbo codes exhibit a sharp threshold: below a critical SNR, iterative decoding fails to converge, and BER is poor. Above the threshold, BER drops sharply. This is the turbo cliff. For system design, SNR must be safely above the cliff; operating just below causes catastrophic failure despite seeming SNR.
How many iterations are needed?
Typical turbo decoders use 8-20 iterations; more iterations improve performance but increase latency. At high SNR, few iterations suffice (convergence is fast). At low SNR, more iterations are needed. Adaptive iteration (check convergence, stop early if done) balances latency and performance.
Related methods
Sources
Berrou, C., Glavieux, A., & Thitimajshima, P. (1993). Near Shannon limit error-correcting coding and decoding: Turbo-codes. In Proceedings of the IEEE International Conference on Communications (ICC), 1064-1070.
Open DOI referenceCopied.Richardson, T. J., & Urbanke, R. L. (2002). The capacity of low-density parity-check codes under message-passing decoding. IEEE Transactions on Information Theory, 47(2), 599-618.
Open DOI referenceCopied.