Skip to content

How phase-diffusion QRNG works, from laser to bits

A phase-diffusion QRNG derives randomness from spontaneous emission in a semiconductor laser driven repeatedly through its lasing threshold. Each pulse starts from vacuum fluctuations, so its optical phase is unpredictable, and interfering consecutive pulses converts that phase difference into a measurable voltage that can be digitised.

Why the phase, and not something else

Choosing a physical process for a random number generator is a trade between how well you can model it and how fast it runs. Phase diffusion is attractive on both counts. The randomness originates in spontaneous emission, which is a quantum process with a well-established description, and the resulting phase randomisation happens on a timescale that supports high generation rates.

The alternative approaches each make different trades. Comparing them properly is worth doing before deciding that any one of them is right for a given system.

From laser to bits

The laser is gain-switched: driven below threshold and then above it, repeatedly. Below threshold the optical field decays toward vacuum. Each time the drive current pushes the laser back above threshold, the pulse builds up from spontaneous emission, and its phase bears no relation to the phase of the pulse before it.

That phase is not directly measurable, so consecutive pulses are interfered in an unbalanced interferometer. The interference converts an unpredictable phase difference into an intensity difference, a photodiode converts intensity to current, and an ADC digitises it. What arrives at the digital side is a voltage whose distribution depends on a phase that was unpredictable at source.

What the model is conditional on

The entropy estimate for this architecture is not a property of the laser alone. It depends on the drive conditions holding the laser genuinely below threshold between pulses, on the interferometer delay matching the pulse period, and on the classical noise contributed by the detection chain being accounted for rather than counted as entropy.

This is the part that separates a physical entropy model from a claim. Each of those dependencies is a quantity that can be measured while the device is running, which is what makes the source observable rather than merely characterised. The conditioning stage that follows then has to be sized against the entropy the model actually supports, not against the raw bit rate.

Nothing here establishes that the output bits are random. What it establishes is a model of where unpredictability enters the system, an estimate conditional on that model, and a set of quantities you can monitor to detect the model ceasing to hold.