Measurement-Based Quantum Computing
In the circuit model
you build a computation by doing things to qubits — firing gate after gate, weaving
entanglement as you go. Measurement-based quantum computing (MBQC) turns that picture
inside out. First you lay down one enormous, fixed sheet of
entanglement — a
cluster state — before the program even starts. Then you compute by doing the one thing
the circuit model saves for last: you measure, one qubit at a time, in cleverly chosen
directions. No gates run during the computation at all. Discovered by Raussendorf and Briegel in 2001,
this is the one-way quantum computer, and — astonishingly — it is exactly as powerful
as the circuit model.
The recipe: entangle first, then measure
MBQC splits every computation into two clean phases:
- Prepare the resource. Take a big lattice of qubits, put every one into
the state |{+}\rangle = \tfrac{1}{\sqrt2}(|0\rangle + |1\rangle), and apply
a controlled-Z
gate between every pair of neighbours. This one-off step spends all the entanglement up
front and produces a fixed, problem-independent cluster state.
- Drive with measurements. Now sweep across the lattice performing
single-qubit measurements.
The direction (basis angle) you measure each qubit in is the program. Crucially the angles
are adaptive: each choice depends on the random outcomes you have already seen — a
classical feed-forward loop running alongside the quantum hardware.
That is the whole machine. During phase 2 there are no two-qubit gates — the
entanglement was all made in phase 1, and measurement gradually consumes it. Because measurement is
irreversible, the resource is used up as you go and can never be rewound: hence one-way.
What a cluster state is, precisely
A graph state (a cluster state is the special case on a regular lattice) is defined by
a graph G=(V,E): put a qubit on every vertex, initialise each to
|{+}\rangle, and apply a controlled-Z along every edge:
|G\rangle \;=\; \Bigg(\prod_{(i,j)\in E} \mathrm{CZ}_{ij}\Bigg)\, |{+}\rangle^{\otimes |V|}.
The controlled-Z gate is symmetric and diagonal — it just flips the sign of the
|11\rangle amplitude:
\mathrm{CZ} \;=\; \begin{pmatrix} 1&0&0&0\\ 0&1&0&0\\ 0&0&1&0\\ 0&0&0&-1 \end{pmatrix},
\qquad \mathrm{CZ}\,|{+}{+}\rangle = \tfrac12\big(|00\rangle+|01\rangle+|10\rangle-|11\rangle\big).
All the \mathrm{CZ}s commute (they are diagonal), so the order of edges does
not matter — the resource depends only on the graph, not on how you built it.
The computation flows across the sheet
Picture the lattice as a fabric of entanglement. Measuring the leftmost column does not destroy the
computation — it pushes the logical information one step to the right, into the still-entangled
remainder, while imprinting a chosen operation on it. Column by column, the answer marches across the
sheet until it lands on the final, unmeasured column. Step through it:
The graph chosen for the resource, together with the pattern of
measurement angles, is what encodes a particular algorithm.
Worked example 1: one measurement = one gate
Here is the primitive that makes the whole model tick. Take a data qubit in an
unknown state |\psi\rangle = a|0\rangle + b|1\rangle and a fresh
resource qubit in |{+}\rangle. Entangle them with one controlled-Z:
\mathrm{CZ}\,\big(|\psi\rangle \otimes |{+}\rangle\big)
= \tfrac{1}{\sqrt2}\Big[\,a|0\rangle\big(|0\rangle+|1\rangle\big) + b|1\rangle\big(|0\rangle-|1\rangle\big)\Big].
Now measure the first qubit in the tilted basis
|{\pm_\varphi}\rangle = \tfrac{1}{\sqrt2}\big(|0\rangle \pm e^{-i\varphi}|1\rangle\big),
getting outcome s\in\{0,1\} (the + result is
s=0). Projecting and normalising, the second qubit is left in
|\psi_{\text{out}}\rangle \;=\; X^{s}\, H\, R_z(\varphi)\, |\psi\rangle,
\qquad R_z(\varphi) = \begin{pmatrix} 1 & 0 \\ 0 & e^{i\varphi} \end{pmatrix}.
Read what just happened. The data teleported one node to the right, and in transit it
was rotated by the single-qubit gate H\,R_z(\varphi) — and
you picked \varphi by choosing the measurement angle. One
measurement, one programmable gate. The only blemish is the random Pauli byproduct
X^{s}: half the time (s=1) an extra
X is stuck on the front, entirely determined by the coin-flip outcome. This
is a close cousin of
quantum teleportation —
MBQC is teleportation harnessed to compute.
Worked example 2: a concrete rotation and its byproduct
Let us feed a definite state through the primitive. Take |\psi\rangle = |{+}\rangle
(so a=b=\tfrac{1}{\sqrt2}) and measure the first qubit in the plain
X basis, i.e. angle \varphi = 0 so
|{\pm_\varphi}\rangle = |{\pm}\rangle. The two-qubit cluster is
\mathrm{CZ}\,|{+}{+}\rangle = \tfrac12\big(|00\rangle + |01\rangle + |10\rangle - |11\rangle\big).
Plug \varphi=0 into the general result. With
R_z(0)=I we get output
X^{s}H|{+}\rangle = X^{s}|0\rangle:
s=0:\ \ |\psi_{\text{out}}\rangle = |0\rangle,
\qquad\qquad s=1:\ \ |\psi_{\text{out}}\rangle = X|0\rangle = |1\rangle.
Each outcome occurs with probability \tfrac12. Notice the two jobs the single
measurement did at once: it rotated the logical state by H
(the chosen gate for \varphi=0) and transferred it to the
next qubit — but it also injected a random X byproduct that flips the answer
between |0\rangle and |1\rangle. To build a longer
gate sequence, the next qubit's measurement angle is flipped to
(-1)^{s}\varphi to pre-compensate for that known byproduct. That
angle-flip is the adaptive feed-forward: later bases genuinely depend on earlier outcomes.
Just as powerful as a circuit — and made for photons
MBQC is not a curiosity or a weaker relative of the circuit model. It is
polynomially equivalent to it: any quantum circuit on
n qubits can be compiled into a cluster state (of size polynomial in
the circuit's) plus a fixed measurement pattern, and vice versa. Universality is inherited entirely from
the resource state and the choice of measurement angles — a suitable graph plus adaptive single-qubit
measurements reproduces any
universal gate set.
Why does anyone care about a second, harder-to-picture model? Because it matches
hardware. On photonic
platforms, deterministic two-qubit entangling gates are notoriously difficult, but generating
entangled resource states offline and performing fast single-qubit measurements is comparatively easy.
MBQC plays exactly to those strengths — build the cluster once, then compute with measurements — which
is why it is the leading paradigm for photonic quantum computing. More broadly, it cleanly
separates the two ingredients of quantum computation: entanglement is the pre-made
resource, and measurement is the driver.
- computation runs in two phases: prepare a cluster/graph state
(|{+}\rangle on every node, \mathrm{CZ} on
every edge), then drive it with single-qubit measurements;
- no two-qubit gates run during the computation — all entanglement is created up
front, and measurement irreversibly consumes it (hence one-way);
- each measurement teleports the logical state one step and applies a chosen single-qubit rotation
H R_z(\varphi), up to a random Pauli byproduct
X^{s};
- measurement angles must be adaptive — later bases depend on earlier outcomes via
classical feed-forward — to steer past the random byproducts;
- MBQC is polynomially equivalent to the circuit model, and is the natural model
for photonic hardware.
The one-way computer runs on a strange kind of thrift. You begin by weaving a single, expensive sheet
of entanglement — a cluster state you will never mend or reuse — and then you compute by
tearing pieces off it, one measurement at a time. Every measurement is a tiny act of
demolition: it collapses that qubit forever, but in doing so it nudges the whole computation one step
forward, printing your chosen rotation onto the survivors. By the time the answer emerges on the far
edge, the sheet is gone. It is the opposite of the reversible, gate-by-gate circuit picture: here the
computation is what remains as the resource is consumed, like a sculpture that appears only as the
marble is chipped away.
Two traps. First, do not read "computing by measurement" as some cheap, measurement-only magic that
must be weaker than a real gate-based machine — MBQC is exactly as powerful as the
circuit model, provably equivalent up to polynomial overhead. Its power is not conjured from nothing:
it is banked in the entanglement of the cluster state, all of which is created up
front before a single measurement. Second — and this is the subtle one — the measurements are
not a fixed script. They must be adaptive: because every measurement leaves a random
Pauli byproduct, the angle of each later measurement has to be adjusted based on the outcomes already
seen (classical feed-forward). Strip out the adaptivity and measure every qubit in a pre-set basis, and
the byproducts never cancel — you get a random mess, not a computation.