Nearest Is Not Safest: Fluid-Antenna Port Selection for Over-the-Air Backpropagation in Urban Multipath
Authors: Xianzhong Li, Yuzhi Yang, Tianhao Guo
Organizations: College of Physics and Electronic Engineering, Shanxi University, Taiyuan 030006, China · College of Information Science, Japan Advanced Institute of Science and Technology, Nomi 9231292, Japan
Over-the-air backpropagation returns a loss gradient through the reciprocal channel, which requires the return port to see the channel of the forward port. If the terminal moves in between, a fluid antenna system can return through the port nearest to the forward position. In ray-traced Chicago streets, with a pedestrian displacement of about half a wavelength and a street-level base station, this rule reverses 20 percent of the gradients at a port spacing of 0.7 wavelengths and 49 percent at 0.9 wavelengths, while one of the four nearest ports stays aligned in all but 2 and 9 percent. We derive a certificate for the sign of the mean correlation from the second moment of a planar angular spectrum. It reduces to a linear program with three variables and infinitely many constraints, is tight, and holds over a region of second moments. In a static Gaussian model, observations collected along a line cannot tell an angular spectrum from its mirror image, so the sign of the cross moment must come from other information. At unsurveyed positions of the same city, a knowledge map that transports surveyed paths through virtual sources and fits them to the causal channel history lowers the reversal rate at 0.9 wavelengths from 49 to 27 percent. Its gain over the nearest port survived every tested error, while its gain over a geometry rule stayed significant at the tested single-error levels of 5 degrees of phase drift per step and 2 mm of position error but not at the next tested levels. Without any survey, across five cities and ten base stations, the nearest port reverses 37 percent of the returns at 0.9 wavelengths and a rule based on the direction and elevation of the base station 20.5 percent. The multilayer perceptrons we tested did not beat these physical rules by the margin fixed in advance in the primary settings.
Figures & tables
Fig. 1: Region of second-moment error ε and position error η in which the certificate ( 4 ) keeps the mean correlation of the nearer example port negative. The dashed line bounds the region in which the farther port keeps a positive mean correlation.
Street level (31 m)
Rooftop (108 m)
d/λ
0.5
0.7
0.9
0.5
0.7
0.9
Nearest
0.122
0.196
0.485
0.004
0.002
0.058
LoS with elevation
0.165
0.257
0.347
0.002
0.003
0.005
Quadratic rule
0.102
0.159
0.273
0.002
0.001
0.003
Spec (reference)
0.100
0.148
0.240
0.002
0.001
0.004
Orc (reference)
0.002
0.016
0.092
0
0
0.001
TABLE I: Reversal rate with pedestrian motion and a finite aperture in Chicago (1000 queries per site). The quadratic rule minimizes rTCr with the horizontal second moment C of the paths at the query position.
Fig. 2: Reversal rate of the nearest port, the full-spectrum reference (Spec) and the best of four candidates (Orc) against the port spacing. Pedestrian motion, finite aperture, 9600 queries per site.
Fig. 3: Reversal rate of the nearest port at d=0.9λ for the street-level site, averaged over 5×5 m cells.
Rule
Full survey
10 m survey
Nearest
0.494
0.494
LoS with elevation
0.366
0.366
Selected map without history
0.319
0.330
Learned without history
0.318
0.318
VS-Hist
0.270
0.265
Learned with history
0.267
0.259
TABLE II: Reversal rate at unsurveyed positions, Chicago street-level site, d=0.9λ .
Rule
Reversal rate
E[sˉ]
Nearest
0.367
0.200
Street axis
0.479
0.022
LoS with elevation
0.205
0.496
Learned, other cities
0.214
0.461
Learned with history, other cities
0.225
0.423
Spec (reference)
0.144
0.603
TABLE III: Zero-survey port selection in five cities, d=0.9λ , all ten sites pooled. E[sˉ] is the mean of s clipped to [−3,3] .
Fig. 4: Zero-survey reversal rate per site at d=0.9λ . Each city is held out in turn and the learned predictor is trained on the other four cities.
Fig. 5: Reversal rate at unsurveyed positions under errors in the history and the positions, Chicago street level, d=0.9λ , full survey. Each panel adds one error, and the other errors are zero.
d/λ
0.7
0.9
1.2
Nearest
0.496
0.031
−0.226
LoS with elevation
0.390
0.240
0.219
Selected map without history
0.515
0.325
0.273
VS-Hist
0.569
0.417
0.306
Learned with history
0.563
0.408
0.299
Spec (reference)
0.618
0.501
0.349
TABLE IV: Mean of the clipped alignment sˉ=min(max(s,−3),3) at unsurveyed positions (Chicago street level, full survey).
Fig. 6: Distribution of the clipped alignment factor of the chosen port at unsurveyed positions, Chicago street level, d=0.9λ , full survey.
Rule
Return 20 dB
Return 10 dB
Cosine
Proj.
Nearest
76.8 [72.2, 81.3]
76.2 [71.6, 80.9]
0.037
0.023
LoS with elevation
89.5 [88.7, 90.3]
89.5 [88.7, 90.4]
0.260
0.225
Selected map
90.3 [89.4, 91.1]
90.4 [89.5, 91.2]
0.357
0.296
VS-Hist
90.5 [90.1, 90.9]
90.8 [90.3, 91.2]
0.446
0.385
Orc (reference)
91.0 [90.4, 91.6]
91.2 [90.6, 91.8]
0.734
0.688
Ideal, hr=hf
90.6 [89.8, 91.4]
91.1 [90.5, 91.7]
0.941
0.978
TABLE V: Training through the scalar link with truncated channel inversion (binary digits task, linear classifier, 2000 steps). Test accuracy in percent, mean and 95 percent interval over ten seeds, and at a return signal-to-noise ratio of 20 dB the mean cosine between the returned and the exact gradient and the median projection coefficient.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
d/λ
ε
Euclid
Quadratic
Certificate
Oracle
Certified
0.5
all
0
0
0
0
2000
0.7
0
0.148
0.047
0.048
0.030
1808
0.7
0.05
0.148
0.050
0.057
0.030
1645
0.7
0.1
0.148
0.057
0.081
0.030
1489
0.9
0
0.340
0.226
0.222
0.144
1284
0.9
0.05
0.340
0.232
0.245
0.144
1090
Appendix
TABLE VI: Two-cluster synthetic spectra, 2000 scenes per row. Fraction of choices with negative mean correlation, and the number of scenes in which the certificate rule chose a candidate with a positive lower bound. No such candidate had a negative true c .
d/λ
Profile
Nearest
M2
M2c4
M4
Best
0.5
A, B, C
0
0
0
0
0
0.5
D, E
0.009, 0.010
0
0
0
0
0.7
A
0.141
0.027
0.026
0.026
0.026
0.7
B
0.042
0.025
0.025
0.025
0.025
0.7
C
0.050
0.025
0.025
0.025
0.025
0.7
D
0.236
0.043
0.043
0.043
0.043
Appendix
TABLE VII: 3GPP CDL profiles reduced to the plane and rotated at random, 1000 geometries per entry. Fraction of choices with negative mean correlation. M2 is the bound rule from the second moment, M2c4 adds E[cos4ϑ] and M4 adds both fourth-order moments.
Site
d/λ
Nearest
LoS
Quadratic
Spec
Orc
Rooftop 108 m
0.9
0.034
0.030
0.005
0.004
0
1.2
0.144
0.464
0.026
0.027
0.023
1.5
0.246
0.589
0.078
0.078
0.072
Street 31 m
0.9
0.352
0.291
0.208
0.185
0.078
1.2
0.498
0.336
0.347
0.233
0.126
1.5
0.552
0.296
0.407
0.182
0.070
Appendix
TABLE VIII: Vehicular displacements in Chicago, unrestricted lattice, 1000 queries per site and spacing. LoS uses the azimuth only. The median numbers of paths are 20, 37 and 21.
Rule
20 m (13)
10 m (31)
5 m (79)
2 m (236)
Full (4517)
Nearest
0.494
0.494
0.494
0.494
0.494
LoS with elevation
0.366
0.366
0.366
0.366
0.366
IDW, score c
0.397
0.324
0.315
0.340
0.328
IDW, score ( 5 )
0.395
0.321
0.316
0.330
0.319
VS, score c
0.346
0.330
0.286
0.281
0.316
VS, score ( 5 )
0.345
0.306
0.290
0.280
0.313
Appendix
TABLE IX: Reversal rate at unsurveyed positions, Chicago street-level site, d=0.9λ , 2400 test queries in 22 cells, by survey density (number of survey points in parentheses). The rule selected on the validation set is named in brackets.
Street level
Street, scattering 0.3
Rooftop (108 m)
Rule
0.7λ
1.2λ
0.9λ
0.9λ
1.2λ
0.9λ , 5 m survey
Nearest
0.200
0.625
0.485
0.071
0.373
0.071
LoS with elevation
0.268
0.382
0.357
0.006
0.095
0.006
Selected without history
0.194
0.353
0.328
0.009
0.097
0.009
VS-Hist
0.180
0.349
0.379
0.006
0.094
0.005
Learned without history
0.193
0.343
0.322
0.018
0.122
0.298
Appendix
TABLE X: Reversal rate at unsurveyed positions for further spacings, the scattering case and the rooftop site, full survey unless stated. The rooftop site has 3120 test queries in 19 cells and a full survey of 3057 points, and its 5 m survey has 63 points.
Rule
0.7λ
0.9λ
1.2λ
Nearest
0.126 (0.607)
0.367 (0.200)
0.573 ( −0.125 )
Street axis
0.292 (0.350)
0.479 (0.022)
0.476 (0.032)
History axial
0.282 (0.367)
0.487 (0.031)
0.485 (0.023)
LoS with elevation
0.132 (0.635)
0.205 (0.496)
0.285 (0.358)
Learned
0.120 (0.636)
0.214 (0.461)
0.305 (0.292)
Learned with history
0.119 (0.630)
0.225 (0.423)
0.316 (0.266)
Appendix
TABLE XI: Zero-survey port selection in five cities, all ten sites pooled, 38184 queries in 3759 clusters. Reversal rate, with the mean clipped alignment E[sˉ] in parentheses.
Nearest
LoS with elevation
Learned
Spec
Orc
Site
0.7λ
0.9λ
1.2λ
0.7λ
0.9λ
1.2λ
0.7λ
0.9λ
1.2λ
0.9λ
0.9λ
Chicago street
0.188
0.495
0.666
0.253
0.362
0.412
0.217
0.371
0.438
0.219
0.102
Etoile street
0.205
0.538
0.663
0.163
0.278
0.323
0.161
0.296
0.358
0.220
0.108
Florence street
0.233
0.499
0.616
0.222
0.340
0.342
0.205
0.348
0.357
0.259
0.120
Munich street
0.221
0.512
0.600
0.210
0.312
0.347
0.191
0.327
0.366
0.232
0.093
San Francisco street
0.175
0.539
0.737
0.142
0.252
0.379
0.144
0.292
0.380
0.212
0.115
Appendix
TABLE XII: Zero-survey reversal rate per site and spacing. Learned is the predictor without history trained on the other four cities.
Setting
Nearest
LoS
Map
VS-Hist
Orc
Nearest − VS-Hist
LoS − VS-Hist
Map − VS-Hist
LoS − Map
No error
0.494
0.366
0.319
0.270
0.090
22.4 [15.0, 28.2]
9.5 [1.3, 20.9]
4.9 [ −0.3 , 11.2]
4.7 [ −0.4 , 10.7]
SNR 20 dB
0.494
0.366
0.319
0.269
0.090
22.5 [15.2, 28.3]
9.7 [1.4, 21.1]
5.0 [ −0.2 , 11.5]
4.7 [ −0.4 , 10.7]
SNR 10 dB
0.494
0.366
0.319
0.276
0.090
21.8 [14.6, 27.4]
9.0 [0.9, 20.2]
4.3 [ −0.8 , 10.5]
4.7 [ −0.4 , 10.7]
SNR 5 dB
0.494
0.366
0.319
0.301
0.090
19.3 [12.3, 24.9]
6.5 [ −1.5 , 17.6]
1.8 [ −3.3 , 7.9]
4.7 [ −0.4 , 10.7]
Drift 5∘
0.494
0.366
0.319
0.278
0.090
21.6 [14.1, 27.5]
8.8 [0.5, 20.0]
4.1 [ −1.2 , 10.5]
4.7 [ −0.4 , 10.7]
Drift 15∘
0.494
0.366
0.319
0.296
0.090
19.8 [13.3, 24.8]
7.0 [ −1.7 , 18.8]
2.3 [ −3.2 , 9.1]
4.7 [ −0.4 , 10.7]
Appendix
TABLE XIII: Error budget at unsurveyed positions, Chicago street level, d=0.9λ , full survey, 2400 test queries in 22 cells. Reversal rates, and the gain of VS-Hist in percentage points with paired cluster-bootstrap 95 percent intervals (2000 resamples).
In physical backpropagation, users return loss derivatives over the time-division duplex uplink, giving the base station the downlink gradient without channel reconstruction. Reciprocity calibration conventionally prioritizes base-station chains, as column-normalized zero-forcing reduces a per-user factor to a stream rotation. This priority can reverse. With ideal base-station chains and no noise, per-user amplitude errors make the returned field a weighted sum-rate gradient. With zero-forcing full service at both optima, the loss is independent of array size. Otherwise the selection penalty, serving the wrong users, carries 86 to 94 percent of the amplitude-induced loss in five of six configurations. Per-user phase errors mainly slow the iteration. Base-station chain errors scale the returned field per antenna, and at 3 dB and 25 degrees they cost one user 1.3 dB of signal-to-noise ratio (SNR) that no array size removes. With 80 users, such errors on both sides, and mismatched pilots, oracle per-user and base-station calibration leave 6.2 and 12.3 percent loss at 256 antennas on clustered channels. This ordering holds at 30 dB SNR and reverses for 1 dB, 10 degree user errors. A capped one-echo estimator comes within 2.4 percentage points of oracle per-user calibration but underperforms no calibration at 10 dB SNR.
Xianzhong Li, Yuzhi Yang, Tianhao Guo
College of Physics and Electronic Engineering, Shanxi University, Taiyuan 030006, China · College of Information Science, Japan Advanced Institute of Science and Technology, Nomi 9231292, Japan
We address the port-selection problem in fluid antenna multiple access (FAMA) systems with multi-port fluid antenna (FA) receivers. Existing methods either achieve near-optimal spectral efficiency (SE) at prohibitive computational cost or sacrifice significant performance for lower complexity. We propose two complementary strategies: (i) GFwd+S, a greedy forward-selection method with swap refinement that consistently outperforms state-of-the-art reference schemes in terms of SE, and (ii) a Transformer-based neural network trained via imitation learning followed by a Reinforce policy-gradient stage, which approaches GFwd+S performance at lower computational cost.
Darian Perez-Adan, Jose P. Gonzalez-Coma, F. Javier Lopez-Martinez +1
Department of Computer Engineering, University of A Coruña, CITIC, A Coruña, Spain · Defense University Center at the Spanish Naval Academy, Marín, Spain · Dept. Signal Theory, Networking and Communications, Research Centre for Information and Communication Technologies (CITIC-UGR), University of Granada, 18071, Granada, Spain
Fluid antenna systems (FASs) provide many candidate ports, but probing all ports for each channel realization incurs substantial online overhead. We study a complementary offline problem: compress the candidate-port space once and reuse the compressed set across future channel realizations. The method collects optimal ports from sampled training realizations, requiring only one scan of the candidate ports per realization. On a DeepMIMO 7-by-7 FAS with 49 candidate ports, 120 training realizations yield 27 representative ports on average while retaining 99.995% of the channel power gain achieved by exhaustive port selection.
Quan Zhou, Ming Zeng
National University of Singapore · Université Laval