Cascadability and the Criteria for Logic

A processor executing one second of your code performs on the order of 10^{15} gate operations, arranged in chains hundreds of gates deep, where the mangled output of each gate is the only input the next one will ever get. No signal is ever re-groomed by a benevolent experimenter; whatever restoration happens must happen inside every single gate, automatically, forever. This is the standard the parts catalogue of the last lesson must be measured against, and it was articulated most sharply by Robert Keyes of IBM — a physicist whose day job, for decades, was asking of every proposed "transistor killer" not does it switch? but can a hundred million of them run each other? His checklist, later sharpened for optics by David Miller, is the most reusable intellectual tool in this module: it applies verbatim to every beyond-CMOS device you will ever be pitched — magnetic, molecular, quantum-dot or photonic.

The checklist

Gain (fan-out): a gate's output must be able to drive more than one subsequent input — signal power greater than input power, supplied from a power source, not from the signal itself;
Level restoration: degraded inputs must map to cleaner outputs, so noise and level errors shrink at every stage instead of accumulating;
Input–output isolation: the output must not talk back — a gate's load, or reflections from it, must not change what the gate's input sees;
Cascadability: the output signal must be of the same physical form (wavelength, level convention, format) as the input of an identical gate, so gates chain indefinitely;
No critical biasing: operation must not depend on holding a parameter (bias power, temperature, wavelength) to within a hair's breadth of a critical value;
Loss-independent logic levels: the meaning of "0" and "1" must survive the attenuation, drift and variation of a real system.

Notice what is not on the list: speed. A slow technology that meets every criterion can be parallelised into a computer; a femtosecond switch that fails restoration cannot compute at any speed, because its errors compound faster than its answers do. The list is a chain of qualifying conditions — fail one and the rest are irrelevant. CMOS passes all six so casually that most engineers never learn the list exists: the transistor has power gain to spare, its output is re-pinned to the supply rails at every stage, the insulating gate isolates input from output, its output voltage is an input voltage, and it tolerates supply and threshold variation that would kill any critically-biased device. That silent, total compliance — not raw switching speed or energy — is what optical logic was actually being asked to match.

The key figure: restoration is a shape

Almost the entire checklist can be read off a single curve: the gate's transfer function, output level versus input level. Chain n identical gates and the signal experiences the n-fold composition f(f(\cdots f(x))) — so the fate of a logic technology is a question about iterating a function. Two candidate technologies are plotted below. One is a restoring gate: an S-shaped (sigmoid) transfer curve with gain, crossing the identity line y = x steeply in the middle and flatly near the ends. The other is a weak optical nonlinearity of the kind the Kerr catalogue actually offers: a gentle bend with net transmission below unity and no steep threshold. Drag the slider to cascade them:

The mathematics is fixed-point iteration, straight from a dynamical-systems toolbox. Where the transfer curve crosses y = x with slope of magnitude less than one, the crossing is a stable fixed point: nearby values are pulled in by each stage. Where it crosses with slope greater than one, the point is unstable — a watershed. A restoring gate has exactly the right arrangement: two stable fixed points (the logic levels "0" and "1") separated by one unstable crossing (the decision threshold). Cascading such a gate improves the signal — a battered 0.7 becomes 0.9, then 0.99: the celebrated noise margin of digital electronics, manufactured by the shape of a curve. The weak nonlinearity has only one stable fixed point, at zero. Iterate it and everything — ones, zeros, and all the information between them — relaxes to darkness.

Worked numbers: watching levels live or die

Run the iteration yourself. The program cascades both transfer functions and prints the logic levels stage by stage — a "1" launched at a degraded 0.75 and a "0" launched at a noisy 0.20:

// A restoring gate: sigmoid with gain ~2.5 at threshold. const restoring = (x: number): number => 1 / (1 + Math.exp(-10 * (x - 0.5))); // A weak, lossy optical nonlinearity: gentle bend, net transmission below 1. const weak = (x: number): number => 0.9775 * x - 0.255 * x * x; function cascade(f: (x: number) => number, x: number, stages: number): string { const out: string[] = [x.toFixed(3)]; for (let i = 0; i < stages; i++) { x = f(x); out.push(x.toFixed(3)); } return out.join(" -> "); } console.log("restoring, '1' in at 0.75: " + cascade(restoring, 0.75, 5)); console.log("restoring, '0' in at 0.20: " + cascade(restoring, 0.20, 5)); console.log("weak, '1' in at 0.75: " + cascade(weak, 0.75, 5)); console.log("weak, '0' in at 0.20: " + cascade(weak, 0.20, 5));

Five stages in, the restoring gate has forced its levels to 0.999 and 0.001 — better than they started, which is the whole miracle: the gate chain is a noise-eating machine. The weak nonlinearity's "1" has meanwhile sagged below 0.5 and keeps falling; a downstream gate reading it against any fixed threshold will soon guess. And this simulation is generous — it ignores fan-out. Ask a gate with no gain to drive two successors and each inherits at most half the power: a further 3 dB of decay per stage, per branch. Without gain above unity, fan-out alone kills a logic family in a dozen stages, regardless of how elegant its switching physics is. Gain, in other words, is not a luxury feature — it is what makes the difference between a signal chain and an obituary.

Robert W. Keyes (1921–2010) spent his IBM career doing something institutionally awkward: rigorously evaluating the exciting device proposals of his own colleagues, and usually concluding that the boring transistor would win. His 1985 essay "What makes a good computer device?" distilled the criteria on this page, and his earlier analyses of optical bistability — then the hottest candidate for optical logic — were blunt: the devices lacked gain, lacked isolation, demanded critical biasing, and dissipated more, not less, than electronics at comparable duty. The optics community's responses ranged from irritation to fury; the passage of time has ranged from vindication to total vindication. The deeper lesson he taught is methodological. Device papers report the metrics their device is good at; a technology evaluation must instead check the metrics a system cannot live without — and those are set by the weakest link, not the headline. Every "the transistor is doomed" article since 1960 has had to get past some version of Keyes' checklist. So far, the checklist is undefeated — a streak examined in detail in the next lesson.

The seductive trap of the 1980s: a nonlinear cavity (a Fabry–Pérot étalon with a Kerr medium inside) can be optically bistable — over a range of input powers it has two stable transmission states, and its input–output curve is a hysteresis loop. Two states! Memory! Surely this is an optical flip-flop, and logic follows? It does not, and the checklist says why. The bistable étalon has no gain: its output is its input, attenuated — it cannot fan out. It has no isolation: it is a resonator, and reflections from downstream re-enter and destabilise it. It is critically biased: the hysteresis loop lives on a knife-edge of holding power, temperature and detuning, so every device on a chip would need individual babysitting. And its two states are defined relative to a holding beam, so its levels are not loss-independent. Bistability gives you a curve with two stable points — but restoration needs those points to be actively regenerated with gain at every stage, which is a property of amplifiers, not of resonators. The general reflex: when shown a device with N stable states, ask who pays to move between them, and who repairs the levels afterwards.

Where this goes next

The checklist is now in hand, and the last two lessons have quietly loaded the dice: optical nonlinearities are weak (so optical gain-with-isolation is expensive), optical resonators are critically biased, χ⁽²⁾ devices change wavelength (breaking cascadability), and intensity-coded levels decay with every decibel of loss. The next lesson runs the full quantitative trial — photon energies against switching charges, diffraction limits against 5-nanometre transistors, and the one asset optics simply does not have: a way to make light stand still.