The Template
A lock-in amplifier recovers signals buried sixty to a hundred decibels below the noise floor. The principle is multiplication. The instrument takes the noisy input — signal plus broadband noise — and multiplies it by a reference signal at the frequency of interest. Any component of the input that matches the reference frequency produces a constant product. Every other frequency produces an oscillating product that, over time, averages to zero. A low-pass filter after the multiplier retains the constant and discards the oscillations. What remains is the amplitude of the input at exactly the reference frequency.
The device was commercialized in the 1960s by Princeton Applied Research, but the principle — phase-sensitive detection — dates to the early days of radio. The multiplication is a correlation: the instrument asks, at every moment, how much does the input resemble the reference? The answer accumulates. Noise, being uncorrelated with the reference, contributes nothing to the accumulation. Signal, being correlated, contributes everything.
The requirement is strict. You must already know the frequency you are looking for. The lock-in amplifier cannot discover signals. It can only confirm the presence of a signal whose frequency is supplied from outside. The prior knowledge is not an aid to detection. It is the detection.
The Global Positioning System transmits on a single frequency — 1575.42 megahertz for the civilian L1 signal. All satellites broadcast simultaneously on this same frequency. The signals arrive at a ground receiver twenty decibels below the thermal noise floor, weaker than the noise by a factor of a hundred. On a spectrum analyzer, nothing is visible.
Each satellite transmits a unique pseudo-random noise code — a binary sequence of 1023 chips repeating every millisecond, generated by a specific pair of feedback shift registers. The codes are designed to be nearly orthogonal: the cross-correlation between any two satellites' codes is close to zero, while the autocorrelation of each code with itself produces a sharp spike at zero lag and near-zero values everywhere else.
The receiver generates a local replica of the expected code and slides it in time against the incoming signal, computing the cross-correlation at each offset. When the replica aligns with the satellite's actual transmission, the correlation spikes. The timing offset at the spike gives the signal's travel time, and the travel time gives the distance to the satellite. Four satellites, four distances, one position.
The architecture is code division multiple access — CDMA. The signals occupy the same bandwidth at the same time. They are separated not by frequency, not by time, but by shape. The receiver's knowledge of the code is the only thing that distinguishes satellite from satellite, and signal from noise. Bradford Parkinson's original design for Project 621B in 1973 exploited this for anti-jamming: a spread-spectrum signal below the noise floor cannot be jammed by an adversary who does not know the code.
On September 14, 2015, at 5:51 AM Eastern time, the LIGO detectors at Hanford, Washington and Livingston, Louisiana recorded a differential arm length change of approximately four times ten to the negative eighteenth meters — one-thousandth the diameter of a proton. The signal lasted two-tenths of a second.
The raw detector output was dominated by seismic vibration, thermal motion in the mirror suspensions, and quantum shot noise in the laser light. The gravitational wave signal was invisible in the time series. It was found by matched filtering: the detector output was cross-correlated with a bank of approximately two hundred and fifty thousand pre-computed waveform templates, each representing a different combination of black hole masses, spins, and orbital parameters. The template that best matched the data — two black holes of thirty-six and twenty-nine solar masses spiraling inward and merging — produced a signal-to-noise ratio of twenty-four. The event, designated GW150914, was detected with a false alarm probability of less than one in two hundred thousand years.
The template bank is computed from general relativity. Each template traces the gravitational wave frequency and amplitude as two compact objects orbit closer, faster, louder — the characteristic chirp. The cross-correlation asks the same question the lock-in amplifier asks: how much does the data resemble this specific shape? The computational cost is large — of order ten to the eighteenth floating-point operations for a year of data — because every template must be tested against every moment.
The constraint is the same as the lock-in amplifier's. LIGO can only detect gravitational waves for which it has templates. A signal from an unanticipated source — a cosmic string, a phase transition in the early universe, a process not yet modeled — would pass through the matched filter undetected, contributing nothing to any correlation. Unmodeled searches exist, but their sensitivity is far lower. The template is the detector.
A reference frequency, a pseudo-random code, a waveform template. In each case, the signal is present in the data but invisible — buried under noise that is stronger by factors of a hundred or a billion. The instrument that recovers it does not filter the noise away. It correlates the data with a known shape, and the correlation amplifies only what matches. The shape must be supplied in advance. The lock-in amplifier needs the frequency. The GPS receiver needs the code. LIGO needs the chirp. Without the template, the signal remains — physically present, causally active, measurably real — and undetectable. The prior knowledge is not a tool used by the detector. It is the detector.