Electronics Guide

The Transmitter Design Problem

Every transmitter forms a modulated waveform, translates it to the assigned carrier, and amplifies it. The order and manner define the architecture. What makes the transmit problem distinct from the receive problem is that the last stage dominates. In a receiver, the first stage sets the noise figure and later stages matter less. In a transmitter, the final power amplifier consumes most of the direct-current power, generates most of the heat, and creates nearly all of the distortion, so the architecture is largely a set of decisions made in service of that one device. The upstream chain exists to hand the amplifier a signal it can reproduce, and often to hand it a signal deliberately distorted so that the amplifier's own errors cancel.

Three quantities drive the compromise. Linearity matters because modern modulation carries information in both amplitude and phase, so any amplitude-dependent gain or phase shift both corrupts the wanted constellation and splashes energy into neighboring channels. Efficiency, the fraction of supplied direct-current power that leaves as radio frequency power, sets battery life in a handset and electricity cost in a network. Spectral purity away from the carrier covers broadband noise, harmonics, mixer products, and converter images, and is policed by regulators and by the transmitter's own co-located receiver.

These pull in opposite directions. An amplifier driven into compression is efficient and nonlinear; backed off, it is linear and wasteful. A frequency plan that removes an unwanted product by filtering adds insertion loss at the output of the chain, where loss is most expensive. Most of what follows describes the methods engineers have found for avoiding all three prices at once.

Heterodyne Transmitters

The two-step heterodyne transmitter mirrors the superheterodyne receiver and retains the same advantage: it keeps the delicate operations at a fixed intermediate frequency, where they can be trimmed once and filtered properly.

Signal Flow

A representative chain begins with in-phase and quadrature baseband signals, a pair of digital-to-analog converters, and reconstruction lowpass filters. A quadrature modulator upconverts the baseband pair to a fixed intermediate frequency, producing a real single-sideband signal, and a bandpass filter cleans up what the modulator produced imperfectly. A second mixer, driven by a tunable local oscillator, translates the result to the assigned channel. A radio frequency filter, a driver, the final power amplifier, and the output filtering follow.

The critical property is that the quadrature modulator works at one frequency only. Its carrier leakage and its in-phase and quadrature mismatches are therefore fixed quantities rather than functions of the tuned channel, so a single trim or a single set of correction coefficients serves the whole tuning range. The same holds for the intermediate-frequency filter, which can be a ceramic, crystal, or surface acoustic wave part manufactured in volume for one of the conventional intermediate frequencies.

The Transmit Image

The final upconverting mixer is a real multiplier, so it produces both the sum and the difference of the intermediate frequency and the local oscillator. If the wanted output sits at the local oscillator plus the intermediate frequency, an unwanted full-strength replica sits at the local oscillator minus the intermediate frequency, separated from the wanted signal by twice the intermediate frequency. This transmit image is the exact counterpart of the receiver's image response, and it must usually be suppressed by sixty decibels or more before it reaches the power amplifier.

The obligation establishes the central trade in transmit frequency planning. A high intermediate frequency places the image far from the wanted channel, easing the radio frequency filter, but forces the intermediate-frequency stage and its filter to work at a higher frequency where parts are dearer. A low intermediate frequency makes the intermediate-frequency stage easy and the radio frequency filter nearly impossible. Designers often resolve the tension by making the second stage a quadrature upconverter as well, so that cancellation suppresses the image by thirty or forty decibels and the filter supplies the rest.

Local Oscillator Leakage and Spurs

Local oscillator energy leaks through every mixer. In a heterodyne transmitter the leak from the final mixer appears at the local oscillator frequency, offset from the carrier by the intermediate frequency, so the same filter that removes the image also attenuates it. This is the architecture's second great advantage, and the direct contrast with direct conversion, where the leak lands exactly on the carrier and no filter can separate it from the wanted signal.

The price of two conversions is a longer spur list. Any harmonic of the intermediate frequency combined with any harmonic of the local oscillator may land somewhere unwelcome, and designers map the combinations on a spur chart before committing to a plan. Products falling inside the transmitter's own band, inside a co-located receiver's band, or inside a protected allocation must be planned away rather than filtered away. The carrier leakage of the intermediate-frequency modulator deserves particular attention, since it upconverts with everything else and emerges as a discrete tone offset from the carrier by the intermediate frequency, where a spectral emission mask may well prohibit it.

The architecture costs an extra mixer, filter, and oscillator, and it resists monolithic integration for the same reason its receive counterpart does. It nonetheless remains standard in cellular base stations, broadcast transmitters, radar exciters, satellite uplinks, and laboratory signal generators, because the local oscillator never runs at the output frequency and therefore cannot be pulled by the power amplifier, and because the image and the leakage are both filterable.

Direct-Conversion Transmitters

The direct-conversion or zero-intermediate-frequency transmitter upconverts baseband in-phase and quadrature signals to the carrier in a single quadrature modulation step. There is no intermediate frequency, no intermediate-frequency filter, and no second mixer. The mirror sideband a quadrature modulator would otherwise produce is suppressed by the quadrature structure rather than by a filter.

The economy is decisive. A direct-conversion transmitter fits comfortably in a complementary metal-oxide-semiconductor process alongside its receiver and its synthesizer, and it has been the dominant handset architecture for two decades. What it buys with that economy is three characteristic impairments, each of which lands on top of the wanted signal and therefore cannot be filtered.

Carrier Leakage

A quadrature modulator multiplies the in-phase signal by the cosine of the carrier and the quadrature signal by the sine, then sums. Any direct-current offset in either baseband path multiplies its carrier phase by a constant and emerges as an unmodulated tone at the carrier frequency, and direct coupling of oscillator energy to the output adds to the same tone. The result, called carrier leakage or local oscillator feedthrough and quoted in decibels relative to the carrier, is a static vector added to every constellation point. It displaces the constellation origin, inflates error vector magnitude, and in a multicarrier system sits in whichever subcarrier occupies the band center.

The remedy is calibration. With the modulation set to zero, the residual output is by definition the leakage, so a detector at the output can guide a search over a pair of digital offset corrections until the residual is minimized. The search must be repeated across the tuning range and over temperature, and production transmitters store a table and refresh it opportunistically. Layout reduces how much correction is needed: differential baseband routing, symmetric placement of the two modulator arms, guard structures around the oscillator distribution, and separation between the oscillator buffer and the modulator output.

Gain and Phase Imbalance

The modulator suppresses the mirror sideband only to the extent that its two arms have equal gain and exactly ninety degrees of phase separation. A gain mismatch or phase error leaks a scaled, conjugated copy of the baseband signal to the output, producing a sideband image reflected about the carrier. In a wideband system that image folds one side of the channel onto the other, and where the channel edge abuts another user it raises adjacent-channel leakage.

The sensitivity is severe. For small errors the image suppression is approximately twenty times the base-ten logarithm of two divided by the root-sum-square of the fractional gain error and the phase error in radians. A one percent gain mismatch combined with a one degree phase error therefore yields only about forty decibels of suppression, which is inadequate for high-order quadrature amplitude modulation; reaching sixty decibels demands roughly ten times better matching than analog layout alone reliably delivers.

Transmitters therefore correct the imbalance digitally, applying a two-by-two matrix to the in-phase and quadrature samples with coefficients derived from a calibration measurement. A single complex coefficient corrects only the frequency-independent part. Mismatch between the two baseband reconstruction filters, and any differential delay between the paths, makes the imbalance a function of frequency, and correcting that requires a short complex filter rather than a scalar. Wideband transmitters almost always need the frequency-selective form.

Local Oscillator Pulling

The third impairment is the most architectural. The oscillator runs at the carrier frequency, and watts of that same frequency emerge from the power amplifier millimeters away. Energy coupled back into the resonator through the substrate, the supply rails, the package, or free space injects into the tank and drags its frequency toward the injected signal. Because that signal is the modulated output, the pulling is modulation-dependent: the oscillator wanders in sympathy with the data, producing a phase error that no static calibration removes and that appears as spectral regrowth and elevated error vector magnitude. It worsens with output power, so it is often invisible on the bench at low power and appears only at full rated output.

The answers are structural. The most common is to run the voltage-controlled oscillator at a frequency the antenna path does not contain and derive the carrier by division or mixing. Running at twice the carrier and dividing by two is popular because the divider yields accurate quadrature phases as a by-product. Running at two-thirds or four-thirds of the carrier and combining division with a mixer, the same trick that produces a sliding intermediate frequency on the receive side, achieves the same isolation. An offset phase-locked loop, or translational loop, goes further: a modulated intermediate-frequency signal becomes the reference for a loop whose oscillator runs at the carrier, so the loop performs the upconversion and actively corrects the oscillator it might otherwise pull. Physical separation, shielding, separate decoupled supply domains, and differential signaling throughout the oscillator and its distribution carry the remainder.

Polar Transmitters

Polar transmitters abandon the Cartesian description of the signal. The modulator computes the instantaneous amplitude and phase of the complex envelope and routes them down separate paths. The phase drives a constant-envelope carrier, which a nonlinear, efficient amplifier reproduces faithfully because it carries no amplitude information to distort. The amplitude is reimposed at the output stage, most often by modulating the amplifier's supply voltage.

The idea is old. Leonard Kahn described it as envelope elimination and restoration in "Single-Sideband Transmission by Envelope Elimination and Restoration," published in the Proceedings of the IRE in July 1952, with the phase component amplified by class C stages and the envelope restored at the final amplifier. Its appeal is that it decouples two conflicting requirements: the phase path needs the kind of fidelity a saturated switching stage provides trivially, and the amplitude path needs only a wide-bandwidth, efficient supply.

Differential Delay

The two paths must arrive at the output stage aligned in time. The amplitude path traverses a supply modulator with substantial group delay; the phase path traverses a modulator or phase-locked loop with an entirely different delay. Any mismatch applies the envelope of one instant to the phase of another, and the reconstructed signal is not the wanted one.

The tolerance is unforgiving, because the error appears as wideband spectral regrowth rather than as a benign in-band degradation. Delay mismatches of a few nanoseconds are enough to violate adjacent-channel requirements on a signal of a few megahertz, and the requirement tightens in proportion to bandwidth. Production polar transmitters therefore include a calibrated adjustable delay in one path, set at manufacture and often adapted in service from an observation of the output spectrum.

Bandwidth Expansion

The second difficulty is fundamental. Amplitude and phase are nonlinear functions of the complex envelope, involving a square root and an arctangent, and neither preserves bandwidth. A signal occupying a given channel yields an amplitude signal several times wider and a phase signal wider still, commonly cited as an order of magnitude for a signal with a large peak-to-average ratio. Both paths must cover their expanded bandwidths or the reconstruction fails.

The expansion is worst where the trajectory passes near the origin of the complex plane. At an envelope null the amplitude must reach zero and the phase must slew nearly one hundred and eighty degrees instantaneously, which no physical supply modulator or phase path achieves. The standard mitigation, sometimes called hole blowing or trajectory shaping, steers the trajectory around a small exclusion zone about the origin, accepting a modest error vector magnitude penalty for a large reduction in required bandwidth. Modulations that avoid the origin by construction, such as offset quadrature phase shift keying and the pi-over-four rotated variants, are correspondingly easier to transmit in polar form.

Modern Practice

A supply-modulated amplifier also changes its phase shift as the supply changes, so amplitude modulation leaks into output phase. That amplitude-to-phase characteristic must be measured and compensated in the phase path, which means a polar transmitter needs predistortion of its own even though its amplifier sees a constant envelope.

Polar techniques found their commercial home in narrowband standards, where the bandwidth expansion is affordable. Enhanced data rates for global evolution, the amplitude-modulated extension of the global system for mobile communications, was widely implemented in polar form, as were many Bluetooth and short-range radios. The modern realization is usually digital: the amplitude word switches a bank of unit amplifier cells on and off, so output amplitude is set by the number of active cells rather than by an analog supply. Such a digital polar transmitter suits a fine-geometry logic process, since it consists largely of switches and logic, but it introduces a new problem, because the amplitude is now a staircase and both quantization noise and cell-to-cell mismatch set the achievable noise and spurious floor.

Outphasing and LINC Transmitters

Outphasing reaches the same destination by a different route. It splits the signal into two constant-amplitude components whose vector sum reproduces the wanted envelope. Both branches run at the maximum amplitude with a phase difference chosen so that their sum has the wanted magnitude, while their average phase carries the wanted phase. In phase, the sum is maximal; as the outphasing angle opens the sum shrinks, and at the extreme the branches cancel entirely.

Henri Chireix described the technique in "High Power Outphasing Modulation," published in the Proceedings of the IRE in November 1935, as a way to amplitude-modulate a broadcast transmitter without modulating the amplifier stages themselves. Donald Cox reintroduced it to the communications community in the 1970s as linear amplification with nonlinear components, producing the acronym LINC by which it is now generally known.

The Combiner Question

Both branch amplifiers see a constant-envelope drive, so both may be strongly nonlinear switching designs running at peak efficiency at all times. That would be a complete solution were it not for the combiner, in which the architecture's entire economics reside.

With an isolating combiner, such as a Wilkinson divider used in reverse or a hybrid with its isolated port terminated, each amplifier sees a constant load regardless of what the other does. Linearity is excellent and the design robust. But the power the branches do not deliver to the load does not disappear; it flows into the isolation resistor as heat. When the outphasing angle is large, which is to say when the wanted envelope is small, almost all the generated power is dissipated there. Averaged over a modulated signal, system efficiency falls in proportion to the ratio of average to peak power, which is precisely the penalty a backed-off linear amplifier suffers. An isolating combiner therefore converts a promising efficiency argument into no argument at all.

The alternative is the non-isolating combiner Chireix originally proposed, with compensating reactances in the two branches. Here the amplifiers are coupled, and as the outphasing angle changes each branch load-modulates the other. Chosen well, the compensating reactances keep both amplifiers efficient over the range of angles the signal actually visits, so the power an isolating combiner would have burned is never generated. The costs are equally real: the amplifiers see varying and partly reactive loads, so gain and phase vary with the outphasing angle and predistortion becomes mandatory; the compensating reactances are correct at one frequency, so the design is narrowband unless considerable effort goes into broadbanding it; and branch mismatch leaks the difference-mode signal to the output exactly as quadrature imbalance leaks a sideband image.

When Outphasing Pays

Outphasing pays when the combiner can be made efficient over the signal's actual envelope distribution and when the resulting nonlinearity is correctable. Multilevel outphasing switches the branch amplitude among a small set of discrete supply levels so that the outphasing angle never opens far, keeping the design where the combiner is efficient. Energy-recovering combiners rectify the power that would otherwise be dissipated and return it to the supply. Both add complexity, and the architecture has consequently remained more common in the literature and in specialist transmitters than in mass-market radios, where the Doherty amplifier and envelope tracking won the same argument with less exotic hardware.

Digital and RF-DAC Transmitters

Rising complementary metal-oxide-semiconductor speed has made it practical to synthesize the modulated carrier in the digital domain and convert it to analog form at or near the carrier frequency, eliminating the analog quadrature modulator and sometimes the mixer entirely. Such designs, variously called digital transmitters, direct radio frequency synthesis, or RF-DAC transmitters, are the transmit counterpart of the direct-sampling receiver.

Forms

The simplest form computes the real passband waveform digitally and presents it to a converter clocked at several gigasamples per second, followed by a reconstruction filter, a driver, and the power amplifier. This is standard in modern instrumentation and infrastructure radios, and integrated devices combining converters of this class with programmable logic, such as the radio frequency system-on-chip families offered by the major programmable-logic vendors, have brought it within reach of medium-volume systems. A second form uses the converter itself as the mixer: in a mix-mode or return-to-zero converter the output waveform is shaped so that its spectrum is strongest in a higher Nyquist zone, placing usable power at a multiple of the sampling clock plus or minus the baseband frequency. A third form dispenses with a conventional converter and builds the amplifier from switched unit cells driven directly by digital words, in polar form with an amplitude word and a phase-modulated clock, or in Cartesian form with two orthogonal cell banks. The output stage is then simultaneously converter, modulator, and amplifier.

Noise and Image Consequences

Digital synthesis moves the impairments from analog mismatch to sampling and quantization. Quantization noise comes first: a converter spreads it across the Nyquist band, and part of that band contains the transmitter's own receive band and the bands of other services. A duplex filter attenuates only by its stopband rejection, so the transmitter's noise density at the duplex offset, quoted in decibels relative to the carrier per hertz, is often the specification that decides whether a design works at all. It drives converter resolution, sampling rate, and any noise shaping applied to move quantization energy away from sensitive offsets.

Sampling images come second. The converter reproduces the signal at every multiple of the sampling frequency plus or minus the baseband frequency, and each image is a full replica of the modulation. The zero-order hold imposes a sinc-shaped envelope that attenuates higher zones, which helps in the first Nyquist zone and hurts when the design deliberately works in a higher one. Either way an analog reconstruction filter must suppress the unwanted zones to the level the spurious limits demand, and its requirements set a floor under the sampling rate.

Clock jitter comes third, translating directly into carrier phase noise, with a penalty that grows with carrier frequency because a fixed timing error is a larger fraction of a shorter period. Finally, mismatch among nominally identical converter elements or amplifier cells produces harmonic and intermodulation spurs. Dynamic element matching, which scrambles the mapping between codes and physical elements from sample to sample, converts those tones into a raised but smooth noise floor, usually the more acceptable failure. In exchange, the digital transmitter offers reconfigurability, multiband operation without duplicated analog hardware, and correction implemented in logic that shrinks with each process node.

Peak-to-Average Power Ratio and Crest Factor Reduction

Everything in the second half of transmitter design follows from one number: the ratio of a signal's peak power to its average power. An amplifier has a definite maximum output beyond which it clips, so reproducing peaks faithfully requires the average output to sit below that maximum by at least the peak-to-average power ratio. The amplifier consequently spends nearly all its time delivering far less than the power it was built for.

The penalty is steep, because efficiency falls with back-off. An idealized class A stage draws constant current regardless of output, so its efficiency falls in direct proportion to output power: a peak of fifty percent becomes five percent at ten decibels of back-off. An idealized class B stage draws current in proportion to output voltage, so its efficiency falls with the square root of output power, and its theoretical peak near seventy-eight and a half percent becomes roughly twenty-five percent at the same back-off. Real devices do worse than the ideal.

Modulation sets the number. Filtered single-carrier modulations are benign, with crest factors under two decibels for pulse-shaped quadrature phase shift keying and under four for filtered sixty-four-state quadrature amplitude modulation. Multicarrier and spread-spectrum signals are far worse, because a sum of many independent components approaches a Gaussian distribution whose peaks are unbounded in principle; orthogonal frequency division multiplexing is commonly quoted near twelve decibels and wideband code division multiple access at about ten and a half decibels on the downlink. Because the peaks are statistical, the relevant figure is the level exceeded with some small probability, read from a complementary cumulative distribution function.

The problem shapes standards, not only circuits. The choice of discrete Fourier transform spread orthogonal frequency division multiplexing on the uplink of long term evolution, and as an option in the fifth-generation new radio, rather than the plain orthogonal frequency division multiplexing used on the downlink, exists specifically to lower the handset's peak-to-average ratio by a few decibels and buy coverage at the cell edge. The base station, which can afford a larger and better-cooled amplifier, keeps the higher-ratio waveform.

Removing the Peaks

Crest factor reduction is the family of digital techniques that deliberately distorts the waveform to lower its peak-to-average ratio before it reaches the amplifier. Hard clipping is cheapest and worst behaved, since a limiter generates broadband intermodulation that violates adjacent-channel requirements at once; clipping is therefore followed by filtering, which restores some peaks, so the process iterates. Peak windowing multiplies the envelope by a smooth window centered on each peak instead of truncating it, concentrating the error inside the channel where it costs error vector magnitude rather than outside where it costs adjacent-channel leakage. Peak cancellation, the most refined common method, subtracts a scaled, phase-rotated cancellation pulse whose spectrum is confined by construction to the transmitted carriers, so the correction adds no out-of-band energy at all.

Whatever the method, the currency is the same: crest factor reduction converts amplifier back-off into error vector magnitude. An operator with an eight percent budget for a sixty-four-state constellation might spend two or three percentage points on crest factor reduction and gain two or three decibels of usable output, which is a large efficiency gain; spending more begins to eat the demodulation margin the link relies on. Ordering matters, because predistortion expands peaks in order to compensate for compression and will partially undo the reduction. Practical chains therefore set the crest factor target with the predistorter's expansion already budgeted.

Digital Predistortion

Digital predistortion attacks the problem from the other side. Rather than reducing the demands on the amplifier, it applies the inverse of the amplifier's nonlinearity to the digital baseband signal, so the cascade of predistorter and amplifier is linear even though the amplifier alone is not. The amplifier can then run much closer to saturation and still meet spectral requirements. Predistortion has displaced the analog linearization techniques that preceded it and is now standard in essentially every base station.

Memory Effects

The simplest predistorter treats the amplifier as memoryless, describing it by two curves: amplitude-to-amplitude, giving output magnitude against input magnitude, and amplitude-to-phase, giving output phase shift against input magnitude. Inverting both and storing them as lookup tables indexed by instantaneous envelope magnitude works well for narrowband signals.

Real amplifiers have memory, and the memoryless model fails as bandwidth grows. Electrical memory arises from the impedance the bias networks present at envelope frequencies and from the frequency-dependent group delay of the matching networks, so the operating point at a given instant depends on the envelope nanoseconds earlier. Thermal memory arises from self-heating, since junction temperature follows the envelope with time constants from microseconds to milliseconds and gain depends on temperature. The two present differently: electrical memory produces asymmetry between the upper and lower adjacent channels, while thermal memory produces a slow drift the adaptation loop must chase.

Predistorters therefore use models with memory. The Volterra series is the general representation of a nonlinear system with memory, but its coefficient count grows explosively with order and depth, so practical designs prune it. The memory polynomial keeps only the diagonal terms, expressing the output as a sum over powers of the envelope magnitude multiplied by delayed inputs. The generalized memory polynomial adds cross terms in which the envelope magnitude at one delay multiplies the signal at another, capturing lagging and leading envelope effects. Neural-network predistorters replace the fixed basis with a learned one and have appeared more recently in commercial equipment.

The Observation Receiver

No predistorter can be built without measuring what the amplifier actually does, so every adaptive system includes a feedback path: a directional coupler at the output, an attenuator, a downconverter, and an analog-to-digital converter that returns the transmitted signal to the digital domain for comparison with the reference. Its requirements are demanding in three ways.

Its bandwidth must cover not only the transmitted channel but the spectral regrowth on either side, since the regrowth is what the predistorter is trying to cancel; three to five times the signal bandwidth is a common requirement, which for a wideband carrier is a considerable specification in its own right. Its linearity and noise must be better than those of the corrected transmitter, because any distortion it generates is indistinguishable from the amplifier's and will be dutifully learned into the model. And the reference and the observation must be aligned in time to a small fraction of a sample, and in carrier phase and gain, before the parameter estimation converges. Estimation is usually performed by an indirect learning architecture, in which the model is identified as the post-inverse of the amplifier from the observed output and the coefficients are copied into the pre-inverse.

Limits and Alternatives

Predistortion cannot recover information destroyed by hard clipping, because saturation is not invertible and beyond peak power there is no output to command. It cannot follow an amplifier whose characteristic changes faster than the adaptation loop tracks. It consumes digital power and silicon area, which is why handsets often use reduced forms or none, and why handset specifications are written looser than base station ones. And it interacts with everything upstream, since crest factor reduction, supply modulation, and bias all change the characteristic it must invert.

Two older architectures survive where digital predistortion does not fit. Cartesian feedback downconverts the output, compares it with the baseband reference in the in-phase and quadrature domain, and closes an analog loop around the whole transmitter; it is elegant and self-correcting, but loop stability confines it to a few hundred kilohertz of bandwidth and therefore to narrowband professional and land mobile radio. Feedforward, once universal in third-generation base stations, extracts an error signal by subtracting a scaled sample of the input from a sample of the output and amplifies that error separately to cancel the distortion. It is broadband and needs no significant adaptation, but the error amplifier and output coupler waste enough power that digital predistortion displaced it as soon as digital signal processing became cheap enough.

Supply Modulation: Envelope and Average Power Tracking

Predistortion makes a backed-off amplifier linear; it does not make it efficient. A backed-off amplifier wastes power because its supply is set for the peak while its output swing is usually far below the peak, so the difference between the rail and the actual device voltage is dropped across the device as heat. Supply modulation removes that waste by lowering the supply when the signal is small.

Envelope Tracking

Envelope tracking varies the supply continuously so that it follows the instantaneous envelope, keeping the device near compression at all output levels. Unlike a polar transmitter, the amplifier still receives the full modulated signal at its input; the supply merely rides along to eliminate wasted headroom, so the stage remains a linear amplifier and predistortion remains available.

The supply is not commanded to equal the envelope exactly. A shaping table maps envelope magnitude to supply voltage along a trajectory chosen to keep the amplifier's gain constant, since a supply that tracked the envelope literally would push the device into compression at low levels and produce severe amplitude-to-amplitude and amplitude-to-phase distortion. Designers speak of an iso-gain trajectory, derived from measured gain contours against supply voltage and output power.

The hard part is the supply modulator, which must deliver amperes at envelope rates with high efficiency of its own, since system efficiency is the product of the amplifier and modulator efficiencies. The standard solution is a hybrid: a switching converter supplies the low-frequency, high-current bulk at high efficiency while a linear amplifier supplies the high-frequency residue, and the two currents sum at the output. Bandwidth is the wall. The envelope is wider than the signal, for the same reason the polar amplitude path is wider, so a modulator for a channel of a given width must handle several times that width. As carrier bandwidths grew from the few megahertz of third-generation systems to the tens and hundreds of megahertz of the fifth generation, the modulator became the bottleneck.

Average Power Tracking

Average power tracking is the pragmatic retreat. The supply is set from the average transmit power over a slot or a transmission and held constant within it. Almost all of the benefit at low average power survives, because the largest single waste in a handset transmitter is running from a full rail while transmitting at a fraction of maximum power, which is the usual condition; what is given up is the instantaneous saving within a single high-power burst. Because the supply changes only at slot boundaries, the modulator needs bandwidth measured in tens or hundreds of kilohertz rather than hundreds of megahertz, and an ordinary switching converter suffices. Modern handset front ends implement both modes and choose between them at run time according to bandwidth and power level.

Doherty Amplifiers and Load Modulation

Supply modulation improves efficiency by changing the voltage the device works against. Load modulation improves it by changing the impedance, and it has the advantage of requiring no fast, high-current auxiliary supply.

The Classical Doherty

William Doherty described the arrangement that bears his name in "A New High Efficiency Power Amplifier for Modulated Waves," published in the Proceedings of the IRE in September 1936. Two amplifiers share the load. The main, or carrier, amplifier is biased for linear class AB operation and runs at all levels. The peaking, or auxiliary, amplifier is biased in class C and stays cut off until the drive exceeds a threshold. The two combine through a quarter-wavelength line acting as an impedance inverter.

At low levels the peaking device is off and presents a high impedance to the combining node, so the inverter makes the main device see a higher resistance than its full-power optimum. The main device therefore reaches maximum voltage swing, and peak efficiency, at an output well below the pair's maximum, which is the first efficiency peak. As the drive passes the peaking device's turn-on point, that device injects current into the combining node, raising the apparent impedance there, and the inverter correspondingly lowers the impedance the main device sees, letting it deliver more current at the same voltage swing. The main device stays saturated while total output rises to the maximum, where the second peak occurs. A symmetric Doherty built from two identical devices places its back-off efficiency peak six decibels below saturation.

Asymmetric, Multi-Way, and Balanced Variants

Signals with peak-to-average ratios of eight or nine decibels want the back-off peak deeper than six decibels. An asymmetric Doherty uses a peaking device larger than the main device, shifting the peak further down at the cost of a harder driving problem. Three-way and N-way designs cascade the principle with additional peaking stages, flattening average efficiency over a wider range at considerable cost in complexity, bandwidth, and the difficulty of turning every stage on at the right moment.

Two limitations constrain the family. The quarter-wave inverter is correct at one frequency, so the classical Doherty is inherently narrowband. And the class C peaking device turns on gradually, so the composite characteristic has a pronounced kink and a gain that expands before it compresses. The Doherty is therefore not a linear amplifier; it is an efficient amplifier that digital predistortion makes linear, and the two techniques arrived in base stations together. Dual-input, or digital, Doherty designs drive the main and peaking paths from separate chains so that the relative amplitude and phase of the two drives can be optimized continuously rather than fixed by a passive splitter.

The load-modulated balanced amplifier, first reported by researchers at Cardiff University in 2016, attacks the bandwidth limitation directly. It begins with a conventional balanced pair, two devices driven in quadrature through a hybrid coupler and recombined through a second one, a configuration that is inherently broadband and well matched. The innovation is to inject a signal from a third control amplifier into the isolated port of the output coupler. That injection changes the impedance both balanced devices see, and it does so across the coupler's full bandwidth rather than at a single design frequency, so controlling the amplitude and phase of the injected signal load-modulates the pair over a fractional bandwidth a classical Doherty cannot reach.

Transmitter Specifications and What Each Protects

A transmitter is judged against a list of numbers that seem arbitrary until one asks what each protects. Every one exists because somebody is harmed when it is violated.

Error Vector Magnitude

Error vector magnitude measures the transmitter's own accuracy. Each transmitted symbol is compared with the ideal constellation point, the magnitude of the difference vector is computed, and the result is normalized and expressed as a percentage or in decibels. It aggregates every impairment that lands inside the channel: carrier leakage, quadrature imbalance, phase noise, residual compression after predistortion, quantization noise, and crest factor reduction error.

What it protects is the link's own demodulation margin. It behaves as a noise floor the receiver cannot escape, so it caps the signal-to-noise ratio the link can achieve regardless of transmit power, and therefore caps the usable constellation order. The 3GPP specifications for the fifth-generation new radio make the relationship explicit by setting a limit per modulation order: seventeen and a half percent for quadrature phase shift keying, twelve and a half percent for sixteen-state quadrature amplitude modulation, eight percent for sixty-four-state, and three and a half percent for two-hundred-fifty-six-state. The last corresponds to a signal-to-distortion ratio of about twenty-nine decibels, the margin a dense constellation needs before the receiver's own noise is even considered.

Adjacent Channel Leakage Ratio

Adjacent channel leakage ratio, sometimes written as adjacent channel power ratio, is the ratio of power in the assigned channel to power spilling into a neighboring channel, both measured through a specified filter over a specified bandwidth. It is the primary measure of spectral regrowth, and it is the number amplifier nonlinearity moves most directly, since third-order and fifth-order intermodulation among the components of a modulated signal lands immediately outside the occupied band. What it protects is the operator on the next channel, who may be a competitor with a base station on the same tower. 3GPP requires at least forty-five decibels for new radio base stations; handset requirements are considerably looser, because a phone has neither the supply voltage, the die area, nor the thermal budget of a base station.

Spectral Emission Mask and Spurious Emissions

Where adjacent channel leakage ratio is one integrated number, the spectral emission mask is a shape: an absolute limit on emitted power in a specified measurement bandwidth, plotted against offset from the channel edge, beneath which the measured spectrum must lie everywhere. The mask catches what the integrated ratio averages away, such as a narrow spur just outside the channel edge that contributes little to a wideband integration but is intolerable to a narrowband neighbor. Masks also encode regulatory geography, tightening sharply at a band edge, so a transmitter on the outermost channel of its band faces a harder problem than the same transmitter in the middle.

Spurious emissions are unwanted outputs far from the carrier: harmonics, mixer products, oscillator leakage, converter images, digital clock harmonics that reach the antenna through the supply or substrate, and switching noise from a supply modulator. Limits are absolute, specified in power per measurement bandwidth across a wide sweep, with much tighter values inside the receive bands of other services and inside protected allocations such as radio astronomy and aeronautical safety bands. Harmonics deserve particular vigilance because arithmetic conspires against designers: the second harmonic of a transmitter in the upper seven-hundred megahertz range falls very close to 1575.42 MHz, the L1 carrier of the global positioning system, which is very likely receiving in the same handset.

Transmit and Passive Intermodulation

The final specification concerns what happens when another transmitter's signal finds its way into this one's output. Energy entering the antenna port from a nearby transmitter reaches the final device, mixes with the wanted carrier in that device's nonlinearity, and radiates back out as products at frequencies neither transmitter occupies. 3GPP specifies this as a transmit intermodulation requirement, injecting an interfering signal at a defined level and offset and limiting the resulting products. It matters most at co-sited installations where several operators share a structure.

Passive intermodulation is the same phenomenon generated without any active device. Metal-to-metal junctions with thin oxide layers, loose or corroded connectors, ferromagnetic hardware in the signal path, and rusted structures near the antenna all behave as weak nonlinearities, and at base station power levels weak is sufficient. A third-order product from two downlink carriers can land squarely in the uplink band, desensitizing the receiver of the very site that generated it, and because the source is mechanical the fault is invisible to electrical design review. The remedies are mechanical: non-ferrous plated connectors, controlled torque, clean junctions, avoidance of dissimilar metals, and separation from suspect structures.

Duplexing, Isolation, Ruggedness, and the Power Budget

The architecture and its linearization sit inside a system whose remaining constraints are no less binding.

Duplexing and Transmit Noise

A radio that transmits and receives on separate frequencies simultaneously needs a duplexer, a pair of filters sharing an antenna port that passes each band in its direction and isolates the two. A radio that alternates in time needs only a switch, which is smaller, cheaper, and lower in loss.

The demanding requirement in frequency division duplex is not rejection of the transmit carrier, which is strong but narrow and which a good filter handles. It is rejection of the transmitter's broadband noise floor at the receive-band offset. Every stage in the transmit chain contributes noise across a wide bandwidth, and a portion of it falls exactly where the receiver is listening, inside the receiver's own passband, where no receiver design helps. The transmitter's noise density at the duplex offset therefore becomes a hard specification, quoted in decibels relative to the carrier per hertz, and it constrains converter resolution, synthesizer phase noise far from the carrier, amplifier noise figure, and duplexer stopband depth together. Handsets rely on acoustic technologies, chiefly surface acoustic wave, temperature-compensated surface acoustic wave, and bulk acoustic wave devices, for skirts steep enough in a package small enough; base stations use cavity filters, which are large but low in loss and tolerant of power.

Time division duplex avoids the noise problem and substitutes timing problems. The switch must operate within a guard period measured in microseconds, the transmit power must ramp within an on-off time mask so the switching transient does not itself violate spectral requirements, and the receiver must survive and recover from whatever leakage reaches it during transmit. Full-duplex operation on a single frequency remains largely a research architecture, because the self-interference exceeds the wanted signal by a hundred decibels or more and suppressing it demands antenna isolation, analog cancellation, and digital cancellation stacked together.

Load Mismatch and Ruggedness

An antenna is not a fixed fifty-ohm load. A hand covering a phone, a nearby metal surface, ice on a mast, or a damaged feeder changes the impedance the amplifier drives, and the reflection sets up a standing wave. Because the phase of the reflection depends on the electrical length to the discontinuity, the amplifier may see any impedance around a circle of constant reflection coefficient on the Smith chart. Some points raise the device's peak voltage beyond its nominal value, risking breakdown; others raise the current, risking thermal failure. A requirement to survive a total mismatch, an open or short circuit at any phase, at full rated power is common in infrastructure equipment.

Protection takes several forms. An isolator or circulator absorbs the reflection before it reaches the device, standard in base stations and impractical in handsets on grounds of size and loss. A directional coupler with a reflected-power detector lets the control loop fold back drive as the mismatch worsens, trading output power for survival, and bias clamps limit the excursion directly. Device technology contributes as well: the high breakdown field of gallium nitride is one reason gallium nitride amplifiers tolerate mismatch that would destroy a comparable silicon laterally diffused metal-oxide-semiconductor device.

Thermal and Power Budget

Whatever direct-current power does not leave as radio frequency power leaves as heat. A transmitter delivering forty watts at thirty percent efficiency draws about one hundred and thirty-three watts and dissipates about ninety-three, and each of those watts must cross a thermal path from junction to case to heatsink to air without letting the junction exceed its rating. Semiconductor failure mechanisms accelerate sharply with temperature, so junction temperature margin is reliability margin, and derating rules for mean time to failure are written in those terms.

The constraint expresses itself differently by application. A remote radio unit on a mast is sealed against weather and cooled by conduction to a finned enclosure, which caps dissipation and therefore caps output power for a given efficiency. A handset has no heatsink but its own case, whose surface temperature is limited by what a user tolerates, so a phone transmitting at maximum power throttles its own duty cycle. A satellite transmitter radiates to space through a fixed radiator area with no margin to spare. These constraints are why efficiency is not merely an engineering preference: in a mobile network the radio access equipment consumes a large share of the operator's electricity and the power amplifier a large share of that, so a few percentage points of amplifier efficiency, aggregated across many sites, appears in financial statements.

Array Transmitters and Per-Element Calibration

Millimeter-wave systems and modern massive multiple-input multiple-output base stations do not have one transmitter; they have many, one behind each radiating element, and the interesting engineering moves from the individual chain to the relationships among them.

A coherent array concentrates its energy in a beam, and effective isotropic radiated power rises faster than total transmitted power, because the array both sums the element powers and focuses them into a narrower solid angle. Per-element power can therefore be small: a millimeter-wave array reaches a useful link budget with element amplifiers of tens or hundreds of milliwatts, a level integrated silicon or silicon-germanium devices deliver comfortably, which is one reason such arrays became economically feasible. Distributed generation also distributes the heat.

Beamforming Partitions

Where the beam is formed determines how many transmit chains are needed. Analog beamforming places phase shifters, and usually variable gain, at radio frequency after a single upconversion, so one modulator and one converter feed the array and it forms one beam at a time. Digital beamforming gives every element its own converter and upconverter, allowing arbitrary and simultaneous multiple beams at a cost in converters, interconnect bandwidth, and power that scales with element count. Hybrid beamforming partitions the array into subarrays, each with a digital chain and internal analog phase shifting, and is the usual compromise in commercial millimeter-wave equipment.

Calibration Becomes the Hard Part

An array steers by controlling the relative amplitude and phase of its elements, so beam accuracy is exactly the accuracy of those relative quantities. Manufacturing tolerance, temperature gradients across a panel, differences in feed length, and gain variation among nominally identical amplifiers all corrupt them, and errors that would be trivial in a single chain raise sidelobes, fill nulls, mispoint the beam, and reduce array gain. Calibration is therefore a prerequisite, not a refinement. A built-in coupling network can sample each element and route it to a common measurement receiver; mutual-coupling calibration dispenses with the extra network by transmitting on one element and receiving on its neighbors; and factory characterization over frequency and temperature, stored and interpolated at run time, handles the systematic component. Most production systems combine a stored characterization with a periodic in-service check.

Two further effects make array transmitters genuinely different rather than merely repetitive. Mutual coupling means the impedance each element presents changes as the beam steers, so every amplifier is load-modulated by the beamforming itself and its gain, phase, and efficiency vary with scan angle; a design must either desensitize the amplifier to that variation or track it in calibration. And distortion beamforms too, since intermodulation products generated coherently in every element combine in space as the wanted signal does, so the array radiates a distortion pattern whose shape depends on how the nonlinearity correlates across elements. Per-element predistortion is expensive, so array-level schemes that observe a combined far-field or coupled sum and correct all elements together are an active area of product development.

Because an active antenna array has no single connector where the composite signal exists, the specifications themselves change. 3GPP defines over-the-air requirements for such equipment, declaring error vector magnitude, adjacent channel leakage, and spurious emissions as radiated quantities in specified directions rather than as conducted power at a port. Testing moves into anechoic chambers and compact antenna test ranges, and the measurement uncertainty budget becomes a serious part of the compliance argument.

Selecting a Transmitter Architecture

The choice follows from the modulation's peak-to-average ratio, the required spectral purity, the power level, and the degree of integration the product demands. High-power infrastructure and instrumentation still favor heterodyne chains with Doherty output stages and digital predistortion. Integrated handset and consumer transceivers favor direct conversion with supply modulation, accepting calibrated impairments for silicon economy. Narrowband, high-efficiency applications favor polar and digital polar designs. Millimeter-wave systems favor arrays of modest elements in which calibration replaces per-element perfection.

In practice the categories combine. A modern base station is a heterodyne or digital transmitter with crest factor reduction, digital predistortion, and a Doherty output stage, replicated per element behind an active antenna. A modern handset is a direct-conversion transmitter with supply modulation, modest predistortion, and a stored calibration table for carrier leakage and quadrature imbalance. The architecture names describe the frequency plan; the engineering effort goes into the efficiency and linearity subsystem wrapped around it.

Summary

Transmitter architecture decides where modulation occurs, how many translations separate baseband from the carrier, and how the final amplifier is kept efficient without becoming intolerable to its neighbors. The two-step heterodyne transmitter modulates at a fixed intermediate frequency, where impairments are constant and where both the transmit image and the oscillator leak fall away from the carrier and can be filtered, at the price of extra hardware and a longer spur list. Direct conversion collapses the chain into a single quadrature modulation at the carrier, buying integration in exchange for three impairments that sit on top of the signal: carrier leakage from baseband offsets and oscillator coupling, a sideband image set by gain and phase imbalance, and oscillator pulling by the power amplifier. Each has an answer, respectively a calibrated offset correction, a digital correction matrix that must become frequency-selective as bandwidth grows, and the architectural decision to run the oscillator at a frequency the antenna path does not contain.

Polar transmitters split amplitude from phase so that a nonlinear amplifier can handle a constant-envelope drive, and they are limited by the delay that must be matched between the paths and by the bandwidth expansion the polar transformation causes. Outphasing splits the signal into two constant-envelope branches, and its viability rests entirely on the combiner: an isolating combiner burns the difference-mode power and surrenders the efficiency argument, while a Chireix combiner recovers it by letting the branches load-modulate each other, at the cost of nonlinearity and narrow bandwidth. Digital and RF-DAC transmitters synthesize the carrier outright, trading analog mismatch for quantization noise at the duplex offset, sampling images, and clock jitter.

The efficiency and linearity subsystem is where most transmit engineering happens. Peak-to-average power ratio forces back-off, back-off destroys efficiency, and the countermeasures divide into reducing the peaks, correcting the distortion, and changing the amplifier's operating conditions. Crest factor reduction buys back-off with error vector magnitude. Digital predistortion inverts the amplifier's characteristic, needs a model with memory as bandwidth grows, and needs an observation receiver several times wider than the signal whose own linearity exceeds the corrected transmitter's. Envelope tracking and average power tracking lower the supply when the signal is small, the first following the instantaneous envelope until modulator bandwidth becomes the wall, the second retreating to per-slot adjustment. Doherty amplifiers and their asymmetric, multi-way, and load-modulated balanced relatives change the impedance instead of the voltage, and depend on predistortion to be linear at all.

The specifications that judge the result each protect somebody: error vector magnitude protects the link's own demodulation margin, adjacent channel leakage ratio and the spectral emission mask protect neighbors in frequency, spurious emission limits protect services in unrelated bands, and transmit and passive intermodulation requirements protect co-sited equipment. Around all of it sit the constraints no architecture escapes: duplexer isolation and the transmit noise floor at the receive-band offset, harmonic filtering, survival into a mismatched antenna, and a thermal budget that converts every point of lost efficiency into heat somebody must remove. In array transmitters all of these persist per element, and the relationships among elements, established and maintained by calibration, become the design's hardest problem.

Related Topics