Diagonally Reconstructed Channel Estimation for MIMO-AFDM with Inter-Doppler Interference in Doubly Selective Channels

I. INTRODUCTION

• The introduction motivates AFDM through high-mobility channels, where Doppler undermines OFDM subcarrier orthogonality and OTFS embedded-pilot estimation has heavy guard-symbol overhead.Existing AFDM estimation approaches are constrained by channel assumptions or computational complexity in MIMO-AFDM.Source: §I · pp. 2–3

A. AFDM modulation

• AFDM multiplexes information symbols on orthogonal chirps through inverse DAFT, and its modulation and demodulation can use an OFDM modulator/demodulator as an inner core.A chirp-periodic prefix is added before transmission to address multipath propagation.Source: §I · p. 2, A. AFDM modulation · pp. 5–6

THREE CASES OF DELAY-DOPPLER PROFILE IN DOUBLY SELECTIVE CHANNEL

• MIMO-AFDM can achieve full diversity in doubly selective channels, with diversity order equal to receive antennas times paths separable in delay or Doppler.This extends the SISO-AFDM diversity analysis to MIMO-AFDM with perfect CSI.Source: THREE CASES OF DELAY-DOPPLER PROFILE… · p. 4
• Fractional Doppler creates four pilot-interference types, with IDoI the most serious and impossible to eliminate when estimating each path’s gain, delay, and Doppler.The estimation challenge is especially relevant because prior EPA-AML has high computation complexity and is unsuitable for MIMO-AFDM.Source: §I · p. 3, THREE CASES OF DELAY-DOPPLER PROFILE… · p. 4
• EPA-DR uses iterative, precalculable transform factors to diagonally reconstruct the AFDM subchannel matrix and directly estimate the effective channel, avoiding IDoI.The method is described as low-complexity.Source: THREE CASES OF DELAY-DOPPLER PROFILE… · p. 4
• The proposed AFDMA resource allocation uses differences between base-station and user delay-Doppler profiles to reduce guard-symbol overhead.The scheme is introduced for single-antenna users and a multi-antenna base station.Source: THREE CASES OF DELAY-DOPPLER PROFILE… · p. 4

B. Channel model

• The doubly selective channel is modeled with P paths, each having a gain, integer delay, and normalized Doppler shift; the maximum delay and Doppler are assumed known.The Doppler shift is expressed as integer and fractional components.Source: B. Channel model · p. 6

C. AFDM demodulation

• AFDM demodulation applies an N-point DAFT to the received time-domain signal, producing DAFT-domain symbols and noise.The effective time-domain channel is a sum of path subchannels, each formed from a Doppler frequency-shift matrix and a cyclic-shift matrix.Source: C. AFDM demodulation · p. 7

D. Input-output relationship

• In the DAFT domain, the input-output relationship combines the path subchannel matrices into an effective channel matrix weighted by path gains.The spacing factor k_ν is used to combat fractional Doppler; c2 is constrained to an arbitrary irrational value or a sufficiently small rational value.Source: D. Input-output relationship · pp. 7–8

E. MIMO-AFDM System

• The input-output relationship combines each transmit antenna’s signal through its channel matrix at each receive antenna, with additive noise.The effective MIMO channel matrix is formed from the channel matrices between receive- and transmit-antenna pairs.Source: E. MIMO-AFDM System · pp. 8–9
• With c1 satisfying (15), MIMO-AFDM achieves full diversity order PNr for a linear time-varying channel with maximum normalized delay lmax and Doppler αmax.Source: E. MIMO-AFDM System · p. 9

A. Effective channnel matrix analysis

• With fractional Doppler, subchannel-matrix entries are largest at a central point in each row and decay as entries move away, distributing nonzero values diagonally.For integer Doppler, each row has only one nonzero central entry.Source: A. Effective channnel matrix analysis · pp. 9–10
• A path’s central-point index depends jointly on its delay-Doppler profile, αmax, and kν; possible integer-profile central points do not overlap and form delay blocks with protection bands.Each delay block has size 2αmax + 1, with protection bands of size kν at its head and tail.Source: A. Effective channnel matrix analysis · p. 10
• When L + 1 ≪ N, each row and column of the effective channel matrix has a large-value nonzero band of size L + 1.Source: A. Effective channnel matrix analysis · p. 10
• The entries’ indices and values in each large-value band reflect propagation-path delay-Doppler profiles and channel gains, respectively, so each band contains all CSI.Source: A. Effective channnel matrix analysis · p. 11

B. EPA channel estimation in MIMO-AFDM

• EPA embeds one pilot per transmit antenna and places zero-guard bands of size L beside the pilots to approximately separate the extracted channel bands at each receive antenna.Source: B. EPA channel estimation in MIMO-AFDM · p. 11

C. Interference analysis among received pilot symbols

• Estimating a path’s channel gain from received pilots assumes its delay-Doppler profile is known, but separating its desired component requires knowledge of other paths’ gains.The residual received-pilot components therefore interfere with estimation of the target path’s gain.Source: C. Interference analysis among… · pp. 12–13
• IDoI occurs between paths with the same delay and different Doppler shifts, whose subchannel matrices overlap severely and whose interference can be comparable to the desired pilot component.Source: C. Interference analysis among… · p. 13
• IDI between different-delay paths is more moderate than IDoI, and increasing kν can further reduce their overlap and suppress IDI.Source: C. Interference analysis among… · pp. 13–14
• Other antennas’ pilots contribute IPI, which can become comparable to IDI where adjacent large-value bands meet; data symbols contribute less-significant IPDI, also mitigated by increasing kν.Source: C. Interference analysis among… · p. 14
• Fractional Doppler produces IDoI, IDI, IPI, and IPDI; among these, IDoI is the most serious and cannot be eliminated when estimating each path’s gain, delay, and Doppler.Source: C. Interference analysis among… · p. 14

IV. EPA-DR CHANNEL ESTIMATION FOR MIMO-AFDM

• The proposed low-complexity EPA-DR method estimates the MIMO effective channel matrix based on diagonal reconstructability of the AFDM subchannel matrix.Source: §IV · p. 14

A. Diagonal reconstructability of AFDM subchannel matrix

• The transform factor linking subchannel-matrix entries at neighboring diagonal positions does not depend on Doppler shift.The paper therefore writes it as T(l, m, m′).Source: A. Diagonal reconstructability of AFDM… · p. 15
• Once an entry H(l,j)[m, m′] is obtained, the next diagonal entry is calculated by multiplying it by T(l, m, m′).Source: A. Diagonal reconstructability of AFDM… · p. 15
• Acquiring any one column of H(l,j) allows the whole subchannel matrix to be reconstructed diagonally.Source: A. Diagonal reconstructability of AFDM… · p. 16

B. EPA-DR channel estimation scheme

• Diagonal reconstruction uses the dominant component within a delay block to estimate the next diagonal entry, instead of treating that component as IDoI.The neighboring coordinate remains in the same delay block, so the dominant component supports the next estimate.Source: B. EPA-DR channel estimation scheme · p. 16
• After diagonal reconstruction, the residual IDI, IPI, IPDI, and noise are modeled as additive noise; the severe IDoI is absent.The residual terms are described as much smaller in energy than the main component.Source: B. EPA-DR channel estimation scheme · p. 17
• Iterative estimation does not accumulate inherited error because the transform factors have unit energy and the inherited errors retain the same magnitude.The method uses previously estimated diagonal entries successively to reconstruct later entries.Source: B. EPA-DR channel estimation scheme · p. 17
• EPA-DR obtains large-valued nonzero bands from received pilots, applies threshold-based magnitude detection, and iteratively reconstructs the effective channel matrix.The procedure is repeated across receive–transmit antenna pairs, whose estimates are assembled into the MIMO effective channel matrix.Source: B. EPA-DR channel estimation scheme · pp. 17–18

C. Performance assessment of EPA-DR algorithm

• MIMO-AFDM’s pilot and guard overhead is nearly half of MIMO-OTFS’s, preserving AFDM’s channel-estimation overhead advantage in MIMO.The passage gives the MIMO-AFDM overhead expression as (N_t + 1)L + N_t and states that the OTFS overhead is nearly twice as large.Source: C. Performance assessment of EPA-DR… · p. 18
• EPA-DR’s transform factors can be precomputed once the AFDM parameters are fixed, with 4L + 1 factors needed in practice.The factors remain reusable despite changing channels.Source: C. Performance assessment of EPA-DR… · p. 19
• EPA-DR avoids EPA-AML’s path-count requirement and brute-force path-parameter search by estimating the effective matrix directly, reducing computation complexity.EPA-AML estimates delay and Doppler shifts and channel gains before calculating the effective matrix; the comparison also identifies its serious IDoI.Source: C. Performance assessment of EPA-DR… · p. 19

V. ORTHOGONAL RESOURCE ALLOCATION AND CHANNEL ESTIMATION FOR AFDMA

• The AFDMA design considers N_U single-antenna users and a base station with N_BS antennas, exploiting differences in their delay–Doppler profiles.The base station is assumed to know L_max and every user’s L_u, while each user knows L_max and its own L_u.Source: §V · pp. 19–20

A. Downlink

• In downlink, guard-symbol blocks separate users; each pilot and data block needs surrounding guards sized at least L_Nu and L_u, respectively.When these conditions hold, users can separate their received data symbols from those of other users.Source: A. Downlink · p. 20
• Downlink data-block sizes can be tuned to users’ throughput demands, subject to the stated sum constraint.Each user uses associated received symbols for EPA-DR channel estimation and data detection.Source: A. Downlink · p. 20
• Changing the order of two downlink resource blocks can require enlarging a guard block from L_2 to L_3, leaving fewer data symbols.The example notes that overlapping user data symbols are acceptable when the target user’s symbols remain uncontaminated.Source: A. Downlink · p. 20

B. Uplink

• In uplink, users receive orthogonal resource blocks according to throughput demands, with guard-symbol blocks placed to avoid inter-user interference.The base station can separate users’ received symbols and perform EPA-DR estimation and data detection; user ordering by L_u is unnecessary.Source: B. Uplink · p. 21
• The downlink and uplink arrangements reduce guard-symbol use by exploiting differences in users’ effective-channel nonzero-band sizes.The paper characterizes the resulting AFDMA system as efficient and easy to implement, with flexibility for differing communication requirements.Source: B. Uplink · pp. 21–22

VI. SIMULATION RESULTS

• The MIMO-AFDM simulations consider integer and fractional Doppler, independent symbols per transmit antenna, and path gains distributed as CN(0, 1/P).For large frame sizes, the evaluation adopts a message-passing detector with complexity order O(n_iter N N_r S |A|).Source: §VI · p. 22

A. Perfect CSI

• For two and three paths, the reported diversity orders are 2 and 3 for SISO-AFDM and 4 and 6 for 2 × 2 MIMO-AFDM; Theorem 1 also holds for fractional Doppler.The integer-Doppler results are presented as verifying Theorem 1.Source: A. Perfect CSI · pp. 22–23
• With integer Doppler, MIMO-AFDM significantly outperforms MIMO-OFDM in BER and achieves nearly the same BER as OTFS in SISO and MIMO.The paper attributes the MIMO-OFDM gap to severe inter-carrier interference under heavy Doppler, and AFDM’s near-OTFS BER to its equivalent delay–Doppler channel representation in the DAFT domain.Source: A. Perfect CSI · p. 23
• 2 × 2 MIMO-AFDM uses 134 pilot-and-guard slots (13.08% of its frame), versus MIMO-OTFS’s 238 slots (23.24%).The paper links the overhead difference to a spectral-efficiency gap shown in Fig. 11.Source: Figure 11 · p. 23, A. Perfect CSI · p. 24

SIMULATION PARAMETERS FOR IMPERFECT CSI

• At SNRp = 30 dB, estimated-CSI BER shows only marginal degradation versus perfect CSI, while EPA-DR provides nearly the same channel-estimation performance with much lower complexity.The conclusion reports BER performance very close to ideal CSI.Source: SIMULATION PARAMETERS FOR IMPERFECT CSI · p. 25, §VII · p. 27
• Increasing kν improves BER by widening protection bands and reducing IDI, IPI, and IPDI, but raises overhead to 6.05%, 11.33%, and 18.40% for kν = 1, 4, and 8.The corresponding pilot and guard overheads are 62, 116, and 188 symbols.Source: SIMULATION PARAMETERS FOR IMPERFECT CSI · p. 25
• The threshold has an optimum around 6N0; values above it omit low-energy pilot CSI, while lower values retain high-energy noise interference.BER becomes less sensitive to threshold variation as SNRp increases.Source: SIMULATION PARAMETERS FOR IMPERFECT CSI · p. 26
• With fractional Doppler, BER first improves and then degrades as SNRp rises because stronger pilots reduce estimation noise but magnify IPDI in received data symbols.This creates a tradeoff between channel-estimation accuracy and pilot-data separability.Source: SIMULATION PARAMETERS FOR IMPERFECT CSI · p. 26

VII. CONCLUSION

• MIMO-AFDM achieves full diversity, BER similar to MIMO-OTFS and significantly better than MIMO-OFDM, and a spectral-efficiency advantage over MIMO-OTFS.The diversity result is stated for doubly selective channels.Source: §VII · pp. 26–27
• EPA-DR uses diagonal reconstructability to calculate the effective channel matrix directly, eliminating severe fractional-Doppler IDoI; the paper also proposes orthogonal resource allocation for AFDMA.Source: §VII · p. 27

• The derivation assumes integer Doppler for convenience, while simulations in Section VI-A show the derived diversity order also holds for fractional Doppler.Source: APPENDIX · p. 27
• Each row of HMIMO has PNt nonzero elements; the vectorized relation uses received and transmitted symbol matrices of sizes N × Nr and N × Nt.Source: APPENDIX · p. 28
• The high-SNR PEP exponent is kNr, and the minimum rank k across distinct transmitted-symbol pairs determines the BER diversity order.Source: APPENDIX · p. 29
• When N ≫ (lmax + 1)(2αmax + 1) and AFDM parameters are tuned as specified, the diversity order is PNr, giving full diversity.Source: APPENDIX · p. 29