Phase-Sensitive Detection in Lock-In Amplifiers

Phase-sensitive detection is the principle behind a lock-in amplifier. It recovers a weak signal by comparing the measurement with a known reference frequency and phase. Components that are correlated with the reference become a slowly varying output. Components that are not correlated with the reference mostly average toward zero after low-pass filtering.

The method is useful when the quantity of interest can be modulated or naturally occurs at a known frequency: chopped optical power, bridge excitation, magnetic susceptibility, impedance spectroscopy, vibration response, resonator readout, low-level photodiode current, or a sensor response buried below broadband noise.

The key engineering idea is not “amplify more.” It is to move the measurement to a known frequency, multiply by a matched reference, reject out-of-band and uncorrelated content, and report the result with a stated bandwidth, phase convention, and validation test.

Measurement Boundary and Assumptions

A lock-in measurement is a chain, not only an instrument setting. The boundary includes the excitation source, reference distribution, sensor or sample, analog front end, anti-alias filtering, demodulator, low-pass filter, scaling convention, data logger, and validation fixture. The result is trustworthy only if the coherent component at the reference frequency represents the physical effect being measured.

The basic assumptions are:

• the desired response is linear enough that modulation does not change the measurand;
• the reference frequency is stable during the averaging window;
• the front end is not saturated before demodulation;
• coherent pickup from the reference path is smaller than the required uncertainty;
• filter settling time is long enough before a value is accepted;
• the reported amplitude convention is explicit: peak, RMS, one-sided amplitude, or engineering-unit response.

This boundary matters because lock-in amplifiers can produce precise numbers from invalid setups. A clean X/Y output does not prove that the optical beam reached the detector, that the bridge was excited correctly, or that a thermal sample was at steady state. It proves only that part of the measured signal is correlated with the reference waveform under the chosen bandwidth.

Signal Model

A simplified measured signal is:

where:

• A is the signal amplitude at angular frequency \omega_0;
• \phi is phase relative to the reference;
• n(t) is noise, interference, drift, or other unwanted content.

The lock-in multiplies x(t) by a reference:

The product is:

Using the trigonometric identity:

the signal part becomes:

A low-pass filter removes the term near 2\omega_0 and keeps the slowly varying term:

This is the in-phase component. A second reference in quadrature:

gives a quadrature component:

depending on sign convention. The magnitude can be estimated from both channels:

The phase is:

The factor of two appears because multiplying two sinusoids at the same frequency produces a DC term equal to half the original amplitude under this peak-amplitude convention.

Scaling Convention

Different instruments report the demodulated result differently. Some display X and Y directly, some multiply by two internally, some report RMS-equivalent amplitude, and some apply calibration factors to convert volts into optical power, displacement, impedance, strain, or another engineering unit.

For a sinusoid with peak amplitude A_{pk}:

If the desired reporting convention is RMS amplitude:

The report must state which convention is used. Otherwise two correct instruments can appear to disagree by factors of 2 or \sqrt{2}. This is a common source of failed correlation between laboratory measurements, production testers, and embedded demodulation firmware.

Orthogonality and Averaging

The rejection mechanism depends on orthogonality over the averaging interval. A signal at the reference frequency produces a stable low-frequency term after multiplication. A broadband or off-frequency component tends to average toward zero, but only after enough cycles and only if it is not phase-locked to the same reference source.

This is why the observation time matters. A nominally narrow bandwidth does not help if the operator records the output before filter settling, changes the sample during the averaging window, or compares readings taken with different time constants.

Why It Improves Weak-Signal Measurements

Noise power depends on bandwidth. If the desired signal is measured over a wide bandwidth, broadband noise can dominate the display even when the signal is repeatable. A lock-in amplifier narrows the measurement bandwidth around the reference by demodulating the signal to near DC and then applying a low-pass filter.

For a first-order low-pass filter with time constant \tau, a common equivalent noise bandwidth screen is:

This is not a universal law for all digital filters or roll-off settings. It is a useful first-order estimate. The actual equivalent noise bandwidth depends on filter order, implementation, settling requirement, and whether the instrument reports RMS, peak, or averaged values.

Longer time constant means lower noise bandwidth and better averaging, but slower response. Shorter time constant gives faster tracking but lets in more noise.

Bandwidth and Settling Tradeoff

The time constant is not just a noise knob. It also defines how long the output needs to settle after a step in signal, phase, reference frequency, sample state, or gain range. A first-order filter reaches about 63\% of a final step after one time constant and approaches final value only after several time constants. Higher-order filters can reduce noise more sharply but may settle more slowly or with different transient shape.

An engineering report should therefore pair every measured value with:

• reference frequency;
• time constant;
• filter order or equivalent noise bandwidth;
• settling delay before reading;
• number of averaged samples or records;
• reason the measurand was stable over that interval.

If a thermal, chemical, optical, or mechanical experiment changes during the averaging window, the lock-in output becomes a filtered history rather than an instantaneous state. That can be useful for controlled trend measurements, but it is not valid for a fast acceptance test unless the settling rule is built into the test method.

Worked Example: Chopped Photodiode Signal

A photodiode measurement uses a mechanical chopper at:

The transimpedance output contains a small sinusoidal component at the chopper frequency:

as peak amplitude. The broadband input-referred noise density around the measurement band is estimated as:

If the signal were viewed over a 1.0\ \text{kHz} bandwidth, the approximate RMS noise would be:

A rough amplitude-to-noise screen is:

The signal is smaller than the broadband noise displayed over that bandwidth.

Now use phase-sensitive detection with:

The first-order equivalent noise bandwidth estimate is:

The filtered noise scale is approximately:

For a signal in phase with the reference:

The demodulated SNR screen is:

The recovered amplitude estimate is:

The result is not magic gain. The improvement comes from measuring only the part of the signal coherent with the reference and narrowing the effective noise bandwidth from about 1000\ \text{Hz} to about 0.125\ \text{Hz}.

Interpreting the Example

The example also shows why a lock-in amplifier can hide setup mistakes. If the photodiode front end saturates from ambient light before the chopped component is demodulated, the computed SNR_{lockin} is irrelevant. If the chopper driver couples electrically into the transimpedance input, the output may improve when the optical path is blocked, which is the opposite of a valid result. If the optical source drifts during a long time constant, the final value may lag the actual source state.

The correct interpretation is conditional: the recovered 20\ \mu\text{V} amplitude is meaningful only after front-end headroom, dark/background condition, reference leakage, phase convention, and settling have been checked.

Phase Error

If only the in-phase channel is used, phase error reduces the measured signal:

For a 20^\circ phase error:

The in-phase reading is about 6.0\% low.

For a 60^\circ phase error:

The in-phase reading is half the correct value. Quadrature detection avoids this loss by measuring both X and Y and computing magnitude, but it still requires a stable reference frequency and careful sign, scaling, and calibration conventions.

When Phase Is the Measurement

In some tests, phase is not an error to be minimized; it is the measured quantity. Examples include impedance spectroscopy, resonator tracking, thermal diffusivity tests, eddy-current inspection, and frequency-response measurement. In those cases the instrument must preserve phase calibration through cables, filters, amplifiers, sample fixtures, and digital processing.

The phase reference should be defined physically. “Zero phase” might mean source voltage, source current, optical chopper blade position, shaker drive voltage, bridge excitation, or a sampled digital timing marker. Each choice answers a different engineering question. A phase number without a reference definition is not reproducible.

Choosing the Reference Frequency

The reference frequency should be chosen so that:

1. the sensor and actuator can respond at that frequency;
2. the analog front end has adequate bandwidth and phase margin;
3. the reference is away from dominant interference such as mains frequency, switching regulators, mechanical vibration, or acoustic pickup;
4. the modulation does not change the physical quantity being measured;
5. sampling, anti-alias filtering, and digital demodulation have adequate timing margin;
6. the result can settle within the required test time.

For optical measurements, a chopper or modulated LED can move the signal away from low-frequency drift and ambient-light variation. For bridge measurements, AC excitation can separate true bridge response from DC offset and thermoelectric effects. For impedance measurements, the reference frequency must also respect the device under test and parasitic capacitance or inductance.

Reference-Frequency Selection Matrix

Constraint	Good sign	Warning sign
sensor dynamics	sensor response is flat or calibrated at f_0	response rolls off or changes phase unpredictably
front-end bandwidth	amplifier and filters preserve amplitude and phase	anti-alias or analog filters distort the modulated component
interference environment	f_0 avoids mains, switching, vibration, and acoustic peaks	coherent interference sits near the reference
sample physics	modulation is small enough to avoid heating, depletion, fatigue, or nonlinear response	modulation changes the state being measured
test duration	settling time fits the available measurement window	required averaging makes the test impractically slow
digital implementation	sample rate, coherent record length, and numerical precision are adequate	aliasing or spectral leakage appears before demodulation

The best reference frequency is often a compromise. Moving to a higher frequency can avoid drift but increase phase lag, capacitance effects, actuator limitations, or front-end noise. Moving lower can improve physical response but expose the measurement to 1/f noise, drift, thermal gradients, and slow interference.

What the Lock-In Does Not Fix

Phase-sensitive detection can recover a weak coherent signal, but it does not make every measurement valid.

It does not fix:

• front-end saturation before demodulation;
• wrong optical, mechanical, thermal, or electrical boundary conditions;
• reference leakage that creates a false coherent signal;
• phase drift from cables, filters, amplifiers, or sample dynamics;
• interference exactly coherent with the reference;
• aliasing before digital demodulation;
• bandwidth that is too narrow to follow the real measurand;
• calibration errors in gain, reference amplitude, or scaling factor;
• noise that changes during the averaging window.

The most dangerous failure is a convincing lock-in number from the wrong coherent signal. A mechanical chopper, LED driver, or excitation source can capacitively, magnetically, thermally, or optically leak into the measurement path and create a reference-synchronous artifact.

Artifact Controls

The strongest artifact controls deliberately break the physical path while leaving parts of the reference system active. For a chopped optical measurement, block the beam but keep the chopper, LED driver, and reference cable running. For a bridge measurement, substitute a precision dummy bridge or shorted input under the same excitation. For a vibration test, disconnect mechanical drive transmission while keeping the electrical reference active if the fixture permits it safely.

The goal is to prove that the coherent output disappears when the physical effect is absent. If the output remains, the lock-in has detected a synchronous artifact, not the desired measurand. The artifact can come from cable capacitance, magnetic pickup, shared grounds, power-supply modulation, scattered light, acoustic coupling, mechanical vibration, thermal cycling, software timing leakage, or digital crosstalk.

Validation Checks

A defensible lock-in measurement should include these checks:

1. Reference-off test: with the modulated source disabled, the coherent output should fall to the expected residual level.
2. Phase sweep: in-phase and quadrature channels should rotate as expected while magnitude remains stable.
3. Frequency sweep: the recovered signal should follow the sensor, sample, and front-end frequency response.
4. Time-constant sweep: noise should fall with narrower bandwidth, while real signal should remain consistent after settling.
5. Linearity test: recovered amplitude should scale with known input amplitude over the intended range.
6. Artifact test: block the physical signal while leaving the reference electronics active to detect pickup or leakage.
7. Bandwidth statement: report time constant, filter order or equivalent noise bandwidth, settling time, and reference frequency with the result.

Validation Package and Reporting Requirements

For a production tester, research setup, or commissioning measurement, validation should produce a package that another engineer can repeat. The package should identify the signal path, reference path, physical stimulus, calibration chain, amplitude convention, uncertainty contributors, and acceptance rules.

Minimum reporting fields are:

Field	Why it matters
reference frequency and source	defines the coherent component being measured
phase-zero definition	makes X, Y, magnitude, and phase reproducible
amplitude convention	avoids peak/RMS/factor-of-two disputes
input range and front-end headroom	proves demodulation did not follow saturation
time constant, filter order, ENBW, and settling delay	defines the noise bandwidth and response time
artifact-control result	proves the reference did not leak directly into the signal
calibration stimulus and traceability	links the demodulated value to engineering units
uncertainty budget	combines noise, gain, phase, calibration, drift, and setup effects

The validation package should also define retest triggers. Changing cable routing, reference distribution, chopper driver, LED driver, excitation amplitude, fixture grounding, front-end filter, firmware demodulator, sampling clock, or shielding can alter a lock-in result even when the sample has not changed.

Acceptance Logic

A measured lock-in value is releasable only when the signal remains stable under reasonable variations in phase, reference frequency, time constant, and physical blocking tests. If a result appears only at one phase setting, one cable routing, one gain range, or one unusually long averaging interval, it should be treated as a diagnostic clue rather than released data.

Engineering Interpretation

A lock-in amplifier is best understood as a correlation measurement. It asks: how much of the measured signal looks like this reference waveform, at this frequency and phase, over this bandwidth?

That question is powerful for weak signals because random broadband noise is mostly not coherent with the reference. But the result is only trustworthy when the reference represents the physical effect of interest, the front end remains linear, the bandwidth is appropriate, and validation rules out coherent artifacts.