What survives the link budget

A radar loses signal with the fourth power of distance

A phone receives a base station’s signal after one trip, so the power falls with the square of the distance. A radar receives its own signal back after two trips and a reflection in between, so the power falls with the fourth power. Everything that follows comes from that.

The difference compounds quickly. Doubling the distance costs a communications link 6 dB and a radar 12 dB, which is another way of saying that doubling a radar’s range takes sixteen times the energy on the target. Put one radio through both equations and the gap is plain (Exhibit 1). A 320 W radio behind a 26.1 dBi panel at 3.85 GHz delivers a phone 1 km away a signal of −23.0 dBm in free space. The echo it gets back from a small drone at the same distance is −80.8 dBm, about 58 dB weaker. A car, being a much bigger reflector, sends back −56.7 dBm.

Integration closes most of that gap, at least against noise. A 1 ms look across 100 MHz adds about 50 dB of processing gain, which is why the noise-only ranges below run to kilometres. What it cannot do is close the gap against anything that integrates along with the echo: the transmitter’s own leakage, reflections from buildings and the ground, and whatever else in the beam happens to be moving.

Every input to the calculation is on a public record

We have not had to assume much. Most of the inputs are 3GPP’s own evaluation assumptions, and the rest are the FCC’s power limits and two preprints (Exhibit 2).

One thing the table makes clear is that neither the regulator nor the noise floor is the limit. The FCC allows 82.1 dBm in 100 MHz and a 320 W radio reaches 81.1 dBm, so the licence leaves headroom to spare. The cap that binds is 3GPP’s: 37 or 52 dBm while listening, set by receiver saturation and isolation, which part 3 works through.

On noise alone a C-band macro would see a small drone kilometres away

Run on those inputs, the radar equation is generous. It gives ranges of kilometres for every target at n77 and of a kilometre or more at 28 GHz (Exhibit 3).

Two measurements give us a check on the method against the real world. In May 2026 Nokia Bell Labs reported detecting a DJI Air 3 with unmodified commercial 5G radio units at 27.6 GHz, “in over 500 meters distance in a challenging radio environment rich of strong clutter”, with a median range error of 23 cm against satellite positioning. Researchers at the University of Science and Technology of China report “a sensing range of up to 1,000 m” and a mean localisation error of 4.9 m in their BSense preprint. Their site is a 4.9 GHz Huawei 5G-Advanced base station with 128 channels and a combined OFDM and chirp waveform. Those are the only two measured single-site ranges on the record, and both fall well inside the noise-only figures.

Five constraints take 7.6 km on paper to about 1.4 km

Between 7.6 km on paper and a kilometre in the field lie five constraints that the noise budget leaves out, and each of them has a number on the record. Applied one after another, as an illustration, they bring the n77 range down to about 1.4 km before elevation is even counted (Exhibit 4).

1. The transmitter. 3GPP capped monostatic power at 37 or 52 dBm because a receiver saturates at −28 dBm and isolation was assumed at 65 or 80 dB. Whatever leaks past the cancellation raises the noise floor. The companies in 3GPP’s evaluation modelled that leakage at 0, 5 or 10 dB above thermal noise; at 5 dB the floor rises by 3 dB and range falls by 16%. The alternative, transmitting at full power and listening afterwards, leaves the site deaf while it transmits. One 30 kHz OFDM symbol lasts about 35.7 µs, which blinds the site to everything within about 5.35 km. The rural results that met 3GPP’s objectives, all from one source, did both: 37 dBm while listening, plus 56 dBm transmissions while the receiver was off.
2. Clutter. In Nokia’s model the echo of one building 30 m away raises the interference floor by 2.9 dB through transmitter intermodulation, and that alone takes 797 m to 675 m. The BSense team give a vivid picture of what clutter looks like at the receiver: in one frame “only a single point corresponds to the UAV, while 174 points are noise”. They averaged 168 false detections in every 640 ms frame and got that down to 0.04 only by filtering at the point, object and trajectory levels. In 3GPP’s evaluation, one single-site result with moving clutter reported false alarms of 91%.
3. The search. A narrow beam has to be pointed somewhere, and a drone could be anywhere. A wide transmit beam covers the sky but gives up about 18 dB of gain, which divides range by 2.8. The millimetre-wave array 3GPP evaluated is analogue and has to sweep. The one result of seven that met every objective used 220 transmit beams with observations of up to 10 ms each, which is about 2.2 s to come back round to every beam.
4. The receive window. OFDM’s cyclic prefix is about one fourteenth of a symbol. At 30 kHz spacing an echo from beyond about 357 m arrives outside it, and the unshifted receive window ends near 5.35 km. At 120 kHz the figures are 89 m and 1,338 m, which match Nokia’s own. In Nokia’s model the window takes 675 m to 540 m at 27.6 GHz; at 30 kHz it takes our n77 step-down from 1.6 km to 1.4 km.
5. Elevation. A macro panel is built to cover the ground, and drones are not on the ground. Seen from a 25 m mast, a drone at 120 m is 43.5° up at 100 m, 17.6° at 300 m and 2.7° at 2 km. 3GPP’s evaluation offered panels with no tilt or with 12° of downtilt, whereas China Mobile advertises its sensing sites with an opening angle of at least 60°, which is a vertical pattern built for the purpose. A hovering drone is a separate problem again. Nokia’s missed detections came at radial speeds under 5 m/s, where the processing that removes static clutter removes the drone along with the buildings.

Thermal noise, in other words, is not what limits a sensing network. Five constraints, applied in turn as an illustration, take a 7.6 km noise-only range to about 1.4 km before elevation is counted. They are a wide search beam, a fluctuating target, one building’s clutter as Nokia Bell Labs modelled it, transmitter leakage 5 dB above thermal noise, and the receive window. That lands much closer to the two published measurements, 500 m and 1,000 m, than to the noise budget. The levers that move range are isolation, clutter suppression and vertical coverage. Transmit power is not one of them.

The claims above 2 km publish no conditions

Five range figures are on the record, and they are not all of the same kind: two measurements that publish their conditions, two claims that do not, and one multi-site demonstration (Exhibit 5).

In May 2023 China Mobile and Huawei claimed that a Shenzhen 5G-A base station could sense targets of 0.01 m², or −20 dBsm, at “单站探测距离超2KM”, a single-site range of more than 2 km. China Mobile Shandong and ZTE repeated both claims in January 2024. Is that plausible? On the noise budget with a wide search beam, a −20 dBsm target is seen at 3.5 km with 52 dBm and a 16 ms dwell, and at 0.75 km with 37 dBm and 1 ms. So the claims fall inside the first case and outside the second, and since none of them publishes its conditions there is no way to tell which applies. The AT&T and Ericsson multistatic demonstration points the other way. Ericsson Federal’s chief executive told RCR Wireless that “It takes at least three for ‘reasonable tracking’”, which is exactly the conclusion 3GPP’s evaluation reached with monostatic sites. One site on its own is rarely enough.

Which uses survive

Set against the ranges, the evaluation and the measurements, the candidate uses sort themselves by how far they need to see and by what limits them first (Exhibit 6).

The pattern is that sensing survives where the network is dense and the customer is fixed in place. A city grid with sites 500 m apart puts every drone within a few hundred metres of three sites, which is the geometry 3GPP’s evaluation passed in. A stadium or a prison is a fixed perimeter that a carrier, or a private network, can surround with sites chosen for the purpose. The use the roadmaps most often draw, a national drone-surveillance layer from existing rural macro sites, does not survive on today’s evidence. Its sites are too far apart for one to see a small drone reliably, and too loosely synchronised for several to combine, as part 3 showed. That one has to wait for the bistatic, multistatic 6G design.

The terms, briefly

• Radar cross-section (dBsm). How large an object looks to a radar, in decibels relative to one square metre. A small drone is about −13 dBsm, a twentieth of a square metre; a car is about +11 dBsm.
• EIRP. Effective isotropic radiated power: transmit power plus antenna gain, the figure the FCC limits.
• Dwell. The transmit time integrated on one target; range grows with its fourth root.
• Swerling 1. A standard model of a target whose echo fluctuates from look to look. It needs about 8 dB more signal than a steady one for the same detection probability.
• Cyclic prefix. The guard copied onto each OFDM symbol. An echo delayed beyond it spills into the next symbol unless the receiver shifts its window.