The Lock-In Amplifier: Pulling a Signal Out of Pure Noise

The core trick: multiply, then average

Suppose the quantity you care about is encoded as a sinusoid V_sig(t) = V_s·sin(ω_r t + θ), where ω_r = 2πf_r is a reference frequency you control (you chopped a light beam, dithered a voltage, or drove a bridge with an oscillator). The lock-in generates its own clean reference V_ref(t) = sin(ω_r t) and forms the product in a mixer (the phase-sensitive detector, PSD):

• V_psd = V_s·sin(ω_r t + θ)·sin(ω_r t) = ½V_s·cos(θ) − ½V_s·cos(2ω_r t + θ)

The product-to-sum identity splits the result into a DC term (½V_s·cos θ) and a term at 2ω_r. Pass this through a low-pass filter with time constant τ and the 2ω_r component — and everything else not at exactly ω_r — is averaged away, leaving the DC term. That surviving DC voltage is proportional to the signal amplitude V_s and to the cosine of the phase difference θ.

Now feed in noise instead of signal. Broadband noise at frequency ω_n mixes to components at (ω_n − ω_r) and (ω_n + ω_r). Only noise within Δf of ω_r survives the low-pass filter; the rest is rejected. The lock-in has become a band-pass filter centered on ω_r with a bandwidth set entirely by the output filter — not by any resonant tank. That is the whole idea: the reference defines a razor-thin passband, and phase-sensitive detection slides it to exactly where the signal lives.

Two channels in quadrature: recovering X, Y, R and θ

The single-phase detector above has a problem: its output ½V_s·cos θ vanishes when the signal is 90° out of phase (θ = 90°), and it drifts if θ wanders. Every real lock-in solves this with a dual-phase architecture — two PSDs driven by references 90° apart:

• In-phase channel: multiply by sin(ω_r t) → X = ½V_s·cos θ
• Quadrature channel: multiply by cos(ω_r t) → Y = ½V_s·sin θ

From the two orthogonal outputs the instrument computes the phase-independent magnitude and the phase:

• R = √(X² + Y²) = ½V_s — independent of θ, so no phase-tuning needed
• θ = arctan(Y / X) — the signal's phase relative to the reference

This is exactly quadrature (I/Q) demodulation, the same math used in radio receivers. Because R is insensitive to phase drift, the dual-phase lock-in gives a rock-steady amplitude reading even when cable lengths, temperature, or the device under test shift the phase. The phase channel θ is itself a measurement — in impedance spectroscopy it separates the resistive (in-phase) from the reactive (quadrature) response of a capacitor or coating.

Why the noise vanishes: equivalent noise bandwidth

The noise rejection is governed by the output low-pass filter. Its equivalent noise bandwidth (ENBW) is what determines how much of the broadband noise spectrum leaks through. For a single-pole RC filter of time constant τ:

• ENBW = 1/(4τ)

So a τ = 1 s filter gives ENBW = 250 mHz; τ = 100 ms gives 2.5 Hz. Real lock-ins offer selectable filter slopes — 6, 12, 18, 24 dB/octave (1 to 4 poles). A 4-pole filter with τ = 100 ms has ENBW ≈ 5/(64τ) ≈ 0.78 Hz, sharper roll-off for the same settling behavior. The output noise scales as the square root of ENBW, so the RMS noise is v_n = e_n·√(ENBW), where e_n is the input noise density in V/√Hz.

A worked number: a signal buried in broadband noise of density e_n = 10 nV/√Hz, measured with τ = 1 s (ENBW = 0.25 Hz), yields output noise v_n = 10 nV/√Hz × √0.25 Hz = 5 nV RMS. Compared to a 100 kHz-wide ordinary amplifier (√100000 ≈ 316), that is a noise reduction of ~630× (56 dB). Push τ to 100 s and you gain another factor of 10 in voltage — at the cost of a ~500 s settling time. The eternal lock-in trade-off is bandwidth vs. speed: narrower ENBW means lower noise but proportionally longer measurement time (settling ≈ 5τ to 10τ).

Modulation: climbing out of the 1/f swamp

A lock-in only helps if you can put your signal at a frequency where the noise is low. The dominant enemy at DC is 1/f (flicker) noise, which rises without bound as frequency drops — amplifier drift, thermal EMFs, mechanical creep, and light-source flicker all live near DC. The strategy is modulation: encode the DC quantity of interest onto an AC carrier at a frequency f_r chosen to sit in the amplifier's quiet, white-noise floor (often a few hundred Hz to tens of kHz, above the 1/f corner but below any pickup lines).

• Optical chopper: a slotted wheel interrupts a beam at f_r (typically 100 Hz–4 kHz). The photocurrent now carries the signal at f_r; DC dark current and slow drift are ignored.
• Bridge excitation: a Wheatstone bridge driven by an AC source at f_r converts a tiny resistance change (a strain gauge, a bolometer) into an amplitude-modulated carrier the lock-in demodulates.
• Frequency/wavelength dither: in spectroscopy, dithering a laser wavelength or a magnetic field produces a signal at the modulation harmonic proportional to the derivative of the lineshape — the basis of derivative and second-harmonic (2f) detection.

Pick the reference frequency deliberately. Avoid 50/60 Hz and their harmonics, avoid switching-supply frequencies, and use a non-round number (e.g., 137 Hz rather than 100 Hz or 120 Hz) so that no interfering line falls exactly at f_r or its aliases. A good modulation choice can move the effective noise floor down by 20–40 dB before the lock-in does any filtering at all.

Analog vs. digital, and the dynamic-reserve budget

Classic analog lock-ins (e.g., the EG&G/PARC and early SRS instruments) used an analog multiplier or a switching demodulator — the reference squared up to a ±1 square wave that simply flips the sign of the input. That is elegant (a chopper is a perfect multiplier by ±1) but it also detects the odd harmonics of f_r (3f, 5f, 7f…) with weights 1/3, 1/5, 1/7, since a square wave contains those harmonics. Modern digital lock-ins (SRS SR860, Zurich Instruments MFLI/UHFLI) digitize the input with a fast ADC — 16 to 18 bits at tens of MSa/s — and multiply by a pure numerically-generated sine, eliminating harmonic sensitivity and drift entirely. The whole PSD becomes DSP math.

The key figure of merit is dynamic reserve — the ratio of the largest tolerable noise/interference at the input to the full-scale signal, in dB:

• Dynamic reserve (dB) = 20·log₁₀(V_noise,max / V_full-scale)

A 100 dB reserve means the instrument can recover a full-scale reading with interfering signals 100,000× larger present. In an analog lock-in, high reserve pushes gain after the PSD, trading off against DC output stability (drift and 1/f of the DC amplifiers). Digital lock-ins largely dissolve this trade-off — the reserve is set by the ADC's number of bits and its own noise floor, so 100–120 dB reserve with low drift is achievable simultaneously. The remaining ceiling is ADC saturation: interference big enough to clip the converter destroys the measurement no matter how narrow the filter, which is why a well-chosen input range and analog pre-filter still matter.

Where it earns its keep: real hardware and applications

Anywhere a small signal can be modulated at a known frequency, a lock-in is the tool of choice:

• Scanning-probe microscopy (AFM/KPFM): the cantilever is driven near its resonance (often 50–350 kHz); the lock-in measures amplitude and phase of the deflection to sub-picometer resolution, feeding the topography feedback loop.
• Electrochemical impedance spectroscopy (EIS): a potentiostat applies a small AC perturbation (5–10 mV) over 1 mHz–1 MHz; the lock-in's X/Y outputs give the complex impedance Z = Z′ + jZ″ used for battery, fuel-cell, and corrosion diagnostics.
• Optical detection: chopped-beam absorption, photoluminescence, Faraday/Kerr rotation, and pump-probe spectroscopy all recover µV-to-nV photocurrents against ambient light.
• Deep-level transient spectroscopy (DLTS) and Hall measurements on semiconductors, where signals sit far below the flicker floor.
• Metrology: AC resistance bridges, capacitance sensors, and gravitational-wave prototype interferometers all use lock-in readout.

Representative instruments: the Stanford Research SR830/SR860 (DC–102 kHz / DC–500 kHz digital), Zurich Instruments MFLI (5 MHz) and UHFLI (600 MHz), and the venerable Signal Recovery 7265. On a budget, the same math runs on a microcontroller or FPGA: an ADC, a sine table, two multipliers, and an IIR low-pass filter make a perfectly good lock-in for a few dollars of silicon.

Failure modes, limits, and best practice

The math is exact, but the physics has teeth. Common ways a lock-in measurement goes wrong:

• Reference phase locking failure: if the internal PLL can't cleanly lock to a noisy or jittery external reference, X and Y wander. Use the cleanest reference edge available (a TTL sync from the chopper controller, not the demodulated optical signal) and check the phase-noise spec.
• Overload / ADC clipping: a large out-of-band interferer clips the front end before filtering ever happens. Watch the overload indicator, drop the sensitivity range, or add an analog notch/pre-filter for the offending line.
• Harmonic contamination: square-wave (chopper) modulation and analog switching demodulators respond to 3f, 5f… A pure-sine digital lock-in and sine modulation avoid this; otherwise ensure no interference or nonlinearity injects energy at odd harmonics of f_r.
• Insufficient settling: reading the output before ~5τ–10τ after a step gives a low, drifting value. In a scanning measurement, dwell time per pixel must exceed the filter settling time or the image smears.
• Ground loops and pickup at f_r: the worst case is interference synchronous with the reference (e.g., electrical crosstalk from the chopper motor). This mimics real signal and cannot be filtered out — fix it with shielding, differential inputs, and physical separation of the modulation source.

Best-practice checklist: choose f_r above the 1/f corner and off every line harmonic; drive with a low-jitter reference; set the input range so the largest interferer never clips; pick the longest τ your measurement time allows; and always sanity-check R against the expected magnitude and θ against the expected phase. Done right, a lock-in delivers the textbook √(ENBW) noise scaling — a signal-to-noise improvement of 40–60 dB that turns an unmeasurable whisper into a clean, four-digit number.

Lock-in (synchronous) detection vs. a conventional tuned band-pass amplifier for extracting a small AC signal

Property	Lock-in amplifier	Analog band-pass amp
Effective bandwidth	Δf ≈ 1/(4τ), down to ~1 mHz	Q-limited, ~f₀/Q, rarely < 1 Hz
Center frequency	Tracks reference exactly (no drift)	Set by LC/RC, drifts with temp
Phase information	Yes — X, Y, R, θ all recovered	Amplitude only
Noise rejection (typ.)	60–120 dB dynamic reserve	30–50 dB, limited by filter Q
1/f noise handling	Excellent — modulate up to quiet band	Poor if signal sits at low f
Cost/complexity	Instrument ($1k–$25k) or DSP block	A few op-amps