Phase Shifters

The last lesson ended on a cliffhanger: the MZI converts a phase \varphi into a power — but something has to set \varphi, on demand, under electronic control. That something is the phase shifter, and it is where the clean linear algebra of interferometers collides with messy materials physics. Every phase shifter plays the same trick: change the effective refractive index n_{\text{eff}} seen by the guided mode over a length L, so the accumulated phase \varphi \;=\; \frac{2\pi}{\lambda}\, \Delta n_{\text{eff}}\, L moves by however much you ask. The whole game is in how you move \Delta n_{\text{eff}}: with heat, with charge, or with an electric field acting on the crystal itself. Three mechanisms, three utterly different personalities — one slow and clean, one fast and dirty, one fast and clean but picky about its material — and every photonic processor you will meet in this course is shaped by which of the three its designers could afford.

Heat: the thermo-optic shifter

Silicon's refractive index rises with temperature at dn/dT \approx 1.86\times 10^{-4}\ \mathrm{K^{-1}} — a big coefficient, as materials go. Deposit a thin metal resistor a micron or two above the waveguide, drive a current through it, and the warmed silicon underneath picks up extra index. The numbers are friendly. For a heater of length L, a full \pi shift needs

\Delta T \;=\; \frac{\lambda}{2\,L\,(dn/dT)} \;=\; \frac{1.55\ \mu\mathrm{m}}{2 \times (100\ \mu\mathrm{m}) \times 1.86\times 10^{-4}\ \mathrm{K^{-1}}} \;\approx\; 42\ \mathrm{K},

an entirely practical temperature swing, typically bought for P_\pi \approx 10\text{–}20\ \mathrm{mW} of steady heater power (a few mW with thermal-isolation trenches). The shifter adds essentially zero optical loss — the light never touches the metal — and its phase is pure, with no amplitude side effects. The price is time and power: heat diffuses on microsecond timescales, so the thermo-optic shifter tops out around a megahertz, and it burns its milliwatts continuously, just to hold a phase still. A mesh with a thousand of them dedicates watts to standing perfectly still. Slow, low-loss, power-hungry: the thermo-optic shifter is the workhorse for configuring a circuit, hopeless for signalling through one.

Charge: the plasma-dispersion shifter

Silicon may be poor at emitting light, but as a semiconductor-style trick it has one great gift: you can flood it with free carriers, and free electrons and holes both lower the refractive index (the plasma-dispersion effect, quantified for silicon by Soref and Bennett in 1987). Build the waveguide as a p–n junction and you can steer carrier density electrically, two ways:

Injection — forward-bias a p-i-n diode and pour carriers into the core. The index change is large (\Delta n \sim 10^{-3}), but you must wait for minority carriers to recombine before the phase relaxes: nanoseconds, capping speed near a gigahertz.
Depletion — reverse-bias a p–n junction and sweep carriers out of the mode. The change is ten times weaker, but there is no recombination to wait for: the junction responds in picoseconds, and depletion shifters run at 25–50 GHz and beyond. This is the engine of nearly every silicon modulator shipping today.

The catch is written into the same physics: free carriers do not just shift the phase of light, they absorb it. Index and loss move together, so a plasma-dispersion shifter is unavoidably lossy (order 1 dB for a \pi), and its loss changes as it switches, smearing unwanted amplitude modulation and chirp onto a signal that was supposed to be pure phase. Fast, efficient, compromised: the classic engineering bargain.

Field: the Pockels shifter

The aristocrat of the family uses the Pockels effect: in certain crystals an applied electric field changes the refractive index directly and linearly, \Delta n \;=\; -\tfrac{1}{2}\, n^3\, r\, E, with r the electro-optic coefficient — the field distorts the electron cloud of the lattice itself, a response covered by the physics of fields in matter. No heat flows, no carriers move: the effect is femtosecond-fast (real devices are limited only by their electrodes, beyond 100 GHz), adds no absorption, and consumes only the tiny energy of charging a capacitor. The catch is the guest list. The Pockels effect exists only in crystals without inversion symmetry, and the star performer is lithium niobate, \mathrm{LiNbO_3}, with r_{33} = 31\ \mathrm{pm/V} — long the heart of fibre-optic modulators as bulky diced crystals, and now available as thin-film lithium niobate (TFLN) bonded onto oxide, where the tight optical mode drops drive voltages by an order of magnitude. Newer contenders such as barium titanate (r \sim 900\ \mathrm{pm/V}) push further. Fast, clean, low-power — but never in plain silicon, so you must bring the material to the chip.

Because of a symmetry, not a shortage of trying. The Pockels effect is a second-order nonlinearity: the polarisation response contains a term \chi^{(2)} E^2, and flipping the field (E \to -E) must flip the physical response. In a crystal with inversion symmetry — where the lattice looks identical when reflected through a point — the flipped configuration is indistinguishable from the original, so the response must satisfy \chi^{(2)} = -\chi^{(2)} = 0. Silicon's diamond lattice is exactly such a centrosymmetric crystal, so its Pockels coefficient is identically zero: a pure group-theory veto. Engineers have attacked the veto at its premise — straining the lattice with a silicon-nitride overlayer to break the symmetry (a measurable but small effect), or applying a DC field to bootstrap the third-order Kerr nonlinearity into an "effective" Pockels response. Nothing yet competes with simply bonding a sliver of lithium niobate where you need it — which is why "which material hosts the phase shifter" is one of the defining platform wars of commercial photonics.

Figures of merit: V_π and V_π·L

Put a phase shifter in one arm of an MZI and drive it with voltage V. The voltage that produces a \pi shift — enough to swing the interferometer from fully bright to fully dark — is V_\pi, the single most-quoted number on any modulator datasheet. But V_\pi alone is a half-truth: since phase accumulates along the device, doubling the length halves the voltage. The honest, length-independent figure of merit is the product V_\pi \cdot L \quad [\mathrm{V\cdot cm}], a property of the mechanism and cross-section, not of one device. Silicon depletion junctions manage 1–2 V·cm; thin-film lithium niobate about 2 V·cm with far lower loss. A worked example: a TFLN shifter with V_\pi\cdot L = 2\ \mathrm{V\cdot cm} built 5 mm long has V_\pi = 2/0.5 = 4\ \mathrm{V} — driveable by (nearly) ordinary electronics. Below, the MZI transfer curve against drive voltage: drag V_\pi and watch the whole personality of the device stretch.

Now place the three mechanisms on one map. Every shifter wants to sit in all three corners at once — fast, low-power, low-loss — and on plain silicon no mechanism manages more than two:

Mechanism	Speed	Efficiency	Excess loss	The catch
Thermo-optic	~1 MHz (µs)	P_π ≈ 2–20 mW, held continuously	≈ 0 dB	slow; static power; thermal crosstalk
Plasma dispersion (depletion)	25–50+ GHz	V_π·L ≈ 1–2 V·cm	~1 dB, and it varies with drive	loss–phase coupling, chirp
Pockels (TFLN, BTO)	> 100 GHz	V_π·L ≈ 2 V·cm (TFLN)	very low	not silicon — heterogeneous material

Worked numbers, live

The three design equations of this lesson fit in a dozen lines. Run them, then change the numbers to the platform of your choice:

const lambda = 1.55e-6; // wavelength, m const dndT = 1.86e-4; // silicon thermo-optic coefficient, 1/K // 1) Thermo-optic: temperature swing for a pi shift, vs heater length for (const Lum of [50, 100, 200]) { const dT = lambda / (2 * (Lum * 1e-6) * dndT); console.log("L = " + Lum + " um -> dT for pi = " + dT.toFixed(1) + " K"); } // 2) Vpi from the Vpi·L figure of merit function vpi(vpiL_Vcm: number, L_mm: number): number { return vpiL_Vcm / (L_mm / 10); // convert mm to cm } console.log("Si depletion, 1.4 V·cm, 4 mm: Vpi = " + vpi(1.4, 4).toFixed(1) + " V"); console.log("TFLN, 2.0 V·cm, 5 mm: Vpi = " + vpi(2.0, 5).toFixed(1) + " V"); // 3) The energy ledger: holding a phase for 1 s const thermoHold = 15e-3 * 1.0; // 15 mW heater, held 1 s const pockelsHold = 0.5 * 100e-15 * 4 * 4; // CV²/2, C = 100 fF charged once to 4 V console.log("thermo-optic hold for 1 s: " + (thermoHold * 1e3).toFixed(1) + " mJ"); console.log("Pockels hold for 1 s: " + (pockelsHold * 1e15).toFixed(1) + " fJ");

That last comparison — millijoules against femtojoules to hold the same phase — is nine orders of magnitude, and it is why the energy accounting of photonic processors in Module 5 hinges so heavily on which phase shifter sits in the mesh.

A datasheet boasting "V_π = 1 V!" tells you almost nothing until you read the next line. Because phase accumulates along the device, any mechanism can reach any V_\pi you like — just make it longer. A 1 V silicon shifter with V_\pi\cdot L = 2\ \mathrm{V\cdot cm} is simply 2 cm long, and it has paid for its low voltage three times over: in optical loss (loss scales with length, so dB stack up), in footprint (2 cm of serpentine waveguide dwarfs the rest of the circuit), and in capacitance/transit time (longer electrodes are harder to drive fast). Comparing shifters by V_\pi alone is comparing engines by their gear ratio. The transferable habits: compare mechanisms by V_\pi\cdot L together with the loss per \pi (dB·V or dB per π), and remember that for thermo-optic shifters the analogous honest number is P_\pi, the held power for a π — a continuing bill, not a one-off charge.

Where this goes next