Direct pinching-location tuning is high-dimensional and non-convex
Pinching-antenna systems (PASSs) let a transmitter move pinching elements along dielectric waveguides, so the antenna array geometry itself becomes a design variable that can be adapted to the propagation environment and the spatial distribution of users. For integrated sensing and communications (ISAC), this geometric freedom can reshape user-facing channels and steer the sensing beam at the same time.
The catch is the tuning problem. The paper states that direct location optimization in PASSs is high-dimensional and non-convex, and that PASS gains materialize only if pinching locations and digital beamforming are jointly tuned at the system update rate, that is, once within a coherence time. The design problem is non-convex on three counts: the communication sum-rate is non-concave in the precoder, the sensing mutual information is non-concave in the precoder, and both the user channels and the transmit array response depend non-linearly on pinching locations through an exponential kernel.
Clusters replace element coordinates with a weight allocation
Instead of optimizing every pinching location, the authors partition the K users into C clusters (with C at most K) using k-means, and each cluster, together with the sensing target, defines a projection point on the waveguide axis. The target projection is p0 = x_t, and the cluster centroids give the remaining points, so the number of projection points is G_c = C + 1.
For each transmit waveguide, the design variable becomes an allocation vector over the probability simplex: how many of that waveguide's elements serve each projection point. The continuous element count is approximated as N_m,i ≈ N_tx α_i^(m), with the closest integer fixing the exact number of elements, and the allocated elements are placed around the corresponding projection point. Because consecutive elements in a group are spaced by the guided wavelength λ_g = λ_0 / n_eff, they share a common guided phase and combine coherently, which is why the authors model each group as a segment rather than averaging the response over it.
Precoding and location tuning are solved in alternation
The multi-objective problem is first scalarized into a weighted sum w_c R_c + w_s I_s^d with w_c + w_s = 1, where R_c is the communication sum-rate and I_s^d is the sensing mutual information. Sweeping the weight pair over the unit simplex samples a trade-off curve that inner-approximates the true Pareto frontier.
The resulting problem is then solved by alternating optimization between two marginal tasks. For digital precoding, the authors derive an equivalent weighted minimum mean squared error (WMMSE) reformulation whose update has a closed form, obtained via a rank-one (Sherman-Morrison) update, with the dual variable chosen by bisection to meet the power constraint. For location tuning, the receive-side locations admit a single-step closed-form solution: the sensing metric depends on them only through the receive array response norm, which is maximized when all receive elements on each waveguide are symmetrically placed around the target with guided-wavelength spacing. The transmit-side locations are then updated through the clustering-based weight allocation and projection.
Complexity scales with clusters rather than element count
The paper compares the cost of the two parameterizations. A direct location update that treats all M_tx N_tx transmit element coordinates as optimization variables requires M_tx N_tx + 1 exact objective evaluations per gradient and scales as O(I_p M_tx^2 N_tx^2 K). The proposed clustering-based parameterization scales as O(I_p M_tx^2 G_c^2 Q K), where G_c is the number of projection points and Q is the number of clusters or targets involved.
For fixed G_c and Q, the leading-order complexity reduction therefore grows as N_tx^2 / (G_c^2 Q). This is the mechanism behind the paper's claim of low-complexity tuning: the design variable is the number of elements assigned to each projection point, not the coordinate of every element.
The weighted objective saturates within a few iterations
In the simulation setup, the transmitter uses M_tx = 2 waveguides and the receiver uses M_rx = 1, all with N_tx = N_rx = 20 pinching elements. The carrier frequency is 28 GHz with effective refractive index n_eff = 1.4, the transmit power budget is P = 1, the downlink noise variance is σ_u^2 = 0.05, the sensing echo noise variance is σ_rx^2 = 0.01, and the frame length is T = 5. Four users are distributed in two clusters drawn from a centroid-first disk model with radius 0.5.
At the balanced weighting (w_c, w_s) = (0.5, 0.5), the proposed PASS (L+W) scheme starts from the same uniform initialization as Baseline A, performs one grouped re-initialization, and then refines through five alternating-optimization updates. The grouped-initialization step alone yields a substantial improvement, and two further updates close most of the remaining gap, after which the trajectory is essentially flat. The authors note that this rapid saturation is what makes the design usable at the system update rate, since the whole loop must fit within one coherence interval. The two baselines produce constant trajectories because they only optimize the precoder for fixed element positions.
The proposed design dominates both fixed-location baselines
The achievable communication-sensing region is obtained by sweeping w_s from 0 to 1 with w_c = 1 − w_s. Across this entire sweep, the proposed design strictly dominates both baselines in the Pareto sense, containing both baselines entirely. The proposed scheme is also nearly flat over a wide range of sensing levels, which the authors read as sensing requirements being met with minimal communication penalty.
The two benchmarks are Baseline A, where transmit pinching elements are uniformly distributed over the aperture while the precoder and receive locations are optimized, and Baseline B, a fixed-location MIMO-PASS benchmark with transmit elements centered at half the aperture with guided-wavelength spacing, where only the precoder is optimized. The comparison is therefore between reallocating elements near scene-dependent projection points and leaving element positions fixed.
Aperture growth favors reallocation over fixed spacing
When the transmit and receive waveguide lengths are jointly swept while element counts are held at N_tx = N_rx = 20, the fixed baselines degrade as the aperture grows, whereas PASS remains robust by reallocating elements near the scene-dependent projection points. The authors attribute this scalability advantage directly to treating element locations as a design variable rather than a fixed geometric constant.
A separate sweep varies the element counts with waveguide lengths held at L_tx = L_rx = 20. PASS benefits from additional elements up to a saturation point, because the allocation variables direct them toward useful user or target projections, until the growing physical extent of a group offsets the array gain. Reported curves are averaged over 40 seeds in the trade-off figure and 10 seeds in the convergence, aperture, and element-budget figures.
The gain depends on how well clusters match the scene
The paper's own framing is that the clustering trick is an extension of the bi-partitioning idea used in earlier single-user, single-waveguide PASS-ISAC work, and that it bridges the gap to a generic multiuser PASS-assisted ISAC setting with multiple transmit and receive waveguides. The evidence here is numerical: the results come from simulations under the stated geometry, frequency, noise, and user-distribution assumptions, not from a hardware testbed or field trial.
That scope matters for reading the trade-off curves. The reported dominance is measured against two specific fixed-location baselines under those simulation conditions, and the complexity reduction is an order-of-growth statement in N_tx, G_c, and Q rather than a wall-clock measurement. The most decision-relevant point for a reader is the comparison the paper actually makes: when the aperture grows with a fixed element budget, reallocating elements toward cluster and target projection points keeps the weighted objective robust, while fixed-spacing designs degrade. Whether that advantage holds depends on how well the chosen number of clusters and their projection points track the actual user and target geometry.
Sources used in this answer
Fast Tuning of Pinching Antenna-Assisted ISAC Systems via User Clustering
A clustering-based PASS-ISAC design reduces pinching-location tuning to a low-dimensional weight allocation solved by projected gradient descent, and in simulations it strictly Pareto-dominates two fixed-location baselines on the communication-sensing trade-off at modest computational cost.