A digital chip, once fabricated, simply works: a transistor that was a switch on Monday is a switch on Friday. A photonic processor enjoys no such peace. Its computational state lives in thousands of analog phases, and everything moves them. Silicon's refractive index shifts with temperature at dn/dT \approx 1.86 \times 10^{-4}\,\mathrm{K}^{-1} — a blessing when you want a thermo-optic phase shifter, a curse the rest of the time, because a phase shifter is also an exquisitely sensitive thermometer. The air conditioning cycles, a neighbouring core wakes up, a heater two rows over changes duty cycle — and your carefully programmed unitary wanders off. This lesson is about the unglamorous machinery that keeps an analog computer being the computer you programmed: monitors, dithers, lock-in detection, feedback loops, and a calibration problem that turns out to be a textbook ill-posed inverse problem. It ends with an awkward bill: the electrical power spent holding still can exceed the power spent computing.

How fast things fall apart

Work the number once and you will never trust a bare mesh again. A waveguide section of length L drifting by \Delta T picks up a phase error

for L = 100\ \mu\mathrm{m} at \lambda = 1550 nm. Now recall what precision demands. Programming a weight to B bits means holding its phase to roughly \pi/2^{B}: about 49 mrad for 6 bits, 12 mrad for 8 bits. Twelve milliradians is 0.16 kelvin — on a die dissipating watts, next to lasers and drivers, inside a rack whose inlet air swings by degrees. Passive stability is hopeless; every phase must be actively supervised. And the enemy is not only the environment: thermal crosstalk means each heater warms its neighbours, so commanding one phase perturbs a whole neighbourhood — the mesh's control knobs are not independent, they are coupled through the silicon itself. (Slower villains lurk behind temperature: mechanical stress relaxation, humidity, and charge trapping that makes some phase shifters wander over days.)

Dither and lock-in: how to see a phase

You cannot correct what you cannot measure, and no detector reads phase directly. The standard trick has three parts. First, monitor photodiodes: tiny tap couplers bleed off ~1% of the light at chosen points so the controller can watch powers without disturbing the signal path. Second, a dither: wiggle the phase under control by a small \delta\sin\omega t (a few millirad, at some kilohertz). The monitored power responds as

so the component at the dither frequency has amplitude proportional to the slope P'(\varphi_0) — a signed error signal that says not just "you're off" but which way. Third, lock-in detection: multiply the detector output by \sin\omega t and low-pass filter, which surgically extracts that one component while rejecting the DC drift and 1/f noise that swamp any naive measurement. Feed the recovered slope into an integrator driving the heater and the loop parks itself where the error vanishes — at a fringe extremum, where P' = 0 and only the 2\omega harmonic remains. Give each of the mesh's phases its own dither frequency and one shared photodiode can supervise dozens of them simultaneously, each loop deaf to the others' tones.

Note the trade baked into the slider: a bigger dither gives a stronger, cleaner error signal — and a bigger deliberate wobble on your computation. Production systems run the dither just above the noise floor and average longer instead.

Calibration is an inverse problem

Feedback holds a phase where you told it to go; calibration is the harder art of learning what "where" means. The chip you receive from the fab is not the chip you designed: every phase shifter has a fabrication offset, every heater its own efficiency, and the crosstalk matrix couples them all. Formally, you observe output powers P = f(\theta_{\text{cmd}};\, \theta_{\text{offset}}, C) for probe inputs and command vectors \theta_{\text{cmd}}, and must infer the offsets and the crosstalk matrix C. This is an inverse problem with all the pathologies the ill-posedness lesson warned about: the map from parameters to observations is many-to-one (interferometric powers are blind to global phases; distinct offset patterns produce nearly identical outputs), and measurement noise gets amplified wherever sensitivity is low — precisely at the fringe extrema where the loops like to sit. The practical playbook is regularisation by structure: calibrate element by element (routing light through one MZI at a time, sweeping its phase, fitting its fringe), then correct pairwise crosstalk, then close global feedback to mop up the residual — rather than inverting one heroic, ill-conditioned model of the whole chip. Self-configuring meshes push further still, arranging the mesh to calibrate itself layer by layer from its own monitor readings, an idea developed in the mesh-programming lesson.

The bill for holding still

Now the systems punchline. A thermo-optic phase shifter needs P_\pi \approx 10\text{–}20 mW to hold a π shift, and holding is continuous — this is not switching energy, it is standing power. A Clements mesh for N = 64 has N(N-1)/2 = 2016 MZIs and N(N-1) = 4032 phase shifters. At an average of even 5 mW each, the mesh burns 20 watts doing nothing — before a single MAC, before the laser, before any converter. Against the mesh's 2N^2 f \approx 41 TOPS at 5 GHz, tuning alone caps the system at about 2 TOPS/W: a respectable digital accelerator's figure, spent entirely on keeping the analog computer honest. This single line item motivates a small industry of alternatives — undercut and suspended heaters (milliwatts, but slower), MEMS and NOEMS actuators (nanowatts of hold power), and non-volatile phase-change weights that hold their state with the power off. Whenever you evaluate a photonic accelerator claim, ask for the tuning power first: it is the least advertised number and, remarkably often, the largest.