Geometric Power from the Resolvent: Fast G-SNR for Drifted, Bounded, Unsynchronized Molecular Channels

In molecular communication, a transmitter encodes bits in the moment it releases a particle. The particle then diffuses to a receiver. Its travel time is random and often heavy-tailed: in free diffusion it follows a Lévy distribution, which has infinite variance. Ordinary signal-to-noise ratio is then undefined. The field's answer is the geometric signal-to-noise ratio (G-SNR), which replaces noise variance with "geometric power," the exponential of the average log-magnitude of the noise. For Lévy noise, the bit error rate is exactly constant at fixed G-SNR (On the Impact of Time-Synchronization in Molecular Timing Channels, Communication System Design and Analysis for Asynchronous Molecular Timing Channels). That result holds only in one-dimensional, unbounded, drift-free space. Real devices are microfluidic channels with flow, walls and imperfect receptors. The receiver's clock is itself set by diffusing "clock molecules" that arrive with jitter (Variance-constrained capacity of the molecular timing channel with synchronization error). A chemical receiver also adds its own reaction delay (Closing the Implementation Gap in MC: Fully Chemical Synchronization and Detection for Cellular Receivers). In those settings nobody currently knows how to compute G-SNR quickly, or whether it still predicts error rate at all.

This note proposes a way to compute geometric power, including synchronization jitter and receiver delay, without simulating particles or forming probability densities. The log-moment of the arrival time is a one-dimensional integral of the arrival time's Laplace transform. That transform is the solution of a steady, elliptic boundary-value problem, and each value of the transform variable costs one sparse linear solve. The operator in that problem, drift plus diffusion, is non-self-adjoint, or "non-Hermitian" in physics language. Non-Hermitian spectral theory then explains why the obvious fast alternative, expanding in eigenmodes, becomes badly conditioned at high flow rates, and why solving directly in the transform domain avoids the problem. With a fast and exact evaluator in hand, the scientific question becomes answerable: where does G-SNR stop predicting bit error rate once geometry, flow and clock jitter enter?

Core Claim

Let $T$ be the first arrival time at the receiver of a particle released at position $x$. Let $L(x,s)=\mathbb{E}_x[e^{-sT}]$ be its Laplace transform, where $s>0$ is the transform variable. A Frullani-type identity, $\int_0^\infty (e^{-s}-e^{-sT})\,ds/s=\log T$, gives

$$\log S_0(T)=\mathbb{E}_x[\log T]=\int_0^\infty \frac{e^{-s}-L(x,s)}{s}\,ds ,$$

where $S_0$ denotes geometric power. For a real-valued noise $X$, the analogous identity uses the characteristic function $\varphi_X(u)=\mathbb{E}[e^{iuX}]$:

$$\mathbb{E}[\log|X|]=\int_0^\infty \frac{e^{-u}-\mathrm{Re}\,\varphi_X(u)}{u}\,du .$$

The second identity is what brings in synchronization. Suppose the effective noise is the information particle's arrival minus an independent clock-molecule arrival, plus an independent chemical-receiver delay. Its characteristic function is then the product of the three individual characteristic functions. Jitter and receiver latency therefore enter as multiplications, with no convolutions of heavy-tailed densities that lack closed forms. Finally, $L(x,s)$ solves $(\mathcal{A}-s)L=0$, where $\mathcal{A}=\mathbf{v}\cdot\nabla+D\nabla^2$ is the drift-diffusion generator used by Characterizing First Arrival Position Channels. The boundary conditions are $L=1$ on the absorbing receiver, reflection on walls, and a Robin condition on partially reactive surfaces. One sparse solve gives $L$ at every release position simultaneously.

The claim has two parts. The first is computational: this "resolvent route" matches particle simulation's G-SNR to within its statistical error, and does so orders of magnitude faster, in geometries where no closed form exists. The second is scientific: using the route across flow strength, confinement and jitter will show where BER invariance at fixed G-SNR breaks. Farsad et al.'s invariance comes from the Lévy law being a scale family. Drifted and confined arrival times are not scale families, so we expect the invariance to degrade gradually rather than hold exactly. Here G-SNR means the metric defined in the Farsad et al. papers above, built from the symbol spacing and the noise's geometric power.

Research Program

What the non-Hermitian lens contributes. Discretized drift-diffusion is an asymmetric hopping lattice of Hatano–Nelson type (Symmetry breaking and spectral structure of the interacting Hatano-Nelson model). With open boundaries its eigenvectors pile up at the downstream wall. This is the skin effect of Exceptional topological insulators, which in fluid terms is the Péclet boundary layer. The Péclet number measures how strongly flow dominates diffusion.

We want to be precise about what this lens adds, because much of the relevant fact is textbook. For constant drift, the substitution $L=e^{-\mathbf{v}\cdot\mathbf{x}/2D}w$ turns the operator into a symmetric one; physicists call this an imaginary gauge transform. The substitution is exact but numerically dangerous, because the factor is exponential in the Péclet number. For non-potential flows, such as the parabolic Poiseuille profile in a microfluidic channel, no such transform exists and the operator is genuinely non-normal: its eigenvectors are far from orthogonal.

Non-Hermitian theory quantifies the consequences. Pseudospectra, the regions of the complex plane where small perturbations can move eigenvalues, are large, and eigenvector condition numbers grow with the Péclet number. These are exactly the sensitivities the file flags for open-boundary non-Hermitian operators. They predict that truncated eigenmode expansions of arrival-time densities will show large transient errors precisely at early arrival times. Early arrivals are the regime that governs the first-arrival detector of Communication over Diffusion-Based Molecular Timing Channels. The resolvent route never diagonalizes the operator, and for real $s>0$ the shifted operator is well posed. We will measure pseudospectra and eigenvector conditioning directly and test these predictions. In this program, the non-Hermitian material is a diagnostic tool, not decoration.

Building the evaluator. We will build a finite-element solver for $(\mathcal{A}-s)L=0$ with streamline stabilization. Values of $s$ will be taken at 40–80 log-spaced quadrature nodes. The small-$s$ behavior of $L$, which scales like $1-c\sqrt{s}$ for Lévy-type tails, will be handled analytically so that the integral converges.

For the synchronization and delay factors: clock-molecule jitter uses the same solver on the master-to-receiver geometry, following Egan et al.'s physical model. Chemical-receiver delay uses characteristic functions of reaction-chain waiting times, calibrated against stochastic simulation of Heinlein et al.'s circuits. The same transform values, passed through Gil-Pelaez inversion (a standard formula recovering a distribution function from its characteristic function), give the error probabilities needed for maximum-likelihood thresholds. This means BER comes out of the same computation.

If G-SNR fails as a predictor. In that case we will test a two-number descriptor: the log-mean together with the log-variance. The log-variance needs only one more integral of the same transform, since $\mathbb{E}[(\log T)^2]$ has an analogous identity. We will test whether this pair restores BER prediction.

Baselines

The naive reference is the variance-based SNR computed from particle simulations truncated at an observation window. We run it at three truncation windows on the same trajectories. It counts as beaten if its predicted BER ordering across geometries is less consistent with simulated BER than G-SNR's ordering is. Farsad et al.'s truncation argument predicts it will fail.

The analytic reference is the closed-form one-dimensional Lévy G-SNR of Communication System Design and Analysis for Asynchronous Molecular Timing Channels, applied with the true distance and diffusion coefficient but ignoring flow and walls. It also serves as a correctness check: in free one-dimensional diffusion our method must reproduce it to quadrature precision.

The strongest competitor for accuracy is direct particle-based simulation, following the protocol of Characterizing First Arrival Position Channels. That work used $10^6$ trajectories and reports agreement with theory at an error of about $10^{-3}$. We treat that as the achievable accuracy target and run the same protocol ourselves. We will measure its wall-clock cost on our hardware rather than assume a figure. We compare at matched wall-clock budgets, and we count our method as winning if it reaches the simulation's $10^6$-trajectory accuracy on $\log S_0$ with at least a hundredfold lower latency.

The strongest competitor for speed is a learned surrogate in the style of RadioUNet. That model maps geometry rasters to fields and reports roughly 1 dB RMSE at $10^{-3}$–$10^{-2}$ s per map on RadioMapSeer. We retrain the same architecture, and a MaxViT hybrid following RMTransformer, to map channel geometry and flow to per-position log geometric power. Training data come from particle simulation under a fixed simulation budget, and our method receives no extra tuning. No reported result was found on Lacuna for any surrogate predicting G-SNR or molecular arrival fields, so the comparison number will be ours. Beating the surrogate means lower log-geometric-power error on held-out geometries, especially out-of-distribution flow speeds, at comparable latency.

Two ablations isolate the key ingredients. The first is an eigenmode expansion of the arrival density, truncated at matched cost. It isolates the value of avoiding diagonalization, and we expect it to degrade with the Péclet number. The second is our method with the jitter and delay factors removed. It isolates what synchronization modeling changes in the design conclusions.

Why not just…?

Why not just run particle simulation and average the log arrival time? This is the real threat, because $\log T$ has finite variance, so the Monte Carlo error falls as $1/\sqrt{n}$. The catch is cost per trajectory rather than sample count. In free diffusion the mean arrival time is infinite, so the expected simulation time of a single trajectory is unbounded unless it is cut off. Cutting it off biases the estimate of $\mathbb{E}[\log T]$ by an amount that depends on the cutoff. In weakly flowing, confined channels, long-surviving trajectories dominate the compute. Monte Carlo also gives no smooth gradients for design optimization, and it resolves the small BERs we care about only slowly. The matched-wall-clock experiment below settles the question. If truncated simulation with $10^4$ samples matches our accuracy at similar cost across the sweep, the computational claim fails, and only the scientific study survives.

Why not just train a neural surrogate on simulated data? The research file's claim that L2 training "diverges" on heavy-tailed fields does not apply here. Log geometric power is a bounded, smooth target, and a U-Net could learn it. What the surrogate cannot do is give guarantees outside its training geometries. Its training labels also inherit Monte Carlo's truncation bias. The sim-to-real and generalization gaps stressed in Overview of the First Pathloss Radio Map Prediction Challenge suggest held-out flow regimes will hurt it. The held-out-Péclet test measures this directly.

Why not just use variance-based SNR with a physical cutoff, or the one-dimensional Lévy formula? Farsad et al. show that truncated variance scales with an arbitrary cutoff, while BER tracks geometric power. Diversity Gain of One-Shot Communication over Molecular Timing Channels shows that variance intuitions mislead even about whether extra particles help. The Lévy formula is exact only without drift or walls. Our sweep reports how far off it is as confinement grows.

Why not just diagonalize the operator and use its eigenmodes? For self-adjoint problems this would be the right choice. Here the operator is non-normal, and the non-Hermitian literature predicts ill-conditioning exactly in the flowing, bounded regime of interest. The eigenmode ablation tests this prediction rather than assuming it.

Proposed Experiments

We build a benchmark with five geometry families. Free one-dimensional diffusion reproduces Farsad et al. One-dimensional diffusion with drift gives inverse-Gaussian arrival times. Lee et al.'s two- and three-dimensional planar receivers with vertical drift form the third family. Two-dimensional Poiseuille microchannels with absorbing receptor patches form the fourth. The fifth is the same channels with partially reactive walls. Ground truth everywhere is particle simulation with $10^6$ trajectories and long windows, following Lee et al.'s protocol.

The first decisive check is correctness. The error in $\log S_0$ against ground truth should be below Monte Carlo standard error in every family, and should match the closed form in free one-dimensional diffusion.

The second is the latency–accuracy frontier against truncated simulation, the neural surrogates and the eigenmode ablation, swept over Péclet numbers from 0.1 to 100.

The third is the scientific test. For each family, we fix G-SNR and vary distance, diffusion coefficient and flow, then simulate BER for the synchronized design (System A) and for asynchronous designs with indistinguishable and distinguishable particle pairs (Systems B and C). The spread of BER at fixed G-SNR measures the breakdown of invariance. We then test whether adding log-variance collapses that spread.

The fourth adds Egan-style clock-molecule jitter and Heinlein-style reaction delays, calibrated with the modified next-reaction stochastic simulator. The question is at what receiver latency the distinguishable-particle design loses its advantage over a jittered synchronized one. This is the asynchronous-versus-chemical-latency tension the file identifies, which no cited work resolves. A negative result on the third check, meaning G-SNR fails to organize BER in bounded flowing channels, would be as useful as a positive one: it would tell the field which regime its central metric covers.