PaperlayerView exact revision on arXiv

Source-linked AI summary

Key Reconciliation with Low-Density Parity-Check Codes for Long-Distance Quantum Cryptography

Mario Milicevic, Chen Feng, Lei M. Zhang, P. Glenn Gulak

arXiv:1702.07740v2quant-phcs.CRcs.ITmath.IT

TL;DR

This work addresses the hardware challenge of long-block LDPC decoding in long-distance CV-QKD by introducing multi-edge QC codes and a GPU decoder with multidimensional reconciliation. With 8-dimensional reconciliation, information throughput is up to 8.03× the lossy-channel secret-key-rate upper bound, and maximum distance extends from 100km to 160km under the assumed channel parameters.

  • Problem

    Implementing LDPC decoders for block lengths on the order of 10^6 bits is challenging because belief-propagation hardware must handle complex CN–VN interconnects.

    Source: C. Quasi-Cyclic LDPC Codes · p. 11

  • Method

    The paper introduces multi-edge QC-LDPC codes and a GPU decoder paired with multidimensional reconciliation to accelerate long-distance CV-QKD key reconciliation.

    Source: §I · p. 2, §V · p. 20

  • Results

    8.03× higher information throughput than the lossy-channel secret-key-rate upper bound is achieved with an 8-dimensional scheme and 1MHz source; maximum distance extends from 100km to 160km.

    Source: §V · p. 20

  • Takeaways & Limitations

    LDPC decoding is no longer the computational bottleneck; secret-key performance remains limited by quantum-channel parameters and privacy-amplification latency.

    Source: §I · p. 3

    The 100km-to-160km distance extension is based on the quantum-channel parameters assumed in this work.

    Source: §V · p. 20

Abstract

from arXiv · show

The speed at which two remote parties can exchange secret keys over a fixed-length fiber-optic cable in continuous-variable quantum key distribution (CV-QKD) is currently limited by the computational complexity of post-processing algorithms for key reconciliation. Multi-edge low-density parity-check (LDPC) codes with low code rates and long block lengths were proposed for CV-QKD, in order to extend the maximum reconciliation distance between the two remote parties. Key reconciliation over multiple dimensions has been shown to further improve the error-correction performance of multi-edge LDPC codes in CV-QKD, thereby increasing both the secret key rate and distance. However, the computational complexity of LDPC decoding for long block lengths on the order of 10^6 bits remains a challenge. This work introduces a quasi-cyclic (QC) code construction for multi-edge LDPC codes that is highly suitable for hardware-accelerated decoding on a modern graphics processing unit (GPU). When combined with an 8-dimensional reconciliation scheme, the LDPC decoder achieves a raw decoding throughput of 1.72Mbit/s and an information throughput of 7.16Kbit/s using an NVIDIA GeForce GTX 1080 GPU at a maximum distance of 160km with a secret key rate of 4.10x10^{-7} bits/pulse for a rate 0.02 multi-edge code with block length of 10^6 bits when finite-size effects are considered. This work extends the previous maximum CV-QKD distance of 100km to 160km, while delivering between 1.07x and 8.03x higher decoded information throughput over the upper bound on the secret key rate for a lossy channel. The GPU-based QC-LDPC decoder achieves a 1.29x improvement in throughput over the best existing GPU decoder implementation for a rate 1/10 multi-edge LDPC code with block length of 2^{20} bits. These results show that LDPC decoding is no longer the computational bottleneck in long-distance CV-QKD.

I. INTRODUCTION

The introduction frames long-distance CV-QKD reconciliation as a low-SNR, long-block decoding challenge and presents GPU-based QC-LDPC decoding as a way to extend distance and throughput. Background and channel assumptions describe how CV-QKD signals are transmitted and measured.

II. BACKGROUND

  • CV-QKD encodes information in coherent-state amplitude and phase quadratures, whereas DV-QKD uses single-photon polarization; CV-QKD can use room-temperature telecommunications detectors.QKD uses quantum and authenticated public channels to establish a shared key in the presence of an eavesdropper.Source: §I · p. 1, §II · p. 3

I. INTRODUCTION

  • Low-SNR optical channels, typically below −15 dB, require low-rate LDPC codes with block lengths around 10^6 bits and hundreds of decoding iterations, leaving decoding complexity a challenge.Multi-dimensional reconciliation had been shown to improve multi-edge code performance at low SNR, but did not remove the long-block decoding challenge.Source: §I · p. 2
  • 1.72 Mbit/s raw decoding and 7.16 Kbit/s information throughput at 160 km, with a 4.10×10^-7 bits/pulse secret key rate under finite-size effects, extend CV-QKD reach from 100 km.The reported result uses an NVIDIA GeForce GTX 1080 GPU and an 8-dimensional reconciliation scheme.Source: §I · pp. 2–3
  • The work introduces a quasi-cyclic construction for multi-edge LDPC codes and an optimized GPU decoder to address decoding complexity for long blocks.The authors argue GPUs suit this setting because of their on-chip memory and floating-point precision, while long-distance ASIC implementations may require prohibitively large die area.Source: §I · pp. 2–3

II. BACKGROUND

A. Quantum Transmission and Measurement

  • Channel characterization assumes 0.005 shot-noise-unit excess noise, 0.041 electronic noise, 0.606 homodyne efficiency, and 0.2 dB/km fiber loss, using previously published experimental values.The optical experimental setup is outside the scope of this work.Source: A. Quantum Transmission and Measurement · p. 4

B. Sifting

After sifting, Alice and Bob reconcile correlated measurements using reverse reconciliation and LDPC decoding over a virtual binary-input Gaussian channel. The section also explains multidimensional reconciliation and how reconciliation efficiency relates to channel capacity.

B. Sifting

  • In sifting, Bob tells Alice which quadrature he measured so she can discard the unused values, leaving correlated sequences for reconciliation and privacy amplification.The resulting sequences are modeled as independent, identically distributed jointly Gaussian pairs.Source: B. Sifting · p. 4

C. Reconciliation

  • Reverse reconciliation supports longer-distance operation better than direct reconciliation, which is unsuitable beyond 100 km; strong error correction is still needed to limit information revealed to Eve.The text reports that direct reconciliation is limited to about 15 km and that two-way interactive protocols incur large latency and communication overhead.Source: C. Reconciliation · p. 4
  • For reconciliation, Alice and Bob select and normalize n measurements to form correlated Gaussian sequences X and Y, with Y = X + Z, then Bob sends a sign-modulated LDPC codeword for Alice to decode.Alice uses her sequence X to recover Bob’s codeword from the authenticated public message.Source: 2) Reverse Reconciliation Algorithm… · p. 5
  • Multidimensional reconciliation groups consecutive transmissions in dimensions d ∈ {1, 2, 4, 8} and maps each group to a virtual binary-input AWGN channel with uniform noise variance across dimensions.The construction uses normed division spaces over the reals, complex numbers, quaternions, and octonions.Source: 3) Multi-Dimensional Reconciliation · p. 5
  • The channel-noise derivations for d = 4 and d = 8 are much longer and are not included.The discussion therefore does not provide those derivations in this section.Source: 3) Multi-Dimensional Reconciliation · p. 6
  • Reconciliation efficiency β relates the LDPC code rate to the mutual information, or Shannon capacity, of the channel; for a fixed code rate, it depends on channel SNR.For a fixed transmission distance, the text says SNR can be optimized by tuning Alice’s modulation variance.Source: 4) Reconciliation Efficiency · p. 6

D. Privacy Amplification

Alice and Bob discard incorrectly decoded messages, concatenate the correctly decoded strings, and independently apply universal hashing to reduce Eve’s knowledge. The passage also notes prior work on faster privacy amplification.

D. Privacy Amplification

  • Alice discards incorrectly decoded messages; Alice and Bob concatenate the remaining decoded strings and independently apply universal hashing to reduce Eve’s knowledge.The resulting shared key can be used with one-time-pad encryption.Source: D. Privacy Amplification · p. 6
  • Universal hashing complexity can be reduced from O(n^2) to O(n log^2 n) using FFT or NTT on a Toeplitz matrix.The passage notes published privacy-amplification speeds of 100Mb/s for Nprivacy = 10^8 bits.Source: D. Privacy Amplification · p. 6
  • Privacy amplification also uses Nquantum − Nprivacy bits for security-parameter estimation, but its full treatment is outside this work’s scope.Source: D. Privacy Amplification · p. 6

E. Maximizing Secret Key Rate with Collective Attacks

The section defines theoretical secret-key performance under collective attacks and examines how modulation variance and reconciliation efficiency affect it. It also distinguishes this ideal rate from practical performance, where frame errors and system imperfections matter.

E. Maximizing Secret Key Rate with Collective Attacks

  • Kopt is an asymptotic theoretical limit and does not account for practical imperfections such as finite-size effects, excess noise, or discretized modulation.The section considers finite-size effects but places other practical imperfections beyond its scope.Source: E. Maximizing Secret Key Rate with… · p. 7
  • The lossy-channel upper bound is determined by fiber transmittance, while the plotted Kopt curves use optimal VA across β-efficiencies from 0.8 to 0.99.The section assumes fiber loss α = 0.2dB/km when defining transmittance.Source: Figure 5 · p. 7, E. Maximizing Secret Key Rate with… · pp. 7–8
  • Optimal modulation variance is selected by distance and β-efficiency to maximize SNR and the corresponding theoretical secret-key rate under perfect error correction.Fig. 4 plots optimal VA versus distance, and Fig. 5 plots the resulting Kopt against distance alongside the lossy-channel limit.Source: E. Maximizing Secret Key Rate with… · p. 7, Figure 4 · p. 7, Figure 5 · p. 7
  • At long distances, maximizing β alone may not maximize usable key rate because Kopt omits reconciliation speed and uncorrectable errors, so frame error rate also matters.Source: E. Maximizing Secret Key Rate with… · p. 8

F. Frame Error Rate for Reverse Reconciliation

Reverse reconciliation must distinguish detected decoding failures from valid codewords that decode to the wrong message. Parity checks and a CRC identify errors, but a rare CRC failure can discard a correct message.

F. Frame Error Rate for Reverse Reconciliation

  • A parity-check failure identifies a decoding error; when the check passes, a CRC compares Bob’s message with Alice’s decoded message before key use.The CRC is skipped when the parity check fails.Source: F. Frame Error Rate for Reverse… · p. 8
  • Long block codes have exponentially many possible codewords, so decoding can converge to a valid codeword with an incorrect message, an undetected error.Such an error is problematic because both parties need the same message for privacy amplification.Source: F. Frame Error Rate for Reverse… · p. 8
  • Frame error rate counts both detected and undetected errors, where a frame error means the decoded message differs from the original.The passage introduces separate probabilities for detected and undetected errors.Source: F. Frame Error Rate for Reverse… · p. 8
  • A rare parity-pass, CRC-fail case can discard a correct message and reduce the secret-key rate.This case is not included by convention in communication theory.Source: F. Frame Error Rate for Reverse… · p. 8

G. Impact of Reconciliation Error and Efficiency on Key Rate

The effective secret-key rate accounts for discarded frames through the frame error rate. For a fixed code rate, reconciliation efficiency and FER jointly shape the distance–key-rate trade-off.

G. Impact of Reconciliation Error and Efficiency on Key Rate

  • The effective secret-key rate scales the asymptotic rate by FER because frames with known errors must be discarded.The analysis assumes fixed physical channel parameters and optimally set modulation variance for each distance and β-efficiency.Source: G. Impact of Reconciliation Error and… · p. 8
  • With a 32-bit CRC, the study empirically finds Pundetected error = 0 and simplifies total decoding FER to Pe = Pdetected error.This simplified FER expression is assumed for the remainder of the work.Source: G. Impact of Reconciliation Error and… · p. 8
  • For a fixed LDPC code rate, a given SNR permits either longer distance with lower key rate or shorter distance with higher key rate.The code can be designed to target FER at a particular SNR and trade β against FER.Source: G. Impact of Reconciliation Error and… · p. 9

H. Secret Key Rate with Finite-Size Effects

Finite-size effects reduce the secret key rate, while code length affects reconciliation efficiency and frame-error rate rather than appearing directly in the finite-size expression. The section also motivates low-rate, long-block LDPC codes and describes the work’s focus on rate-0.02 codes and their general design.

H. Secret Key Rate with Finite-Size Effects

  • The finite-size key-rate expression includes an offset and a scaling by N_privacy/N_quantum; LDPC block length affects the rate indirectly through β and FER.A conservative security parameter is ε = 10^-10.Source: H. Secret Key Rate with Finite-Size… · p. 9

III. LDPC CODES FOR RECONCILIATION

  • High-efficiency reconciliation and LDPC codes with block lengths on the order of 10^6 bits remain computationally challenging for long-distance CV-QKD.Advanced reconciliation can reach 95% efficiency but has computational complexity.Source: §III · p. 9
  • The study designs rate-0.02 multi-edge codes with block lengths near 10^6 bits, focusing on one low-rate code across transmission distances rather than multi-rate programmability.It examines how β-efficiency and FER affect achievable secret-key rate and reconciliation distance.Source: §III · p. 10

A. General Construction of LDPC Codes

  • General LDPC design selects degree distributions and then samples a Tanner graph, but this approach is non-trivial for low-rate codes near Shannon capacity at low SNR.The graph is sampled to satisfy the selected degree distributions, up to rounding error.Source: A. General Construction of LDPC Codes · p. 10

B. Multi-Edge LDPC Codes

Multi-edge LDPC codes support near-capacity performance at low rates, and quasi-cyclic structure makes their large parity-check matrices more suitable for hardware decoding. The tested q = 50 code has the best overall FER, while larger expansion factors degrade error-correction performance.

B. Multi-Edge LDPC Codes

  • Multi-edge LDPC codes offer near-Shannon-capacity performance for low-rate codes, which is crucial for achieving high reconciliation efficiency in long-distance CV-QKD.Their low error-floor advantage for high-rate codes is less important here because the operating FER is about 10^-1.Source: B. Multi-Edge LDPC Codes · p. 10
  • The designed rate-0.02 multi-edge matrices use 10^6-bit blocks and a 9.8 × 10^5 by 10^6 parity-check matrix; imposing QC structure can reduce decoder complexity with minimal performance degradation.The random matrices are designed for the waterfall region, where high FER makes error floors less significant.Source: B. Multi-Edge LDPC Codes · p. 11

C. Quasi-Cyclic LDPC Codes

  • QC matrices regularize decoder interconnects, addressing the unordered memory access and complex routing that make random-code decoders difficult to scale in hardware.QC parity-check matrices use cyclically shifted identity or zero submatrices; larger q reduces decoder complexity but can degrade error-correction performance.Source: C. Quasi-Cyclic LDPC Codes · p. 11
  • The multi-edge QC construction samples a base graph at length n/q, assigns shifts to its nonzero entries, then expands them into circulant permutation matrices.Each zero entry expands to a q × q all-zero matrix.Source: C. Quasi-Cyclic LDPC Codes · p. 11
  • Only q = 21 and q = 50 codes are retained: q = 100, 500, and 1000 perform significantly worse than a random code with the same degree distribution.The q = 21 code uses n = 1.008 × 10^6 bits at rate 0.02; q = 50 uses n = 10^6 bits at rate 0.01995.Source: C. Quasi-Cyclic LDPC Codes · pp. 11–12
  • Degree-1 variable nodes along the QC matrix diagonal need VN-to-CN messages only in the first iteration, speeding decoding; the lower-triangular parity component also permits forward-substitution encoding.The latter simplifies encoding but is not claimed to reduce decoding complexity or improve error-correction performance.Source: C. Quasi-Cyclic LDPC Codes · p. 12

D. LDPC Decoding: Belief Propagation Algorithm

The decoder uses Sum-Product belief propagation to iteratively pass probabilistic messages between variable and check nodes. It stops when the parity-check condition is satisfied or the maximum iteration count is reached.

D. LDPC Decoding: Belief Propagation Algorithm

  • Belief propagation iteratively exchanges probabilistic updates between variable and check nodes until the parity-check condition is satisfied or the iteration limit is reached.The work uses the Sum-Product variant.Source: D. LDPC Decoding: Belief Propagation… · p. 12
  • Each variable node is initialized with a channel log-likelihood ratio derived from Alice’s correlated Gaussian sequence and channel noise variance.The channel-noise expression differs for d = 2, 4, and 8 reconciliation, while the rest of the algorithm remains the same.Source: D. LDPC Decoding: Belief Propagation… · p. 12
  • Sum-Product decoding is used because Min-Sum, though faster to implement, does not perform well at low SNR and is unsuitable for long-distance CV-QKD.The Sum-Product updates pass messages along Tanner-graph edges and test the syndrome for early termination.Source: D. LDPC Decoding: Belief Propagation… · p. 12

E. Error-Correction Performance of Multi-Edge QC Codes

The q = 50 code has the strongest overall FER performance among the tested codes, and higher-dimensional reconciliation improves coding gain in the waterfall region. Invalid decoding remains a measurable source of frame errors, while key-rate optimization requires trading off reconciliation efficiency and FER.

E. Error-Correction Performance of Multi-Edge QC Codes

F. Finite Secret Key Rate

Finite-size secret-key rates are evaluated across reconciliation dimensions, LDPC codes, and privacy-amplification block sizes. The d = 8 scheme reaches 160 km with Nprivacy = 10^12 bits, while the GPU section frames decoder throughput at maximum-distance operating points.

F. Finite Secret Key Rate

  • 18 km of additional maximum distance is obtained for d = 8 at β = 0.99 when Nprivacy increases from 10^10 to 10^12 bits.The text identifies privacy-amplification block size as important to maximum transmission distance.Source: F. Finite Secret Key Rate · p. 14, Figure 19 · p. 15
  • In each of the Nprivacy = 1010 and Nprivacy = 1012 cases, the three LDPC codes achieve similar finite secret-key rates.Source: F. Finite Secret Key Rate · p. 14
  • 160 km is the maximum distance for d = 8 reconciliation with Nprivacy = 10^12 bits; d = 1 reaches 132 km with the same block size.At Nprivacy = 10^10 bits, the respective maxima are 142 km and 124 km.Source: F. Finite Secret Key Rate · p. 15, Figure 19 · p. 15, Figure 17 · p. 15
  • For d = 8, β = 0.99 is achievable, whereas d = 1 reaches only β = 0.96 before FER makes the finite key rate zero.At β = 0.95 and β = 0.96, d = 8 also achieves higher secret-key rates than d = 1 for all three codes.Source: F. Finite Secret Key Rate · p. 15, Figure 17 · p. 15
  • The d = 2 and d = 4 schemes reach β = 0.97, with maximum distances of 129 km and 138 km for Nprivacy = 10^10 and 10^12 bits, respectively.These results continue the comparison of the intermediate reconciliation dimensions.Source: F. Finite Secret Key Rate · p. 15

IV. GPU-ACCELERATED LDPC DECODING

  • The following GPU section compares decoder throughput at maximum CV-QKD distances for d = 1, 2, 4, and 8 against achievable secret-key rates.It also compares the implementation with prior LDPC work and non-LDPC codes.Source: F. Finite Secret Key Rate · p. 15, §IV · p. 16

A. GPU-based LDPC Decoder Implementation

The GPU implementation reorganizes Sum-Product decoding around four parallel kernels and uses quasi-cyclic structure to reduce data-permutation and memory-access costs. Throughput depends on iteration latency and count, while early termination is scheduled using empirically measured iteration averages.

A. GPU-based LDPC Decoder Implementation

  • The parallelized Sum-Product decoder uses four multithreaded kernels, with operations reordered to maximize thread-level parallelism and minimize per-iteration latency.The implementation runs on a single NVIDIA GeForce GTX 1080 GPU; overall throughput depends on iteration count, per-iteration latency, and block length.Source: A. GPU-based LDPC Decoder Implementation · p. 16, Figure 20 · p. 16
  • For 10^6-bit codes, data permutation and message passing account for 25%–50% of GPU runtime per decoding iteration, making memory-access patterns important to throughput.Unordered global-memory accesses are costly, whereas coalesced accesses allow concurrent thread execution.Source: A. GPU-based LDPC Decoder Implementation · p. 16
  • When early termination is disabled, the long-diagonal parity-check construction allows fewer threads in message-passing and variable-node update kernels, yielding marginal per-iteration speedup.With early termination enabled, the kernels require threads for all relevant connections or variable nodes.Source: A. GPU-based LDPC Decoder Implementation · p. 17
  • Early termination is checked only after a code- and SNR-specific average iteration count determined through FER simulation and stored in a lookup table.The check adds parity computations and message-passing work; a lookup table selects when to enable it based on SNR.Source: A. GPU-based LDPC Decoder Implementation · p. 17
  • Quasi-cyclic structure reduces permutation and memory-access complexity, with faster data-permutation kernels than random codes because of fewer check-node–variable-node edge connections.The parity-check matrices use compressed edge storage to reduce message-passing memory-read latency.Source: A. GPU-based LDPC Decoder Implementation · p. 17, Table II · p. 17
  • Information throughput scales raw GPU throughput by FER and code rate to account for discarded frames and parity bits.The raw rate is block length divided by per-iteration latency times the number of iterations; Table III is identified as covering throughput at maximum reconciliation distance.Source: A. GPU-based LDPC Decoder Implementation · p. 17, Table III · p. 18

B. Information Throughput Results

The GPU decoder’s information throughput exceeds the lossy-channel secret-key-rate limit in the reported comparisons, while QC codes improve throughput over purely random codes. These results hold across long-distance settings, including comparisons at higher source repetition rates.

B. Information Throughput Results

  • Approximately 3× higher raw decoding throughput and 5.1× to 12.8× higher information throughput were achieved by the q = 21 and q = 50 QC codes over the purely-random code.The comparison is at the maximum-distance point for each β-efficiency and includes d = 1, 2, 4, and 8 reconciliation.Source: B. Information Throughput Results · p. 18
  • Beyond 130 km, the operating secret key rate K′finite is 2176× to 57112× lower than the fundamental limit K′lim under the assumed 1 MHz source rate.The source repetition rate is treated as realistic for the comparison.Source: B. Information Throughput Results · p. 19
  • 1.07× to 8.03× higher information throughput K′GPU than the fundamental secret key rate limit K′lim was achieved across the reported comparisons.Table III reports the comparison at maximum-distance points for the reconciliation schemes.Source: B. Information Throughput Results · p. 19
  • The d = 1 and d = 2 schemes are not well-suited to long-distance CV-QKD because K′lim is less than 1×, while QC codes have lower decoding latency than purely-random codes at long distances.The lower-latency comparison supports reverse reconciliation at high β efficiencies.Source: B. Information Throughput Results · p. 19
  • Even at a 1 GHz source repetition rate, GPU information throughput exceeds the operating secret key rate beyond 130 km by 1.8× and 18.7× under the same channel parameters.The reported comparison supports GPU viability at those distances and source rate.Source: B. Information Throughput Results · p. 19
  • For the q = 50 QC-LDPC code with d = 8 reconciliation, K′GPU exceeds the lossy-channel upper bound K′lim from β = 0.8 to β = 0.99.The comparison assumes a 1 MHz source; Fig. 22 plots the GPU information throughput against the key-rate limit.Source: B. Information Throughput Results · pp. 18–19, Figure 22 · p. 18

C. Comparison to Other CV-QKD Implementations

The comparison reports a throughput advantage for the q = 512 QC code, while the conclusion finds that GPU-based reconciliation is no longer a computational bottleneck and extends the reported maximum CV-QKD distance to 160 km. The authors also identify remaining boundaries in alternative-code comparisons and privacy-amplification latency.

C. Comparison to Other CV-QKD Implementations

  • At the time of writing, only one reported CV-QKD implementation targeted low-SNR, long-distance reconciliation.Polar-code comparison was unavailable for the rate 0.02 QC-LDPC codes; RA-code performance had not been investigated beyond 70 km, and no hardware throughput comparison was available for RA or raptor codes.Source: C. Comparison to Other CV-QKD… · pp. 19–20
  • The q = 512 QC code achieved 1.29× the 7.1 Mb/s throughput reported by Jouguet and Kunz-Jacques and about 3× lower per-iteration latency than this work's random code.The comparison targets SNR = 0.161 on BIAWGNC with d = 8 and FER = 0.04 for rate 1/10 codes.Source: Table IV · p. 19, C. Comparison to Other CV-QKD… · pp. 19–20
  • The comparison's three codes achieved 83.8 km under the assumed quantum-channel parameters, and per-iteration latency includes CPU–GPU data-transfer time.The throughput comparison used the prior GPU implementation and two newly designed codes: a random code and a q = 512 QC code.Source: C. Comparison to Other CV-QKD… · p. 19
  • The q = 512 QC code reached 9.17 Mb/s maximum raw throughput, versus 2.78 Mb/s for the designed random rate 1/10 code.The q = 512 design's lower connection count reduces GPU message-passing memory accesses.Source: C. Comparison to Other CV-QKD… · p. 19

V. CONCLUSION

  • With 8-dimensional reconciliation, information throughput reached up to 8.03× the lossy-channel secret-key-rate upper bound, and maximum CV-QKD distance increased from 100 km to 160 km.These results use the quantum-channel parameters assumed in the work.Source: §V · p. 20
  • The authors propose extending the QC-LDPC and GPU techniques to free-space and satellite QKD, integrated QKD chips, and DV-QKD forward error correction.The paper notes that the demonstrated computational speedup was achieved on a single GPU.Source: §V · p. 20
  • The authors identify reducing privacy-amplification latency for N_privacy ≥ 10^12 bits as necessary for secret-key exchange beyond 100 km.They also point to excess-noise reduction and future study of low-SNR LDPC decoding under non-Gaussian noise.Source: §V · p. 20
Loading 1702.07740v2…