Make your paper interactiveCreate

Text from the interactive explanation

Near-Field Low-Complexity Sensing for Reconfigurable Holographic Surfaces with Hardware Constraints

Yinuo Dong, Qingchao Li, Soon Xin Ng, Mohammed El-Hajjar

Original paper

6G sensing · holographic surfaces

Could the mast that carries your calls also see where you are, using only the radio waves your phone already sends?

Future 6G networks aim to combine communication with sensing, which engineers call ISAC (integrated sensing and communication). A recent paper imagines a base station wrapped in a reconfigurable holographic surface: thousands of tiny tunable antenna elements on several flat faces. Nearby devices are so close that their waves arrive curved, and practical elements are imperfect. Below, a toy version of that tower is already at work.

Toy simulation, seen from above. The tower is drawn much larger than a real one (not to scale). Drag a device to move it.

Why one facet can’t tell how far

Radio waves spread out from a phone like ripples from a pebble. Close to the pebble the ripples are tightly curved; far away they look almost straight. If an antenna is wide enough to notice the curve, the amount of bending reveals how far away the source is. Engineers call that zone the near field. Beyond it, in the far field, the wave looks flat across the antenna and only its direction can be read.

The paper’s surface works at 30 GHz, where one wavelength is 1 centimetre. The tower as a whole is large compared with that, so phones a few metres away are in its near field. But the surface is split into flat facets, and each facet on its own is small.

Predict: can one facet tell a phone 6 m away from one 60 m away?

Blue lines are wave crests drawn to true scale around the aperture (the vertical bar). The amber dots on the right show how much later each element hears a crest than the middle one does. A curved wave gives a curved row of dots; a flat wave gives a straight column inside the green “flat enough” band. Drag sideways across the picture to move the source.

The usual boundary is the Rayleigh distance, 2D²/λ, where D is the aperture size and λ the wavelength: about 1.3 m for an 8 cm facet and about 50 m for a 50 cm surface. Inside it the edge-to-middle delay exceeds one sixteenth of a cycle.

So a single facet is effectively in the far field even for a phone a few metres away: it can say which way, not how far. The paper turns this into a design choice. Each facet reports only a direction, and distance is never scanned at all. Instead, the position comes from where the facets’ direction rays cross , the way your two eyes judge depth.

Two eyes, one catch

Two neighbouring facets each draw a ray toward the same person. Where the rays cross is the estimated position. Now give one facet a small angle error.

The blue ray is correct; the amber ray has the error (its correct direction is shown faint). The pink segment is how far the crossing has moved. Tap or drag in the picture to move the person.

This is the warning the paper stresses: because the facets sit close together, their rays cross at a shallow angle, so small angle errors are magnified into large distance errors, more so the farther away the person is. On a real surface the facets are even closer together than drawn here, which makes the effect stronger.

Getting each facet’s angle right therefore matters a lot. And that is exactly where practical hardware gets in the way.

A hum that sounds like a ghost

To listen in a particular direction, each element of the surface shifts the timing of the wave it receives, its phase, so that waves from that direction add up. An ideal element could choose any phase at full strength: any point on the dashed circle below. Practical metamaterial elements can’t. In the paper’s hardware model (a Lorentzian-type response), every response an element can actually produce lies on a small circle that passes through zero. Ask for one phase and you get a different phase, at reduced strength.

Drag around the dial or use the slider. The realised response is ψ = (j + ejζ)/2, where ζ is the phase you ask for (paper, Eq. 10).

Here is the key step in the paper’s analysis. Any point on the small circle is the sum of two arrows: a fixed pink arrow, 0.5j, that stays the same whatever you ask for, plus a turning blue arrow of half length that carries the phase you wanted (paper, Eq. 15). The paper calls the fixed part a constant component. We’ll call it the hum .

Every element hums the same note

Now look at all 16 elements across one facet hearing a wave from an angle. With ideal pixels the arrows turn step by step along the row, and that turning pattern encodes the direction. With realistic pixels every element also carries the same pink hum. An identical signal on every element is exactly what a wave from straight ahead looks like, because such a wave reaches all elements at the same moment.

Listening for directions

Facets pick out directions with an algorithm called MUSIC (multiple signal classification). It splits the received data into a part that contains the signals and a part that is only noise, then sweeps every look direction and scores how badly that direction’s pattern fits the noise-only part. Directions that don’t fit the noise at all score high, and those peaks are the sources. Below, real MUSIC runs on fresh simulated data for one 16-element facet about four times a second.

Before you switch to realistic pixels: what will the hum do to the spectrum?

The facet sits at the bottom, looking up. Lettered markers on the rim are the transmitting phones (drag them, or tap a direction). The blue shape is the MUSIC score in every direction; amber dots mark the peaks matched to phones, pink marks the ghost. Toy simulation: 16 elements, 64 snapshots.

This matches what the paper proves for the full surface: MUSIC still peaks near the true directions, but the constant component adds extra, deterministic peaks (pseudo-peaks) and nudges estimates off target (angular bias).

Try the ghost merge : with phone A just 4° from straight ahead, the ghost and the real peak fuse into one, roughly halfway between. The facet reports about 2° instead of 4°, and nothing in its own data says it is wrong. Remember this spot; chapter 4 needs it.

Don’t search the whole sky

Scoring every direction with MUSIC is expensive, and the paper wants each facet to be cheap. So a facet first takes a coarse look with a few wide beams, keeps only the promising slices of sky, splits them, and repeats. The coarse screening uses regularized orthogonal matching pursuit (ROMP), a greedy method that keeps groups of comparably strong candidates instead of betting on a single one. MUSIC then refines only inside the surviving windows.

Bar length = how strong each wide beam is. Kept beams are bright, discarded ones grey; kept windows from earlier layers stay as a band on the rim. Toy simulation with realistic pixels: two phones, 16 elements, signal-to-noise ratio 10 dB.

In this simplified screening a beam survives if it is at least half as strong as the best one, keeping at most twice as many beams as there are phones. Each wide beam counts as one scored direction.

Notice that the ghost at 0° survives screening too. The hum really is a strong signal from straight ahead, and one facet alone cannot tell it apart from a real phone. But the ghost always points straight out of whichever facet hears it, so different facets put it in different places, while a real phone sits where their rays agree. That is the opening for the next chapter.

What the paper found for one facet paper’s simulations

Values read from the paper’s Fig. 9 (approximate). Average angle error of one facet against signal-to-noise ratio, in the paper’s simulations. Hover or tap the chart to read values; arrow keys work when it is focused.

In the paper’s simulations, coarse-to-fine ROMP-MUSIC has the lowest error at medium and high signal-to-noise ratios. At low signal-to-noise, exhaustive ROMP is slightly better: a wrong coarse decision can throw away the true direction before MUSIC ever looks there. The paper describes this as the intended trade-off between complexity and robustness. Plain MUSIC with exhaustive search stays in the middle: without the sparsity step, it can lock onto the hardware’s pseudo-peaks.

Not every eye deserves a vote

Each facet now hands in one direction. Before the tower ever runs, it builds a trust map offline: for every direction, many simulated trials (Monte Carlo trials) measure how often a facet’s answer lands within 1° of the truth. That probability is the facet’s confidence for that direction. The paper reports that its estimator keeps the average angle error below 1° over a wide central region; the trouble is concentrated in particular directions, which is exactly what a map can capture.

The paper then combines the rays in four steps (its Fig. 6). Step through them:

All rays. Find the point closest to all rays at once, a plain least-squares (LS) intersection. Quick, and a starting point.

Drop low-trust rays. A facet whose reported direction falls where its trust map is weak (confidence below a threshold, α1) is set aside.

Drop disagreeing rays. Any remaining ray that points away from where the other trusted rays meet by more than a second threshold, α2, is set aside too.

Trust-weighted refinement. Starting from the survivors’ intersection, the position is nudged step by step to fit the remaining rays best, with more trustworthy rays pulling harder (weight √c). The paper uses the Levenberg–Marquardt method, a standard way to solve this kind of fit.

The coloured ring around each listening facet is its trust map: green = usually right, pink = often wrong. The dot where a ray crosses its ring shows how much that report is trusted. Notice the pink notch straight ahead (the ghost drags answers onto the facet’s centre line) and the dim edges (grazing angles). Rays: white = used, green = kept, pink dashed = low trust, coral dotted = disagrees. The faint white crosshair is the plain intersection, for comparison. Drag the person, tap to move them, or focus the picture and use the arrow keys.

Trust maps have two kinds of weak spot. The ghost spot puts the person 4° off a facet’s centre line: the ghost drags that facet’s report about halfway toward its centre line, into the pink notch, so in stage ② its ray is set aside. Because that ray runs through the middle of the bundle, it only nudges the plain fix sideways. The grazing spot is where plain intersection really suffers: two of the four listening facets see the person almost edge-on (about 78° and 84° from straight ahead), where their answers scatter by degrees. Those rays cross the others at a shallow angle and often drag the plain fix by more than half a metre; the trust-weighted fix sets them aside. Compare the misses at stages ① and ④.

If fewer than two facets survive the checks, there is no confident fix at all, and the tower says so instead of guessing. That is a real trade-off, and the paper measures it (see “The price of being strict” below).

The toy’s trust map

This map comes from our own offline Monte Carlo run of one 16-element toy facet, read by the angle the facet reports (across) and the signal-to-noise ratio (down). With ideal pixels almost every direction is trustworthy except grazing angles near the edges. With realistic pixels a pink stripe appears just around straight ahead, where the ghost drags answers. The paper builds the same kind of map over two angles (azimuth and elevation), with 103 Monte Carlo trials per grid point, all offline, so it costs nothing while the tower is running.

Deeper note: in our enlarged toy tower the agreement test compares each ray with the point where the other trusted rays meet. The paper compares each facet’s direction with the mean direction of the others, which works because real facets sit centimetres apart. The thresholds here (α1 = 0.8, α2 = 1°) are tuned for the toy; the paper does not print its values.

What the simulations showed

The paper tests the full method in Monte Carlo computer simulations at 30 GHz, with 3072 elements in total shared among L facets (16 elements across each facet, so more facets means fewer elements per facet). Users are placed at random across the tower’s near-field coverage area. Everything in this chapter comes from those simulations, not from measurements. Each chart is redrawn from the paper’s figures with values read by eye.

Position error as more users share the tower

Values read from the paper’s Fig. 11 (approximate). RMSE (root-mean-square error) is the typical distance between the estimated and the true position. Distance-MUSIC is a baseline that scans distance as well as angle with MUSIC; “single facet, L = 2” is the paper’s fair-comparison baseline with the same 3072 elements.

In the paper’s simulations, plain LS and distance-MUSIC get worse as facets are added: more inputs, but more of them biased by the hardware and nothing to filter them. The trust-weighted method gets better with more facets and levels off between about 1.2 and 1.8 m once 15 or more users are present. The authors summarise it as reducing the RMSE by more than 80% compared with conventional LS- and MUSIC-based schemes, while maintaining meter-level accuracy.

The price of being strict

Values read from the paper’s Fig. 12 (approximate). Success rate = how often a method returns a valid position. The paper shows no trust-weighted point at L = 4.

The trust-weighted method only answers when at least two trustworthy facets agree (in the paper, within 3°). With 6 facets that fails often; with 12 almost never. Plain LS and distance-MUSIC nearly always output a position, right or wrong.

Seeing better, serving better

Values read from the paper’s Fig. 13 (approximate). Sum rate = combined data throughput of all users when the tower aims its beams using the estimated positions; “ideal” is the paper’s fully digital reference.

In the paper’s simulations, the trust-weighted method reaches approximately 90% of the ideal sum rate, while LS falls behind because its beams stay misaligned. Adding facets helps, with diminishing returns beyond L = 8.

Settles in a few steps

Values read from the paper’s Fig. 7 (approximate). L = 12 facets, K = 10 users.

In the paper’s simulations, the refinement settles within about five steps, even from a badly perturbed start.

Cheaper than scanning distance

The paper’s complexity estimate for the multi-facet schemes is about 1.92 × 10⁷ operations, against about 3.60 × 10⁷ for distance-MUSIC, almost flat across facet counts (values read from its Fig. 5). The trust map is built offline, so it adds nothing at run time.

Too few facets give too little diversity; too many shrink each facet and add control and training overhead.

Robust to echoes

With weak-to-moderate reflections added (multipath), a single facet’s angle error rises smoothly as reflections get stronger, with no abrupt failure (paper’s Fig. 10, simulations).

Meter-level accuracy, the paper notes, is already enough for beam management, grouping users and locating IoT devices.

Aiming at where you are

Back to the opening question. A tower that knows roughly where you are can aim a focused beam of data at you instead of spreading energy everywhere. Here each device gets a beam aimed at its estimated position. Switch between the two ways of combining rays and watch who stays inside their beam.

Illustration built on the same toy simulation: realistic pixels, 8 facets, beam spots 0.3 m in radius. Green ring = inside its beam; pink dashed ring = missed or no confident fix.

In the paper’s simulations, better positions translate into higher data rates, and meter-level accuracy already supports beam management, user clustering and IoT device localization. The authors name broadband extensions and cooperative multi-panel configurations as next steps.

So, can the mast see where you are? In these simulations, to within a metre or two, using only the signals your devices already send, and despite imperfect pixels, by letting each facet judge only direction and trusting only the facets that have earned it.

Where this comes from

The paper

Yinuo Dong, Qingchao Li, Soon Xin Ng and Mohammed El-Hajjar. Near-Field Low-Complexity Sensing for Reconfigurable Holographic Surfaces with Hardware Constraints. Author manuscript, University of Southampton, 2026. Read the paper (PDF)

All results attributed to the paper come from its Monte Carlo simulations (30 GHz, 3072 elements, 16 elements across each facet). The charts are redrawn from the paper’s Figs. 5, 7, 9, 11, 12 and 13 with values read by eye, so they are approximate; the paper’s own figure images are not reproduced. All visuals on this page are original code.

What the interactive scenes are

A flat, top-down (2-D) model. Each facet is a line of 16 elements half a wavelength apart; the paper’s facets are 2-D panels that also estimate elevation.

The Lorentzian element model is the paper’s (ψ = (j + ejζ)/2); MUSIC uses an ideal codebook, as in the paper’s mismatch analysis.

Each facet’s angle errors are replayed from our own offline Monte Carlo run (60 trials per angle and signal-to-noise level, one interfering user, coarse window ±7°), and the same run gives the toy trust map.

The tower is drawn far larger than a real one (facets 1.2 m from the centre) so that the rays are visible. Per-facet signal strength falls with distance and with the number of facets.

Numbers labelled “in this toy” describe this model only and are not comparable to the paper’s results. The thresholds α1 = 0.8 and α2 = 1° are tuned for the toy.

The beam scene in chapter 6 is an illustration, not a data-rate calculation.

Topics in this iPaper

Engineering
we consider near-field multi-user localization as a key ISAC function

Abstract, page 1

Computer science
ROMP operates as an iterative greedy algorithm to approximate the sparse representation vector

Section III-B, page 8

integrated sensing and communication
Integrated sensing and communication (ISAC) has emerged as a key application scenario for forthcoming sixth-generation (6G) communications

Section I, page 1

reconfigurable holographic surfaces
We consider a multisector RHS architecture, which is physically implemented as a prism-shaped metasurface mounted on a base station.

Section II-A, page 4

near field localization
In this paper, we focus on near-field multi-user localization within the broader ISAC framework.

Section I, page 2

direction of arrival estimation
enabling robust multi-user angle-of-arrival (AoA) estimation under limited-resolution hardware

Abstract, page 1

multiple signal classification
we provide a rigorous mathematical proof to elucidate the emergence of pseudo-peaks in the MUSIC algorithm under model mismatch conditions

Section III-A, page 6

compressed sensing
we employ ROMP as an enhanced alternative to conventional OMP

Section III-B, page 8

metasurfaces
we adopt the Lorentzian phase constraint model [41] [42] [43] [44],which characterizes the feasible set of element response coefficients

Section II-C, page 5

Near-Field Low-Complexity Sensing for Reconfigurable Holographic Surfaces with Hardware Constraints — Interactive explanation | iPaper.ai