Digital Communications

Confidence Value

Q: How does a confidence value differ from a hard decision?

A hard decision collapses each received sample to a single bit (0 or 1) before decoding, discarding all information about how reliable that decision was. A confidence value, or soft decision, retains both the most likely bit and a magnitude that expresses certainty, usually as a log-likelihood ratio (LLR). A large positive LLR means a strong vote for bit 0, a large negative LLR a strong vote for bit 1, and an LLR near zero means the bit is essentially a coin flip. Decoders that accept confidence values typically gain about 2 dB of coding gain over hard-decision decoders at a bit error rate of 1e-5, because the decoder can weight ambiguous bits less heavily than confident ones. This is why turbo and LDPC systems are built around soft information end to end.

Q: How many bits are needed to quantize a confidence value?

In hardware, confidence values (LLRs) are quantized to a fixed number of bits to keep memory and routing manageable. Most practical LDPC and turbo decoders use 4 to 6 bits per LLR: typically 1 sign bit plus 3 to 5 magnitude bits. Simulation studies show that 5-bit (one sign, four magnitude) quantization with a well-chosen saturation level loses only about 0.1 to 0.2 dB versus floating-point soft decisions, while 3-bit quantization can cost 0.5 dB or more. The choice of the clipping or saturation threshold matters as much as the bit width: clipping LLRs too aggressively flattens the dynamic range and slows decoder convergence, while too high a threshold wastes resolution on rarely seen extreme values. 5G NR and DVB-S2 receivers commonly settle on 6-bit LLRs as the performance-versus-area sweet spot.

Q: How is the channel LLR computed for a BPSK signal in AWGN?

For BPSK over an additive white Gaussian noise channel with mapping 0 to +1 and 1 to -1, the channel confidence value is L(y) = 2y / sigma^2, where y is the matched-filter output sample and sigma^2 is the noise variance per dimension. This makes the LLR directly proportional to the received amplitude and inversely proportional to noise power, so strong samples in a quiet channel produce large-magnitude confidence values while weak samples in a noisy channel produce small ones. The factor 2/sigma^2 is the channel reliability factor Lc, often written Lc = 4 times (Es/N0), equivalently 4 times the code rate times (Eb/N0), for normalized constellations. Estimating sigma^2 accurately is important: an SNR estimate off by 1 to 2 dB scales every LLR incorrectly and can degrade iterative decoder performance by several tenths of a dB.

/KON-fih-dens VAL-yoo/

Within a digital receiver, this is the soft-decision reliability metric attached to every demodulated bit, expressing not just which bit is most likely but how certain that estimate is. Most often it takes the form of a log-likelihood ratio (LLR), where the sign indicates the favored bit (positive for 0, negative for 1) and the magnitude indicates strength of belief. These confidence values are the soft information that iterative decoders for turbo codes and LDPC codes exchange and refine over successive iterations. Compared with hard-decision decoding, carrying confidence values through the decoder recovers roughly 2 dB of coding gain at a bit error rate of 1×10^-5, which is why modern standards such as 5G NR and DVB-S2 quantize LLRs to 6 bits and propagate them end to end.

Category: Digital Communications

Typical form: Log-likelihood ratio

HW quantization: 4 to 6 bits

How Soft Information Drives Iterative Decoders

The shift from hard-decision to soft-decision receivers was one of the practical breakthroughs that let coded systems approach the Shannon capacity limit. A hard decision throws away everything except the sign of each received sample, so a bit that just barely crossed the decision threshold is treated identically to one that landed far from it. A confidence value preserves that distance. The decoder can then trust unambiguous bits, discount marginal ones, and reconcile conflicting evidence from the channel and from the code's parity structure. This weighted reasoning is exactly what message-passing decoders for LDPC codes and the BCJR-based component decoders inside turbo codes rely on.

Concretely, the confidence value is almost always represented as a log-likelihood ratio. The LLR for a bit is the natural log of the ratio of the probability the bit is 0 to the probability it is 1, conditioned on the observation. A value of +5 means the bit is about e⁵, or roughly 148 times, more likely to be 0 than 1; a value near 0 means the receiver has essentially no information. Working in the log domain turns probability products into sums, which keeps the arithmetic numerically stable and lets hardware replace multipliers with adders and a simple correction term.

An iterative decoder updates these values rather than the bits themselves. Each iteration produces extrinsic confidence, the new belief about a bit derived from everything except that bit's own channel input, and feeds it forward as a priori information to the next stage. Over 8 to 20 iterations the magnitudes typically grow as the decoder gains certainty, and decoding stops when all parity checks are satisfied or an iteration cap is reached. Tracking how the average confidence magnitude evolves is the basis of EXIT-chart analysis used to design the codes themselves.

Channel LLR and the Box-Plus Update

Log-likelihood ratio (definition):
L(b) = ln[ P(b = 0 | y) / P(b = 1 | y) ]

Channel LLR, BPSK in AWGN:
L(y) = 2y / σ² with reliability factor L_c = 4 × (E_s/N₀) = 4 × R × (E_b/N₀)

Combining two independent confidences (box-plus):
L₁ ⊞ L₂ = 2·tanh^-1[ tanh(L₁/2) × tanh(L₂/2) ]
≈ sign(L₁)·sign(L₂) × min(|L₁|, |L₂|)

Where y = matched-filter sample, σ² = noise variance per dimension, E_s/N₀ = SNR per coded symbol, E_b/N₀ = SNR per information bit, R = code rate (so E_s = R·E_b). The min approximation is the min-sum rule used in hardware LDPC decoders, typically with a scaling factor of 0.75 to 0.875.

Hard Decision vs. Soft Confidence Values

Attribute	Hard Decision	2-bit Soft	3-bit Confidence	6-bit Confidence	Float LLR
Info retained	Bit only	Sign + 1 mag	Sign + 2 mag	Sign + 5 mag	Full
Coding gain vs. hard	0 dB (ref)	~1.0 dB	~1.5 dB	~1.9 dB	~2.0 dB
Loss vs. float	~2.0 dB	~1.0 dB	0.4 to 0.6 dB	0.1 to 0.2 dB	0 dB (ref)
Decoder memory	Lowest	Very low	Low	Moderate	Impractical
Typical use	Legacy BCH	Coarse soft	Low-area IoT	5G NR, DVB-S2	Simulation

Common Questions

Frequently Asked Questions

How does a confidence value differ from a hard decision?

A hard decision collapses each sample to a single bit before decoding, discarding how reliable that choice was. A confidence value keeps both the favored bit and a magnitude of certainty, usually an LLR: large positive votes for 0, large negative for 1, near zero is a coin flip. Decoders using confidence values gain roughly 2 dB over hard decisions at a BER of 1×10^-5, which is why turbo and LDPC systems carry soft information end to end.

How many bits are needed to quantize a confidence value?

Most LDPC and turbo decoders use 4 to 6 bits per LLR, typically 1 sign bit plus 3 to 5 magnitude bits. A 5-bit (1 sign, 4 magnitude) representation with a well-chosen saturation level loses only about 0.1 to 0.2 dB versus floating point, while 3-bit can cost 0.5 dB. The clipping threshold matters as much as bit width: clip too hard and convergence slows; set it too high and resolution is wasted. 5G NR and DVB-S2 typically settle on 6-bit LLRs.

How is the channel LLR computed for a BPSK signal in AWGN?

For BPSK in AWGN with 0 to +1 and 1 to -1, the channel confidence value is L(y) = 2y / σ², where y is the matched-filter sample and σ² is the noise variance per dimension. The LLR scales with received amplitude and inversely with noise power. The factor 2/σ² is the reliability factor L_c. An SNR estimate off by 1 to 2 dB scales every LLR wrong and can cost several tenths of a dB in iterative decoding.

Related Terms

← Conductor Thickness Configuration Item →