The basic two-transistor current mirror we covered previously has a dirty secret: its output current doesn't quite equal the reference current. Base currents siphon away a small but annoying fraction, and the output impedance is mediocre (just the Early-effect resistance r_o of one transistor). The Wilson current mirror fixes both problems with one extra transistor and a clever feedback trick.

The topology: Three matched BJTs. Q1 and Q2 form a standard diode-connected mirror pair at the bottom. Q3 sits on top, with its emitter tied to Q2's collector (which is also the mirror's output node before Q3), and its base tied to Q1's collector. Q3's collector is the actual output. The reference current I_REF flows into Q1's collector through a resistor from the supply.

Why it works: If the output current tries to rise, Q3's emitter current rises, which forces Q2's collector current up. That pulls Q1's V_BE reference down — but Q1 and Q2 are mirrored, so Q1's collector current must follow. Since I_REF is fixed by the external resistor, Q1's collector voltage drops, which drops Q3's base voltage, which throttles the output back down. Negative feedback regulates the output current to match the reference.

The numbers: The basic mirror has an error of roughly 2/β (about 1% for β=200). The Wilson mirror's error drops to approximately 2/β² — for β=200, that's 0.005%. The output impedance jumps from r_o (maybe 100 kΩ) to roughly βr_o/2 (around 10 MΩ). That's a 100× improvement in both metrics.

Real-world example: Precision IC bias networks in instrumentation op-amps like the AD8221 use Wilson mirrors (or their MOSFET cascode cousins) to set tail currents in input differential pairs. The high output impedance keeps the differential pair's common-mode rejection ratio (CMRR) above 100 dB even as the common-mode voltage swings across the supply rails. A plain mirror would let the tail current modulate with V_CM, destroying CMRR.

Rule of thumb: Use a Wilson mirror when you need better than 0.1% current matching or output impedance above 1 MΩ, and your headroom budget allows two V_BE drops (about 1.4 V) from output to ground. If headroom is tight, consider the improved Wilson variant with a fourth transistor that equalizes the V_CE across Q1 and Q2, eliminating systematic error from Early-effect mismatch.

One caveat: the loop has gain, so it can ring or oscillate with capacitive loads. Add a small compensation cap (10–100 pF) from Q3's base to ground if you see ringing on the output current.