Squeezing Clean Power From a Hungry Amplifier Without Backing Off the Drive

Every operator who has watched an SSB transmitter splatter across the band knows the uncomfortable trade behind it. Push the power amplifier toward its rated output and the spectrum smears into the neighbors. Pull the drive back into the safe linear zone and half the gain evaporates, the heat sink runs cool, and the rig delivers a fraction of what the final could give. The conventional wisdom says you pick one side of that bargain. Envelope feedback quietly refuses the choice, and it does so with parts most builders already have on the bench.

The idea is old enough to predate solid-state finals, yet it keeps resurfacing because the physics never changed. A power amplifier distorts because its gain is not constant across the output swing. Near the top of the range the device starts to compress, the transfer curve bends, and the bend manufactures intermodulation products that land squarely in the adjacent channel. Envelope feedback watches the actual amplitude coming out, compares it to the amplitude going in, and corrects the difference in real time. The correction loop works on the shape of the modulation rather than the radio frequency carrier, which is exactly what makes it practical at high power.

Why ordinary negative feedback collapses at radio frequencies

The instinct of any analog designer faced with distortion is to wrap a negative feedback loop around the offending stage. At audio that instinct is correct and nearly automatic. At 14 MHz or 144 MHz it falls apart. Feedback works by sending a sample of the output back to the input with inverted phase, so the loop subtracts the error. The catch is phase. A signal travels through the amplifier, through the sampling network, and back to the summing point, and that journey takes time.

The phase shift accumulated around the loop depends on that delay and the frequency. For a propagation delay td around the loop, the phase rotation at frequency f is

phi = 360 f td (degrees)

Put numbers to it. A loop physical path of just 10 centimeters, with signal velocity roughly two thirds the speed of light, gives a delay of

td = 0.10 / (0.67 * 3e8) = 5.0e-10 s = 0.5 ns

At 1 MHz that delay rotates the returning signal by

phi = 360 1e6 0.5e-9 = 0.18 degrees

which is utterly harmless. At 144 MHz the same path rotates it by

phi = 360 144e6 0.5e-9 = 25.9 degrees

and that is only the wiring. Add the transistor's own internal delay and the storage time of the devices, and the total easily reaches the 180 degrees at which negative feedback becomes positive feedback and the amplifier oscillates. The Barkhausen condition for oscillation is that the loop gain magnitude reaches or exceeds unity at the frequency where the total phase shift hits 360 degrees of inversion. The loop gain a designer needs for distortion reduction is the same loop gain that drives the circuit into that condition, which is why direct radio frequency feedback is a tightrope few amateur builders walk twice.

Envelope feedback sidesteps the whole problem with one move. It does not feed the radio frequency signal back at all. It detects the envelope, the slow amplitude variation riding on the carrier, and feeds that back instead. A voice SSB envelope occupies only a few kilohertz. Recompute the phase rotation at the envelope band edge, say 3 kHz, for the same 0.5 ns path:

phi = 360 3e3 0.5e-9 = 0.00054 degrees

The phase budget is effectively infinite. A correction loop operating over a few kilohertz behaves like a tame audio circuit, carries large loop gain, and never threatens to oscillate from wiring delay.

How the detector and the control loop flatten the gain curve

The architecture follows once the principle clicks. Two envelope detectors do the sensing. One samples a portion of the radio frequency drive at the amplifier input and recovers its envelope. The second samples the output through an attenuator and recovers that envelope. A comparison amplifier looks at the two and produces an error voltage proportional to how far the output amplitude has strayed from a faithful scaled copy of the input.

The loop forces the amplifier gain to equal the inverse of the feedback attenuation. Call the attenuation in the feedback path beta and the raw amplifier gain A. Closed-loop gain follows the classic feedback relation

Gcl = A / (1 + A * beta)

When the loop gain A*beta is large, this collapses to

Gcl = 1 / beta

so the overall gain is set by a stable passive attenuator rather than by the temperamental transfer curve of a power transistor. The distortion reduction follows the same factor. If the open-loop gain varies by some amount due to compression, the closed loop divides that variation by (1 + A*beta):

(distortion out) = (distortion open-loop) / (1 + A * beta)

Expressed in decibels, a loop gain of L dB reduces the gain ripple by L dB.

A worked example of the distortion improvement

Walk one final through the arithmetic. Suppose the gain compresses by 1.5 dB between the quiescent point and peak envelope power. In linear terms a 1.5 dB ripple is a gain ratio swing of

10^(1.5/20) = 1.189

so the gain wanders about 19 percent across the swing. Left uncorrected, a gain ripple of this size drives third order intermodulation products to roughly 25 to 28 dB below the tones, a level that earns complaints.

Now apply a loop with 15 dB of loop gain across the modulation band. The linear loop gain is

A * beta = 10^(15/20) = 5.62

The residual gain ripple shrinks to

1.5 dB / (1 + 5.62) = 1.5 / 6.62 = 0.227 dB

A quarter of a decibel of residual ripple drops the third order products into the 40 dB region. The improvement in decibels is exactly the loop gain:

improvement = 20 log10(1 + Abeta) = 20 * log10(6.62) = 16.4 dB

Reported laboratory results for envelope and related feedback methods land in exactly this neighborhood, with adjacent channel improvements on the order of 10 to 15 dB.

The efficiency dividend that comes for free

Here is the part that turns a distortion fix into a genuine bargain. The usual way to clean up an amplifier is backoff, operating well below the compression point. Run a final 6 dB below rated output and the spectrum cleans up, but efficiency craters because the device idles in its least efficient region while delivering a quarter of the power. The drain efficiency of a class B stage scales with the output voltage swing. Relative to the peak efficiency eta_max near saturation, the efficiency at a backoff of b dB falls roughly as

eta(b) = eta_max * 10^(-b/20)

so 6 dB of backoff multiplies efficiency by

10^(-6/20) = 0.50

cutting it in half. Envelope feedback lets the amplifier run close to compression because the loop cleans up the distortion that compression would create, so the device stays where it is efficient. Published measurements on a power feedback variant at 850 MHz captured the effect: for the same adjacent channel interference, efficiency climbed from 35 percent to 48 percent once the correction loop was added, a thirteen point gain with no change in spectral cleanliness. The gain is not lost because the loop corrects the gain rather than reducing it.

Where the loop bites back and how builders tame it

The first limit is detector matching. The two detectors must track across temperature and level, because any mismatch shows up as residual distortion. A diode detector has a curved law near low levels; its output relates to input roughly as

Vout = Vin^2 / (Vin + Vt)

approaching square-law at small signals and linear at large ones. If the input and output detectors sit on different parts of that curve, the loop corrects toward a wrong target. Builders use matched pairs on a common thermal mass so both drift together.

The second limit is loop bandwidth versus stability. The detectors and comparison amplifier add their own delay, and the loop must roll off before its phase reaches inversion. The compensation network sets a dominant pole at

fp = 1 / (2 pi R * C)

and a practical rule ties this below the lowest modulation frequency the amplifier must pass, so

1 / (R C) < 2 pi * f_mod_min

This keeps the loop from chasing components it cannot faithfully correct and from injecting memory effects that themselves create distortion.

The third consideration is that envelope feedback corrects amplitude but not phase. A final near compression distorts both amplitude and phase, a coupling called AM-to-PM conversion. A pure envelope loop straightens amplitude and leaves the phase error in place. For voice this is usually tolerable; for demanding digital modes it becomes the limiting factor, which is why elaborate schemes add a phase path or move to Cartesian feedback that handles both axes.

Setting the loop gain with numbers rather than guesswork

Loop gain is the product of every gain and loss the error signal sees around the circle: the comparison amplifier gain Aerr, the gain control sensitivity Kc in dB per volt, the detector slope, and the feedback attenuation. The distortion reduction equals (1 + A*beta) as shown above, so a builder budgets the chain to hit a target loop gain. A comparison amplifier voltage gain of 20, feeding a control element delivering 1 dB of amplitude change per error volt, with unity-scale detectors, assembles a loop gain near

20 * log10(20) = 26 dB at low modulation frequencies

falling as the modulation climbs because the compensation pole and the detector both attenuate the faster components. The frequency dependence means correction is strongest where the envelope moves slowly and weakest on the sharpest transients, which are exactly what generate the widest splatter, so holding loop gain up across the full envelope band is what cleans the band edges. The practical optimum for an amateur build sits near 15 dB: generous enough to drop intermodulation by an order of magnitude, conservative enough to keep phase margin.

Comparing the families of correction so the choice is informed

Feedforward samples the distortion directly, amplifies it in a separate path, and subtracts it with a matched delay. It is unconditionally stable because nothing circles back, but it demands two amplifiers and exquisite delay matching, and it drifts with temperature because cancellation depends on amplitude and phase staying balanced to a fraction of a decibel. Predistortion distorts the input in the inverse shape of the amplifier curve. Analog predistortion is cheap and broadband but approximate; digital predistortion measures the output and computes the inverse continuously, reaching 20 dB and more, but it needs a processor, fast converters, and an adaptation loop. Envelope feedback threads between these. It is closed loop, so it self-corrects for drift, yet it operates only on the slow envelope, so it keeps the stability that radio frequency feedback throws away. The cost is that it corrects amplitude alone. For a voice SSB final built from simple, drift-tolerant parts, envelope feedback is the efficient choice; for demanding digital modulation where phase error dominates, the design must step up to a scheme that corrects both axes.

A technique that rewards the builder who measures

The honest verdict is that envelope feedback is neither magic nor obsolete. It is far simpler than digital predistortion and far more stable than direct radio frequency feedback, and it delivers real, measurable improvement a builder can verify on a spectrum analyzer in an afternoon. The path to a working loop runs through measurement. Characterize the gain compression first, because that number sets the loop gain needed. Measure the detector responses and match them. Watch the loop on a spectrum analyzer while sweeping the modulation, looking for the rise in noise floor that signals the loop edging toward instability.

The deeper lesson reaches past any single circuit. Distortion in a power amplifier is not a fixed property of the device. It is a behavior that can be observed, sampled, and corrected if the correction acts on the right variable at the right speed. Envelope feedback found that variable in the modulation envelope and that speed in the audio range, and in doing so it turned an apparently fundamental trade between power and cleanliness into an engineering problem with a tidy solution.