Dirty paper coding (DPC) is a way to transmit to several users at once. The transmitter already knows the interference one user's signal will cause another, so it pre-cancels that interference before sending, much as you would write on paper whose stains you already know. The cancellation is sequential. The user encoded first sees interference from everyone. The user encoded last sees none. The encoding order is therefore a design choice. In joint radar–communication transmitters, which must also produce a prescribed radar beam pattern, current designs fix this order in advance and optimize only the continuous beamforming matrices.
This note proposes to treat the order as a genuine optimization variable. It asks two things. First, does the order actually matter once the radar constraint is imposed? Second, if it does, can a cheap learned model of "how good is this order" find near-best orders among the $K!$ possibilities, without running the full beamforming optimizer for each one? Think of the problem as expensive single-objective optimization over permutations. The scalar objective is the converged spectral efficiency, and each evaluation is a complete iterative beamformer design.
Core Claim
Joint Transmit and Receive Beamforming Design for DPC-Based MIMO DFRC Systems maximizes weighted sum rate in a dual-function radar–communication system. It uses DPC and imposes a hard radar covariance equality constraint $\mathbf{F}\mathbf{F}^H=\mathbf{R}_x$, where $\mathbf{F}$ is the transmit beamformer and $\mathbf{R}_x$ the prescribed sensing covariance. The paper solves the problem by block coordinate descent with majorization–minimization steps on a matrix manifold, but holds the encoding order fixed.
In the unconstrained Gaussian broadcast channel, classical theory makes order largely irrelevant. Sum rate is order-invariant when transmit covariances are optimized, and for weighted sum rate the order follows the weights. That theory is outside this file, and I state it as background. The fixed covariance $\mathbf{R}_x$, a fixed number of streams per user, and the iterative solver's local optima all remove the freedom that makes order irrelevant.
We therefore hypothesize that under the radar constraint, the best order and a fixed order can differ in spectral efficiency by a margin that matters. We further hypothesize that this gap grows with the number of users $K$ and with how poorly the radar subspace aligns with the users' channels.
The second claim is methodological. We expect a surrogate-guided permutation search to come within 1–2% of the best-order optimum while making only a few dozen full solver calls instead of $K!$. In that search, a probabilistic model over rankings proposes orders, a cheap model predicts their value, and only the most promising orders get full evaluations.
The idea is not "apply an optimizer to DPC." Both cited communication papers already do that for continuous variables. The idea is to expose a discrete variable those papers fix or set by hand, and to measure whether searching over it is worth the compute.
Research Program
The program has three stages, and the first can kill the project.
Stage one: map the order landscape where exhaustive search is affordable. For $K\le 6$ users, we run the DPC–DFRC solver from the paper above to convergence for every permutation. We record the full distribution of converged spectral efficiency over orders. The sweep covers SNR, the radar beamwidth encoded in $\mathbf{R}_x$, and channel–radar alignment. If the spread between median and best order is negligible everywhere, the problem does not exist, and we report that as the finding.
Stage two: build the search method. The expensive evaluation is a full solver run. The cheap signals come from two sources:
- Low-fidelity proxy. We run the same solver truncated to two or three iterations, warm-started from a shared solution. The original paper reports that a significant portion of converged spectral efficiency is reached within about ten iterations, which makes a truncated run a plausible proxy.
- Learned model. An XGBoost or random-forest regressor is fit on features of each order: precedence indicators for user pairs, channel norms by position, and each user's channel projection onto the radar subspace.
Orders are proposed by an estimation-of-distribution algorithm over permutations, using a Plackett–Luce ranking model. We borrow the population management of Un-evaluated Solutions May Be Valuable in Expensive Optimization. That paper fixes the "standing still" collapse of surrogate-assisted evolution, where only one elite per generation is evaluated, by keeping the top half of surrogate-ranked but unevaluated candidates in the breeding pool. Here, only one or two orders per generation receive a full solver run, so the same stagnation risk applies directly. Precedence features give a natural meaning to "partial ranking" information, which that paper flags as an open issue outside scalar fitness.
Stage three: scale and stress. We move to $K=8$ (40,320 orders), where exhaustive search is impractical inside a coherence interval. We also port the search to the Tomlinson–Harashima precoding (THP) transceiver of MIMO-Aided Nonlinear Hybrid Transceiver Design for Multiuser mmWave Systems Relying on Tomlinson-Harashima Precoding. THP is the practical modulo-arithmetic version of DPC, and that paper's greedy ordering, built on an MSE lower bound, is the main ordering competitor. We then repeat the comparison under the 5 ms CSI feedback delay that paper studies, to see whether a searched order is more fragile than a greedy one.
Baselines
The naive reference is the fixed order used in the DPC–DFRC paper, with random order as a second floor. The headline result of that paper, against which our gains will be framed, is "at least 73% higher spectral efficiency compared to existing methods in a high SNR regime" and a "~20-fold increase in spectral efficiency at 30 dB SNR over a benchmark scheme without DPC". Neither figure involves order optimization. I found no reported result on Lacuna for encoding-order optimization in DPC–DFRC.
The strongest existing ordering method is the greedy sort derived from the MSE lower bound in the THP mmWave paper. That paper reports "a substantial ~3dB gain at a Symbol Error Rate (SER) of $2 \times 10^{-7}$ compared to random cancellation ordering" with 32 antennas, 8 RF chains, and 4 users. We reimplement it in that exact configuration and confirm we reproduce roughly that 3 dB before using it elsewhere. In the DFRC setting we adapt it by computing the bound on the radar-constrained effective channel. We also add a weight-descending order, which theory suggests in the unconstrained case.
The oracle is exhaustive search over all orders for $K\le 6$.
The fairness rule is a matched budget: every method gets the same number of full solver calls and the same channel realizations. Hyperparameters are tuned on separate channel draws.
Three ablations isolate the key ingredients:
- the search with the learned model but without retaining unevaluated candidates, the analogue of USEA-NS in the USEA paper;
- the search with the truncated-solver proxy alone;
- a GP-based Bayesian optimizer over permutation embeddings, as a stronger generic surrogate.
Beating the field means two things. At $K=6$ we must close at least half of the greedy-to-oracle gap with 30 or fewer full calls. At $K=8$ we must beat greedy ordering by a statistically significant margin across channel draws, at no more than twice its wall-clock cost.
Why not just…?
Why not just use the greedy MSE-bound ordering? This is the real threat. If the greedy order is within a fraction of a percent of the exhaustive optimum, search is pointless. The THP paper only compares greedy ordering with random ordering, never with exhaustive search, and it describes the bound as an approximation without a global optimality guarantee. Stage one measures exactly this gap. If the gap is small across the swept regimes, we will report that the greedy order suffices and drop the search component.
Why not just run every order with a truncated solver and keep the best? At $K\le 6$ this cheap baseline may well match our method, and we include it as an ablation. At $K=8$ it still costs 40,320 truncated runs. Its accuracy also depends on early iterations ranking orders the same way converged runs do. The DPC–DFRC paper notes that fast early convergence holds especially at lower SNR, which suggests the ranking may be less faithful at high SNR, where DPC gains are largest. We measure the rank correlation between truncated and converged values directly.
Why not just drop DPC for linear precoding, where order does not exist? The file's evidence is strongly against this. The DPC–DFRC paper reports the roughly 20-fold spectral-efficiency gap at 30 dB SNR. Nonlinear Precoding in the RIS-Aided MIMO Broadcast Channel shows that linear precoding saturates when channels are poorly conditioned, while THP keeps scaling. We include a zero-forcing reference so that the value of order search can be read against the larger DPC-versus-linear gap.
Proposed Experiments
All experiments use the simulated downlink multi-user DFRC model of the DPC–DFRC paper and the mmWave hybrid THP model of the THP paper.
| Metric | Where measured |
|---|---|
| Converged weighted spectral efficiency, as a gap to the exhaustive optimum | DFRC model |
| Number of full solver calls and wall-clock time | Both models |
| Rank correlation between surrogate predictions and true values | Both models |
| SER curves | THP model only |
Three checks decide the outcome:
- Stage-one landscape spread at $K=4$ to $6$. If the best order beats greedy by less than 1% in nearly all regimes, the project stops.
- Matched-budget comparison at $K=8$. This is where search must beat both greedy ordering and truncated-solver selection.
- Delayed-CSI test. This checks whether a searched order keeps its advantage when the channel used for search is stale.
A positive result would establish encoding order as a missing design variable in radar–communication DPC. A negative result would justify the field's habit of fixing the order.
References
- Joint Transmit and Receive Beamforming Design for DPC-Based MIMO DFRC Systems
- Un-evaluated solutions may be valuable in expensive optimization
- MIMO-Aided Nonlinear Hybrid Transceiver Design for Multiuser Mmwave Systems Relying on Tomlinson-Harashima Precoding
- Nonlinear Precoding in the RIS-Aided MIMO Broadcast Channel