Implementation of Physical-layer Network Coding

• The paper presents the first PNC implementation on a software-radio platform and describes prototyping as a step toward realizing PNC.The implementation effort also raises issues for further research.Source: §1 · pp. 1–2
• In a two-way relay channel, two end nodes send simultaneously to a relay, which broadcasts a network-coded packet for self-information-based decoding.PNC can reduce the exchange from four time slots to two compared with a conventional relay system.Source: §1 · p. 2
• Earlier implementations used analog network coding (ANC), which is simple but amplifies noise along with the signal before forwarding.Source: §1 · p. 2
• Symbol-synchronous and symbol-asynchronous FPNC have nearly the same BER performance for both channel-coded and unchannel-coded operation.Source: §1 · p. 2

• The section briefly overviews PNC challenges and the implementation approaches used to tackle them.Source: Challenges · p. 3

• For high-speed synchronous PNC, tight symbol and carrier-phase alignment is challenging, and timely uplink CSI collection is difficult in fast fading.Source: Asynchrony · p. 3
• Asynchronous PNC avoids tight end-node synchronization and precoding or uplink CSI, but leaves symbol and carrier-phase misalignments for the relay to handle.Source: Asynchrony · p. 3
• FPNC performs independent XOR mapping per OFDM subcarrier; when relative delay is within the CP, DFT aligns the time-misaligned samples in frequency.OFDM lengthens symbol duration per subcarrier and reduces the relative effect of multipath delay spread.Source: Asynchrony · p. 3

• The relay needs uplink CSI and must estimate two channels from simultaneous signals and preambles, unlike point-to-point OFDM estimation.Source: Channel Estimation · p. 3
• Overlapping training symbols and pilots can complicate channel estimation, so the implementation assigns orthogonal training symbols and pilots to the end nodes.Source: Channel Estimation · pp. 3–4
• Because fully removing one end node’s CFO interference increases it for the other, the implementation compensates for the mean of the two CFOs.Source: Channel Estimation · p. 4

Joint Channel Decoding and Network Coding

• Link-by-link channel-coded PNC maps overlapped channel-coded symbols to the XOR of source symbols, then re-encodes that XOR for forwarding.The design uses the 802.11a/g convolution code.Source: Joint Channel Decoding and Network… · pp. 4–5
• Link-by-link channel-coded PNC is reported to perform better than end-to-end channel-coded PNC.Source: Joint Channel Decoding and Network… · p. 4
• CNC combines channel decoding with network coding to recover the XOR of source information, not the individual messages.The paper identifies CNC as critical to link-by-link channel-coded PNC performance.Source: Joint Channel Decoding and Network… · p. 4

2. Effect of Delay Asynchrony in Frequency Domain

• The delay-asynchrony derivation extends to multipath, with the relative timing represented through the nodes’ path delays.The model allows A and B to transmit at different times by adding the lag to one node’s path delays.Source: §2 · p. 5, §2.1 · p. 5

2.1. Effective Discrete-time Channel Gains

• The analysis derives the effective discrete-time channel gains by incorporating pulse shaping and matched-filtering-and-sampling into the channel model.The equivalent discrete-time model represents the channel gains and noise after those operations.Source: §2.1 · pp. 5–6, Figure 3 · p. 6

2.2. Delay-Spread-Within-CP Requirement

• Adding a cyclic prefix turns the channel’s linear convolution into a circular convolution over the OFDM symbol, enabling the frequency-domain expression.The CP consists of the last C samples of each length-N input sequence, appended to its beginning.Source: §2.2 · p. 8
• When FPNC delay spread is no greater than the CP length, the received subcarrier signal is Y[k] = HA[k]XA[k] + HB[k]XB[k] + W[k], with different subcarriers isolated.The time-domain delay spread is incorporated into HA[k] and HB[k], and XOR mapping is performed within each subcarrier.Source: §2.2 · p. 9
• The frequency-domain result assumes no carrier frequency offset; with CFO, inter-carrier interference may occur.Source: §2.2 · p. 9

3. FPNC Frame Format

• A polling frame can trigger near-simultaneous transmissions; the frame adapts the 802.11a/g preamble, with sample misalignment constrained to the CP length.Source: §3 · pp. 9–10

3.1. FPNC Short Training Symbol

• FPNC adopts the 160-sample 802.11 STS: ten identical 16-sample units used by the relay for sample-timing recovery.Source: §3.1 · p. 10
• Cross-correlation identifies the following LTS boundaries; when unsynchronized, the last two of 20 peaks identify A and B if delay spread stays within the CP.Source: §3.1 · p. 10

3.2. FPNC Long Training Symbol

• The 802.11 LTS is a 160-sample sequence with a CP and two identical units, used for channel estimation and CFO compensation.Source: §3.2 · p. 11
• FPNC doubles the LTS length and shortens its CP from 32 to 16 samples, avoiding overlap with following data when delay spread is below 16 samples.Source: §3.2 · p. 11

3.3. FPNC Pilot

• Two-pilot and four-pilot designs had comparable BER for BPSK and QPSK with linear-interpolation tracking in point-to-point experiments.Source: §3.3 · p. 12
• Nulling selected pilots makes nodes A and B orthogonal for separate channel-gain tracking, while pilots track channel changes after the initial LTS estimate.Source: §3.3 · p. 12

4.1.1. CFO Estimation

• Orthogonal LTSs allow separate estimation of the two CFOs; the phase estimate uses paired LTS samples and their median to reduce the effect of outliers.The normalized CFO represents the additional phase advance from one sample to the next.Source: §4.1.1 · pp. 12–13

4.1.2. Compensation for Two CFOs

• Median CFO estimates outperform mean estimates, while compensating with the mean of both nodes’ CFOs outperforms compensating with either estimate alone.The median is reported as more stable because some phase samples are outliers; the compensation complexity matches point-to-point communication.Source: §4.1.2 · pp. 13–14, §4.1.1 · p. 13

4.2. FPNC Channel Estimation

• An initial channel estimate comes from the LTS, while non-overlapping pilots track later OFDM-symbol channel changes independently for each end node.The method estimates one end node’s channel and applies the same procedure to the other.Source: §4.2 · p. 15

4.3. FPNC Mapping

• Link-by-link CNC maps the relay’s overlapped channel-coded frames to the XOR of source symbols, then channel-encodes that XOR for broadcast.The paper’s XOR-CD design separates mapping from decoding and uses a convolutional code with a Viterbi decoder.Source: §4.3 · pp. 16–18
• With a linear convolutional code shared by the nodes, the XOR of channel-coded symbols equals the encoding of the XOR source symbols, allowing a conventional Viterbi decoder.The implementation uses a log-max approximation because the stated decision rule has high computational complexity.Source: §4.3 · p. 17
• The XOR decision rule selects the nearest point in a constellation formed from both end nodes’ channel gains and can be used with non-Gaussian noise.The paper presents the mapping for BPSK modulation.Source: §4.3 · pp. 17–18, Table 1 · p. 18

5.1. FPNC Implementation over Software Radio Platform

• The three-node GNU Radio testbed uses commodity PCs and USRPs, and experiments use 802.11g channel 1 at 2.412 GHz with 1,000 packets per power level.SNR ranges from 5 to 20 dB, with 1,500-byte packets.Source: §5.1 · p. 18, §5.2 · p. 19

5.2.1. Time-Synchronous FPNC versus Time-Asynchronous FPNC

• Asynchronous FPNC has essentially the same BER as synchronous FPNC, for both channel-coded and unchannel-coded systems.The tested eight-sample offset is within the 16-sample cyclic prefix; other tested offsets had similar results.Source: §5.2.1 · p. 19, Figure 13 · p. 20

5.2.2. FPNC versus Other Approaches for TWRC

• The scheme comparison measures frame-error rates for channel-coded systems and derives throughput per direction from those rates.The compared approaches are FPNC, straightforward network coding, and traditional scheduling.Source: §5.2.2 · pp. 20–21
• Above 19 dB, FPNC throughput is approximately 99% higher than traditional scheduling and 49% higher than straightforward network coding.These gains are reported as essentially matching the ideal slot-counting gains of 100% and 50%, respectively.Source: §5.2.2 · p. 21, Figure 14 · p. 22
• Below about 17 dB, FPNC performs worse than straightforward network coding and traditional scheduling in this implementation.The authors attribute this likely to XOR-CD losing information before channel decoding; a joint CNC scheme could potentially improve performance at higher implementation complexity.Source: §5.2.2 · p. 21

6. Conclusion and Future Work

• FPNC performs XOR mapping on simultaneously received signals in OFDM's frequency domain, and the relay handles subcarrier signals separately when delay spread is shorter than the cyclic prefix.After the discrete Fourier transform, the subcarrier signals are isolated under that delay-spread condition.Source: §6 · p. 22
• The paper reports its first implementation of relay-side XOR mapping on simultaneously received signals, addressing implementation challenges including carrier-frequency-offset compensation and channel estimation.The implementation also addresses FPNC mapping.Source: §6 · p. 22
• Time-domain symbol asynchrony does not cause performance degradation in FPNC, according to the experimental results.The paper reports this result as validation of FPNC's advantage.Source: §6 · p. 23
• 99% versus traditional scheduling (TS) and 49% versus strait-forward network coding (SNC) throughput gains are reported above 19 dB SNR.These gains are reported in the high-SNR regime.Source: §6 · p. 23
• The implementation prioritizes simplicity: XOR-CD underperforms known methods at low SNR, CFO compensation uses the mean of the two end nodes' CFOs, and 802.11 constraints may leave better alternatives unexplored.The authors say better CFO compensation methods await further theoretical study.Source: §6 · p. 23