Episode 2 — OFDM, the Cyclic Prefix, and 5G Numerology

"How to Make a Brutal Channel Behave Like a Spreadsheet"

Pillar: Signal Processing & Communications · Runtime ~13 min · ~2,000 words · Elena & Marcus

MARCUS: Back for episode two. Last time we taught the grid to fake inertia. Today we cross into communications, and the goal is to make you genuinely understand the waveform underneath Wi-Fi, LTE, and 5G — not the acronym, the actual mechanism. Elena, set the objective.

ELENA: By the end of this episode you should be able to explain why OFDM turns the worst property of a wireless channel into a triviality, what the cyclic prefix is actually doing — and it's doing two distinct jobs people constantly conflate — and what 5G changed by making the waveform reconfigurable. We'll re-derive enough that it sticks.

MARCUS: And the one-sentence hook?

ELENA: OFDM is the trick that takes a horrible, echo-filled multipath channel and turns equalizing it into one division per subcarrier. We're going to make a hard channel behave like a spreadsheet — one cell, one number, done.

The Refresher: Why Wideband Channels Are Painful

ELENA: Let's rebuild the problem first, because OFDM only makes sense as a solution to a specific pain. You transmit a signal; it reflects off buildings, the ground, a truck. The receiver hears the direct path plus a bunch of delayed, attenuated copies. In the time domain, that's the channel impulse response — several taps spread out over the delay spread of the environment. And the received signal is your transmitted signal convolved with that impulse response.

MARCUS: And convolution in time is the villain.

ELENA: It's the villain in two ways. First, intersymbol interference. If you're sending symbols fast — short symbol period — and the channel smears energy across a delay spread longer than that symbol period, then each symbol's echoes land on top of the next symbols. They bleed into each other. The faster you transmit, the worse it gets, which is a miserable trade because speed is the whole point.

MARCUS: And the second way?

ELENA: Frequency-selective fading. Take that multipath channel into the frequency domain and it is not flat — it has deep notches where reflections cancel and peaks where they add. A wideband single-carrier signal sees a channel whose gain varies wildly across its own bandwidth. To undo that at the receiver you need an equalizer that inverts the channel, and for a long, nasty impulse response that's a long, expensive adaptive filter — many taps, constantly retrained, numerically touchy. In the 90s that equalizer was the modem, and it was the hard part. So hold the pain in mind: wideband single carrier means costly time-domain equalization and ISI that worsens with speed.

What Changed: Divide and Conquer the Spectrum

ELENA: Now the core idea, and it's beautiful. Instead of sending one wideband high-rate stream that sees a frequency-selective channel, split the bandwidth into a large number of narrow subcarriers, and send a slower stream on each. Here's the payoff: if each subcarrier is narrow enough, the channel across that one subcarrier is essentially flat. No notch within it — just a single complex gain. One number.

MARCUS: So you've traded one hard wideband problem for hundreds of trivial narrowband ones.

ELENA: Exactly, and that's the conceptual heart of OFDM — orthogonal frequency-division multiplexing. A frequency-selective channel, chopped finely enough, becomes a stack of flat channels. The brutal convolution-and-equalize problem collapses into: for each subcarrier, divide out one complex number to undo its gain and phase. That's the spreadsheet. Each cell is independent. The catch historically was cost — hundreds of subcarriers used to mean hundreds of oscillators and filters, which was absurd. The thing that made OFDM practical is that the whole bank of subcarriers is generated and torn down by a single FFT.

The Mechanism: The IFFT Is the Modulator

ELENA: Let me teach the part that turns a vague picture into real understanding. You have a block of data symbols — say QAM constellation points. You assign one to each subcarrier: this is the frequency-domain description of the signal you want. Subcarrier zero gets this QAM point, subcarrier one gets that one, and so on. Now — how do you produce the time-domain waveform that has exactly that spectrum?

MARCUS: You inverse-transform it.

ELENA: You take the inverse FFT. That is the entire modulator. The IFFT takes your block of frequency-domain symbols and hands you the time-domain samples to transmit. At the receiver, you do the forward FFT and you're back in the frequency domain, reading one symbol per subcarrier. The IFFT is the OFDM modulator; the FFT is the demodulator. No bank of oscillators — one efficient transform that scales as N log N. That's the engineering miracle that made the whole thing economical.

MARCUS: Where does the "orthogonal" come from, precisely?

ELENA: From the spacing. The subcarriers are spaced exactly 1 over the symbol period apart — Δf = 1/T. With that spacing, over one symbol period each subcarrier completes an integer number of cycles, so when you correlate to recover one subcarrier, every other subcarrier integrates to exactly zero. They're orthogonal. That's why you can pack them so tightly that their spectra overlap and yet not interfere — the orthogonality, not a guard band, is what keeps them separable. It's the discrete Fourier basis doing what it always does, applied to a radio.

The Cyclic Prefix: The Part Everyone Gets Half-Right

ELENA: Now the piece I most want you to walk away owning, because it's the most misunderstood. The cyclic prefix. After the IFFT produces your block of time-domain samples, you take a chunk from the end of the block and paste a copy onto the front. You transmit that copy, then the block. That's the cyclic prefix. And it does two distinct jobs.

MARCUS: Give me the first one.

ELENA: Job one is the easy one: it's a guard interval. The channel's multipath smears each block out by the delay spread. If the cyclic prefix is longer than that delay spread, then all the destructive intersymbol interference — the echoes of the previous block — lands entirely within the prefix. At the receiver you simply discard the prefix, and with it you throw away every echo from the previous symbol. ISI, gone. The condition is just: make the cyclic prefix longer than the channel's delay spread.

MARCUS: That much I think people get. What's the second job?

ELENA: The second job is subtle and it's the elegant one. Remember the channel applies linear convolution to your signal. But the clean per-subcarrier division we want — the spreadsheet — only works if the channel acts as circular convolution, because circular convolution in time is exactly point-by-point multiplication in the DFT domain. Linear convolution is not. So here's the trick: by prepending a cyclic copy of the block's tail, you make the transmitted block look periodic over the receiver's FFT window. And when a linear-convolution channel acts on a signal that looks periodic across that window, the result, inside the window, is mathematically identical to circular convolution.

MARCUS: So the cyclic prefix tricks a real, linear channel into behaving like a circular one.

ELENA: That's exactly it, and that's the deep reason it's cyclic and not just a blank guard interval. A blank guard would kill ISI — job one — but it would leave you with linear convolution and no clean diagonalization. The cyclic content is what converts the channel into a diagonal operator in the FFT domain, so each subcarrier becomes that single complex multiply, and equalization becomes one division per subcarrier. One-tap equalization. The cyclic prefix is the bridge between "I killed the echoes" and "I made the channel diagonal." Two jobs, one elegant move.

MARCUS: Nothing's free, though. What does the cyclic prefix cost?

ELENA: Overhead, directly. The prefix carries no new data — it's a repeat. So your efficiency is reduced by the ratio of prefix length to total symbol length. A longer prefix buys robustness to longer delay spreads but costs more throughput. That tension — prefix length versus efficiency — is a knob, and remembering that it's a knob is the setup for the modern part. There's a second cost too: because the time-domain signal is a sum of many independent subcarriers, occasionally they all align in phase and you get a big peak. High peak-to-average power ratio, PAPR, which forces you to back off the power amplifier or apply mitigation. That's OFDM's standing tax.

Industry Standard Today: 5G Numerology

ELENA: Now the catch-up. OFDM itself isn't new — it's in Wi-Fi and it's in 4G LTE. What changed in 5G NR is that the waveform's parameters became scalable. In LTE, the data subcarrier spacing was essentially fixed at 15 kilohertz. 5G NR introduces numerology: the subcarrier spacing can be 15 kilohertz times two to the power μ — so 15, 30, 60, or 120 kilohertz for data, with 240 used for synchronization signals and even wider spacings added in later releases for the highest bands — and the system picks the numerology to fit the job.

MARCUS: Why would you want different spacings? What's the trade?

ELENA: This is where everything we built pays off, so let's reason it out instead of memorizing. Recall Δf = 1/T. If you widen the subcarrier spacing, you shorten the symbol period — they're reciprocals. Short symbols mean lower latency, which you want for ultra-reliable low-latency applications. Wide spacing also means each subcarrier is more tolerant of frequency errors and phase noise, which matters enormously at millimeter-wave frequencies where oscillators are noisier and Doppler from motion is larger. So for mmWave 5G, you want wide subcarrier spacing — 120 kilohertz.

MARCUS: But there has to be a downside, because otherwise everyone would use the widest.

ELENA: And you already have the tool to find it. Shorter symbol period means a shorter cyclic prefix, because the prefix is a fraction of the symbol. And a shorter cyclic prefix covers a shorter delay spread. So wide-spacing numerologies only work where the channel's delay spread is small — which is fine at mmWave with short range and few long echoes, but not for a large outdoor cell with kilometer-scale multipath. There you want narrow spacing — 15 or 30 kilohertz — a longer symbol, and a longer cyclic prefix to swallow the bigger delay spread, accepting higher latency in return.

MARCUS: So numerology is just exposing the cyclic-prefix-versus-delay-spread knob as a system parameter.

ELENA: That's the whole insight, and you just stated it better than most textbooks. 5G took the fixed trade-off baked into LTE and made it a dial the network sets per deployment — even mixing numerologies in the same band for different services at the same time. Massive outdoor cell on narrow spacing, low-latency factory link on wide spacing, mmWave hotspot on the widest, all running the same fundamental OFDM machinery you now understand. One more practical note: 5G NR uses cyclic-prefix OFDM in both directions, with an optional DFT-spread variant — DFT-s-OFDM — on the uplink to tame that PAPR problem for coverage-limited handsets. Same waveform family, one extra transform to cut the peaks.

Self-Test

MARCUS: Convert it to skill. What should the listener actually work through?

ELENA: Two exercises. The first is a real design calculation. Pick a deployment — say an outdoor cell with a delay spread of a couple of microseconds. Your cyclic prefix has to exceed that delay spread, so compute roughly how long the prefix must be, then ask which numerology gives a symbol period long enough that a sensible fractional prefix covers it. You'll find the narrow 15 or 30 kilohertz numerology fits and the 120 kilohertz one doesn't — and you'll have derived the deployment rule instead of looked it up. Then compute the cyclic-prefix overhead as the prefix-to-symbol ratio so you can see the throughput you paid.

MARCUS: And the second?

ELENA: The conceptual proof-of-understanding. In your own words, explain why the prefix has to be cyclic rather than just a blank guard interval. If you can say that a blank guard would kill ISI but would leave you with linear convolution, while the cyclic copy makes the channel act as a circular convolution and therefore diagonal in the FFT domain — enabling one-tap equalization — then you have understood the single most important idea in modern digital communications. That's the episode.

MARCUS: Next time, embedded and control: field-oriented control of permanent-magnet motors — how Clarke and Park transforms turn a screaming three-phase machine into something as easy to command as a DC motor. It's the same spirit as today, actually.

ELENA: It really is. Today we made a channel diagonal with a Fourier transform. Next time we make a motor simple with a rotating transform. Same move, different frame.