mmWave UE Positioning Under Blockage: Limits for ISAC-Assisted Handover

This work has been submitted to the IEEE Conference for possible publication. Copyright may be transferred without notice, after which this version may no longer be accessible.

mmWave UE Positioning Under Blockage: Limits for ISAC-Assisted Handover Seyed Bagher Hashemi Natanzi∗ , Bo Tang† ∗† Department of Electrical and Computer Engineering, Worcester Polytechnic Institute (WPI), USA

arXiv:2611.00453v1 [cs.NI] 30 Oct 2026

Email: {snatanzi, btang1}@wpi.edu

Abstract—Proactive handover can save a millimeter-wave (mmWave) link before a passing body blocks it, but only if the network knows where the user is. We ask whether the radio units that serve a user can localize it accurately enough from its uplink sounding reference signals. We derive position error bounds that include synchronization and array calibration errors, with and without reflected paths known from a digital twin, and test two estimators in a ray-traced street canyon on held-out scenarios. Three findings stand out. First, with line of sight to both radio units, array calibration, not bandwidth, limits the bound. Second, blockage degrades the line-of-sight-only bound sharply, whereas known reflections keep the map-aided bound low. Third, a simple dominant-path estimator fails where it matters most: its errors grow sharply around blockage, and proactive handover becomes worse than the 3GPP baseline. A map-aided multipath tracker cuts the median error from 12 to 3.7 cm and removes this significant penalty at link margins of 15 dB and above, but does not yet beat the baseline. Index Terms—mmWave positioning, position error bound, uplink SRS, blockage, proactive handover, digital twin, ISAC.

I. I NTRODUCTION Millimeter-wave (mmWave) bands provide the bandwidth that 5G and 6G networks need for high data rates, but their links are fragile: a pedestrian or vehicle crossing the line of sight (LoS) attenuates the signal by tens of decibels within milliseconds [1], [2]. Reactive handover cannot avoid such outages. Integrated sensing and communication (ISAC) and other sensing-aided approaches therefore aim to predict blockage and hand the user over in advance [3]–[5]. Predicting a blockage means predicting whether a body will cross the line between a radio unit and the user. The prediction can thus be no better than the network’s knowledge of where the user is. In a companion study [6], we found that a blockage-aware handover planner outperforms the 3GPP A5 event, a reactive trigger that hands over when the serving cell falls below one threshold and a neighbor exceeds another, only if the user position is known with an error of about 8–13 cm (per horizontal axis), and that a meter-level error removes the benefit entirely. Can the network meet this requirement with the infrastructure it already has? 5G NR supports uplink positioning from sounding reference signals (SRS) [7], and recent street-scale studies with several base stations report errors below a meter [8]. Position error bounds for mmWave systems suggest that *Corresponding author: snatanzi@wpi.edu

much better accuracy is possible in principle [9], [10], but they typically consider ideal hardware or a single impairment in static scenes [11]. It remains unclear which effect limits decimeter accuracy in practice, how blockage changes it, how close practical estimators come, and whether their errors are small enough when handover decisions are made. This paper addresses these questions in an urban street canyon, a street lined with buildings on both sides, served by two O-RAN radio units (O-RUs). Our contributions are: A positioning analysis for blockage-aware handover: a single-epoch equivalent Fisher information analysis with synchronization and array calibration errors as Bayesian priors, two models of a blocked LoS, and map-aided information from reflected paths with known geometry. • Two practical uplink estimators: a dominant-path estimator with an extended Kalman filter (EKF), whose single-epoch front end is compared with the bound by decomposing its error into bias and random spread, and a map-aided multipath tracker that associates extracted paths with virtual anchors of the digital twin. • An analysis of positioning errors around blockage onsets and a closing experiment that feeds estimated and synthetic bound-level positions into a blockage-aware handover planner. All code and the scripts that generate every number are open source.1 •

II. R ELATED W ORK Bounds for mmWave positioning. Fisher information analyses show that the angle and delay resolution of mmWave arrays support sub-decimeter accuracy [9], [10], building on the equivalent Fisher information framework for wideband localization [12]. Multipath components with known geometry carry position information as virtual anchors [13], [14], and multipath tracking associates them with such anchors probabilistically [15]. We add hardware errors as priors and evaluate bounds and estimators under traffic-driven blockage. Uplink SRS positioning. 3GPP specifies uplink time- and angle-based positioning [7]. Dong et al. [8] combine LoS/nonLoS gating, dominant-path angle and delay estimation, and multi-base-station EKF fusion, reach sub-meter errors in a raytraced campus, and point to reflections as virtual anchors for future work. 1 https://github.com/CLIS-WPI/sense2connect

Sensing-aided handover. Vision-, radar-, and ISAC-aided handover schemes predict blockage or user motion [4], [5]; our companion study [6] identified the user position as a binding requirement. III. S YSTEM M ODEL AND P ROBLEM F ORMULATION A. Signal Model A single-antenna user equipment (UE) at horizontal position p at known height (distances and directions are evaluated at the 3D point [pT , zUE ]T ) transmits an uplink SRS that is received by a set C of O-RUs, each with an M -element array. At ORU c, subcarrier frequency fn , and element m, the received sample is X yc,n,m = βc,ℓ e−j2πfn (τc,ℓ +δc +b) ejψc,m sm (uc,ℓ )+wc,n,m , ℓ

(1) where path ℓ has complex gain βc,ℓ (including blockage loss), delay τc,ℓ , and arrival direction uc,ℓ with steering response sm ; δc is the time offset of O-RU c, b the UE clock bias, ψc,m the calibration phase error of element m, and wc,n,m ∼ 2 CN (0, σw ) independent thermal noise. Hardware errors are 2 ) and ψc,m ∼ N (0, σϕ2 ). random with δc ∼ N (0, σsync B. Geometry and Blockage The LoS delay and angles are functions of p and the O-RU position pc , e.g., τc,0 = ∥p − pc ∥/c0 . A specular reflection on a known surface is an LoS path from the virtual anchor p̃c,ℓ , the image of pc in that surface, so that τc,ℓ = ∥p − p̃c,ℓ ∥/c0 [13]. Moving bodies attenuate every path segment according to a blockage model; we use 3GPP model B, which treats each body as a screen with knife-edge diffraction at its edges [16]. An O-RU is blocked for a UE when the loss of its LoS path exceeds a threshold. Model B reduces the power of a blocked LoS but keeps its geometry, whereas the field that actually arrives diffracts around the blocker. We therefore consider two models of a blocked LoS: kept, in which its delay and angles still follow p (optimistic), and biased, in which they carry an unknown bias. In the biased model, a bound treats them as free parameters, and an estimator observes a blocked LoS whose delay and direction deviate from the geometry of p. C. Positioning Requirement of Blockage-Aware Handover A blockage-aware handover policy π predicts the SNR of each cell from the tracked blockers and the reported UE position p̂t = pt +et , and selects the serving cell accordingly. Its outage time J(π; {p̂t }) counts the time below the SNR required for the service rate, including the interruptions of its handovers. Since the policy aims to outperform a reactive reference policy π0 (here, the 3GPP A5 event), the positioning requirement is that the error process {et } keeps
J π; {p̂t } ≤ J(π0 ). (2) This requirement depends on the error at the times the policy acts, not only on its overall statistics. We therefore examine (i) what limits the position error bound, (ii) how blockage and map knowledge change it, (iii) how a practical estimator compares, and (iv) whether its errors satisfy (2).

IV. P OSITION E RROR B OUND AND E STIMATOR A. Single-Epoch Position Error Bound Collect the channel parameters of all O-RUs at one epoch in η (delays, angles, complex gains) and the hardware and timing errors in ν = [{δc }, b, {ψc,m }]. The Fisher information of η follows from (1) by the Slepian–Bangs formula; we compute it exactly by factorizing its frequency and antenna sums. With T = ∂η/∂p, the equivalent Fisher information of the position is
−1 Jp = TT Jη T − TT Jην Jνν + Λ Jνη T, (3) where all blocks are those remaining after the channel nuisance parameters have been eliminated jointly by a Schur complement: the complex gains of all paths; in the LoS-only bound, the delays and angles of all reflected paths; and in the biased model, the delay and angles of a blocked LoS. The second term eliminates the timing and hardware errors in ν. Λ holds the prior precisions of the O-RU time offsets (σsync ) and of the element phases (σϕ ); the UE clock bias has no prior (time difference of arrival, TDoA). In the mapaided bound, the delays and angles of the reflected paths are not nuisance parameters: their geometry is known from the digital twin, and their entries of T follow analytically from the virtual anchors. Finite differences at 1000 random positions, an independent NumPy implementation, and a brute-force Fisher matrix confirm the derivatives and (3) to relative errors of −6 −10 3.2 × 10−6 , 2.9 × 10 q , and 6.6 × 10 . The position error tr(J−1 bound is PEB = p ). Because the hardware errors enter through priors, it is a hybrid Cramér–Rao bound: it bounds the root-mean-square (RMS) error of unbiased singleepoch estimates for a given geometry, averaged over the noise and the hardware priors; it is neither a median error nor a guarantee that an estimator attains it, a biased estimator can have a smaller spread, and it excludes the temporal information of tracking. B. Dominant-Path Estimator Per O-RU and epoch, a zero-padded two-dimensional DFT over the array selects the strongest beam; a Hann-windowed delay transform of the beamformed signal gives the delay of the dominant path, refined by parabolic interpolation and followed by an angle refinement at that delay. A peak-tosecond-peak test flags reflected dominant paths. Because a planar array of isotropic elements cannot distinguish the two sides of its plane, each angle yields two direction hypotheses; the estimator keeps the joint hypothesis that is consistent across O-RUs, lies in the walkable area of the digital-twin map, and passes the gate of a constant-velocity EKF, which fuses weighted least-squares fixes over time. C. Map-Aided Multipath Tracker Per O-RU and epoch, a sparse extraction with continuous refinement returns up to eight multipath components with variances from the local Fisher information. The sources are the LoS and every single- and double-bounce virtual anchor of the street-facing building faces and the ground

z [m]

buildings

O-RU 0

north sidewalk O-RU 0 4 westbound 18.2 m

2

bus UE track

Reported UE

O-RU 1

eastbound

UE bus

car distance from UE [m]

(b) Side view along the serving link O-RU 0 (facade, 8 m)

south sidewalk

True UE 40 m

O-RU 1 Actual link

Predicted link

Other-cell link

Reflected path

(a) Top view, lamppost deployment

40 m (c) Facade deployment

Fig. 1. Scenario geometry (dimensions indicated; axes scaled independently). (a) Two neighboring cells share the street; O-RU 0 and O-RU 1 are on 5 m lampposts, 40 m apart along the street. The true serving link (solid red) crosses the bus, whereas the link predicted from the reported UE position (dotted) passes it, so a position error turns into a missed blockage prediction; green: link to the other cell and one of its facade reflections (dashed). (b) The LoS passes through a bus but above a car. (c) In the facade deployment, O-RU 0 is mounted on the north facade at 8 m.

that is geometrically valid at the predicted position. Loopy belief propagation computes the probabilities that a source generated a component or that the component is clutter, with constant visibility priors and a blocked mode that models a detoured LoS by an inflated covariance [15]; a probabilisticdata-association EKF with a constant-velocity model then updates the position, velocity, UE clock bias, and O-RU time offsets, and projects the position onto the walkable area. A grid search over the walkable area initializes the track and re-initializes it after a loss of association. V. E XPERIMENTS AND R ESULTS A. Setup Scenario. We use the street canyon of [6] (Fig. 1): buildings 18.2 m apart, two 3.5-m lanes, and sidewalks on both sides; actors and UEs move on a periodic section of 80 m. Two ORUs with 8 × 8 half-wavelength arrays at 28 GHz form two neighboring cells that share this section: O-RU 0 on the north side, either on a lamppost (5 m high) or on a facade (8 m), and O-RU 1 on a 5-m lamppost on the south sidewalk, 40 m further along the street. Pedestrian UEs at 1.5 m walk along the south sidewalk at 1.0–1.2 m/s. Each 60-s run has low or high traffic: 37.5 or 100 vehicles per lane and km (cars, buses, trucks at 6–12 m/s) and 6 or 12 pedestrians on the sidewalks plus 2 or 4 crossing. Paths are traced with Sionna RT [17], [18] (specular, order up to 2, at most 8 paths per link). Service model. The SNR of a cell adds, over all traced paths, the path power with the full array gain after its model-B loss; this is an optimistic approximation, since no single beam collects all paths with full gain. The UE is in outage when this total remaining power gives an SNR below γreq = 9.3 dB, the value needed for 400 Mbit/s. As in [6], the link margin is the

median unblocked best-cell SNR above γreq (3GPP budget: 35.9 dB, Ref.), swept from 0 to 30 dB by a common extra path loss. During 10 dB LoS events, the strongest surviving path of the same cell lies a median 26–33 dB below the unblocked LoS, so the margin decides whether the same cell can still serve the UE. Handover model. The policy is the planner of [6]: every 0.1 s it receives the blocker tracks from ISAC sensing and the reported UE position, predicts the SNR of each cell by applying model B on the line between each O-RU and the reported position, and chooses the cell sequence that minimizes the predicted outage over a horizon of 0.5–3 s, with an interruption of 20 ms per handover; decisions take effect after a control-loop delay of 20 ms, and the sensing overhead of 2.3% of the resources is charged to it. The A5 thresholds and time-to-trigger are tuned per margin on the tuning seeds; the tuned time-to-trigger, 40 ms, is the smallest value of the grid. With perfect blocker tracks and a white Gaussian UE position error, the planner satisfies (2) only up to a per-axis standard deviation of about 8–13 cm, i.e., an RMS distance error of about 11–18 cm [6]. Blocked LoS. An O-RU is blocked for a UE at an epoch when the model-B loss of its LoS path is at least 10 dB. In the biased model, the signal fed to the estimator replaces the blocked LoS by a detour around its dominant blocker (the blocker with the largest model-B loss): with q the shorter of the detour points over the top and past either side of the blocker box, next to the point of the LoS closest to the blocker, the path keeps the attenuated complex gain of the LoS but has delay (∥q − pc ∥ + ∥p − q∥)/c0 and arrival direction along q − pc ; reflected paths are only attenuated. Signals and hardware. The UE transmits one SRS symbol

TABLE I S INGLE - EPOCH ESTIMATOR ERROR AT FIXED GEOMETRIES (500 DRAWS EACH , MAIN HARDWARE , TD OA, NO EKF) VERSUS THE SINGLE - EPOCH L O S- ONLY PEB ( BIASED MODEL FOR A BLOCKED L O S); ALL VALUES IN CM ; RMSE2 = BIAS2 + CENTERED RMS2 (E Q . (4)).

Mount lamppost lamppost lamppost facade facade facade

LoS-only bound

LoS state UE x [m] both LoS -20.0 both LoS 20.0 blocked (O-RU 0) 21.9 both LoS -20.0 both LoS 20.0 blocked (O-RU 0) 21.9

Map-aided bound

Median PEB / position error [m]

(a) LoS to both O-RUs

10−1

400 MHz PEB bias centered RMSE 1.5 1.3 1.7 2.2 4.2 1496.1 55.4 1497.1 46.3 301.3 4.9 301.3 3.3 31.9 4.8 32.2 6.7 88.6 6.1 88.8 90.8 189.2 7.0 189.3

Estimator

(b) O-RU(s) blocked

10 cm

10 cm

100

100

10−2

200 400 Bandwidth [MHz]

200 400 Bandwidth [MHz]

Fig. 2. Median single-epoch position error bound (LoS-only, map-aided) and median dominant-path estimator error versus bandwidth, for epochs (a) with LoS to both O-RUs and (b) with at least one blocked O-RU; main configuration (95% CI). The bound and the error are different quantities (Section IV). The LoS-only bound barely changes with bandwidth because array calibration limits it (Section V-B).

per 0.1-s epoch at 23 dBm, spread uniformly over the active subcarriers; the O-RU noise figure is 7 dB. Numerology 3 (120 kHz) with 100, 200, and 400 MHz carriers uses 792, 1584, and 3168 active subcarriers [19]. We sweep σsync of 0, 1 and 3 ns (drawn once per run, a bias that the EKF cannot average out; 40 realizations over the held-out runs) and σϕ of 0, 2 and 5 deg (fixed per O-RU and run). The main configuration uses TDoA, 400 MHz, σsync = 1 ns, σϕ = 2 deg, and the biased blocked-LoS model. Protocol. Estimator parameters were tuned on seeds 101– 105 and methods fixed on development seeds 1001–1010. An earlier held-out set was consumed before the front/back disambiguation was added, so all results come from a second, never-used set of ten seeds (3001–3010), each simulated for two mounts and two traffic densities. The seed is the statistical unit: we average its runs, compare conditions with the exact two-sided Wilcoxon signed-rank test over the ten seeds, and give 95% bootstrap intervals over seeds. The mapaided tracker was developed after the dominant-path estimator, on the same tuning and development seeds, and was run once on the held-out seeds with unchanged parameters.

100 MHz PEB bias centered RMSE 1.5 7.7 7.4 10.7 4.5 875.6 345.0 941.2 91.5 286.1 9.3 286.3 3.4 30.0 16.1 34.1 7.1 94.0 9.4 94.5 208.0 185.0 10.4 185.3

B. What Limits the Bound Fig. 2 shows the median bounds versus bandwidth. With LoS to both O-RUs, the LoS-only PEB is below 10 cm in 84% of the epochs at 400 MHz and still in 81% at 100 MHz. Relaxing one limit at a time at 200 MHz scales the median LoS-only PEB (4.8 cm) by 0.97 for twice the bandwidth, 0.94 for ten times the SNR, 0.72 for perfect synchronization, and 0.07 for perfect calibration. On these short links, the delay and angle measurements are precise enough that the element phase errors, which bias the angles, dominate; the map-aided bound, in contrast, does improve with bandwidth (Fig. 2). Without hardware errors, the LoS-only PEB is below 10 cm in 92% of the epochs; with 3 ns and 5 deg errors, in 44%. The map-aided PEB is far less sensitive (99% and 98%), because reflections from known surfaces add independent geometric information. C. Blockage and the Digital Twin In the biased model, the median LoS-only PEB rises by a factor of 2.8 during blockage (p = 0.002), and only 47% of the blocked epochs remain below 10 cm. If the attenuated LoS kept its geometry, the factor would be only 1.2, which shows how optimistic it is to apply model B’s power loss without its geometric consequence. Map-aided information lowers the PEB during blockage to 0.09 times the LoS-only value (p = 0.002). Along the street, the LoS-only PEB varies with the geometry of the two O-RUs from about a centimeter to over a decimeter, while the map-aided PEB stays mostly in the millimeter range. A low map-aided PEB indicates that the information is present in the signal. D. Practical Estimators and the Bound At 400 MHz, the dominant-path estimator reaches a median error of 12 cm (10 cm with LoS to both O-RUs, 27 cm during blockage); 45% of its errors are below 10 cm, and its per-seed 90th percentile is 2.10 m on average. At 100 MHz, the median grows to 24 cm. To compare the estimator with the bound on equal terms, we ran 500 independent noise and hardware draws through the single-epoch estimator (without EKF) at 6 fixed geometries and decomposed its error as   RMSE2 = ∥E[p̂] − p∥2 + E ∥p̂ − E[p̂]∥2 , (4)

TABLE II D OMINANT- PATH ESTIMATOR ERROR AROUND REFERENCE - CELL BLOCKAGE EVENTS (10 D B, MAIN CONFIGURATION ). Window Median 90th perc. >0.1 m >0.25 m 1.0–0.5 s before onset 12 cm 4.24 m 54% 30% 0.5–0 s before onset 11 cm 4.49 m 53% 29% During the event 87 cm 3.26 m 88% 72% Elsewhere 10 cm 1.59 m 49% 29%

i.e., squared bias plus squared centered RMS (Table I). The hardware errors are redrawn in every draw, so their contribution appears in the centered RMS, whereas the bias contains only deterministic errors such as unresolved multipath and the detoured LoS; within a run, a hardware realization is fixed and acts as a bias. The squared bias accounts for most of the mean squared error in five of the six geometries at 400 MHz (98–100% of it); in the remaining one (lamppost, both LoS, x = −20 m), the squared bias is 38% of it and the random spread is larger. The bias is caused by interference of unresolved multipath, and in one geometry by a front/back mirror solution. The centered RMS is geometry dependent (0.077–13 times the PEB at 400 MHz): close to the bound at some geometries and an order of magnitude above it at others. In blocked geometries, the centered RMS falls below the PEB, which is possible because the estimator is biased: it uses the detoured LoS precisely but in the wrong place. Over the held-out runs, front/back mirror solutions remain in 5% of the epochs and dominate the squared error (RMS over all epochs 6.36 m), while reflected dominant paths set the 90th percentile. A variant that used reflections as virtual anchors did not reduce the tail on the development seeds (90th percentile 3.52 vs. 1.87 m) and was not pursued. The map-aided tracker reaches a median error of 3.7 cm (4.5 cm with at least one O-RU blocked) and a 90th percentile of 0.25 m, significantly better than the dominant-path estimator (all p = 0.002). Its largest errors occur in the 1 s after a UE reenters the periodic street section, a simulation artifact in which the track keeps the pre-jump position until it re-initializes; these windows contain 1.7% of the epochs (not excluded). E. Positioning Errors Around Blockage Onsets Table II shows the estimator error around the 752 blockage events of the reference cell, defined per UE as the cell with the larger unblocked received power; an event is an interval with LoS loss of at least 10 dB, with gaps below 0.5 s merged; each epoch is counted in one window, with priority to the event, and the first 2 s of each run are excluded. In the second before onset, a window relevant to proactive handover, the median error and the share of errors above 0.25 m are about as elsewhere, but the large errors become larger: the 90th percentile rises from 1.59 to 4.49 m. During the event, most of the distribution deteriorates, with the median growing to 87 cm. The overall median hides both effects. For the mapaided tracker, the 90th percentile does not rise before onset (0.20 vs. 0.35 m elsewhere), and its median during the event

TABLE III P LANNER OUTAGE MINUS A5 [ S /UE- MIN ] BY SOURCE OF THE UE POSITION , WITH PERFECT BLOCKER TRACKS . N EGATIVE VALUES ( SHADED ) MEAN LESS OUTAGE THAN A5; † : NOT SIGNIFICANT ( SEED - LEVEL W ILCOXON p ≥ 0.05; POINTWISE TESTS , EXPLORATORY ). R EF. = 3GPP REFERENCE BUDGET (35.9 D B MARGIN ). ∗ O RACLE DIAGNOSTIC : MIRROR SOLUTIONS REPLACED BY THE SINGLE - EPOCH HYPOTHESIS CLOSEST TO THE TRUE POSITION . UE position

15 dB

Link margin 20 dB 25 dB

30 dB

Ref.

Reference inputs exact −0.108 −0.066† −0.061 −0.007† −0.025 white 1 m error +1.402 +0.824 +0.476 +0.290 +0.117 Synthetic, bound level (full covariance) map-aided bound −0.099 −0.049† −0.044 +0.021† −0.005† LoS-only bound +0.162 +0.301 +0.175 +0.151 +0.097 LoS-only, no fix held +0.071† +0.142† +0.058† +0.073† +0.036† Estimated map-aided tracker +0.009† +0.031† +0.072† +0.015† −0.002† dominant-path +0.652 +0.507 +0.284 +0.118 +0.052 ideal hardware +0.584 +0.483 +0.295 +0.121 +0.054 capped at p90 +0.525 +0.434 +0.251 +0.118 +0.051 mirror-corr.∗ +0.461 +0.443 +0.260 +0.115 +0.048

is 6.4 cm. The windows refer to the blockage events of a reference cell, not to the decision times of a particular planner. F. Implications for Handover Table III feeds positions into the planner, with perfect blocker tracks so that the user position is the only source of error. With exact positions, the planner has a lower outage than A5 from 15 dB margin upward (significantly at 15 and 25 dB and at the reference). As a
sensitivity test, we feed synthetic errors et ∼ N 0, Jp (t)−1 , drawn independently per epoch from the single-epoch bound of the respective model, which is anisotropic (median ratio of principal standard deviations 3.0). Errors at the map-aided bound preserve a lower mean outage than A5 at most tested margins, significantly at 15 dB (at 25 dB only without correction); errors at the LoS-only bound, which grow during blockage, make it worse than A5 at every margin. With the positions of the dominant-path estimator, the planner is significantly worse than A5 at every margin, also with ideal hardware; with those of the map-aided tracker, it is not significantly different from A5 at any margin from 15 dB up (p ≥ 0.084), although at 25 dB its bootstrap interval lies above zero; at 10 dB it is still significantly worse (+0.414 s/UE-min, p = 0.0059), as at all lower margins (p ≤ 0.0059). Capping each error of the EKF output at the pooled 90th percentile (1.90 m) or replacing its mirror solutions by the single-epoch hypothesis closest to the true position (an oracle diagnostic, not refiltered) reduces the loss but leaves it significant, so the largest errors are not its only cause. With Holm correction across the margins of each row, all these significant results remain significant except the mapaided bound at 25 dB (adjusted p = 0.055). Synthetic bound-level errors are an optimistic model; they show what positioning would have to achieve, not that a practical estimator achieves it. In the 3.0% of epochs in which

both O-RUs are blocked, the LoS-only bound is singular; these epochs are treated as having no fix (100 m standard deviation). Holding the last valid position instead keeps the LoS-only row above A5 at every margin but removes its significance, so that row depends mainly on how missing fixes are handled. The synthetic errors are independent across epochs, whereas hardware errors persist over a run; the effect of temporal correlation on the planner is not evaluated here. The absolute gains at stake are small: at 15 dB margin, A5 has an outage of 1.88 s/UE-min, and the relative change is −5.7% with exact positions and −5.2% with bound-level map-aided positions; at the reference, A5 has 0.058 s/UE-min, and the changes are −43.1% and −8.2%† , respectively; the application-level impact is not evaluated here. VI. D ISCUSSION AND L IMITATIONS For blockage-aware handover, what matters is the error around blockage onsets, not the overall median. The dominantpath estimator falls short of the bound mainly because of systematic multipath errors, which a simple virtual-anchor variant did not remove. The map-aided tracker resolves several paths and associates them probabilistically with known surfaces; it cuts the median error about threefold but remains well above the map-aided bound. With its positions, proactive handover is no longer significantly worse than A5 from 15 dB up, yet it does not reach the small gain of exact positions. Whether this gap comes from the tracker’s residual errors, including those after the UE jumps, remains open. The study has limitations. It considers one street canyon with two O-RUs, specular paths of order at most 2 without diffuse scattering, and fully digital arrays read out in one symbol; their isotropic elements cause a front/back ambiguity that a back plane would remove [8]. The map is exact in both the map-aided bound and the tracker, which was added after the first evaluation. A blocked LoS follows our biased model, and bodies are opaque knife-edge screens, without window transmission, vehicle-body reflections, or self-blockage by the user. Finally, the bounds are single-epoch and ignore the temporal information of tracking. VII. C ONCLUSION We asked whether the radio units that serve a user can localize it accurately enough for blockage-aware handover, in a short-range street canyon with two O-RUs. With LoS to both O-RUs, array calibration rather than bandwidth limits the LoS-only bound. Blockage degrades this bound substantially, whereas known reflected paths keep the map-aided bound much lower. A dominant-path estimator makes mostly systematic multipath errors that worsen around blockage, and its positions make proactive handover worse than reactive handover. A map-aided multipath tracker cuts the median error to a few centimeters and keeps it low around blockage onsets; with its positions, proactive handover is no longer significantly worse than the 3GPP event from 15 dB up. Reaching the small gain that exact positions provide remains open.

ACKNOWLEDGMENT This material is based upon work supported in part by NSF under Awards CNS-2120442 and IIS-2325863, and NTIA under Award No. 51-60-IF007. Any opinions, findings, and conclusions or recommendations expressed in this publication are those of the author(s) and do not necessarily reflect the views of the NSF and NTIA. AI assisted with coding and enhancing the text; the authors verified all content. R EFERENCES [1] T. S. Rappaport et al., “Millimeter wave mobile communications for 5G cellular: It will work!” IEEE Access, vol. 1, pp. 335–349, 2013. [2] G. R. MacCartney, T. S. Rappaport, and S. Rangan, “Rapid fading due to human blockage in pedestrian crowds at 5G millimeter-wave frequencies,” in Proc. IEEE GLOBECOM, 2017. [3] F. Liu et al., “Integrated sensing and communications: Toward dualfunctional wireless networks for 6G and beyond,” IEEE J. Sel. Areas Commun., vol. 40, no. 6, pp. 1728–1767, 2022. [4] G. Charan, M. Alrabeiah, and A. Alkhateeb, “Vision-aided 6G wireless communications: Blockage prediction and proactive handoff,” IEEE Trans. Veh. Technol., vol. 70, no. 10, pp. 10 193–10 208, 2021. [5] Y. Li, F. Liu, H. Xiong, Z. Wang, Narengerile, and C. Masouros, “Look before switch: Sensing-assisted handover in 5G NR V2I networks,” IEEE Trans. Mobile Comput., vol. 25, no. 6, pp. 8063–8076, Jun. 2026, doi: 10.1109/TMC.2025.3647295. [6] S. B. Hashemi Natanzi and B. Tang, “ISAC-assisted proactive handover in mmWave networks: Value of foresight and localization requirements,” submitted to IEEE WCNC, 2027. [7] 3GPP, “Study on NR positioning support,” 3GPP, Technical Report TR 38.855, Release 16, 2019. [8] W. Dong, Z. Ju, H. Jiao, R. Chen, and L. Chen, “Deployable uplink SRSbased 5G NR positioning in mixed LoS/NLoS environments,” Expert Syst. Appl., vol. 328, p. 132887, 2026, doi: 10.1016/j.eswa.2026.132887. [9] A. Shahmansoori, G. E. Garcia, G. Destino, G. Seco-Granados, and H. Wymeersch, “Position and orientation estimation through millimeterwave MIMO in 5G systems,” IEEE Trans. Wireless Commun., vol. 17, no. 3, pp. 1822–1835, 2018. [10] Z. Abu-Shaban, X. Zhou, T. Abhayapala, G. Seco-Granados, and H. Wymeersch, “Error bounds for uplink and downlink 3D localization in 5G millimeter wave systems,” IEEE Trans. Wireless Commun., vol. 17, no. 8, pp. 4939–4954, 2018. [11] M. F. Keskin, F. Jiang, F. Munier, G. Seco-Granados, and H. Wymeersch, “Optimal spatial signal design for mmWave positioning under imperfect synchronization,” IEEE Trans. Veh. Technol., vol. 71, no. 5, pp. 5558– 5563, May 2022, doi: 10.1109/TVT.2022.3149974. [12] Y. Shen and M. Z. Win, “Fundamental limits of wideband localization— Part I: A general framework,” IEEE Trans. Inf. Theory, vol. 56, no. 10, pp. 4956–4980, 2010. [13] E. Leitinger, P. Meissner, C. Rüdisser, G. Dumphart, and K. Witrisal, “Evaluation of position-related information in multipath components for indoor positioning,” IEEE J. Sel. Areas Commun., vol. 33, no. 11, pp. 2313–2328, 2015. [14] K. Witrisal et al., “High-accuracy localization for assisted living: 5G systems will turn multipath channels from foe to friend,” IEEE Signal Process. Mag., vol. 33, no. 2, pp. 59–70, 2016. [15] F. Meyer, T. Kropfreiter, J. L. Williams, R. Lau, F. Hlawatsch, P. Braca, and M. Z. Win, “Message passing algorithms for scalable multitarget tracking,” Proc. IEEE, vol. 106, no. 2, pp. 221–259, 2018. [16] 3GPP, “Study on channel model for frequencies from 0.5 to 100 GHz,” 3GPP, Technical Report TR 38.901, Release 19, 2025. [17] J. Hoydis et al., “Sionna RT: Differentiable ray tracing for radio propagation modeling,” in Proc. IEEE GLOBECOM Workshops, 2023. [18] J. Hoydis et al., “Sionna, version 2.2.0,” https://nvlabs.github.io/sionna/, 2022. [19] 3GPP, “NR; user equipment (UE) radio transmission and reception; part 2: Range 2 standalone,” 3GPP, Technical Specification TS 38.101-2, 2024.

Related documents

• mmWave UE Positioning Under Blockage: Limits for ISAC-Assisted Handover#1536283
• Multi-User mmWave Beam and Rate Adaptation via Combinatorial Satisficing Bandits#1156328
• Multi-User mmWave Beam and Rate Adaptation via Combinatorial Satisficing Bandits#19035
• Rate-Splitting-Inspired Uplink ISAC: A Rate-Region Analysis#925174
• Rate-Splitting-Inspired Uplink ISAC: A Rate-Region Analysis#920939