High Performance Error Correction for Quantum Key Distribution using Polar Codes

1 Introduction

• DVQKD and CVQKD transmit raw-key bits over noisy quantum channels, and errors must be corrected for Alice and Bob to obtain the same key.The channel error distributions differ between the two QKD families.Source: §1 · pp. 1–2
• Error correction reveals information that must be subtracted from the final key, while limited throughput can keep the key rate below what the optics allow.Cost, power consumption, and latency are described as less important constraints in QKD than in typical error-correction applications.Source: §1 · p. 2
• The paper evaluates polar codes for QKD by examining their distance to Shannon bounds and decoding speed.Polar codes were introduced by Arıkan.Source: §1 · p. 2
• Polar decoding is deterministic and scales in O(N log N), but high QKD efficiencies require very large blocks on the BSC or BIAWGNC.The decoder's recursive structure is identified as a source of good practical performance.Source: §1 · p. 2

2.1.1 Key rate and distance of error correction to Shannon bounds

• Imperfect reconciliation reduces secret-key rate and range, with its effect most severe at large distances.The paper attributes that severity to the faster distance-dependent decrease of I(x, y) − S(x : E) than of I(x, y) and S(x : E) individually.Source: §2.1.1 · p. 3

2.1.2 Key rate and error correction computation time

• In a reported system, 63,000 symbols per second of correction throughput left only 18% of available symbols usable for secret keys.The raw optical repetition rate was 500 kHz, while the data rate entering error correction was 350 kHz.Source: §2.1.2 · p. 3

2.1.3 Key rate and error correction frame error rate

• QKD can discard incorrectly decoded blocks, so the key rate is reduced by (1 − FER); the study targets a FER of 0.1.The passage identifies accepting a high FER as a way to enable faster error correction.Source: §2.1.3 · pp. 3–4

2.2 Previous work

• Cascade is highly interactive but exceeds 96% efficiency; LDPC requires one exchange, while binary-input AWGN codes extend achievable CVQKD distance.High-efficiency LDPC decoding speed had received little investigation outside CVQKD, where reported 40 kb/s on CPU and 60 kb/s on GPU.Source: §2.2 · p. 4

3 Polar codes for QKD: efficiency vs. block sizes

• Prior polar-code work addresses wiretap channels, quantum-information transmission, and optical-channel capacity, but QKD bounds Eve’s information without characterizing an Alice–Eve channel.The wiretap setting assumes that channel is characterized.Source: §3 · p. 5
• Polar codes support rateless operation as channel noise changes and non-interactive error correction, unlike two-way protocols such as Cascade.The section examines the block sizes needed for QKD efficiency because block size affects practical decoding throughput.Source: §3 · p. 5
• Channel polarization recursively combines symmetric binary discrete memoryless channels until bit-channels approach noiseless or fully noisy states, with the noiseless fraction equal to channel capacity asymptotically.At finite block lengths, the construction uses bit-channels with almost-zero error probability.Source: §3 · p. 5
• BSC polar-code efficiency exceeds 95% over almost the full 0–0.11 error range at block length 2^24; the figure shows the block-size curves.Smaller blocks can suffice when the full error range need not be covered.Source: §3 · p. 5, Figure 1 · p. 6
• Frozen bits are selected using Density Evolution to estimate bit-channel capacities and bound block FER at 0.1, allowing some blocks to be discarded at verification.The construction method is described in.Source: §3 · p. 5
• For the studied AWGNC SNRs, polar codes reach 90% efficiency over almost the full range at block length 2^27, versus about 95% for LDPC codes.Long-distance CVQKD remains possible with polar codes; the authors note polarization speed might improve with a different channel-combining recursion.Source: §3 · pp. 5–6

4 Decoding speed: numerical results

• Polar successive-cancellation decoding reaches about 10 Mb/s on modern CPUs; LDPC decoding reaches 10 Mb/s on an AMD Tahiti GPU, versus about 800 kb/s for LDPC on one modern CPU core.The polar simulations used an Intel Core i5 670 at 3.47 GHz.Source: §4 · pp. 6, 8
• The polar decoder uses fixed-point arithmetic and a lookup table for ϕ(x) to update LLRs, while its recursive successive-cancellation structure is sequential.Look-ahead techniques are reported to reduce successive-cancellation latency by 50% in.Source: §4 · pp. 6, 8
• For the studied AWGNC SNRs, block length 2^27 gives polar-code efficiency above the 90% target over almost the entire range.The figure compares block sizes from 2^17 through 2^27.Source: Figure 2 · p. 7
• The LDPC GPU benchmark uses a floating-point flood decoder with fixed iterations and internal and external parallelism; the CPU benchmark uses shuffle decoding with early termination.The speed tables cover BSC and BIAWGNC cases; the LDPC iteration limits differ by table and code.Source: §4 · p. 8, Table 2 · p. 8, Table 1 · p. 8
• Polar decoding throughput is sufficient for state-of-the-art QKD implementations, despite the large blocks required for the needed efficiencies.The paper gives the best reported QKD key rate as about 1 Mb/s.Source: §4 · pp. 8–9

• Polar codes provide error correction for both DVQKD and CVQKD, achieving good efficiencies at noise levels compatible with QKD.Source: §5 · p. 9
• Because polarization is slower for the BIAWGNC than for the BSC, CVQKD requires higher block sizes and is less practical than DVQKD.Source: §5 · p. 9
• Recursive decoding runs at CPU speeds similar to LDPC decoding on GPUs, and the work reports the first practical application of polar codes.The paper identifies decoding as a frequent bottleneck in recent QKD implementations.Source: §5 · p. 9