OFDM-ISAC drops the either-or: joint power and constellation optimization eases the sensing-communication trade-off in tracking
Synopsis
In OFDM-ISAC, communications favor waterfilling power with Gaussian-like inputs while sensing favors uniform power and constant-modulus constellations, creating a trade-off in frequency-selective channels. This work treats per-subcarrier power and kurtosis as joint variables and uses an autoencoder to train offline a family of finite constellations conditioned on kurtosis, generated online for each allocation. Simulations show joint optimization clearly beats optimizing power or constellation alone at intermediate SNR, and the pretrained family nearly realizes the predicted allocation gains.
Interpretation
The sensing-communication trade-off in OFDM-ISAC no longer has to be eased by tuning power alone or constellation alone; power and constellation are solved together in one allocation problem. Earlier joint approaches either selected constellations from a finite set of conventional modulation formats, which underestimates achievable performance, or designed continuous input distributions, which give an upper benchmark practical finite constellations cannot reach, and both focused mainly on the high-SNR regime where the power-induced trade-off largely vanishes. In simulation comparisons, optimizing only power or only kurtosis is communication-limited or infeasible under strict sensing constraints, while the joint optimization achieves the best simultaneous sensing and communication performance.
Constellations are no longer optimized one at a time for a fixed sensing constraint; a family of finite constellations takes a target kurtosis as input and is generated on demand without retraining. Previous work optimized individual constellations for fixed kurtosis constraints, which becomes impractical when continuously varying kurtosis is required. Across the full kurtosis range tested, the polar parameterization accurately produces the target kurtosis, whereas the Cartesian parameterization fails to produce constant-modulus constellations and kurtosis-specific constellations partly undershoot the target.
The sensing constraint is written according to the task: target tracking constrains only a small delay neighborhood around the predicted position instead of averaging over all delays. Earlier work used the delay-averaged sidelobe level as the sensing constraint, corresponding to an initial detection mode in which a target may appear at any delay, and discarded the deterministic structure of the power-allocation-induced sidelobes. Comparing detection and tracking shows that tracking alleviates the sensing-communication trade-off, especially at small kurtosis, with performance approaching the detection case as kurtosis grows.
Perspective
The conclusions rest on simulations of a monostatic OFDM-ISAC system in which the communication side assumes known channel state information, the sensing channel is unknown and estimated from echoes, and the sensing channel contains one dominant self-interference component and one target of interest. Results cover power-kurtosis allocation and constellation generation under both detection and tracking sensing constraints; the authors state that the proposed constellation family can generate constellations on the fly for per-subcarrier target kurtosis after training, avoiding storage of separately optimized constellations.
Communication performance uses a mutual-information upper bound, which the authors state is not necessarily achievable because the entropy-maximizing output distribution is not guaranteed to be realizable under the system model. In simulation, predicted curves and achieved markers show a small gap at low-kurtosis operating points, which the authors attribute to the bound's known overestimate of achievable mutual information. The authors also note that the neural network architectures were not optimized for minimal complexity, leaving the complexity-performance trade-off for further study.