Skip to content
Documentation Background

Wireless link math

This is an optional deep dive. Wireless fundamentals explains what these quantities mean. Here the numbers are worked out, so link rates and coverage claims can be checked.


A decibel is ten times the base-10 logarithm of a power ratio.

RatiodB
10
2about 3
1010
10020
100030
0.5about -3
0.1-10

Because logs turn multiplication into addition, gains and losses along a link simply add. Absolute power uses dBm, decibels relative to 1 milliwatt: 0 dBm is 1 mW, 20 dBm is 100 mW, and -30 dBm is 0.001 mW.

Noise floor. Thermal noise power is -174 dBm per hertz at room temperature. For a channel of width B, add 10 log10(B). For 20 MHz that is 10 log10(20,000,000) = 73 dB, so the floor is -174 + 73 = -101 dBm. Receiver electronics add a noise figure, commonly 5 to 8 dB, so a practical floor is about -95 dBm.

SNR. A received signal of -60 dBm over a -95 dBm floor has an SNR of 35 dB. At -80 dBm the SNR is 15 dB. A 40 MHz channel has 3 dB more noise power than a 20 MHz channel, which lowers SNR by 3 dB for the same signal.


C = B x log2(1 + SNR) (SNR as a ratio, not dB)

20 MHz channel:

SNR (dB)SNR ratiolog2(1 + SNR)Capacity
10103.4669 Mbit/s
201006.66133 Mbit/s
3010009.97199 Mbit/s

Going from 10 dB to 30 dB is a hundredfold increase in signal power for under three times the capacity. That is the logarithm at work.

Now compare two ways to improve the 20 dB case:

ChangeNew SNRCapacity
Baseline20 dB (100)133 Mbit/s
Double the transmit power (+3 dB)23 dB (about 200)20 MHz x log2(201) = 153 Mbit/s
Double the bandwidth to 40 MHz, same total power17 dB (about 50), since noise doubles40 MHz x log2(51) = 227 Mbit/s
Double the bandwidth and the power20 dB (100)40 MHz x log2(101) = 266 Mbit/s

Doubling power adds about 15 percent. Doubling the bandwidth adds about 70 percent even after the noise penalty. Spectrum is what raises capacity. This is why newer standards add wide channels and bands.

Shannon’s figure is an upper bound that needs ideal coding. Real systems reach a sizable fraction of it, and the gap closes with better codes and higher-order modulation.


In free space, received power falls with (f x d)^2. In decibels, the loss grows by 10 x n x log10(d) for a path loss exponent n.

ChangeExtra loss
Double the distance, free space (n = 2)6 dB
Double the distance, indoors (n = 4)12 dB
Move from 2.4 GHz to 5 GHz, same distance20 log10(5/2.4) = 6.4 dB
Move from 5 GHz to 50 GHz20 dB, a factor of 100 in power

Example: an AP at 2.4 GHz gives a client -60 dBm at 10 m. At 20 m indoors with n = 4 the loss is 12 dB more, so -72 dBm. The same AP at 5 GHz would add 6.4 dB, putting the client at about -78 dBm at 20 m. Walls add more (several dB each). These are the margins that decide whether a client holds a high-order modulation.


Higher-order modulation needs a higher SNR for the same error rate. An example table for a target bit error rate below 10^-4:

SNRModulationBits per symbol
Below 13 dBQPSK2
13 to 17 dB16-QAM4
Above 17 dB64-QAM6

The exact thresholds depend on the coding rate and the receiver. The pattern is what holds: raw rate triples from QPSK to 64-QAM, but only if the SNR rises by about 5 dB or more. 1024-QAM, used by recent Wi-Fi, needs a clean link in the low 30s dB.


An LTE or 5G resource block, in its 15 kHz spacing, spans 12 subcarriers (180 kHz). In a 1 ms scheduling interval there are 14 OFDM symbols, so one block holds 12 x 14 = 168 resource elements, each carrying one symbol.

For a 20 MHz carrier, which has 100 blocks:

resource elements per ms = 100 x 168 = 16,800
at 64-QAM (6 bits) = 16,800 x 6 = 100,800 bits per ms = 100.8 Mbit/s raw
at 256-QAM (8 bits) = 134.4 Mbit/s raw

These are raw figures. Reference signals, control channels, and the redundancy of the error-correcting code consume a large share, so delivered rates are lower. Using two spatial streams doubles the figure. The result sits in the same range as the Shannon numbers above for a 20 MHz channel, which is the check that the arithmetic is sensible.

Wi-Fi works the same way with a 78.125 kHz subcarrier spacing, giving a 12.8 microsecond symbol, against 66.7 microseconds for the 15 kHz spacing. The symbol time is the reciprocal of the spacing. A cyclic prefix is added to each symbol so multipath echoes that arrive late land in the guard time instead of smearing the next symbol.


Two users share a cell. Each millisecond the scheduler gives all resource blocks to one of them. User A has a strong channel and user B a weak one. The expected rates d (in Mbit/s) if scheduled in four consecutive intervals:

IntervalA’s rateB’s rate
1102
283
3102
483

Assume the long-run average rates are R_A = 9 and R_B = 2.5. Three policies:

Maximum throughput always picks the larger d: A in all four intervals. Total delivered: 10 + 8 + 10 + 8 = 36. B receives nothing.

Proportional fair picks the largest d / R:

IntervalA: d/RB: d/RPick
110/9 = 1.112/2.5 = 0.80A
28/9 = 0.893/2.5 = 1.20B
31.110.80A
40.891.20B

Total delivered: A gets 10 + 10 = 20, B gets 3 + 3 = 6, total 26. B is served exactly when its channel is at its best, and A when its own is.

Round robin alternates by turn, regardless of the channel. Starting with B: B gets 2 + 2, A gets 8 + 8, total 20. Starting with A: A gets 10 + 10, B gets 3 + 3, total 26. The outcome depends on luck, and PF avoids that.

PolicyTotal deliveredB’s share
Maximum throughput360
Proportional fair266
Round robin20 to 264 to 6

This simplification freezes the averages. In a real scheduler R is an exponentially weighted moving average that updates every interval, R(t) = b x r(t) + (1 - b) x R(t-1), so a user who has not been served sees their ratio rise until they are scheduled. That is what prevents starvation.


Light travels about 299,792 km per second. The geostationary altitude is 35,786 km. A request from a terminal goes up to the satellite and down to the gateway. The reply makes the same trip back. That is four traversals of the altitude:

GEO: 4 x 35,786 km = 143,144 km -> 143,144 / 299,792 = 0.477 s (477 ms minimum RTT)
LEO at 550 km: 4 x 550 km = 2,200 km -> 2,200 / 299,792 = 7.3 ms (minimum, satellite overhead)

The GEO figure is a hard floor, with processing and the terrestrial path added on top, which is why real GEO round trips measure 500 to 800 ms. The LEO minimum assumes a satellite directly overhead. In practice the satellite is at an angle, so the slant range is longer, and ground segments and queues add more, bringing typical figures to the 25 to 40 ms range. Inter-satellite links in vacuum have a speed-of-light advantage over fiber (about 50 percent faster than light in glass), which is why long-distance LEO paths can beat terrestrial ones.

For TCP, the bandwidth-delay product at GEO is large. A 50 Mbit/s link with a 600 ms RTT needs 50,000,000 x 0.6 / 8 = 3.75 MB in flight to be full, so window scaling is essential and slow start takes many round trips.


A GPS signal travels at the speed of light, so a timing error converts directly to distance error: 1 microsecond is about 300 meters. A receiver measures the arrival time of each satellite’s signal against its own clock. The receiver’s clock is inexpensive and slightly wrong, so every measured distance is off by the same unknown amount.

That makes four unknowns: x, y, z, and the clock offset. Four satellites give four equations, enough to solve for all of them. The clock offset falls out of the solution, which is why a cheap receiver can have the timing accuracy of an atomic clock as a by-product.