Photonics

Choose the optical mode that carries the qubit#

Define path/time-bin/polarization encodings, state preparation, and conversion assumptions.

A photonic quantum computer stores information in properties of individual photons — which path a photon occupies, when it arrives, or how it is polarized — and processes that information by letting photons interfere in networks of beam splitters and phase shifters. Photons are appealing carriers for three reasons: they move fast, they barely notice their environment, and they travel down the same fibers and waveguides that classical communications already uses. A qubit that tolerates room temperature and rides existing optical infrastructure is a genuine engineering prize.

Photonic quantum technologies encode and process information in optical modes using interference, sources, circuits, and detectors. [photonic-review] [photonic-many-qubit-chip]

Define the photon's role before pricing the path#

A photon can carry a computational mode, distribute entanglement between modules, connect chips, or serve as the measured output of a sampler or sensor. Those roles impose different requirements. A networking photon may be valuable when it is heralded and its arrival time is known; a computational architecture may require many indistinguishable photons to survive a synchronized circuit. Write the role, encoding, success event, clock model, and validation rule before comparing components.

Encoding also changes the resource unit. A qubit may occupy path, polarization, time-bin, frequency-bin, or a more elaborate multimode code. Count occupied optical modes, vacuum and ancilla modes, sources, switches, delay lines, detectors, and feed-forward resources. Calling a spatial mode a “qubit” without its source and detector hides the components that determine delivered rate.

Sensitivity rule: Perturb every efficiency in both linear units and its native measurement uncertainty, then rank its effect on accepted results per second. Include repetition count and multiplicity, because a moderately lossy switch crossed ten times can matter more than the least efficient component crossed once. Check that the proposed improvement does not reduce indistinguishability, raise multiphoton noise, exceed detector rate, or lengthen feed-forward delay. Component optimization is valid only inside the complete path.

Interference requires indistinguishable photons#

Connect source purity, timing, spectral matching, phase stability, and beam-splitter operations.

The same aloofness is the catch. Photons do not interact with each other on their own, so two-qubit logic has to be engineered indirectly — typically through interference plus measurement, which makes gates probabilistic. And photons are easily lost: absorbed in a waveguide, missed by an inefficient detector, or filtered out by imperfect optics. A lost photon is not a flipped bit; it is missing information.

Weak direct photon-photon interaction motivates measurement, ancilla, multiplexing, or other architecture-specific resources for logic. [photonic-review]

Multiply efficiencies in linear units#

For a path with independent stages, one-photon survival is the product of source delivery, coupling, propagation, switching, filtering, and detector efficiencies. Convert a loss of $L$ decibels to transmission $10^{-L/10}$ before multiplying. If $n$ photons must all arrive, a simple independent model raises the single-path factor to the relevant multiplicity. That exponent is why modest component loss becomes architectural.

Independence is an assumption. Source brightness can trade against multiphoton contamination; detector efficiency can trade against dark counts; switching loss can depend on route; thermal drift can correlate many modes. The budget should contain both component measurements and an end-to-end check. If their product predicts a success rate inconsistent with observation, investigate coupling definitions, conditional denominators, synchronization, and correlated loss rather than tuning an unexplained factor.

Measurement and feed-forward make logic conditional#

Trace detectors, success heralds, classical latency, switching, and multiplexing for a probabilistic primitive.

So the first diligence question about any photonics announcement is which role is being claimed. Photonics appears in the quantum world as a full computing platform, as a networking layer connecting other qubit types, as an interconnect inside a larger machine, as a sensing or communication system, or as a component business selling sources and detectors. Each role has its own evidence standard, and a strong demonstration in one role proves almost nothing about the others.

Loss and indistinguishability are central system variables whose component contributions compound along an optical path. [photonic-review] [full-stack-review]

Indistinguishability is part of the interference primitive#

Photons interfere only to the extent that their spectral, temporal, spatial, and polarization modes overlap. A source record should therefore include brightness, purity, multiphoton probability, indistinguishability under the relevant separation, repetition rate, and collection efficiency. Quoting the best value from each of different operating points constructs a source that was never operated.

Interferometer meshes add phase stability and calibration requirements. Small phase errors accumulate across depth, and thermal or electro-optic tuning consumes control channels and power. Validate the programmed transformation with coherent-light calibration where appropriate, then test the single-photon regime. Programmable many-photon chip experiments show substantial integrated control under specific devices and tasks [photonic-many-qubit-chip]; their evidence should not be generalized to an unreported source or detector stack.

Multiply the complete transmission path#

Build a dimensionally correct end-to-end survival budget and show the assumptions behind independent factors.

The right mental model for a photonic architecture is a compounding budget. If a computation needs many photons to survive generation, coupling into a chip, an interferometer mesh, switching, and detection, then the end-to-end success probability is roughly $P_{\mathrm{success}}\approx\prod_g(1-\varepsilon_g)^{n_g}$, where $\varepsilon_g$ is the loss or error per component of type $g$ and $n_g$ is how many such components sit in the path. Small per-component losses sound harmless until the exponents grow: a chain that loses two percent at each of one hundred stages keeps about thirteen percent of its photons.

A platform evaluation must distinguish computation, communication, and implementation criteria. [photonic-programmable-advantage] [photonic-review]

Measurement-induced logic consumes probability and time#

Because direct photon-photon interactions are weak, architectures often use interference, ancilla photons, measurement, and feed-forward to implement conditional operations. A successful detector pattern may herald the desired transformation. Count every attempted resource state, accepted event, discarded event, detector, and classical decision. Postselection can demonstrate a circuit relation but does not provide an on-demand gate unless the success mechanism is incorporated into the architecture.

Multiplexing trades hardware for delivery probability. Several sources or preparation attempts run in parallel; switches route a successful one onward; delay lines align it with other events. The resulting rate depends on source repetition, heralding latency, switch depth, storage loss, detector recovery, and controller schedule. Report accepted outputs per second and per consumed source attempt, not only conditional fidelity.

Compute, network, and component evidence diverge#

Create separate evidence contracts for delivered entanglement, integrated components, and computation.

This is why photonics people obsess over indistinguishability — photons must be identical in every degree of freedom or they fail to interfere — and over per-decibel loss numbers that engineers in other modalities never have to quote. It is also why many photonic computing schemes lean on feedforward: measure some photons early, compute the correction classically, and reconfigure later optics in real time. Feedforward turns a passive optical network into a computer, at the price of adding detectors, logic, and latency to the loss budget.

Fault-tolerant photonic-computing claims require an explicit loss/error model and correction architecture rather than component transmission alone. [photonic-manufacturable-platform] [photonic-review]

Feed-forward makes the optical circuit a timed system#

A detection event must be digitized, classified, processed, and converted into a switch or phase-setting command before the affected photon reaches the control point. The optical delay required to wait for that decision adds propagation loss and footprint. Trace detector timestamp to actuator settling and give a worst-case deadline. A demonstration with offline post-processing is not evidence for real-time feed-forward.

Detectors bring efficiency, dark counts, timing jitter, dead time, number resolution, and cryogenic requirements. Match each metric to the clock window and photon number expected by the protocol. A low dark-count rate integrated over a long or numerous set of windows can still matter. Detector saturation can cap throughput before source brightness does.

A useful optical stack is an integrated result#

Identify a proof experiment combining source, circuit, detector, and feed-forward rather than separate best components.

For memo discipline, score claims the way this book scores every hardware claim: decision score = claim quality + proof progress − risk − kill criteria pressure . The score forces a photonic pitch to declare which layer it has actually proven — sources, integrated optics, loss management, detection, networking value, or application-level utility — instead of letting a strong layer borrow credibility from a weak one.

Current quantitative source, detector, feed-forward, and integrated-system records need primary papers beyond the registered review. [photonic-review]

Separate sampling evidence from universal-computing evidence#

A programmable photonic processor can perform a sampling task whose validation and classical comparison are carefully defined [photonic-programmable-advantage]. That can be a major systems result without demonstrating a universal fault-tolerant gate set. Record whether the claim concerns sampling complexity, a logical operation, delivered entanglement, or component manufacture. Each needs a different baseline.

A manufacturable integrated platform addresses another layer: process design, component libraries, wafer-scale consistency, packaging, and a route to assembly [photonic-manufacturable-platform]. Manufacturing evidence does not by itself solve source yield, loss correction, or feed-forward, but it can remove an important scaling uncertainty. A good dossier gives credit at the demonstrated layer and leaves the remaining layers visible.

Audit every conditional denominator#

1. pump pulses emitted;
2. source events and heralds;
3. photons coupled into the declared reference plane;
4. circuit attempts and detector windows;
5. accepted patterns and independently validated results.

For a path experiment, count pump pulses, source heralds, photons coupled, circuit attempts, detector windows, accepted patterns, and validated results. A conditional fidelity among accepted patterns and an unconditional accepted-result probability answer different questions. Postselection belongs in the numerator and denominator record; it cannot disappear when the result is translated into throughput.

Measure the complete path and compare it with the product model. If a twenty-stage model predicts survival (P) but observed delivery differs, test source correlation, route-dependent switching, timing mismatch, detector saturation, background, and inconsistent reference planes. Each efficiency must state where the photon enters and leaves the measured component. Otherwise two adjacent stages may both claim the same coupling loss or neither may claim it.

Set the proof gate by role. A component must show yield and insertion loss after packaging. A network link must show delivered entanglement rate and quality including heralding. A sampler must show task validity and classical comparison. A universal-computing proposal must show how loss, conditional logic, feed-forward, and correction compose. This role-specific denominator is the defense against photonic evidence transferring farther than it earned.

Jointly budget indistinguishability, loss, and feed-forward#

A product of transmission efficiencies is necessary but not sufficient. Two photons that arrive can still fail to interfere if their spectra, arrival times, polarization, or spatial modes differ. The path ledger should attach an indistinguishability measurement and uncertainty to the source pair and optical configuration used by the circuit. Combining a best detector efficiency from one wavelength with a source visibility from another experiment creates a device that has never existed.

Conditional photonic logic also consumes time. Detection, electrical discrimination, controller decision, and switch settling must complete before the affected mode reaches its routing point, or the architecture needs delay lines and memory. Those additions introduce propagation loss and occupy hardware. Put feed-forward latency in seconds beside optical path length and group velocity, then verify that the commanded switch acts on the intended time bin.

End-to-end evidence reports attempts, heralds, detected patterns, accepted events, logical successes, and wall-clock duration. Each ratio has a distinct denominator. A high conditional fidelity after rare heralding may be valuable, but it does not imply high application throughput. Conversely, a low-loss passive component does not establish a computing system without sources, interference, detection, control, and an error-management path. The integrated result is the closed chain, not the most favorable component row. Repeat the calculation with measured uncertainty and correlation bounds; otherwise multiplying central efficiencies can report spurious precision even when the optical path itself is completely specified.