Eight-Block Region-Preserving Null in DBU-OFDM

• Eight-Block Region-Preserving Null is a precoding scheme in DBU-OFDM that divides data subcarriers into eight blocks, maintaining strict isolation of pilots and nulls.
• The design employs a block-diagonal unitary transform parameterized by Householder reflections and phase matrices to ensure strict unitarity and low-complexity equalization.
• It achieves improved PAPR reduction and enhanced sensing reliability by balancing frequency-domain diversity with computational efficiency in integrated sensing and communication systems.

Searching arXiv for the cited DBU-OFDM paper and closely related context papers. Eight-Block Region-Preserving Null denotes a specific precoding configuration within deep block-unitary precoded OFDM (DBU-OFDM) in which the trainable transform is restricted to eight blocks of data subcarriers, while pilot subcarriers are left unchanged and null subcarriers remain identically zero. In the DBU-OFDM formulation, the construction is explicitly structure-preserving: it retains the DFT-based OFDM signal model, preserves pilot/null resource regions exactly, and remains compatible with low-complexity frequency-domain equalization (Luo et al., 11 Apr 2026). The design is realized by a block-diagonal unitary transform on the data subcarriers, with each block parameterized by recursive Householder reflections and optional diagonal phases, so that strict unitarity is maintained throughout training and deployment.

1. Definition and signal model

In the DBU-OFDM framework, an OFDM system with $N$ subcarriers is partitioned into three disjoint index sets: $I_d$ for data subcarriers, $I_p$ for pilot subcarriers, and $I_0$ for null subcarriers, with $N_{\text{data}} + N_p + N_{\text{null}} = N$ (Luo et al., 11 Apr 2026). The defining “region-preserving null/pilot protection” property is that the trainable precoder acts only on data subcarriers, leaves pilot subcarriers unchanged, and keeps null subcarriers identically zero. This prevents any mixing of data into pilot or null regions and preserves both pilots and zeros exactly.

Let $F_N$ denote the normalized $N \times N$ DFT matrix. The frequency-domain OFDM symbol $x \in \mathbb{C}^N$ is defined by

$$x[I_d] = U_d\, d,\quad x[I_p] = p,\quad x[I_0] = 0,$$

where $d \in \mathbb{C}^{|I_d|}$ are data symbols, $I_d$0 are known pilots, and $I_d$1 is the data-only unitary transform. The time-domain signal is

$I_d$2

The unitarity constraint

$I_d$3

preserves energy, avoids noise enhancement, and maintains OFDM’s DFT-diagonalization structure.

This arrangement places the learning capacity entirely inside the data region. A plausible implication is that the design is intended to adapt waveform properties without sacrificing the interoperability of standard OFDM resource mapping, channel estimation, and guard-band handling.

2. Eight-block block-unitary construction

The “eight-block” aspect is realized by partitioning the data-subcarrier set into $I_d$4 disjoint blocks,

$I_d$5

The prescribed practical choice is to use contiguous blocks aligned to the native frequency order so as to avoid crossing pilot/null regions, while comb pilots remain fixed in $I_d$6 and guard/DC tones remain in $I_d$7 (Luo et al., 11 Apr 2026).

The corresponding data transform is block diagonal:

$I_d$8

with

$I_d$9

Each block mixes only the subcarriers inside its own subset $I_p$0. Consequently, no energy leaks into pilots or nulls, and inter-block boundaries are preserved.

The framework allows blocks to be contiguous or non-contiguous, provided they do not include indices from $I_p$1 or $I_p$2. The stated practical recommendation is contiguous blocks matched to the comb pilot grid, because this ensures structural compatibility and straightforward resource mapping. The paper further states that the eight-block choice balances diversity and complexity: larger blocks yield broader mixing and stronger diversity but higher transform complexity, whereas $I_p$3 offers moderate-range mixing well-suited for integrated sensing and communication (ISAC) and low-complexity equalization (Luo et al., 11 Apr 2026).

3. Householder parameterization and strict unitarity

For each block $I_p$4, DBU-OFDM parameterizes the block transform as a product of Householder reflections, optionally followed by a diagonal phase matrix:

$I_p$5

where

$I_p$6

Each $I_p$7 is Hermitian and unitary. The diagonal phase term is

$I_p$8

with real trainable phases $I_p$9. The equivalent normalized form

$I_0$0

is also given (Luo et al., 11 Apr 2026).

Because these blockwise transforms are assembled into a block-diagonal $I_0$1, pilot and null preservation follows directly from construction: $I_0$2 and $I_0$3 remain unchanged. The parameterization is described as strictly unitary at all times, thereby avoiding re-orthogonalization. Numerical stability and differentiability are attributed to the smooth dependence on $I_0$4 and $I_0$5, while complexity is controlled by the number of reflections $I_0$6 per block.

The practical configuration example gives representative values for $I_0$7: $I_0$8 reflections for mid-complexity, $I_0$9 for stronger shaping, and $N_{\text{data}} + N_p + N_{\text{null}} = N$0 for PAPR-focused training. The parameter count per block is stated as $N_{\text{data}} + N_p + N_{\text{null}} = N$1 complex vectors $N_{\text{data}} + N_p + N_{\text{null}} = N$2 plus $N_{\text{data}} + N_p + N_{\text{null}} = N$3 real phases $N_{\text{data}} + N_p + N_{\text{null}} = N$4, with total parameter count approximately $N_{\text{data}} + N_p + N_{\text{null}} = N$5 (Luo et al., 11 Apr 2026).

4. Compatibility with OFDM equalization and training objectives

After CP removal and DFT, the frequency-selective channel model is

$N_{\text{data}} + N_p + N_{\text{null}} = N$6

where $N_{\text{data}} + N_p + N_{\text{null}} = N$7 and $N_{\text{data}} + N_p + N_{\text{null}} = N$8 is AWGN (Luo et al., 11 Apr 2026). Comb pilots support low-complexity per-subcarrier channel estimation,

$N_{\text{data}} + N_p + N_{\text{null}} = N$9

followed by interpolation to data tones. Data-subcarrier equalization produces $F_N$0, after which symbol recovery is performed by the inverse block-unitary transform,

$F_N$1

Since $F_N$2 acts only on data tones and is block diagonal, standard OFDM equalization remains compatible and low complexity.

The training loss is given as a weighted multi-objective function,

$F_N$3

The examples listed for its terms are supervised bit-wise BCE via soft demapping after equalization and $F_N$4 for $F_N$5, MSE of estimated range/velocity using differentiable delay-Doppler soft-max processing for $F_N$6, and a CCDF-tail proxy for $F_N$7 based on PAPR exceedance above a target threshold $F_N$8:

$F_N$9

with $N \times N$0 controlling penalty strength.

The time-domain PAPR is defined by

$N \times N$1

The stated rationale for block-unitary mixing is that, within each block, frequency-domain diversity spreads each symbol across multiple subcarriers, so post-equalization recovery depends on the aggregate channel over the block rather than a single tone. This reduces sensitivity to deep fades on individual subcarriers (Luo et al., 11 Apr 2026).

5. Sensing formulation and ISAC role

DBU-OFDM is positioned as an integrated sensing and communication waveform, and the eight-block region-preserving null design is part of that broader ISAC objective (Luo et al., 11 Apr 2026). For direct OFDM sensing after equalization, the received frequency-domain observation is modeled as

$N \times N$2

where $N \times N$3 are known effective symbols. Matched demodulation is written as

$N \times N$4

The delay and Doppler steering vectors are

$N \times N$5

$N \times N$6

and the delay-Doppler correlation function is

$N \times N$7

Differentiable estimation is implemented through a soft-max over a delay-Doppler grid,

$N \times N$8

Multipath SIC is then applied to extract $N \times N$9 dominant paths, and the range/velocity estimates are

$x \in \mathbb{C}^N$0

The sensing loss is

$x \in \mathbb{C}^N$1

The paper states that DBU-OFDM improves range and velocity estimation especially in dimension-limited settings, particularly for small $x \in \mathbb{C}^N$2, by producing better-behaved frequency-domain amplitudes and phase structure after block-unitary mixing while preserving pilots and nulls for coexistence (Luo et al., 11 Apr 2026). This suggests that the eight-block restriction is not merely a complexity device; it also constrains the learned transform so that sensing-oriented structure is improved without disrupting OFDM compatibility.

6. Configuration example, complexity, and empirical performance

A concrete configuration is given using the paper’s Config. 3: $x \in \mathbb{C}^N$3 subcarriers, CP length $x \in \mathbb{C}^N$4, guards $x \in \mathbb{C}^N$5 on each edge, $x \in \mathbb{C}^N$6 DC nulls, and $x \in \mathbb{C}^N$7 comb pilots, yielding

$x \in \mathbb{C}^N$8

data subcarriers (Luo et al., 11 Apr 2026). The corresponding index sets are defined by guard/DC indices for $x \in \mathbb{C}^N$9, comb pilots for $$x[I_d] = U_d\, d,\quad x[I_p] = p,\quad x[I_0] = 0,$$0, and the remaining tones for $$x[I_d] = U_d\, d,\quad x[I_p] = p,\quad x[I_0] = 0,$$1. The eight-block partition then splits $$x[I_d] = U_d\, d,\quad x[I_p] = p,\quad x[I_0] = 0,$$2 into contiguous blocks of approximately equal size, for example about $$x[I_d] = U_d\, d,\quad x[I_p] = p,\quad x[I_0] = 0,$$3–$$x[I_d] = U_d\, d,\quad x[I_p] = p,\quad x[I_0] = 0,$$4 tones per block, aligned to pilot positions so that no block contains pilot or null indices.

The implementation workflow is stated explicitly: build $$x[I_d] = U_d\, d,\quad x[I_p] = p,\quad x[I_0] = 0,$$5, $$x[I_d] = U_d\, d,\quad x[I_p] = p,\quad x[I_0] = 0,$$6, and $$x[I_d] = U_d\, d,\quad x[I_p] = p,\quad x[I_0] = 0,$$7; partition $$x[I_d] = U_d\, d,\quad x[I_p] = p,\quad x[I_0] = 0,$$8 into eight contiguous blocks; initialize trainable Householder vectors and phase parameters in each block; form each $$x[I_d] = U_d\, d,\quad x[I_p] = p,\quad x[I_0] = 0,$$9; assemble $d \in \mathbb{C}^{|I_d|}$0; construct $d \in \mathbb{C}^{|I_d|}$1 with data, pilots, and nulls placed in their designated regions; compute $d \in \mathbb{C}^{|I_d|}$2; and train with the multi-objective loss. Low-complexity detection is equally direct: after equalization, apply $d \in \mathbb{C}^{|I_d|}$3 blockwise, then standard demapping and decoding.

The algorithmic complexity per OFDM symbol for the forward transform is stated as follows. For block $d \in \mathbb{C}^{|I_d|}$4, each Householder reflection requires one inner product $d \in \mathbb{C}^{|I_d|}$5 of order $d \in \mathbb{C}^{|I_d|}$6 and one scaled subtraction $d \in \mathbb{C}^{|I_d|}$7 of order $d \in \mathbb{C}^{|I_d|}$8, giving approximately $d \in \mathbb{C}^{|I_d|}$9 complex operations per block and total complexity $I_d$00. The inverse $I_d$01 has identical cost (Luo et al., 11 Apr 2026).

Hardware feasibility is supported by FPGA results for a pipelined Householder cascade with a merged two-stage implementation, reusing the same kernel for forward and inverse transforms. The reported throughput is up to approximately $I_d$02 MS/s, with latency scaling with $I_d$03: for $I_d$04, about $I_d$05s latency and $I_d$06 MS/s; for $I_d$07, about $I_d$08s latency and $I_d$09 MS/s; and for $I_d$10, about $I_d$11s latency and about $I_d$12 MS/s. The associated resource and power figures are also reported, confirming tunable complexity-performance tradeoffs and microsecond-level latency suitability.

Empirically, the paper reports that DBU-OFDM achieves PAPR tail performance close to block-pilot DFT-s-OFDM while retaining comb-type pilots. With oversampling by $I_d$13, it is approximately $I_d$14 dB worse than block-pilot DFT-s-OFDM in the low-PAPR region but offers more than $I_d$15 dB gain over conventional OFDM for CCDF around $I_d$16 to $I_d$17. Over-the-air USRP validation shows average PAPR reduction of about $I_d$18 dB and tail reduction of about $I_d$19 dB. Communication performance shows BER and BLER improvements in frequency-selective fading via within-block diversity, with BLER gains consistent across SNR and stronger gains for larger blocks. For sensing, lower range and velocity MSE than conventional OFDM is reported, especially for small $I_d$20. For $I_d$21QAM at $I_d$22, the measured EVM is approximately $I_d$23 for OFDM and approximately $I_d$24 for DBU-OFDM, indicating nearly unchanged communication quality (Luo et al., 11 Apr 2026).

7. Significance, scope, and best-practice interpretation

Within DBU-OFDM, the eight-block region-preserving null design is presented as a practical intermediate solution between conventional model-based OFDM waveforms and unconstrained neural transceivers (Luo et al., 11 Apr 2026). Its significance lies in the simultaneous enforcement of several structural constraints: pilots and nulls are untouched by learning, CP and DFT continue to diagonalize frequency-selective channels, single-tap frequency-domain equalization remains usable, and the trainable component remains exactly unitary.

The guidance given in the paper is operationally specific. Blocks should be chosen within $I_d$25 so as to avoid $I_d$26 and $I_d$27, preferably with edges aligned to the pilot grid. Moderate $I_d$28 values such as $I_d$29–$I_d$30 are suggested as a starting point, with larger values such as $I_d$31 reserved for PAPR-tail optimization. The weights $I_d$32, $I_d$33, and $I_d$34 should be tuned according to the desired communication, sensing, and PAPR tradeoff. Pilots and nulls should be preserved by construction and never mixed into $I_d$35.

The resulting interpretation is technically narrow and precise. “Region-preserving” refers to exact preservation of pilot and null subcarrier regions; “null” refers to the enforced zero-valued null resources; and “eight-block” refers to the partition of data subcarriers into eight unitary mixing regions. In this form, the method delivers controlled diversity and PAPR shaping while retaining OFDM structure and hardware realizability. A plausible implication is that the approach is designed for deployment scenarios in which strict compatibility constraints rule out end-to-end unconstrained learned waveforms, yet performance gains are still sought through trainable, architecture-level adaptation.