Antennas that can change position escape a limit that fixed antennas cannot. A fixed antenna may sit in a spot where signals from different paths cancel, and it has to stay there. A movable antenna can slide a fraction of a wavelength and land on a spot where those signals add up. The price of that freedom is search. Every candidate spot must be probed with pilot signals before the system knows which one is good. With several elements, the number of joint placements grows exponentially: exhaustive search over a pinching-antenna waveguide needs more than $2^{20}$ training steps (Beam Training for Pinching-Antenna Systems (PASS)).
Protein bioinformatics hit the same exponential wall with sequences and solved it in a specific way. A reduced amino acid alphabet merges the 20 residues into a handful of interchangeable groups, so that statistics over short windows of $k$ residues become estimable, and exact lookup by hashing becomes possible. We propose to do the same for antenna positions. Each element's dense set of candidate positions is merged into a small "position alphabet" of $c$ symbols, grouping positions that are interchangeable for the rate they deliver across many user locations. Joint placements then become short words over that alphabet. A table of how well each word performs can be learned offline from channel statistics and consulted online with a few pilots. The scientific claim is that the best alphabet size $c$ follows a predictable tradeoff between quantization and estimation error, and that it grows with carrier frequency. This parallels how the best amino acid alphabet size shifts between datasets.
Core Claim
The claim has three parts. First, a position alphabet built from channel statistics, with merges chosen to minimize expected rate loss, reaches a larger fraction of the oracle rate than a uniform grid with the same number of symbols. It also needs fewer pilot steps than the multi-stage search in current pinching-antenna work. Second, the gain comes from non-uniformity. Along a waveguide or aperture, the distance over which the channel stays correlated varies with user geometry, so a uniform grid wastes symbols in smooth regions and under-resolves the regions where the phase changes quickly. Third, there is a measurable crossover in alphabet size. Below it, merging destroys the phase resolution needed for near-field focusing. Above it, the joint-placement table cannot be estimated from the available data. The crossover should move to larger $c$ at 28 GHz than at 5 GHz.
The protein literature motivates the third part. In Prediction of Protein–Protein Interactions with Clustered Amino Acids and Weighted Sparse Representation, which combines cluster-profile encodings with a sparse-representation classifier, the best alphabet was 8 clusters on yeast (90.91% accuracy) but 13 clusters on H. pylori (83.04% accuracy). The right reduction depends on the data, not on chemistry alone. The file flags this tension as an open gap: aggressive reduction filters noise in proteins, while pinching-antenna focusing becomes more sensitive to small position shifts at high frequency. No one has characterized where one regime turns into the other. We intend to measure that boundary in the antenna setting, where ground-truth channels can be simulated exactly.
Research Program
The method borrows two specific ideas from the reduced-alphabet literature and translates the rest.
The first borrowed idea is the substitution matrix: a score for how safely one residue can replace another. For candidate positions $p$ and $q$ of one element, we define the cost of replacing $q$ by $p$ as the expected rate lost across a training distribution of user locations $u$:
$$D(p,q) = \mathbb{E}_u\left[\max_{\text{config} \ni q} R_u - \max_{\text{config} \ni p,\ \text{config}\setminus q} R_u\right].$$
Here $R_u$ is the achievable rate for user $u$ with a given placement of all elements. Positions are merged by agglomerative clustering on $D$ until $c$ symbols remain. Each symbol is represented by the member position that minimizes expected loss. This is deliberately task-aware. Two positions whose channels look similar in isolation but interact differently with the other elements' phases will not be merged. The analogy is to reduced alphabets that are tuned for a prediction task rather than defined by physicochemistry alone.
The second borrowed idea is exact lookup over reduced words, as in Clustering huge protein sequence sets in linear time. That work hashes 13-letter reduced $k$-mers to group 1.6 billion sequences in near-linear time. In our setting, offline simulation fills a table indexed by a coarse user signature: the quantized received-signal-strength profile from a single probing antenna, which is the same measurement used in Stage 1 of the three-stage scheme. For each signature, the table stores the top few alphabet words. Online, the system takes the coarse measurement, retrieves candidates, and spends its remaining pilots only on those candidates.
The quantization–estimation tradeoff gives the alphabet-size rule. A position error $\delta$ produces a phase error of about $2\pi\delta/\lambda$, and the rate loss is roughly quadratic in small phase error. With aperture length $L$ split into $c$ cells, quantization loss therefore scales like $(L/(c\lambda))^2$. The estimation error of a table over $c^N$ words from $n$ offline samples scales like $\sqrt{c^N/n}$. The minimizing $c$ increases as $\lambda$ shrinks and as $n$ grows. The precise exponents are a prediction to test, not an assumption. In practice we choose $c$ by held-out rate, and then check whether the chosen values follow this scaling.
The fixed-position constraint enters in two places. It is the naive reference system, and it is also a hardware reality. Pinch points and fluid-antenna ports are already discrete. Reducing their number cuts pilot overhead and also the actuation or switching burden that the file identifies as a limitation of movable designs.
Baselines
The naive reference is the fixed-position system: fixed-interval pinching ("Fixed Pinc.") and a conventional uniform linear array ("Conv."), as defined in Beam Training for Pinching-Antenna Systems (PASS). That paper reports that dynamic pinching consistently beat both on data rate. It gives no numeric rate in the file, so we use these systems only as a floor. Any position alphabet that does not clearly exceed them has failed.
The strongest existing method is the Three-Stage Beam Training scheme (3SBT) from the same paper. It uses a coarse single-antenna strength estimate, then multi-antenna phase matching, then a local search over a scalable spatial codebook. It reaches its alignment in 48 training steps, against more than $2^{20}$ for exhaustive search. No reported data-rate figure for 3SBT was found on Lacuna, so we will reimplement it on the same simulation settings the paper describes: 28 GHz and 5 GHz carriers, a near-field spherical-wave model, 18 activated pinching antennas, −90 dBm noise, and a 3 m waveguide height. We will first reproduce its qualitative ordering over fixed pinching before any comparison. The comparison is at matched pilot budgets of 16, 32, 48, and 96 steps. The position alphabet counts every online measurement against that budget, including the coarse signature. Beating 3SBT means either a higher median fraction of oracle rate at 48 steps, or matching its 48-step rate with at most half the steps, on held-out user positions.
A uniform grid with exactly $c$ symbols per element, using the same lookup table and online procedure, isolates the non-uniform, statistics-driven merging.
A geometric k-means over coordinates, also with $c$ symbols, tests whether channel statistics matter beyond geometry.
An ablation that merges positions by per-position channel correlation, rather than the joint rate-loss cost $D$, isolates the task-aware ingredient.
A second ablation keeps the alphabet but drops the offline table and searches the reduced space greedily online. It isolates the lookup idea.
Every learned method receives the same offline simulation budget and the same held-out users. Tuning budgets are equal and cover $c$ for our method, codebook size for 3SBT, and cluster count for the k-means variants. The upper bound is the oracle placement under full channel knowledge, found by continuous optimization.
Why not just…?
Why not just use a uniform grid at half-wavelength spacing? This is the default spatial codebook, and in far-field, rich-scattering conditions it may be nearly optimal, because channel correlation there is close to uniform in space. The threat is real, and the uniform-grid baseline at equal $c$ tests it directly. Our hypothesis relies on near-field geometry, where the rate of phase change along the waveguide depends on user distance and angle. If the learned alphabet beats the uniform grid by more at 28 GHz and short user distances than at 5 GHz and long distances, non-uniformity is the mechanism. If the two tie everywhere, the idea reduces to choosing grid density, and we will report that.
Why not just design the codebook with standard vector quantization (Lloyd iterations on channel vectors)? This is the closest classical competitor. Our alphabet differs in two ways. It quantizes each element's physical position rather than the joint beam vector, so joint placements factor into words that can be counted and hashed. It also merges by downstream rate loss rather than by channel distance. The per-position-correlation ablation is essentially a Lloyd-style distance criterion applied per position. If it matches the full method, the task-aware cost $D$ is unnecessary. We would then keep only the factorized-word and lookup contribution and say so.
Why not optimize continuous positions by gradient or manifold methods? 6D Movable Antenna Enhanced Multi-Access Point Coordination via Position and Orientation Optimization does this with successive convex approximation and Riemannian optimization. The problem is non-convex, and it assumes channel knowledge across the aperture. The pinching-antenna tutorial (Pinching-Antenna Systems (PASS): A Tutorial) identifies acquiring that knowledge as the main overhead. We use continuous optimization with full channel knowledge as the oracle ceiling, not as a competitor. The honest comparison is continuous optimization fed by channels estimated from the same pilot budget, which we include at each budget.
Why not regress positions directly from pilots with a neural network, instead of discrete words? This is the antenna analogue of replacing reduced alphabets with continuous protein embeddings. In proteins, LGCA-VHPPI found that a discrete 6-class encoding beat a doc2vec embedding predictor by 2–6% in Matthews correlation. That result comes from a different domain, however, so it is weak evidence here. We will train a small regressor from the same coarse signature to positions, with the same offline data. If it matches the table at equal pilots, the case for discreteness rests only on interpretability and actuation savings.
Proposed Experiments
We will build the simulator from the 3SBT setting described above, since the file gives no public channel dataset (no DeepMIMO artifact was found on Lacuna). We will extend it with scattering clusters, so that uniform correlation does not hold by construction. We sample user positions over a 2D grid under the waveguide, and hold out both unseen positions and unseen scatterer layouts.
The primary metric is the fraction of oracle rate as a function of pilot steps. Secondary metrics are sum rate for two to four users under non-orthogonal multiple access, the number of distinct pinch activations (as a proxy for switching cost), and table size.
The decisive check is the alphabet-size sweep. We vary $c$ from 2 to the full candidate count, at both carrier frequencies and at three offline sample sizes $n$, and plot held-out rate against $c$. The proposal's central prediction holds if each curve has an interior maximum, if that maximum shifts to larger $c$ at 28 GHz and at larger $n$, and if the shift roughly follows the scaling derived from phase error. It fails if held-out rate rises monotonically with $c$, because then reduction only saves pilots and never helps estimation.
A second check compares the learned alphabet with the uniform grid, separately for near-field and far-field users. The mechanism claim requires the gap to concentrate in the near field.
A third check concerns transfer. The learned alphabet, frozen, should keep most of its advantage on held-out scatterer layouts. This would mirror how a reduced amino acid alphabet learned on one proteome remains useful on another.
References
- Beam Training for Pinching-Antenna Systems (PASS)
- Prediction of Protein–Protein Interactions with Clustered Amino Acids and Weighted Sparse Representation
- Clustering huge protein sequence sets in linear time
- 6D Movable Antenna Enhanced Multi-Access Point Coordination via Position and Orientation Optimization
- Pinching-Antenna Systems (PASS): A Tutorial
- LGCA-VHPPI: A local-global residue context aware viral-host protein-protein interaction predictor