The last stage: shouting across the gap
At the receive end of a radio link, the signal is a whisper and the job of the low-noise amplifier is to listen without breathing on the microphone. At the transmit end, the job is the opposite: shout. Your phone has to throw enough energy into the air that a cell tower hundreds of metres away — or a satellite hundreds of kilometres up — can still pick the signal out of the cosmic hiss. That shouting is the job of the RF power amplifier, or PA, and it is almost always the very last active block before the antenna.
How loud is a shout? A typical smartphone PA delivers up to roughly +23 dBm — about 200 milliwatts — at the antenna port. A Wi-Fi access point might push +20 dBm. A broadcast FM transmitter or a cellular base station runs from tens of watts to kilowatts. The numbers span seven or eight orders of magnitude, but the underlying tension is always the same, and it is worth stating up front: a PA must turn DC power from the battery into RF power in the antenna efficiently, cleanly, and at high output — and you cannot freely have all three.
Efficiency vs. linearity: the PA's central war
Think of the PA as an engine that converts the battery's DC into RF. Efficiency is simply RF power out divided by DC power in. The waste — the difference — comes out as heat, which drains your battery, cooks the chip, and forces bigger heat sinks. So why not just run at 100%? Because the easy way to be efficient is to let the transistor slam fully on and fully off like a switch, and a switch is brutally nonlinear — it mangles the shape of the signal it is supposed to faithfully amplify.
Why does nonlinearity matter so much? A modern signal — QAM or OFDM — carries information in the precise amplitude and phase of the carrier. If the PA's gain sags when the input gets large (it always does, eventually), the output is a warped copy of the input. Mathematically, a nonlinear transfer characteristic generates harmonics (at 2f, 3f…) and, far more dangerously, intermodulation products. Feed in two tones at f1 and f2 and the cubic term of the nonlinearity produces new tones at 2f1−f2 and 2f2−f1 — third-order products that land *right next to* your signal, inside the channel, where no filter can remove them. They smear energy into adjacent channels (measured as ACPR, adjacent-channel power ratio) and corrupt your own constellation (measured as EVM, error vector magnitude).
Pout (dBm)
^
| ___ saturation (Psat)
| __--
| _-- <- gain drops here
| _-- . <- P1dB (1 dB compression point)
| _-- .'
| _-- .' ideal linear gain (slope = small-signal gain)
| _- .'
|- .'
+----------------------------------> Pin (dBm)
|<-- back-off -->|
Run HERE (backed off) Run HERE for max efficiency
linear, low efficiency compressed, distortedThis is the heart of the trade. Back-off means deliberately operating the PA several dB below its saturated output, in the gentle linear part of the curve, so the signal stays clean — at the cost of efficiency, because you've sized a big, power-hungry transistor and are only using a fraction of its swing. The harder you compress (the closer to saturation), the more efficient you get, but the dirtier the output. A constant-envelope signal like old FM or GMSK doesn't care — you can clamp it into a hard-switching, high-efficiency mode because there's no amplitude information to preserve. But a high-peak-to-average signal like 64-QAM OFDM, with a peak-to-average power ratio (PAPR) of 8–12 dB, must be backed off enormously, which is exactly why your phone gets hot and your battery dies when you're streaming at the edge of coverage.
Taming distortion: feedback, predistortion, and matching
How do engineers fight nonlinearity? The classic tool is negative feedback: sample the output, compare it to the input, and feed back the error to flatten the gain. It works beautifully at audio and IF frequencies, but at GHz the loop's own phase shift turns feedback into oscillation, so a global feedback loop around an RF PA is rarely practical. Instead, RF designers reach for two other levers. Digital predistortion (DPD) measures the PA's exact warping and pre-warps the digital signal with the *inverse* curve, so the two distortions cancel and the output comes out straight — this is how base stations run efficient PAs hard while still meeting spectral masks. And good [[ee-impedance-matching|impedance matching]] at the output makes sure the carefully built RF power actually transfers into the antenna instead of reflecting back as heat.
Matching at the PA output is subtle, because the impedance that maximizes *power* is not the one that maximizes *efficiency* or *linearity*. A PA designer runs a load-pull: they physically (or in simulation) vary the load impedance presented to the transistor and map out contours of constant output power and constant efficiency on a Smith chart, then choose a compromise point. The output matching network is then built — often with bond wires, on-chip inductors, and microstrip — to transform the 50 Ω antenna into that optimum load. Get it wrong and a high VSWR from an antenna detuned by your hand near the phone can throw power back into the PA and, in the worst case, destroy it.
The mixer: a multiplier that moves frequency
A radio has a problem of geography. Your data — voice, video, bits — naturally lives at low ('baseband') frequencies, but it has to travel on a carrier at 0.7, 2.4, 5, or 28 GHz, and the receiver has to bring it back down to where its filters and ADC can handle it. You need a device that slides a signal from one frequency to another without destroying the information riding on it. That device is the mixer, and its secret is breathtakingly simple: it multiplies.
Multiply two sine waves and a trigonometric identity hands you the magic: cos(A)·cos(B) = ½cos(A−B) + ½cos(A+B). Multiply your signal at frequency f_RF by a clean reference tone at f_LO from a local oscillator (LO), and out come two new copies — one at the sum (f_RF + f_LO) and one at the difference (f_RF − f_LO). Pick off whichever one you want with a filter, and you have moved your signal — modulation, phase, amplitude and all — wholesale to a new frequency. That's the entire trick of the radio's 'gearbox'.
DOWN-CONVERSION (receiver) UP-CONVERSION (transmitter)
RF in ----[ X ]---- IF out Baseband --[ X ]-- RF out
| |
f_LO f_LO
(local osc.) (local osc.)
f_IF = | f_RF - f_LO | f_RF = f_BB + f_LO
e.g. 2400 MHz signal e.g. 100 MHz baseband
mixed with 2300 MHz LO mixed with 2300 MHz LO
-> 100 MHz IF (kept) -> 2400 MHz RF (kept)
-> 4700 MHz sum (filtered out) -> 2200 MHz image (filtered out)Crucially, a mixer is fundamentally a nonlinear or time-varying device — it has to be, because a purely linear block can only scale a signal, never create new frequencies. In practice we don't build a clean analog multiplier at GHz; instead the LO is a big square-ish drive that switches transistors on and off, effectively multiplying the RF by a square wave. The classic Gilbert cell does exactly this: a differential pair carrying the RF current is steered by an LO-driven switching quad, producing a 'doubly balanced' mix that suppresses LO and RF leakage to the output. The price of a switching mixer is that it also mixes with the LO's harmonics, so layout, filtering, and a clean LO all matter.
The image problem and the cost of free lunch
The difference operation hides a nasty surprise. Down-conversion uses |f_RF − f_LO|, and the absolute-value bars are the whole problem: two different input frequencies land on the same IF. If your LO is at 2300 MHz and you want the 2400 MHz signal at 100 MHz IF, then a *different* signal sitting at 2200 MHz — exactly as far below the LO as your wanted signal is above it — also produces 100 MHz IF and lands right on top of your wanted channel. That unwanted twin is the image, and once both have been folded onto the same IF, no filter on Earth can separate them.
wanted LO image
| | |
2400 MHz 2300 MHz 2200 MHz
| | |
+--- 100 ---+--- 100 ---+
MHz above MHz below
\ /
\ /
v v
both -> 100 MHz IF (collision!)
Fix #1: RF image-reject filter BEFORE the mixer
Fix #2: image-reject mixer (Hartley/Weaver, I/Q phasing)
Fix #3: zero-IF (f_LO = f_RF) -- the signal IS its own imageThere's a second cost. Unlike an LNA, a typical passive mixer has conversion loss (a few dB), and even active mixers add significant noise — so a mixer placed early in the receive chain hurts the system noise figure badly. This is exactly why the LNA comes *first*: by Friis' formula, the LNA's gain divides down the noise contribution of the lossy mixer behind it. The architecture is a chain of compromises, and the order of the blocks is itself a design decision.
Putting it together: the transceiver
Now stand back and watch all the blocks fall into place. A transceiver is just a receive chain and a transmit chain sharing one antenna. On receive, the faint signal from the antenna is lifted by the LNA, slid down in frequency by a mixer driven by the LO, filtered, and digitized. On transmit, the modulated baseband is mixed *up* to the carrier by another mixer, then handed to the PA — which shoves it, at last, into the antenna. The whole block diagram is a story of a signal climbing up and back down the frequency ladder.
ANTENNA
|
[ duplexer /
T-R switch ]
RECEIVE <----+----> TRANSMIT
| ^
+---------+ +---------+
| LNA | | PA | <- last stage, watts out
+---------+ +---------+
| ^
+---------+ +---------+
| MIXER | | MIXER | <- frequency translate
+----X----+ +----X----+
| |
f_LO o------[ LO / PLL ]---o f_LO <- shared synthesizer
| |
+---------+ +---------+
| IF/BB | | IF/BB |
| filter, | | DAC, |
| ADC | | filter |
+---------+ +---------+
| ^
DIGITAL <--- 0101 ---> DIGITAL
(demod, decode) (mod, encode)- Receive path. Antenna → LNA (set the noise figure, add gain) → mixer (slide RF down to IF/baseband) → channel-select filter → ADC → digital demodulation.
- Transmit path. Digital modulator → DAC → up-mixer (slide baseband up to the carrier) → driver → PA (set the output power and efficiency) → antenna.
- Shared LO. A single PLL/synthesizer generates the clean LO tone both mixers use; its purity (low phase noise) directly limits how cleanly you can place and recover a constellation.
- One antenna, two jobs. A duplexer (frequency split) or T/R switch (time split) hands the antenna back and forth between the PA's shout and the LNA's listen — and keeps the PA's watts from frying the LNA.
With the PA and the mixer in hand, the receive-and-transmit signal chain is complete: you can hear a whisper, translate it to where you can read it, build a reply, lift the reply onto the carrier, and shout it back. Everything before this rung built the *quiet* front end; this rung built the *loud* back end and the gearbox between them. The one piece still missing is the thing that actually launches and catches the wave in free space — the antenna — and that is where the track turns next.