Channel Estimation for OFDM via Delay-Doppler Refinement

I. Introduction

• High Doppler breaks OFDM subcarrier orthogonality and causes intercarrier interference, while reduced coherence time and limited pilots impair conventional TF estimation.DD-domain channel representations offer a quasi-static, sparse view of rapidly time-varying TF channels.Source: §I · p. 1
• DD-domain pilot arrangements can conflict with practical OFDM frames, and an earlier cross-domain method assumed an ICI-free TF channel, causing practical model mismatch and an error floor.The cited approach used TF estimation followed by DD refinement; the paper's prior method instead used DD-domain coarse estimation and TF-domain least squares.Source: §I · p. 1
• The proposed OFDM estimator first applies TF-domain LMMSE estimation, then refines the estimate in DD using threshold-based channel-parameter estimation.It treats the effective DD channel matrix as an aggregation of channel responses with different phase shifts under integer delay and Doppler.Source: §I · p. 1
• The authors report significantly better NMSE than conventional OFDM channel-estimation schemes.The comparison is reported for the proposed method under the paper's numerical evaluation.Source: §I · p. 2

A. OFDM Transmissions over Doubly-Selective Channel

• A mismatch between the DFT sizes for the two TF-domain symbol representations spreads symbols across frequency and introduces pilot-data interference at the receiver.The text attributes this effect to CP appending and removal, and says estimation uses the model in (13) to obtain a structured representation for cross-domain transformation.Source: A. OFDM Transmissions over… · pp. 2–3
• The transmit signal is formed by adding an OFDM cyclic prefix and an additional OTFS cyclic prefix before pulse shaping.The OFDM prefix uses ACP, while the additional prefix is represented by the OTFS CP addition matrix.Source: A. OFDM Transmissions over… · p. 2
• The channel is modeled as a sparse sum of paths with gains, delays, and Doppler shifts, and the analysis assumes zero fractional delay and Doppler indices.The fractional indices are set to zero on the stated assumption of sufficient delay and Doppler resolution.Source: A. OFDM Transmissions over… · p. 2
• After matched filtering and CP removal, the received signal is represented using the effective time-domain channel matrix, then transformed by DFTs to obtain TF-domain observations.The compact received-signal representation follows stacking the received symbols and removing the OTFS CP; OFDM CP removal is also applied.Source: A. OFDM Transmissions over… · p. 2

B. Properties of the TF Domain Effective Channel Matrix

• The effective TF channel matrix has a structured block arrangement, with H_j,0^TF denoting its M × M subchannel blocks for j = 0, …, N − 1.The received signal is expressed block-wise using this structure.Source: B. Properties of the TF Domain… · p. 3
• Doppler makes the diagonal subchannel block dense and produces ICI, whereas the adjacent block accounts for ISI caused by multipath delay.The effective channel matrix is described in terms of M × M subchannel matrices.Source: B. Properties of the TF Domain… · p. 3
• A delayed OFDM symbol extends into the following symbol interval, so its observation has length 2M and the full received signal is assembled by aggregating the per-symbol observations with zero-padding.The observations and transmit symbols are partitioned into N symbol vectors.Source: B. Properties of the TF Domain… · p. 3

A. TF Domain Dense IOR

• The derivation focuses channel estimation on the nonzero elements of H_TF and rewrites the input-output relationship in dense form.The vector relationship is expressed as vec(H_TF) + w = X_TF h_TF + w.Source: A. TF Domain Dense IOR · p. 3
• The block-wise relationship is further rewritten using an equivalent channel vector for each OFDM symbol and its corresponding equivalent transmit matrix.The dense representation uses h_d,TF, a dense format of h_TF.Source: A. TF Domain Dense IOR · pp. 3–4
• The equivalent transmit matrices are arranged as submatrices within the dense transmit matrix.The text identifies X_TF^(i) as a submatrix of X_d.Source: A. TF Domain Dense IOR · p. 4

B. TF Domain Estimation

• A full-size LMMSE estimator uses the received signal, equivalent pilot matrix, channel prior mean, and covariance to obtain a coarse TF-domain estimate.The prior mean and covariance are initialized through Monte Carlo simulations.Source: B. TF Domain Estimation · p. 4
• Applying LMMSE produces an a posteriori channel estimate and its residual error covariance matrix.Source: B. TF Domain Estimation · p. 4
• When the channel estimate is unbiased, the trace of its residual error covariance represents the estimation MSE.Source: B. TF Domain Estimation · p. 4
• The a posteriori TF estimate is reshaped with appropriate zero-padding and transformed into the DD domain as the initial estimate for refinement.The transformation uses the SFFT, whose inverse is also identified in the supplied passages.Source: B. TF Domain Estimation · p. 4

C. DD Domain Channel Matrix Properties

• Each DD-domain channel-matrix column represents an observation of the discretized channel response with a particular location shift and phase shift.Across the matrix, this yields MN observations, which can be viewed as responses to MN uniquely placed parallel pilots.Source: C. DD Domain Channel Matrix Properties · p. 4
• The location shifts across observations arise from varying DD-domain pilot positions and the twisted-convolution effect.Source: C. DD Domain Channel Matrix Properties · pp. 4–5
• The first matrix column corresponds to a pilot at (0, 0), producing a channel response without location shift.Source: C. DD Domain Channel Matrix Properties · p. 4
• A pilot at delay-Doppler location (1, 0), corresponding to HDD’s second column, yields an observed delay-Doppler channel response shifted by one delay unit with corresponding phase shifts.Source: C. DD Domain Channel Matrix Properties · pp. 4–5

D. DD Domain Refinement

• The method suppresses effective DD-domain channel coefficients below a threshold based on the average LMMSE estimation MSE.The fixed threshold 3√N0 is not used because estimation first applies LMMSE in the TF domain.Source: D. DD Domain Refinement · p. 5
• It infers path delay and Doppler indices from row and column indices, compensates phase, then averages MN DD-channel estimates and thresholds them again.The passages describe this averaging and second threshold as potentially reducing both missed estimations and false alarms.Source: D. DD Domain Refinement · p. 5
• Algorithm 1 summarizes the refinement steps: threshold coefficients, infer path indices, compensate phase, and average estimates.Source: D. DD Domain Refinement · p. 5
• Estimated fading gains and delay-Doppler indices reconstruct the DD channel matrix, which is transformed back to TF for the refined estimate.Source: D. DD Domain Refinement · p. 5

IV. Numerical Results

• The simulation uses M = N = 16, three independently resolvable paths, delays from [0, 2], and Doppler indices from [−3, 3].Path coefficients are independently zero-mean complex Gaussian with variance 1/P, and pilots use a lattice-type TF pattern.Source: §IV · p. 5
• ST-LS and ST-LMMSE estimate CSI at pilot locations and use nearest-nonzero-element linear interpolation elsewhere in both TF dimensions.Source: §IV · p. 5
• The STE benchmarks assume an element-wise TF input-output relationship and estimate only diagonal channel components, whereas FS-LMMSE captures ICI effects.Source: §IV · p. 6
• The proposed CE consistently outperforms all benchmarks across SNRs; STE baselines saturate near −6 dB, while DD refinement gives lower NMSE than FS-LMMSE.The reported saturation is attributed to STE's failure to account for ICI.Source: §IV · p. 6, Figure 2 · p. 6

• A full-sized LMMSE estimator produces a coarse TF estimate, which is transformed to DD for thresholding and phase-compensation refinement.The approach exploits the TF channel matrix structure to derive a dense input-output relationship.Source: §V · p. 6
• The scheme consistently outperforms conventional OFDM estimators in NMSE.Source: §V · p. 6
• Estimates from different observations are averaged, then used to reconstruct the DD channel matrix and transform it back to TF for the final estimate.Source: §V · p. 6