Wireless links inside a processor package could let many cores talk at once on one frequency band, provided each transmission lands only at its intended receiver. Time reversal does this in the reverberant metal box of a package. The transmitter sends the time-flipped echo pattern of the channel, so the reflections pile up into a sharp spike at the target and smear into weak noise elsewhere. Full-wave simulations report strong focusing: an 8 dB gain in signal-to-interference-plus-noise ratio (SINR) and three concurrent links at 140 GHz. The weakness is that every one of these filters is hard-coded from a single, perfectly known channel. The same papers concede that heat changes package dimensions and material permittivity, which shifts the echo pattern and undoes the alignment. They also report "interference hotspots," where a node placed symmetrically to the target receives a focused copy of a message meant for someone else.
We propose to stop treating the time-reversed channel as the filter. Instead, we design the filters jointly across all concurrent links so they hold up across a whole family of thermally perturbed channels. The design problem is not convex, but it splits naturally into a convex part minus another convex part. That structure lets us solve it with difference-of-convex programming, a method that repeatedly replaces the troublesome subtracted part with a straight-line bound and solves the resulting convex problem. We borrow the evaluation discipline of rare-event physics, where a detector is judged by how much signal it keeps for a given amount of background it lets through. We also borrow that field's hard lesson that a design tuned only on simulation must be tested on conditions the simulation never saw.
Core Claim
Fixed time-reversal filters in chip packages are fragile in two ways: to channel drift and to symmetric hotspots. Both problems are better posed as a worst-case "focus versus leakage" design over an ensemble of perturbed channels than as a property of the time-reversal operation itself. A difference-of-convex solver over that ensemble should keep the minimum per-link SINR, measured at held-out thermal states, well above both plain time reversal and simple per-link robust precoders. Plain time reversal is the reference; we expect it to degrade sharply off its design point. If the simplest robust alternative closes the gap, we will report that and drop the heavier machinery.
Prior Work and Where This Sits
The motivating evidence comes from the package time-reversal line. Exploration of Time Reversal for Wireless Communications within Computing Packages simulates a 10×10 mm² flip-chip package at 60 GHz in CST Microwave Studio. It shows focusing peaks roughly three orders of magnitude above non-time-reversal signals and a bit error rate of $10^{-15}$ at 17 dB SNR at 10 Gb/s. Collective Communication Patterns Using Time-Reversal Terahertz Links at the Chip Scale extends this to a 140 GHz interposer with 12 monopoles, reporting 45 mW versus 0.2 mW peak received power and three concurrent links. It also identifies symmetry-induced hotspots. Both papers assume a static, pre-characterized channel while acknowledging thermal drift. Neither designs filters for anything but the nominal channel. A MAC Protocol with Time Reversal for Wireless Networks within Computing Packages builds a protocol on top of these links in an architectural simulator and inherits the same assumption. Opportunistic Beamforming in Wireless Network-on-Chip offers the phased-array alternative. It also documents a roughly 15 dB loss from bulk silicon and beam-pattern breakup beyond $5\lambda$ from package reflections. This is evidence that in packages, reverberation must be exploited rather than fought with narrow beams. We therefore build on time-reversal-style temporal filters rather than on geometric beams.
On the optimization side, two papers frame the method. Cooperative Interference Management in Multi-Cell Downlink Beamforming studies full frequency reuse with multiple transmit antennas and single-antenna receivers. It shows that weighted sum-rate maximization cannot reach parts of a non-convex rate region, which pushes us toward a max-min (fairness) objective rather than a weighted sum. 6D Movable Antenna Enhanced Multi-Access Point Coordination via Position and Orientation Optimization handles interference-coupled rate with successive convex approximation and alternating optimization. However, it assumes perfect instantaneous channel knowledge, which is exactly the assumption we relax.
The solver machinery comes from the difference-of-convex literature. DC Proximal Newton for Nonconvex Optimization Problems shows that keeping the convex part intact and linearizing only the subtracted part can be much faster than first-order non-convex solvers, with a nearly 6× speedup over GIST. Distributed Nonconvex Optimization for Sparse Representation likewise finds that preserving convex structure beats full linearization. Neither paper touches wireless filters.
From rare-event physics we take the evaluation lens rather than the method. Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment and Machine learning-powered data cleaning for LEGEND report signal efficiency against background leakage as a pair of numbers, not a single score. Implementation of the Random Forest method for the Imaging Atmospheric Cherenkov Telescope MAGIC warns that discriminators trained on Monte Carlo degrade when real conditions drift outside the simulated range. These papers constrain how we must evaluate.
We are not claiming that difference-of-convex methods for SINR-based beamforming are new as a mathematical idea. The new parts are three. First, the object is temporal focusing filters in a reverberant package. Second, the objective is worst-case over a physically generated thermal ensemble. Third, the formulation adds explicit leakage caps at non-target nodes, which targets the reported symmetric hotspots directly.
Formulation
Transmitter $j$ applies a real filter $w_j$ of $L$ taps. Thermal state $s$ (one of $S$ perturbed package models) yields, for each transmitter–node pair, a convolution channel. From this channel we form a positive semidefinite matrix $Q^{(s)}_{kr}$, the received energy that filter $k$ deposits at node $r$ in the decision sample. We also form $R^{(s)}_{kr}$, the energy it deposits at node $r$ in all other samples, which captures self-interference across time and spillover.
We lift each filter to $W_j = w_j w_j^\top \succeq 0$, so every energy term becomes linear in $W_j$. Let $t(j)$ be the target of link $j$, let $I^{(s)}_j(W)$ denote the interference-plus-noise energy at $t(j)$, and let $\sigma^2$ be the noise power. The design problem is
$$\max_{W_j \succeq 0,\ \mathrm{tr}\,W_j \le P}\ \min_{s}\ \min_{j}\ \Big[\log\big(\mathrm{tr}(Q^{(s)}_{j t(j)} W_j) + I^{(s)}_j(W)\big) - \log I^{(s)}_j(W)\Big] - \rho \sum_j \big(\mathrm{tr}\,W_j - \lambda_{\max}(W_j)\big).$$
The problem is subject to a background cap $\mathrm{tr}(Q^{(s)}_{jr}W_j) \le \beta$ for every non-target node $r$ and every state $s$. Here $P$ is the per-transmitter energy budget, $\rho$ weights a penalty that pushes each $W_j$ back to rank one, and $\beta$ is the allowed focused leakage. The cap plays the role of a background budget. Each rate term is a concave function minus a concave function. The rank penalty is a linear term minus the convex largest-eigenvalue function.
The difference-of-convex step linearizes the subtracted $\log I$ and $\lambda_{\max}$ terms at the current iterate. The latter is linearized through its top eigenvector. Each step is then a convex semidefinite program with an epigraph variable for the max-min. Every step improves a lower bound on the objective, and the iterates converge to a stationary point. That is a local guarantee only, so we use multiple initializations, including the time-reversal filter itself.
Research Program
The first component is a channel ensemble generator. We will rebuild the 60 GHz flip-chip and 140 GHz interposer geometries as described in the two time-reversal papers. We will then produce perturbed versions by scaling dimensions with thermal expansion and shifting substrate and heat-spreader permittivity across a stated temperature range. Each version is a full-wave run, giving channel impulse responses between all antenna pairs. The thermal states are split into a design set and a held-out set. The held-out set includes interpolated states and states that extrapolate beyond the design range.
The second component is the solver. We will implement the lifted program for filter lengths up to a few hundred taps, which is feasible as an offline design computed once per package. We will also test a non-lifted variant that works on $w_j$ directly with the same linearization, to see whether lifting is needed. A proximal-Newton inner solver in the style of DC Proximal Newton is a fallback if generic semidefinite solvers are too slow.
The third component is analysis. We will examine how the guaranteed minimum SINR trades off against the size of the thermal range designed for, how the background cap $\beta$ redistributes energy away from mirror-image nodes, and how many design states $S$ are needed before held-out performance stops improving. That last question is the analogue of a training-set size curve.
Baselines
The naive reference is the hard-coded time-reversal filter computed on the nominal channel, exactly as in Exploration of Time Reversal and Collective Communication Patterns. On the nominal channel the reported figure for the 140 GHz setup is an 8 dB SINR improvement over non-time-reversal transmission, with three concurrent links (Collective Communication Patterns). We will first reproduce something close to this on our rebuilt geometry to validate the pipeline. No reported result was found on Lacuna for time-reversal SINR under thermal perturbation. That number is the first thing we measure.
The strongest simple competitor is a per-link signal-to-leakage precoder. Each filter maximizes target energy divided by total leakage plus noise, using channel matrices averaged over the design ensemble. This is a closed-form generalized eigenvector. It already accounts for drift on average and for leakage at every node, with no iterative solver. No reported result was found on Lacuna for this precoder in chip packages.
A third baseline is time reversal with periodic recalibration. The channel is re-measured and the filter recomputed at each test state. We grant this an oracle channel as an upper bound and then charge it a stated measurement overhead.
All methods share the same rebuilt geometries, design and held-out splits, energy budget, and filter length. Each tunable quantity ($\rho$, $\beta$, and regularization in the eigenvector baseline) receives the same grid-search budget on design states only.
Two ablations isolate the proposal's ingredients. The first removes the worst-case aggregation and replaces it with an ensemble average. This tells us whether robustness comes from the max-min or merely from seeing more channels. The second removes the leakage caps, which tells us whether hotspot suppression needs explicit constraints.
Beating the field means three things on held-out states. The worst-link SINR must exceed the averaged signal-to-leakage precoder by a margin larger than the spread across random restarts. The method must hold three or more concurrent links with no link below the SINR implied by the reported 17 dB, $10^{-15}$ error-rate operating point. And it must stay within a stated gap of the oracle recalibration bound.
Why not just…?
Why not just recalibrate the time-reversal filter whenever the temperature changes? This is a real option, and it serves as our upper-bound baseline. Its cost is what the original papers avoided by assuming a static channel: on-die channel sounding, sampling hardware, and downtime. TR-MAC was designed around zero-overhead filters. The experiment will show how close a fixed robust design comes to the oracle. If the gap is small, recalibration hardware becomes unnecessary. If it is large, the result instead tells designers how often recalibration is needed.
Why not just use the closed-form averaged signal-to-leakage precoder? This is the most serious threat, because it captures drift on average and leakage at all nodes in one eigen-decomposition. It differs from our method in two ways. It optimizes each link separately, ignoring that one link's filter is another link's interference. It also optimizes the average rather than the worst state, which is exactly where a hard-coded design is likely to fail. If held-out worst-link SINR differs by under about 1 dB, we will recommend the closed form and treat the difference-of-convex solver as unnecessary for this regime.
Why not just schedule the symmetric node pairs in different time slots? Avoiding hotspots by scheduling is simple and protocol-compatible. However, it gives up exactly the concurrency that spatial reuse is meant to buy, and TR-MAC shows that a single shared channel can match multi-channel token passing only when links can run concurrently. We will report the number of concurrent links each method sustains, so the cost of scheduling around hotspots is visible.
Why not just use successive convex approximation or weighted MMSE instead of a difference-of-convex method? The step we take is a special case of convex approximation that keeps the convex part exactly. Distributed Nonconvex Optimization for Sparse Representation reports that this beats full linearization. The real question is not the solver label but whether the max-min, the ensemble, and the leakage caps matter. The ablations answer that. We do not claim a solver contribution.
Proposed Experiments and Decisive Checks
The main experiment designs every method on the design thermal states and reports results on held-out states. Following the PandaX-III and LEGEND practice, we report each method as a curve of target-focusing efficiency against worst background leakage at non-target nodes, rather than as a single number. Alongside the curves we report the minimum per-link SINR, the bit error rate at fixed SNR for 10 Gb/s links, and the number of sustainable concurrent links.
The first decisive check is extrapolation. Following the MAGIC warning that discriminators trained on Monte Carlo fail outside simulated conditions, we test at temperatures beyond the design range. A method that wins only by interpolation is reported as such.
The second decisive check is a modeling-fidelity shift. We evaluate filters on channels from a higher-fidelity model than the one used for design, with added surface roughness, finer meshing, and the lossy silicon parameters that Opportunistic Beamforming found cost about 15 dB. This tests whether robustness to one kind of mismatch transfers to another.
The third check targets the hotspot directly. We measure focused energy at the mirror-image node of each target in the symmetric 140 GHz layout, with and without leakage caps.
We state plainly that all of this remains simulation-based. It narrows the reality gap by forcing designs to survive conditions they were not tuned on, but it does not close it. The natural follow-up is a channel-sounding measurement in a scaled metallic cavity.
References
- Exploration of Time Reversal for Wireless Communications within Computing Packages
- Collective Communication Patterns Using Time-Reversal Terahertz Links at the Chip Scale
- A MAC Protocol with Time Reversal for Wireless Networks within Computing Packages
- Opportunistic Beamforming in Wireless Network-on-Chip
- Cooperative Interference Management in Multi-Cell Downlink Beamforming
- 6D Movable Antenna Enhanced Multi-Access Point Coordination via Position and Orientation Optimization
- DC Proximal Newton for Nonconvex Optimization Problems
- Distributed nonconvex optimization for sparse representation
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines
- Implementation of the Random Forest method for the Imaging Atmospheric Cherenkov Telescope MAGIC