betaFeedback

WorksEngineering

Quantum Optimization Algorithm for Pinching-Antenna Activation with Max-Min Fairness

Abstract

In this letter, we propose a training-free quantum approximate optimization algorithm with a linear-ramp schedule (LR-QAOA) for antenna activation in pinching-antenna systems (PASS). Since all active pinching antennas (PAs) share one feed and split its power, maximizing the signal-to-noise ratio (SNR) of the weakest user becomes a combinatorial max-min problem that must be re-solved whenever the users move. We formulate the problem as a two-local Ising Hamiltonian with one qubit per PA and no penalty terms, employing a weightedsum surrogate to prioritize weaker users. Instead of training the circuit with a classical optimizer, the proposed LR-QAOA method tunes its parameters via a linear ramp; the measured patterns are ranked by the exact max-min objective, and the best ones are refined with a short local search using the same number of evaluated patterns. Monte Carlo simulations show that the proposed LR-QAOA method outperforms greedy activation, projection-guided activation, and semidefinite relaxation-based classical baseline methods when all methods test the same number of PA on/off patterns. Experiments on real IBM Quantum hardware further show that the execution time remains nearly constant as the number of PAs grows, whereas exhaustive search grows exponentially. However, gate noise in the NISQ era restricts near-optimal hardware solutions to small numbers of antennas only.

Introduction

[1]-antenna systems (PASS) carry a radio-frequency signal along a low-loss dielectric waveguide and radiate it only where a small dielectric “pinch” touches the guide [1]. Each pinch acts as an antenna, and since the waveguide can extend over many meters, the radiating points can be placed close to the users. PASS thus converts the problem of large-scale path loss into one of selecting good positions, establishing strong line-of-sight links with very simple hardware [2].

Since moving the pinches mechanically is slow, low-complexity deployments pre-install the pinching antennas (PAs) at fixed positions and only switch them on or off [3]–[5]. However, this simplification comes at an algorithmic price. Because all active PAs share a single feed, their signals add coherently while the transmit power is split among them. Consequently, a new PA adds a path but weakens every other one, and a PA that adds in phase at one user may cancel at another. For NN PAs, the problem is combinatorial, with 2N−12^N-1 candidate patterns, and must be solved again whenever the users move.

When one PA pattern serves several users, fairness matters. Maximizing the signal-to-noise ratio (SNR) of the weakest user leads to a max-min fractional 0–1 program [5], which is a quadratic fractional problem even for a single user [6]. Existing approaches trade optimality for speed, e.g., game-theoretic activation [3], projection-guided greedy activation (PGGA) [4], deep learning [6], and trellis-based Viterbi state selection (VSS) [5]. Semidefinite relaxation (SDR) [7] remains the classical general-purpose tool for such binary quadratic problems. As shown in this work, greedy methods tend to stop too early, since the power split makes almost every addition appear harmful, whereas exhaustive search becomes impractical for real-time operation beyond about 20 PAs.

Quantum optimization offers a different route. The quantum approximate optimization algorithm (QAOA) [8], [9] prepares a parameterized quantum state that favors low-energy solutions of an Ising Hamiltonian and samples from it, which makes it a natural fit for binary optimization problems. Prior works have investigated quantum optimization for MIMO detection [10], [11], and QAOA for RIS phase optimization [12] and MIMO-QAM data detection [13]. Despite these works, several research gaps remain. First, to the best of our knowledge, PA activation has not yet been formulated as an Ising problem, and such a formulation is not immediate, since the max-min objective contains both a ratio and a minimum, neither of which is quadratic. Second, standard QAOA tunes 2p2p circuit angles for every instance with a classical optimizer, which is too slow when every channel realization is a new instance. Linear-ramp QAOA (LR-QAOA) avoids this by fixing all angles through a single scale parameter that transfers across instances [14], but has not been applied to PA activation problem yet. Third, gate noise, which grows with circuit size, limits current hardware, so both the execution time and the achieved quality must be evaluated on real quantum devices. This work addresses these three gaps through the following contributions:

  • We formulate max-min PA activation as a two-local Ising Hamiltonian with one qubit per PA and no penalty terms, and analyze how its weighted-sum surrogate ranks the true optimum.

  • We propose a training-free LR-QAOA framework with post-selection on the exact objective and a short local refinement. We compare the proposed LR-QAOA framework with greedy activation, PGGA, and SDR-based classical baseline methods.

  • We validate the proposed QAOA approach on IBM Quantum hardware with up to 120 qubits and report both the quantum processing unit (QPU) time and the solution quality actually achieved.

Together, these contributions show that LR-QAOA finds better activation patterns than greedy, PGGA [4], and SDR [7] in the multi-user case. Moreover, its time complexity scales more favorably than that of exhaustive search, which makes LR-QAOA a scalable candidate for PASS as quantum hardware improves, since its complexity grows polynomially with the number of PAs and linearly with the number of users.

The remainder of this work is organized as follows. § 2 presents the system model and the Ising formulation. § 3 presents the LR-QAOA algorithm, and § 4 presents the simulation and hardware results, while § 5 concludes the paper.

Notations: Matrices and vectors are denoted by boldface uppercase (A\mathbf{A}) and lowercase (a\mathbf{a}) letters, respectively. The transpose and the conjugate transpose of A\mathbf{A} are denoted by AT\mathbf{A}^{T} and AH\mathbf{A}^{H}, respectively. Re⁡(⋅)\operatorname{Re}(\cdot) denotes the real part, ∣⋅∣\lvert\cdot\rvert the absolute value, ∥⋅∥\lVert\cdot\rVert the Euclidean norm, and 1\mathbf{1} the all-ones vector. A complex Gaussian random variable with mean θ\theta and variance σ2\sigma^{2} is denoted by CN(θ,σ2)\mathcal{CN}(\theta,\sigma^{2}). XnX_n and ZnZ_n denote the Pauli-XX and Pauli-ZZ operators acting on qubit nn, ∣0⟩\lvert0\rangle and ∣1⟩\lvert1\rangle the computational basis states, ∣+⟩=(∣0⟩+∣1⟩)/2\lvert+\rangle=(\lvert0\rangle+\lvert1\rangle)/\sqrt{2}, and (⋅)⊗N(\cdot)^{\otimes N} the NN-fold tensor product. The single-qubit rotation gates are RX(θ)=e−iθX/2R_X(\theta)=e^{-i\theta X/2} and RZ(θ)=e−iθZ/2R_Z(\theta)=e^{-i\theta Z/2}, the two-qubit rotation is RZZ(θ)=e−iθZ⊗Z/2R_{ZZ}(\theta)=e^{-i\theta Z\otimes Z/2}, and CNOT denotes the controlled-NOT gate.

System Model and Ising Formulation

PASS Model

We consider the PASS model of [5], as shown in Figure 1. A waveguide of length LL is deployed along the xx-axis at height HH and is fed at ψ0=(−L/2,0,H)\boldsymbol{\psi}_0=(-L/2,0,H). The NN PAs are located at ψn=(−L/2+(2n−1)L/(2N),0,H)\boldsymbol{\psi}_n=(-L/2+(2n-1)L/(2N),0,H), and MM single-antenna users are located at ψm=(xm,ym,0)\boldsymbol{\psi}_m=(x_m,y_m,0), where xmx_m and ymy_m are uniformly distributed in [−L/2,L/2][-L/2,L/2] and are assumed to be known at the base station. Neglecting in-waveguide loss, the effective channel bm,n=hm,ngnb_{m,n}=h_{m,n}g_n from PA nn to user mm is given by

PASS with a waveguide of length $L$ at height $H$ carrying $N$ pre-installed PAs; one activation pattern $\mathbf{a}$ serves $M$ users

Figure 1. PASS with a waveguide of length LL at height HH carrying NN pre-installed PAs; one activation pattern a\mathbf{a} serves MM users.

bm,n=e−j2πλ∥ψm−ψn∥∥ψm−ψn∥⏟hm,ne−j2πλg∥ψn−ψ0∥∥ψn−ψ0∥⏟gn,(1)b_{m,n}= \underbrace{\frac{e^{-j\frac{2\pi}{\lambda}\lVert\boldsymbol{\psi}_m-\boldsymbol{\psi}_n\rVert}}{\lVert\boldsymbol{\psi}_m-\boldsymbol{\psi}_n\rVert}}_{h_{m,n}} \underbrace{\frac{e^{-j\frac{2\pi}{\lambda_g}\lVert\boldsymbol{\psi}_n-\boldsymbol{\psi}_0\rVert}}{\sqrt{\lVert\boldsymbol{\psi}_n-\boldsymbol{\psi}_0\rVert}}}_{g_n}, \tag*{(1)}

where λg=λ/neff\lambda_g=\lambda/n_{\mathrm{eff}} denotes the guided wavelength, and bm=[bm,1,…,bm,N]T\mathbf{b}_m=[b_{m,1},\ldots,b_{m,N}]^{T} denotes the channel vector of the mm-th user. An activation vector a∈{0,1}N\mathbf{a}\in\{0,1\}^{N} switches on 1Ta1^{T}\mathbf{a} PAs, which share the transmit power PP equally. Under time-division access, the SNR of user mm is given by

γm(a)=Pησw2∣aTbm∣21Ta,(2)\gamma_m(\mathbf{a})=\frac{P\eta}{\sigma_w^2}\frac{\lvert\mathbf{a}^{T}\mathbf{b}_m\rvert^2}{\mathbf{1}^{T}\mathbf{a}}, \tag*{(2)}

where η=c2/(16π2fc2)\eta=c^2/(16\pi^2f_c^2) is the free-space path-loss scaling factor, cc denotes the speed of light, fcf_c the carrier frequency, PP the transmit power, and σw2\sigma_w^2 the variance of the additive white Gaussian noise w∼CN(0,σw2)w\sim\mathcal{CN}(0,\sigma_w^2). The corresponding achievable rate is log⁡2(1+γm)\log_2(1+\gamma_m). To ensure fairness across users, the max-min activation problem is given by [5]

max⁡a∈{0,1}N∖{0} Γ(a),Γ(a)=min⁡m∣aTbm∣21Ta.(3)\max_{\mathbf{a}\in\{0,1\}^{N}\setminus\{\mathbf{0}\}}\ \Gamma(\mathbf{a}),\qquad\Gamma(\mathbf{a})=\min_m\frac{\lvert\mathbf{a}^{T}\mathbf{b}_m\rvert^2}{\mathbf{1}^{T}\mathbf{a}}. \tag*{(3)}

Ising Formulation of the Max-Min Problem

To be solved by QAOA, the optimization problem must be expressed as a quadratic function of the binary activation variables, i.e., as an Ising Hamiltonian. However, the problem in (3), contains a ratio and a minimum over the users, neither of which is quadratic. Therefore, we replace it by a quadratic surrogate, which we build in three steps. First, note that for binary a\mathbf{a}, the numerator is a quadratic form given by ∣aTbm∣2=aTRma\lvert\mathbf{a}^{T}\mathbf{b}_m\rvert^2=\mathbf{a}^{T}\mathbf{R}_m\mathbf{a}, where Rm=Re⁡(bmbmH)\mathbf{R}_m=\operatorname{Re}(\mathbf{b}_m\mathbf{b}_m^{H}). The off-diagonal entries of Rm\mathbf{R}_m are positive for PA pairs that arrive in phase at user mm and negative for pairs that cancel. Second, following Dinkelbach’s method [15], Γ(a)≥t\Gamma(\mathbf{a})\geq t holds if and only if aTRma−t1Ta≥0\mathbf{a}^{T}\mathbf{R}_m\mathbf{a}-t\mathbf{1}^{T}\mathbf{a}\geq0 for every user, and hence the ratio becomes a price on each active PA. Instead of iterating on tt, we adopt a fixed price μ\mu. Third, we replace the minimum by a weighted sum with weights wm=∥bm∥−2νw_m=\lVert\mathbf{b}_m\rVert^{-2\nu}, which assigns more weight to users with weak channels. Combining these steps yields the surrogate objective

F(a)=aTRa−μ1Ta,R=∑mwmRm,(4)F(\mathbf{a})=\mathbf{a}^{T}\mathbf{R}\mathbf{a}-\mu\mathbf{1}^{T}\mathbf{a},\qquad\mathbf{R}=\sum_m w_m\mathbf{R}_m, \tag*{(4)}

where μ=κmax⁡nRnn\mu=\kappa\max_n R_{nn} is a fraction of the weighted power of the best single PA. Since an2=ana_n^2=a_n, maximizing FF is equivalent to minimizing C(a)=∑n(μ−Rnn)an−C(\mathbf{a})=\sum_n(\mu-R_{nn})a_n- 2∑n<n′Rnn′anan′2\sum_{n<n'}R_{nn'}a_na_{n'}. Substituting an=(1−zn)/2a_n=(1-z_n)/2 and replacing znz_n by the Pauli operator ZnZ_n yields, up to a constant, the Ising Hamiltonian

HC=∑nhnZn+∑n<n′Jnn′ZnZn′,(5)H_C=\sum_n h_n Z_n+\sum_{n<n'}J_{nn'}Z_nZ_{n'}, \tag*{(5)}

with local fields and pairwise couplings given by

hn=Rnn−μ2+12∑n′≠nRnn′,Jnn′=−Rnn′2,(6)h_n=\frac{R_{nn}-\mu}{2}+\frac{1}{2}\sum_{n'\ne n}R_{nn'},\qquad J_{nn'}=-\frac{R_{nn'}}{2}, \tag*{(6)}

where ∣1⟩n\lvert1\rangle_n indicates that PA nn is active. Since the problem is unconstrained, HCH_C requires no penalty terms, and every measured bitstring except the all-zero one corresponds to a valid pattern.

The surrogate F(a)F(a) is bounded below by an affine function of Γ(a)\Gamma(a) as

F(a)≥((∑mwm)Γ(a)−μ)1Ta,(7)F(a)\geq\left(\left(\sum_m w_m\right)\Gamma(a)-\mu\right)\mathbf{1}^{T}a, \tag*{(7)}

since every user’s gain is at least that of the worst user, with equality for a single user. Hence, a pattern with a good worst user always has a large FF, i.e., a low energy of HCH_C, so the max-min optimum remains among the low-energy patterns that QAOA favors, and post-selection on Γ\Gamma picks it among the sampled patterns. Since a large FF does not guarantee a large Γ\Gamma, this bound loosens as users are added, which is why we add local refinement in § 3. Finally, 0≤κ≤10\leq\kappa\leq1 sets the price of each active PA, such that κ=0\kappa=0 activates too many PAs and κ=1\kappa=1 activates very few or none.

LR-QAOA With Local Refinement

Starting from the initial state ∣+⟩⊗N\lvert+\rangle^{\otimes N}, which is the ground state of the mixer Hamiltonian HM=−∑nXnH_M=-\sum_n X_n and assigns equal probability 2−N2^{-N} to every activation pattern, a depth-pp circuit prepares

∣ψp⟩=e−iβpHMe−iγpHC⋯e−iβ1HMe−iγ1HC∣+⟩⊗N.(8)\lvert\psi_p\rangle=e^{-i\beta_pH_M}e^{-i\gamma_pH_C}\cdots e^{-i\beta_1H_M}e^{-i\gamma_1H_C}\lvert+\rangle^{\otimes N}. \tag*{(8)}

Since all terms of HCH_C in (5) are diagonal in the computational basis and therefore commute, the cost unitary factorizes exactly into single- and two-qubit rotations,

e−iγHC=∏nRZ(2γhn)∏n<n′RZZ(2γJnn′),(9)e^{-i\gamma H_C}=\prod_n R_Z(2\gamma h_n)\prod_{n<n'}R_{ZZ}(2\gamma J_{nn'}), \tag*{(9)}

Algorithm 1 LR-QAOA for Max-Min PA Activation

Require: {bm}\{b_m\}, depth pp, scale Δ\Delta, factors κ,ν\kappa,\nu, number of evaluated patterns SS

Phase 1: Hamiltonian and schedule (classical)

  1. R←∑m∥bm∥−2νRe⁡(bmbmH)R\leftarrow\sum_m\lVert b_m\rVert^{-2\nu}\operatorname{Re}(b_mb_m^H), μ←κmax⁡nRnn\mu\leftarrow\kappa\max_n R_{nn}

  2. Compute hnh_n, Jnn′J_{nn'} by (6) and γℓ\gamma_\ell, βℓ\beta_\ell by (10)

Phase 2: Sampling (quantum)

  1. From ∣+⟩⊗N\lvert+\rangle^{\otimes N}, apply pp layers of cost (RZ,RZZ)(R_Z,R_{ZZ}) and mixer (RX)(R_X)

  2. Measure S/2S/2 shots; A←A\leftarrow distinct nonzero patterns

Phase 3: Post-selection and refinement (classical)

  1. Sort AA by decreasing Γ\Gamma; a⋆←a^\star\leftarrow best of AA

  2. for each a∈Aa\in A, until S/2S/2 more evaluations are used do

  3. &nbsp;&nbsp;&nbsp;&nbsp;repeat flip PAs of aa one at a time in random order; accept the first flip that raises Γ\Gamma

  4. &nbsp;&nbsp;&nbsp;&nbsp;until no flip improves Γ\Gamma

  5. &nbsp;&nbsp;&nbsp;&nbsp;a⋆←arg⁡max⁡{Γ(a⋆),Γ(a)}a^\star\leftarrow\arg\max\{\Gamma(a^\star),\Gamma(a)\}

  6. end for

  7. return a⋆a^\star

where RZZR_{ZZ} acts on qubits nn and n′n'. The cost layer imprints on each pattern a phase proportional to its cost, so that, for the computational basis state ∣a⟩\lvert a\rangle, e−iγHC∣a⟩=e−iγE(a)∣a⟩e^{-i\gamma H_C}\lvert a\rangle=e^{-i\gamma E(a)}\lvert a\rangle, where E(a)E(a) denotes the energy of aa under HCH_C. Similarly, the mixer factorizes into independent single-qubit rotations e−iβHM=∏nRX(−2β)e^{-i\beta H_M}=\prod_n R_X(-2\beta), which let the amplitudes of patterns that differ in one PA interfere. Through this interplay of phases and interference, the amplitudes of low-energy patterns grow, whereas those of high-energy patterns are suppressed. As depicted in Figure 2, each cost layer thus uses one RZR_Z gate per PA and one CNOT–RZR_Z–CNOT block per coupling, whereas each mixer layer uses NN RXR_X rotations. Since RR is generally dense, every pair of PAs is coupled, and one layer contains N(N−1)/2N(N-1)/2 two-qubit blocks.

LR-QAOA circuit with p linear-ramp layers

Figure 2. LR-QAOA circuit with pp linear-ramp layers; measured patterns are post-selected on Γ\Gamma and refined by local search.

Measuring ∣ψp⟩\lvert\psi_p\rangle in the computational basis returns pattern a\mathbf{a} with probability ∣⟨a∣ψp⟩∣2\lvert\langle a\vert\psi_p\rangle\rvert^2, where outcome ∣1⟩n\lvert1\rangle_n is read as an=1a_n=1. QAOA is therefore used here as a sampler rather than as an estimator of ⟨HC⟩\langle H_C\rangle: we only require that good patterns appear among the measured ones, not that the most frequent pattern be optimal. This viewpoint matters because, as shown in § 4-A, the ground state of HCH_C is often not the max-min optimum.

Standard QAOA tunes the 2p2p angles {γℓ,βℓ}\{\gamma_\ell,\beta_\ell\} for each instance with a classical optimizer, which requires many repeated circuit executions per channel realization. Rather than training the angles, we fix them by a linear-ramp schedule, given by [14, 16]

γℓ=ℓpΔ,βℓ=(1−ℓ−1p)Δ,(10)\gamma_\ell=\frac{\ell}{p}\Delta,\qquad\beta_\ell=\left(1-\frac{\ell-1}{p}\right)\Delta, \tag*{(10)}

where the single scale parameter Δ\Delta is selected once offline and applied to all instances. The schedule in (10) mimics a first-order Trotterization of the annealing path H(s)=(1−s)HM+sHCH(s)=(1-s)H_M+sH_C, s∈[0,1]s\in[0,1], with Δ\Delta acting as the time step: the cost angles increase while the mixer angles decrease, so the circuit gradually shifts weight from HMH_M to HCH_C. A small Δ\Delta keeps the state close to ∣+⟩⊗N\lvert+\rangle^{\otimes N}, so that the samples approach uniform sampling, whereas a very large Δ\Delta increases the discretization error. Within a broad intermediate range, the samples concentrate on low-energy patterns without any per-instance training, which makes the schedule suitable when the users move and a new instance arises in every channel realization.

As outlined in Algorithm 1, the S/2S/2 measured patterns are scored with the exact objective Γ\Gamma rather than the surrogate, which corrects for the mismatch between FF and Γ\Gamma discussed in § 2-B. The remaining S/2S/2 evaluations are spent on a single-flip local search started from the best measured patterns, so exactly SS patterns are evaluated in total, as for every sampling baseline. In the remainder of this work, “LR-QAOA” refers to this complete procedure.

Complexity Analysis: The LR-QAOA circuit requires NN qubits and O(pN2)O(pN^2) gate operations, or a depth of O(pN)O(pN) with a swap network [17]. Phases 1 and 3 require O(N2M+SNM)O(N^2M+SNM) operations, and no classical optimization loop is involved. Moreover, since one evaluation of Γ\Gamma takes about 5 μ\mus on a CPU, whereas one shot on current hardware takes 1.4–7.8 ms including overheads, replacing half of the shots by local search halves the number of shots and thus reduces the QPU time. In contrast, exhaustive search requires O(2NM)O(2^NM) operations with Gray-code updates.

Numerical Results

Simulation Results

We consider a simulation scenario by adopting the parameters of [5], i.e., H=3H=3 m, L=50L=50 m, fc=28f_c=28 GHz, neff=1.4n_{\mathrm{eff}}=1.4, a noise power of −90-90 dBm, and P=10P=10 dBm, and use M=3M=3, p=5p=5, and Δ=1.75\Delta=1.75 unless stated otherwise. The scale Δ\Delta is selected once offline on 60 held-out instances with N=12N=12 and kept fixed for all other instances. LR-QAOA is compared with greedy activation, PGGA [4], and SDR with Gaussian randomization [7] using the same number of evaluated patterns SS. The surrogate does not always place the optimum at the ground state of HCH_C, especially when several users are served. Still, the optimum usually lies among the lowest-energy patterns, so post-selection on Γ\Gamma picks it up from the samples. A small κ\kappa favors many active PAs, which the power split penalizes. In practice, the choice of κ\kappa matters little, since the loss is almost the same for any κ\kappa between 0.7 and 1. We therefore fix κ=0.75\kappa=0.75 offline, just like Δ\Delta, and use ν=1\nu=1.

Figure 3(a) shows the worst-user rate versus NN for M=3M=3 and S=1024S=1024, averaged over 1000 instances. It can be observed that the performance of LR-QAOA almost overlaps the exhaustive search upper bound for all NN, even though it looks at only a tiny fraction of the patterns, whereas SDR falls slightly further behind as NN grows. Furthermore, the classical greedy activation and PGGA methods show a much larger performance gap since both of these methods add one PA at a time and stop as soon as no single addition improves Γ\Gamma. Because every new PA takes power away from the others, adding one PA often lowers Γ\Gamma, even when adding several PAs together would raise it. Therefore, these baseline methods stop too early, with fewer active PAs than the optimal pattern.

Three-panel comparison of worst-user rate, rate loss, and execution time

Figure 3. (a) Worst-user rate versus NN, (b) rate loss of LR-QAOA versus SS, and (c) execution time versus NN on ibm_fez quantum processor.

We define the rate loss of a method that returns the pattern a^\hat{\mathbf a} as

L(a^)=log⁡2(1+γˉΓ(aopt))−log⁡2(1+γˉΓ(a^)),(11)\mathcal{L}(\hat{\mathbf a})=\log_2\left(1+\bar{\gamma}\Gamma(\mathbf a^{\mathrm{opt}})\right)-\log_2\left(1+\bar{\gamma}\Gamma(\hat{\mathbf a})\right), \tag*{(11)}

where γˉ=Pη/σw2\bar{\gamma}=P\eta/\sigma_w^2 and aopt\mathbf a^{\mathrm{opt}} is the solution of (3), found by exhaustive search over all 2N−12^N-1 patterns. Figure 3(b) shows the rate loss of LR-QAOA versus SS for N=12N=12, 16, and 20, using post-selection only. It can be observed that for M=1M=1, a few tens of patterns SS are enough to bring the loss below 0.03 bps/Hz (dotted line), whereas M=3M=3 needs a few hundred, and more as NN grows. Hence, more users need more evaluated patterns to reach the same rate loss.

Quantum Hardware Results

Next, we execute LR-QAOA on the 156-qubit IBM Heron processor ibm_fez for N=6N=6–120, with p=5p=5, M=3M=3, and 32 shots per instance, using Pauli twirling and dynamical decoupling. Given QPU access costs, we limit sampling to 32 shots and refine the best measured shot with a short local search of at most 512 evaluations. SDR uses the same number of randomized candidates. The QPU time excludes queuing and transpilation overheads, whereas exhaustive search and SDR run on a CPU. Since exhaustive search cannot be evaluated beyond N=24N=24 on a classical CPU, the runtime results are extrapolated beyond N=24N=24. Exhaustive search and SDR runtimes are reported for a server with two AMD EPYC 9555 64-core processors (128 cores, 256 threads) and 1 TB of RAM, where SDR is solved by CVXPY [18].

Figure 3(c) compares the runtimes of the proposed LR-QAOA on actual quantum hardware with that of the classical exhaustive search and SDR on classical CPU. It can be observed that the QPU run-time of LR-QAOA stays at 2 s for every NN, because it is set by the fixed overhead of each job rather than by the number of qubits. Exhaustive search, on the other hand, doubles its CPU time with every added PA and becomes slower than LR-QAOA beyond N∼26N\sim26. Furthermore, it can be observed that the runtime of SDR keeps growing with NN and is more than that of the quantum runtime for beyond N=24N=24 PAs.

Table I compares the run times and worst user rates of the LR-QAOA on quantum hardware against the classical exhaustive search (Exh.) and SDR-based methods. Here, Shots denotes the worst-user rate of the best of the 32 measured patterns, i.e., LR-QAOA with post-selection on Γ\Gamma but without local search. For N≥32N\ge32, the time of exhaustive search is extrapolated from the 36.4 ns per pattern measured at N=24N=24. The table shows that LR-QAOA keeps its advantage over SDR on real hardware as well. It finds the optimum for every N≤20N\le20 and comes very close at N=24N=24, while SDR misses it at several sizes. For larger NN, the two methods are nearly equal at N=32N=32, and LR-QAOA is clearly better at N=64N=64 and 120. The LR-QAOA based on shots without post selection, however, are far from optimal, because the circuit contains pN(N−1)/2pN(N-1)/2 two-qubit blocks whose errors add up even with error mitigation. Most of the gain therefore comes from the local search post-selection. Therefore, we do not claim a quantum advantage on today’s hardware. However, our quantum hardware results show that the QPU run-time does not grow with NN, so the advantage over exhaustive search will become usable as quantum hardware gate errors come down in the future.

NNTime (s) Exh.Time (s) QPUTime (s) SDRRate (bps/Hz) Exh.Rate (bps/Hz) LR-QAOARate (bps/Hz) ShotsRate (bps/Hz) SDR
69.9⋅10−59.9 \cdot10^{-5}2.00.074.1374.1373.9154.137
89.5⋅10−59.5 \cdot10^{-5}2.00.344.4024.4024.4024.402
101.8⋅10−41.8 \cdot10^{-4}2.00.074.4604.4603.9414.058
124.6⋅10−44.6 \cdot10^{-4}2.00.534.6774.6774.0014.677
167.0⋅10−37.0 \cdot10^{-3}2.00.835.2135.2134.0165.213
205.5⋅10−25.5 \cdot10^{-2}2.00.905.1675.1674.4394.973
240.612.02.655.7375.7124.0575.550
321.6⋅1021.6 \cdot10^{2}2.047.6–5.5853.6055.603
646.7⋅10116.7 \cdot10^{11}2.083.8–6.3303.7695.918
1204.8⋅10284.8 \cdot10^{28}2.0212.8–7.3433.8743.182

Table 1. Runtime and worst-user rate comparison for M=3M = 3 users.

Conclusion

We have proposed a training-free quantum optimization algorithm that selects which pinching antennas to activate in PASS to maximize the worst-user rate. We mapped the max-min SNR optimization problem to a two-local Ising Hamiltonian with one qubit per PA and no penalty terms through a quadratic surrogate that is lower-bounded by the worst-user SNR, with equality for a single user. Since this bound loosens for several users, the LR-QAOA samples are post-selected on the exact objective and refined by a short local search. Our simulation results show that LR-QAOA outperforms greedy activation, PGGA, and SDR-based classical baseline methods when all methods test the same number of PA on/off patterns. Furthermore, on the IBM quantum processor ibm_fez, the LR-QAOA takes only 2 s of QPU run-time for up to 120 qubits, and with local search post-selection, LR-QAOA finds the optimum up to N=20N = 20 and beats SDR for large NN. However, in the NISQ era, due to quantum hardware gate noise, most of the gain of the proposed LR-QAOA comes from the local search post-selection. Future work will focus on investigating shallower hardware-aware circuits, stronger error mitigation, and multi-waveguide PASS systems.

References

  1. [1]A. Fukuda, H. Yamamoto, H. Okazaki, Y. Suzuki, and K. Kawai, “Pinching antenna: Using a dielectric waveguide as an antenna,” NTT DOCOMO Technical Journal, vol. 23, no. 3, pp. 5–12, 2022.
  2. [2]Y. Liu, H. Jiang, X. Xu, Z. Wang, J. Guo, C. Ouyang, X. Mu, Z. Ding, A. Nallanathan, G. K. Karagiannidis, and R. Schober, “Pinching-antenna systems (PASS): A tutorial,” pp. 4881–4918, 2026.DOI
  3. [3]K. Wang, Z. Ding, and G. K. Karagiannidis, “Antenna activation and resource allocation in multi-waveguide pinching-antenna systems,” IEEE Transactions on Wireless Communications, vol. 25, pp. 4070–4082, 2026.arxiv.org/abs/2505.02864
  4. [4]L. Cao, H. Yin, X. Pei, R. Song, and Y. Liu, “Multi-pin and movable-pin enabled discretely-activated pinching antenna systems,” IEEE Wireless Communications Letters, vol. 15, pp. 590–594, 2026.DOI
  5. [5]V. E. Galanopoulou, T. K. Yiokou, G. D. Karagiannidis, S. A. Tegos, and P. D. Diamantoulakis, “Viterbi state selection for discrete pinching antenna systems,” IEEE Communications Letters, vol. 30, pp. 1523–1527, 2026.
  6. [6]O. G. Karagiannidis, V. E. Galanopoulou, P. D. Diamantoulakis, Z. Ding, and O. A. Dobre, “Deep learning optimization of two-state pinching antennas systems,” IEEE Transactions on Cognitive Communications and Networking, vol. 12, pp. 4942–4956, 2026.arxiv.org/abs/2507.06222
  7. [7]Z.-Q. Luo, W.-K. Ma, A. C.-C. So, Y. Ye, and S. Zhang, “Semidefinite relaxation of quadratic optimization problems,” IEEE Signal Processing Magazine, vol. 27, no. 3, pp. 20–34, 2010.DOI
  8. [8]E. Farhi, J. Goldstone, and S. Gutmann, “A quantum approximate optimization algorithm,” arXiv:1411.4028, 2014.arxiv.org/abs/1411.4028
  9. [9]K. Blekos, D. Brand, A. Ceschini, C.-H. Chou, R.-H. Li, K. Pandya, and A. Summer, “A review on quantum approximate optimization algorithm and its variants,” Physics Reports, vol. 1068, pp. 1–66, 2024.DOI
  10. [10]J. Cui, Y. Xiong, S. X. Ng, and L. Hanzo, “Quantum approximate optimization algorithm based maximum likelihood detection,” IEEE Transactions on Communications, vol. 70, no. 8, pp. 5386–5400, 2022.arxiv.org/abs/2107.05020
  11. [11]M. Kim, D. Venturelli, and K. Jamieson, “Leveraging quantum annealing for large MIMO processing in centralized radio access networks,” in Proc. ACM SIGCOMM, 2019, pp. 241–255.DOI
  12. [12]S. Paul and N. K. Kundu, “Efficient quantum algorithm for phase optimization of 1-bit RIS-assisted MIMO communication system,” 2026. [Online]. Available: https://arxiv.org/abs/2607.14700
  13. [13]S. Paul, S. Banerjee, D. Bhowmik, and N. K. Kundu, “Warm-start quantum approximate optimization algorithm for QAM MIMO data detection,” 2026. [Online]. Available: https://arxiv.org/abs/2604.18479
  14. [14]J. A. Montañez-Barrera and K. Michielsen, “Toward a linear-ramp QAOA protocol: Evidence of a scaling advantage in solving some combinatorial optimization problems,” npj Quantum Information, vol. 11, p. 131, 2025.
  15. [15]W. Dinkelbach, “On nonlinear fractional programming,” Management Science, vol. 13, no. 7, pp. 492–498, 1967.DOI
  16. [16]S. H. Sack and M. Serbyn, “Quantum annealing initialization of the quantum approximate optimization algorithm,” Quantum, vol. 5, p. 491, 2021.DOI
  17. [17]B. O’Gorman, W. J. Huggins, E. G. Rieffel, and K. B. Whaley, “Generalized swap networks for near-term quantum computing,” 2019. [Online]. Available: https://arxiv.org/abs/1905.05118
  18. [18]S. Diamond and S. Boyd, “CVXPY: A Python-embedded modeling language for convex optimization,” Journal of Machine Learning Research, vol. 17, no. 83, pp. 1–5, 2016.

Paper details

vlnc:W_8edwmzcs
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents