Billions of conversations at once
The ITU counts more mobile subscriptions in the world than people. Each phone, laptop, watch, car and sensor is a radio transmitter, and nearly all of them share a few hundred megahertz of the spectrum below 7 GHz, the part that passes through walls and travels kilometres. If every device simply transmitted whenever it liked, on whatever frequency it liked, nothing would work.
The previous lessons gave us the ingredients: antennas and link budgets, modulation and OFDM, capacity and coding. This lesson assembles them into systems. Three ideas let the same spectrum be used billions of times over: reuse it in space (cells), share it in time, frequency and code (multiple access), and separate users by direction (antenna arrays and MIMO). Then we look at the systems built from them in 2026 (5G, Wi-Fi 7, Bluetooth, satellite constellations), at who decides which frequencies they may use, and at what researchers are working on for 6G.
The cellular idea
Early mobile telephone services used one powerful transmitter on a hill to cover a whole city, with a few dozen channels for everyone. In 1947 engineers at Bell Laboratories proposed the alternative: cover the area with many small, low-power cells, and reuse the same frequencies in cells far enough apart not to interfere. Capacity then grows with the number of cells, not with the spectrum. The first commercial cellular networks opened in Tokyo (NTT, 1979), the Nordic countries (NMT, 1981) and the United States (AMPS, 1983).
If the available channels are divided into sets and each set is used once in every cluster of neighbouring cells, the geometry of hexagons allows only certain cluster sizes, for whole numbers (1, 3, 4, 7, 9, 12, …), and the distance between cells using the same channels is for cells of radius .
A phone at the edge of its cell receives its own base station from distance about and the six nearest co-channel cells from distance about . Received power falls as distance to the power , the path-loss exponent, which is 2 in free space and typically 3 to 4 on the ground, where the direct wave and reflections from the ground and buildings partly cancel. So the signal-to-interference ratio is roughly
import math
# Hexagonal cells: valid cluster sizes are N = i^2 + i*j + j^2. The nearest cells reusing the
# same channels are D = R * sqrt(3N) away, and there are six of them in the first ring.
clusters = sorted({i * i + i * j + j * j for i in range(5) for j in range(5) if i + j > 0})[:8]
print(" N D/R SIR (gamma = 3.5) SIR (gamma = 4) channels per cell (from 600)")
for n in clusters:
q = math.sqrt(3 * n)
sir = [10 * math.log10(q**g / 6) for g in (3.5, 4.0)]
print(f"{n:3d} {q:6.2f} {sir[0]:12.1f} dB {sir[1]:15.1f} dB {600 // n:14d}")Analog phones needed about 18 dB of SIR for acceptable speech, which pointed to : each cell could use only a seventh of the channels. Digital systems with error-correcting codes tolerate far lower SIR, and the trade swung: GSM used clusters of around 3 to 4, and LTE and 5G use reuse 1, every frequency in every cell, accepting an SIR near 0 dB at the cell edge and relying on coding, adaptive modulation, interference coordination between base stations and beamforming to cope. The users at the cell edge get low rates; the users near the mast get very high ones.
Multiple access
Within one cell, many users share the base station's spectrum. There are four classic ways to divide it, and every generation of mobile networks is named partly by its choice.
Definition— FDMA, TDMA, CDMA and OFDMA
- Frequency-division multiple access (FDMA) gives each user its own frequency channel for the whole call (first-generation analog systems).
- Time-division multiple access (TDMA) gives each user the whole channel for a short, repeating time slot (GSM: 8 slots on each 200 kHz carrier).
- Code-division multiple access (CDMA) lets all users transmit at once over the whole band, each multiplying its data by a different fast spreading code; the receiver separates them by correlating with the wanted code (3G: WCDMA and cdma2000).
- Orthogonal frequency-division multiple access (OFDMA) splits an OFDM signal's subcarriers and symbols into small tiles and assigns tiles to users, changing the assignment every millisecond or less (4G LTE, 5G NR, Wi-Fi 6 and later).
Two related choices complete the picture. Duplexing separates the two directions of a link: frequency-division duplex (FDD) uses a pair of bands, time-division duplex (TDD) alternates in time on one band, which suits the asymmetric traffic of downloads and lets a base station measure the channel on the uplink and use it for the downlink. And space-division multiple access serves several users on the same tile in different directions, using antenna arrays, which is what massive MIMO does.
CDMA deserves a demonstration, because its separation of users looks like magic. Each user gets a row of a Walsh–Hadamard matrix as its code; the rows are orthogonal, so correlating the sum of everyone's signals with one user's code cancels all the others exactly.
import numpy as np
def walsh(n):
"""n x n Walsh-Hadamard matrix (n a power of 2), built by Sylvester's doubling."""
w = np.array([[1]])
while len(w) < n:
w = np.block([[w, w], [w, -w]])
return w
W = walsh(8)
print("rows are orthogonal: W W^T = 8 I ->", np.array_equal(W @ W.T, 8 * np.eye(8, dtype=int)))
rng = np.random.default_rng(3)
users = {"Amara": 1, "Bilal": 2, "Chen": 5} # each user gets one row (code) of W
bits = {name: rng.choice([-1, 1], 6) for name in users} # six bits each, as +-1
amplitude = {"Amara": 1.0, "Bilal": 0.5, "Chen": 2.0} # received at different strengths
# Each bit is multiplied by the user's 8-chip code; the air adds all users' chips together.
air = sum(amplitude[n] * np.kron(bits[n], W[users[n]]) for n in users)
air = air + rng.normal(0, 0.5, air.shape) # plus noise
for name, row in users.items():
chips = air.reshape(-1, 8) # one row of 8 chips per bit
corr = chips @ W[row] / 8 # correlate with this user's code
decided = np.sign(corr).astype(int)
print(f"{name:6} sent {bits[name]} correlator {np.round(corr, 2)} "
f"decided correctly: {np.array_equal(decided, bits[name])}")
unused = air.reshape(-1, 8) @ W[7] / 8
print("correlating with an unused code gives only noise:", np.round(unused, 2))All three users are recovered, even though they share every chip in time and frequency and arrive at different strengths, and correlating with an unused code yields only the noise. The orthogonality is exact only when the users' chips are aligned in time, as on a base station's downlink; on the uplink, and with multipath, the codes are only nearly orthogonal, and 3G networks needed fast, precise power control so that a nearby phone would not drown a distant one.
4G LTE and 5G NR
4G LTE (first deployed in 2009–2010) uses OFDMA on the downlink with 15 kHz subcarriers, 1 ms subframes and carriers up to 20 MHz, which can be aggregated; DFT-spread OFDM on the uplink for lower peak power; turbo codes; and MIMO with up to 4 or 8 layers. Its core network is all-IP, so voice became an application (VoLTE).
5G NR ("New Radio", first specified in 3GPP Release 15 in 2018) keeps OFDM but makes it flexible. The key parameter is the numerology : the subcarrier spacing is kHz, and every symbol and slot shrinks by the same factor.
import math
print("5G NR numerologies (normal cyclic prefix: 14 OFDM symbols per slot)")
print(" mu spacing slot symbol + CP slots per 1 ms subframe")
for mu in range(5):
scs = 15e3 * 2**mu
slot = 1e-3 / 2**mu
print(f" {mu} {scs / 1e3:5.0f} kHz {slot * 1e3:6.4f} ms {slot / 14 * 1e6:8.2f} us {2**mu:10d}")
print(f" a 100 MHz carrier at 30 kHz holds 273 resource blocks = {273 * 12} subcarriers "
f"= {273 * 12 * 30e3 / 1e6:.2f} MHz")
# Satellites: delay and Doppler for a circular orbit, seen from the ground
c, GM, R = 299_792_458.0, 3.986004418e14, 6371e3
def slant_range(h, elev_deg):
e = math.radians(elev_deg)
return math.sqrt((R + h) ** 2 - (R * math.cos(e)) ** 2) - R * math.sin(e)
print("\norbit height speed one-way delay, overhead / 25 deg Doppler at 2 GHz, 25 deg")
for name, h in [("low Earth orbit", 550e3), ("medium Earth orbit", 8062e3), ("geostationary", 35786e3)]:
v = math.sqrt(GM / (R + h)) # circular orbital speed
overhead, low = slant_range(h, 90), slant_range(h, 25)
v_radial = v * R / (R + h) * math.cos(math.radians(25)) # range rate at 25 deg elevation, overhead pass
if name == "geostationary":
v_radial = 0.0 # fixed above the equator relative to the ground
print(f"{name:18} {h / 1e3:6.0f} km {v / 1e3:5.2f} km/s {overhead / c * 1e3:9.1f} ms / {low / c * 1e3:5.1f} ms"
f" {v_radial / c * 2e9 / 1e3:18.1f} kHz")Wider subcarrier spacing gives shorter symbols, which suit high frequencies (where oscillator phase noise and Doppler shifts are larger) and low latency; narrower spacing gives longer cyclic prefixes, which suit large cells with long echoes. So 5G uses 15 or 30 kHz in the bands below 7 GHz and 120 kHz in millimetre-wave bands, and a 100 MHz carrier at 30 kHz is 3276 subcarriers.
5G's other features follow from the earlier lessons:
- Two frequency ranges. FR1 runs from 410 MHz to 7.125 GHz; the workhorse in many countries is the mid-band around 3.3–3.8 GHz, with carriers of up to 100 MHz. FR2 covers millimetre waves from 24.25 GHz upwards (extended to 71 GHz in Release 17), with carriers up to 400 MHz but short range and poor penetration of walls, bodies and foliage. As of 2026 most 5G traffic worldwide is carried in FR1; millimetre-wave networks serve dense hot spots, venues and fixed wireless access.
- Massive MIMO and beamforming. Mid-band base stations commonly use panels of 32 or 64 transceivers, forming narrow beams towards individual users and serving several at once on the same frequencies, as in the array calculation of antennas and radio links. At millimetre waves, beams are essential just to close the link budget, and phones carry small phased arrays.
- New codes and a new core. LDPC and polar codes (the previous lesson), and a cloud-native core network that supports network slicing, separate virtual networks with different guarantees on one infrastructure.
- Evolution. Release 18 (frozen in 2024) began 5G-Advanced, adding AI and machine-learning features in the radio interface, better positioning, reduced-capability devices for wearables and sensors, and integration of satellite access, which Release 17 introduced. Release 19 continued that work.
Wi-Fi 6 and 7, and Bluetooth Low Energy
Wi-Fi uses licence-exempt bands: anyone may transmit, within power limits, so the protocol must share the air politely, listening before transmitting. Recent generations add cellular-style tools for crowded places.
- Wi-Fi 6 (IEEE 802.11ax) brought OFDMA, so an access point can serve many small packets from many devices in one transmission; uplink multi-user MIMO; 1024-QAM; "BSS colouring" to tell its own transmissions from a neighbour's; and target wake time, which lets battery devices sleep until an agreed moment. Wi-Fi 6E is the same technology extended into the 6 GHz band, opened (fully or partly) by regulators in many countries from 2020: the United States, Korea, Brazil and others opened 5.925–7.125 GHz, while Europe and several other regions opened the lower 500 MHz.
- Wi-Fi 7 (IEEE 802.11be), certified by the Wi-Fi Alliance since early 2024, adds 320 MHz channels in the 6 GHz band, 4096-QAM, and multi-link operation, in which a device uses links in two or three bands at once for higher throughput or lower, more predictable latency.
- Wi-Fi 8 is being developed as IEEE 802.11bn, aimed at reliability (fewer stalls, better behaviour when moving between access points and in congested spaces) rather than a higher peak rate; as of 2026 the standard is still in draft.
Bluetooth Low Energy (BLE) is designed for devices that run for months or years on a coin cell. It operates in the 2.4 GHz band with 40 channels 2 MHz wide (three for advertising, 37 for data, with adaptive hopping among them to avoid busy ones), uses Gaussian FSK at 1 or 2 Mbit/s, and adds a long-range "coded" mode that uses error-correcting codes to trade rate for up to about four times the range. Its trick for low power is a very low duty cycle: devices wake, exchange a few packets in a millisecond or two, and sleep. LE Audio brought a new codec and broadcast audio, and Channel Sounding (Bluetooth Core 6.0, 2024) lets two devices measure their distance to tens of centimetres from phase measurements across many frequencies, for digital keys and item finding.
Satellites in low Earth orbit, and direct-to-device
For decades, satellite communication meant geostationary satellites, 35,786 km above the equator, which stay fixed in the sky so that a dish can point at one permanently. The penalty is distance: a round trip through one takes about half a second, and the path loss is huge. Low-Earth-orbit (LEO) satellites at a few hundred to about 1200 km are closer by a factor of 30 to 60. The second half of the program in the 5G section compared three orbits.
Its output shows the trade. A satellite at 550 km is 1.8 ms away overhead, so a LEO link can have latency comparable to terrestrial broadband; but it moves at 7.6 km/s, crosses the sky in a few minutes, and shifts a 2 GHz signal by tens of kilohertz, more than a 5G subcarrier spacing. Covering the Earth continuously then needs hundreds to thousands of satellites, with handovers between them every few minutes, and either steerable phased-array terminals or links between satellites.
As of 2026 several constellations are operating or being built: SpaceX's Starlink, by far the largest, with thousands of satellites in service; Eutelsat OneWeb; Amazon's Project Kuiper, which began launching production satellites in 2025; and China's Guowang and Qianfan constellations, both deploying since 2024. Medium-Earth-orbit systems such as SES's O3b mPOWER sit in between.
Direct-to-device (D2D) service connects ordinary phones to satellites, with no dish. Apple's emergency messaging via Globalstar satellites began in 2022, some phones in China can make calls through the Tiantong geostationary system, and in 2024–2025 Starlink's direct-to-cell satellites (with mobile operators in several countries) and AST SpaceMobile's large-array satellites began service and trials using operators' own mobile spectrum. 3GPP's non-terrestrial-network (NTN) features, from Release 17, standardise how a phone handles the delay and Doppler: it uses its own GNSS position and the satellite's published orbit to pre-compensate its timing and frequency. As of 2026, D2D offers messaging and modest data rates in most services, and is a complement to, not a replacement for, terrestrial networks.
Who decides who transmits where
Radio waves ignore borders, so spectrum is governed by treaty. The International Telecommunication Union (ITU), the United Nations agency for information and communication technologies, maintains the Radio Regulations, a treaty that allocates every band to services (mobile, broadcasting, satellite, radio astronomy, aeronautical, and so on), with differences between three world regions. The allocations are revised at World Radiocommunication Conferences, held every three to four years; WRC-23 met in Dubai, and WRC-27 is preparing decisions on new bands for mobile networks and satellites. The ITU-R also defines the requirements that a system must meet to be called IMT-2020 (5G) or IMT-2030 (6G).
Within each country a national regulator decides who may use each band and on what terms: Ofcom in the United Kingdom, the FCC in the United States, ANATEL in Brazil, the Department of Telecommunications in India, MIIT in China, the ACMA in Australia and the Communications Authority in Kenya, among many others, with regional coordination through bodies such as Europe's CEPT. Licensed bands are usually assigned to mobile operators by auction or administrative award, for exclusive use; licence-exempt bands (2.4, 5 and 6 GHz among them) are open to anyone whose equipment meets the technical rules; and newer shared models, such as the three-tier Citizens Broadband Radio Service in the United States, let new users in where incumbents (for example, naval radars) are not using a band at a given time and place. Equipment must be certified for the rules of each market, which is why a phone's radio supports a different set of bands in different countries.
MIMO: more antennas, more capacity
Shannon's formula says capacity grows only logarithmically with power. Multiple-input multiple-output (MIMO) systems found another way. With transmit and receive antennas in a rich scattering environment, the channel is a matrix of complex gains, and the transmitter can send several independent streams on the same frequencies at the same time; the receiver separates them because each arrives with a different spatial "signature".
Theorem— MIMO capacity
With the channel (an complex matrix) known at the receiver, equal power on each transmit antenna and total signal-to-noise ratio SNR, the capacity is
where are the singular values of .
The second form shows what MIMO does: the singular value decomposition turns the matrix channel into parallel scalar channels (see linear algebra), each with its own Shannon capacity. If the antennas see independent fading, the singular values are all substantial, and capacity grows linearly with , as Gerard Foschini and Michael Gans at Bell Labs and Emre Telatar showed in the late 1990s. The program averages the capacity over 4000 random Rayleigh-fading channels at 20 dB.
import numpy as np
rng = np.random.default_rng(2026)
def mimo_capacity(nt, nr, snr_db, trials=4000):
"""Capacity in bit/s/Hz of an nt x nr link with independent Rayleigh fading,
equal power on each transmit antenna, and the channel known at the receiver:
C = log2 det(I + (SNR / nt) H H^*)."""
snr = 10 ** (snr_db / 10)
h = (rng.normal(size=(trials, nr, nt)) + 1j * rng.normal(size=(trials, nr, nt))) / np.sqrt(2)
m = np.eye(nr) + snr / nt * h @ np.conj(np.transpose(h, (0, 2, 1)))
c = np.log2(np.linalg.det(m).real)
return c.mean(), np.percentile(c, 10)
print(" antennas mean capacity 10 % outage (SNR 20 dB, bit/s/Hz)")
for nt, nr in [(1, 1), (1, 4), (4, 1), (2, 2), (4, 4), (8, 8)]:
mean, outage = mimo_capacity(nt, nr, 20)
print(f" {nt} x {nr} {mean:12.2f} {outage:13.2f}")
print(f" AWGN 1 x 1, no fading: {np.log2(1 + 100):.2f}")
mean44 = mimo_capacity(4, 4, 20)[0]
print(f" in 100 MHz, 4 x 4 at 20 dB averages {mean44 * 100:.0f} Mbit/s against "
f"{mimo_capacity(1, 1, 20)[0] * 100:.0f} Mbit/s for 1 x 1")A single antenna at each end averages 5.9 bit/s/Hz (less than the 6.7 of an unfaded channel, because fading sometimes leaves the link weak), and in 10 % of channels it gets less than 3.6. Four receive antennas (1 × 4) add a few bits by collecting more energy and, more importantly, raise the worst case: diversity. But only multiple antennas at both ends multiply capacity: 2 × 2 roughly doubles it, 4 × 4 nearly quadruples it to 22 bit/s/Hz, and 8 × 8 reaches 44. In a 100 MHz carrier that is the difference between 590 Mbit/s and 2.2 Gbit/s at the same total power.
Towards 6G: research directions as of 2026
The ITU-R's framework for IMT-2030 (Recommendation M.2160, 2023) sets the goals for 6G: it extends 5G's three usage scenarios (enhanced broadband, massive machine communication, and ultra-reliable low latency) into immersive communication, massive communication and hyper-reliable communication, and adds three new ones: ubiquitous connectivity, integrated sensing and communication, and integrated AI and communication. In 3GPP, Release 20 began 6G study items in 2025, and the first 6G specifications are expected in Release 21, with commercial networks targeted for around 2030. Research programmes are running in many regions, including the European Union's Smart Networks and Services Joint Undertaking, the Next G Alliance in North America, Japan's Beyond 5G programme, China's IMT-2030 (6G) Promotion Group and India's Bharat 6G Alliance.
The directions with the broadest support, as of late 2026, are these. All are research and pre-standardisation; which ones reach products, and when, is not yet settled.
- New spectrum. The "upper mid-band" from about 7 to 24 GHz (WRC-27 is studying parts of it, alongside 4.4–4.8 GHz, for mobile use) offers wide channels with better coverage than millimetre waves, with massive MIMO arrays of hundreds of elements. Sub-terahertz bands above 100 GHz are studied for very short, very fast links.
- Integrated sensing and communication. Using the network's own signals as a radar, to detect and track drones, vehicles or people, raising privacy questions alongside technical ones.
- AI in the air interface. Machine-learned channel estimation, beam management and even learned waveforms, with the open question of how to specify and test behaviour that is learned rather than written down.
- Non-terrestrial integration. Satellites, high-altitude platforms and terrestrial cells as one network from the start, building on 5G's NTN work.
- Energy efficiency. Networks already use a noticeable share of operators' costs in electricity; 6G research puts energy per bit, and switching hardware off when traffic is low, among the primary design targets.
The professional view
- Standards are written by companies, together. 3GPP working groups of engineers from equipment makers, operators and chip designers in Europe, Asia and the Americas meet several times a year and agree specifications by consensus. An engineer at a vendor may spend a career implementing a few sections of them, and the specifications themselves are free to download from 3GPP.
- Open and virtualised radio networks. O-RAN interfaces and virtualised base-station software let operators mix vendors and run baseband processing on general-purpose servers with accelerators, a significant change in how networks are built in the 2020s.
- Planning and optimisation. Operators plan coverage with ray-tracing and statistical propagation models, then tune live networks with drive tests and the measurements phones report; self-organising network functions automate much of the tuning.
- Testing and certification. Devices pass conformance tests defined by 3GPP and industry certification programmes before operators accept them, and regulators check emissions and exposure against national rules derived from international guidelines.
Common misconceptions
Misconception
"5G means millimetre waves." Most 5G networks in 2026 run in the bands below 7 GHz, especially around 3.5 GHz, and some in low bands below 1 GHz shared with 4G. Millimetre-wave 5G is a minority used where very high capacity is needed over short distances.
Misconception
"More signal bars means a faster connection." Bars measure received signal strength, but speed depends on the signal-to-interference-plus-noise ratio, on how many other users share the cell, on the bandwidth the operator has, and on the backhaul. A strong signal in a crowded stadium can be slow.
Misconception
"Satellite direct-to-device will replace mobile masts." A satellite beam covers a large area with one share of spectrum, so the capacity per square kilometre is far lower than a terrestrial network's. D2D fills gaps in coverage (oceans, mountains, disasters); it cannot carry a city's traffic.
Misconception
"Each new generation uses frequencies that are more dangerous." All mobile and Wi-Fi frequencies are non-ionising (see the first lesson of this level), and exposure limits based on international guidelines apply across all of them. Higher frequencies are absorbed closer to the skin surface, and the limits are set accordingly.
Exercises
1. Wireless systems in 2026
Warm-upHumanA cellular layout uses clusters of cells, and the path-loss exponent is . Using the lesson's estimate with six equidistant interferers, what is the signal-to-interference ratio at the cell edge, in dB, to the nearest whole number?
2. Wireless systems in 2026
Warm-upHumanA 5G NR millimetre-wave carrier uses numerology . How long is one slot, in microseconds?
3. Wireless systems in 2026
CoreHumanA user terminal sends a packet to a geostationary satellite 35,786 km directly overhead, which relays it to a ground station that also has the satellite directly overhead. How long does the packet take in propagation alone, user to ground station, in milliseconds, to the nearest whole number? Use m/s.
4. Wireless systems in 2026
CoreHumanWhy did 4G and 5G choose OFDMA rather than the CDMA of 3G for wideband channels?
5. Wireless systems in 2026
CoreHumanA new satellite service wants to use a band worldwide. Where are the international allocations of radio bands to services decided?
6. Wireless systems in 2026
CoreComputationalWrite mimo_capacity(H, snr_db), returning the capacity in bit/s/Hz of a known channel matrix H (shape , complex) with equal power on each of the transmit antennas:
The tests use an identity channel (parallel streams), a rank-one channel (all ones), and a single transmit antenna with four receivers.
7. Wireless systems in 2026
StretchComputationalImplement code-division multiple access:
walsh(n)returns the Walsh–Hadamard matrix for a power of 2, built by Sylvester's doubling , starting from .despread(chips, code)cuts the received chips into blocks oflen(code), correlates each block with the code, and returns a decision of +1 or −1 per block.
The tests send three users at very different strengths through noise, all on the same chips, and expect each to be recovered.
Further reading and resources
- 3GPPThe partnership of standards bodies from several regions that writes the LTE, 5G NR and, from Release 20, 6G specifications.
- ITU Radiocommunication Sector (ITU-R)Maintains the Radio Regulations, organises the World Radiocommunication Conferences, and defines the IMT-2020 (5G) and IMT-2030 (6G) requirements.
- IEEE 802.11 Wireless LAN Working GroupWrites the Wi-Fi standards, including 802.11ax (Wi-Fi 6), 802.11be (Wi-Fi 7) and the 802.11bn project that will become Wi-Fi 8.
- Bluetooth SIGPublishes the Bluetooth Core Specification, including Bluetooth Low Energy, LE Audio and Channel Sounding.
- Wi-Fi AllianceThe industry body that certifies interoperable Wi-Fi products and names the generations.
- On Limits of Wireless Communications in a Fading Environment when Using Multiple Antennas · Gerard J. Foschini and Michael J. Gans · 1998Wireless Personal Communications, vol. 6. With Emre Telatar's work at the same time, the result that MIMO capacity grows linearly with the number of antennas.
- OfcomThe United Kingdom's communications regulator, an example of a national spectrum authority.
- Federal Communications CommissionThe United States' communications regulator.