The Deferred Measurement Principle

In short

The deferred measurement principle says that any intermediate measurement whose classical outcome controls a later quantum operation can be replaced by a fully quantum-controlled operation, with the measurement pushed to the very end of the circuit. Measure qubit a now and, if you got 1, apply U to the rest of the circuit — that is exactly equivalent to applying a controlled-U gate with a as the control, then measuring a at the end. Same output statistics, same reduced state on everything else. It is a theoretical simplification (you never have to reason about measurements in the middle of the circuit), a hardware-adaptation trick (every simulator and many older devices only support end-of-circuit measurement), and the single move that makes the textbook presentation of quantum teleportation legible. The price is that the deferred form needs a quantum-controlled gate where the original needed only a classically conditional one — on real hardware, sometimes the mid-circuit version is actually cheaper.

Open any paper on quantum teleportation. In the middle of the circuit, Alice measures her two qubits, gets two classical bits, sends them to Bob, and Bob does or does not apply an X and a Z depending on what those bits say. Clear enough as a protocol. But as an algebraic object, it is a nightmare: you have a quantum state, then a measurement that yields one of four outcomes, then a branch of four possible follow-up unitaries, and a proof of correctness that has to handle each branch separately.

Now open the same paper in its "equivalent deferred" form. Alice applies a CNOT and a Hadamard. Then a pair of controlled gates — controlled-X and controlled-Z — fire from Alice's qubits onto Bob's. Then, and only then, at the very end of the circuit, Alice measures. No classical bits flying around mid-circuit. No branched analysis. One unitary to prove correct.

Both circuits describe the same protocol and both produce the same statistics. The move that takes you from the first to the second is the deferred measurement principle. It says that a measurement mid-circuit that classically controls a later gate can always be re-expressed as a quantum-controlled gate plus a measurement at the end. Not approximately. Not in some limit. Exactly.

This chapter spells the principle out, proves it in a single line of algebra, walks through the teleportation rewrite in both forms side by side, and explains why — despite the mathematical equivalence — real hardware still cares whether you measured mid-circuit or deferred to the end.

What the principle says

Fix a circuit in which, at some point in the middle, you measure a qubit a in the computational basis. Call the outcome b \in \{0, 1\}. Downstream of that measurement, you apply one of two unitaries to the rest of the register: U_0 if b=0, and U_1 if b=1. Then (possibly) more gates run, and at the very end the remaining qubits are measured.

The deferred form replaces this whole arrangement. Instead of measuring qubit a now, apply a quantum-controlled gate with a as the control. The controlled gate fires U_0 on the target when a is |0\rangle and U_1 on the target when a is |1\rangle — call the combined operation C_a(U_0, U_1), a standard controlled-unitary. Leave a alone for the rest of the circuit, and measure it at the end alongside everything else.

Two circuits that produce identical output statistics. Left: measure $a$ mid-circuit, use the classical bit $b$ to choose which unitary to apply to $r$. Right: leave $a$ quantum, apply a controlled-$U$ with $a$ as control, and measure everything at the end.

The claim is a very specific one. Consider any observable you could compute from the outputs — the probability that the final measurement of r gives 0, the joint probability that a and r both give 1, the full output distribution over all the classical bits that come out of the circuit. These are identical for the two circuits. You cannot tell them apart by looking at outputs.

If that is true — and it is, and you will see the proof in the next section — then the mid-measurement circuit and the deferred circuit are genuinely interchangeable. Which one you draw is a question of convenience, not correctness.

The one-line proof

Write the state just before the mid-circuit measurement. Call the control qubit a and bundle everything else as a register r. The pre-measurement state is some superposition

|\psi\rangle_{ar} \;=\; |0\rangle_a |\phi_0\rangle_r + |1\rangle_a |\phi_1\rangle_r,

where |\phi_0\rangle_r and |\phi_1\rangle_r are unnormalised register states (their norms squared are the probabilities of getting 0 and 1 on a, and they can be entangled with a in any way whatsoever).

Why this decomposition is general: any joint state of a and r can be written as a sum over the computational-basis values of a, with whatever register state goes with each. The coefficients and the register states carry all the information; the decomposition loses nothing.

The mid-measurement path

Measure a. With probability p_0 = \langle \phi_0 | \phi_0 \rangle you get outcome 0, and the post-measurement register state is |\phi_0\rangle / \sqrt{p_0}. With probability p_1 = \langle \phi_1 | \phi_1 \rangle you get outcome 1, and the post-measurement register state is |\phi_1\rangle / \sqrt{p_1}.

Now apply U_b conditioned on the outcome. In the b=0 branch, the register becomes U_0 |\phi_0\rangle / \sqrt{p_0}. In the b=1 branch, the register becomes U_1 |\phi_1\rangle / \sqrt{p_1}.

The joint probability that a reported b and the subsequent measurement of r in any orthonormal basis \{|r_k\rangle\} gives outcome k is

P_{\text{mid}}(b, k) \;=\; p_b \cdot \bigl|\langle r_k | U_b |\phi_b\rangle / \sqrt{p_b}\bigr|^2 \;=\; \bigl|\langle r_k | U_b |\phi_b\rangle\bigr|^2.

Why the \sqrt{p_b} cancels: the post-measurement state is renormalised by that factor, so its squared overlap with |r_k\rangle picks up a 1/p_b, which exactly cancels the outer p_b that came from the probability of outcome b. The net result is a formula in which p_b does not explicitly appear.

The deferred path

C_a(U_0,U_1) |\psi\rangle_{ar} \;=\; |0\rangle_a U_0|\phi_0\rangle_r + |1\rangle_a U_1|\phi_1\rangle_r.

Now measure a and r at the end, in the computational basis and \{|r_k\rangle\} respectively. The joint probability that a reports b and r reports k is the squared amplitude on the |b\rangle_a |r_k\rangle_r basis state:

P_{\text{def}}(b, k) \;=\; \bigl|\langle b|_a \langle r_k|_r \bigl(|0\rangle_a U_0|\phi_0\rangle + |1\rangle_a U_1|\phi_1\rangle\bigr)\bigr|^2 \;=\; \bigl|\langle r_k | U_b |\phi_b\rangle\bigr|^2.

Why only one of the two terms survives: \langle b|0\rangle is 1 if b=0 and 0 otherwise, and similarly for \langle b|1\rangle. Exactly one of the two terms in the sum contributes, and it is the term where the controlled unitary fired U_b as demanded.

The identity

The two joint distributions are literally the same:

P_{\text{mid}}(b, k) \;=\; \bigl|\langle r_k | U_b |\phi_b\rangle\bigr|^2 \;=\; P_{\text{def}}(b, k).

Every observable built from these outputs — marginals, conditionals, expectation values, histograms — is therefore the same in both circuits. Done. That is the entire content of the deferred measurement principle.

What the proof actually uses. Three ingredients: (i) postulate 3 of quantum mechanics, the Born rule for computing measurement probabilities; (ii) the definition of a quantum-controlled gate as the diagonal sum |0\rangle\langle 0| \otimes U_0 + |1\rangle\langle 1| \otimes U_1; and (iii) linearity — the controlled gate acts on the whole superposition term by term. No extra postulates, no approximations, no limits. It is an algebraic identity.

Tracking probabilities line by line. The $p_b$ that appears in the mid-measurement bookkeeping cancels after renormalisation; the deferred path never produces it. The two final probabilities are identical term by term.

Teleportation in both forms

The canonical worked example, and the one that makes the principle concrete, is quantum teleportation. Alice has a qubit |\psi\rangle = \alpha|0\rangle + \beta|1\rangle she wants to send to Bob. They share a Bell pair. In the textbook version, Alice does a Bell-basis measurement on her qubit and her half of the Bell pair, gets two classical bits, sends them to Bob, and Bob applies X^{b_0} Z^{b_1} to his half depending on the bits.

The mid-measurement version

The circuit has three qubits. Call them q_A (Alice's to-send qubit, in state |\psi\rangle), e_A (Alice's half of the Bell pair), e_B (Bob's half of the Bell pair, already entangled with e_A). Alice runs a CNOT from q_A to e_A, then a Hadamard on q_A, then measures q_A and e_A in the computational basis, producing two classical bits b_1 (from q_A) and b_0 (from e_A). She sends (b_1, b_0) to Bob. Bob applies X^{b_0} to e_B, then Z^{b_1} to e_B. Bob's qubit is now in state |\psi\rangle.

The deferred version

Replace the two measure-and-classically-control blocks with two quantum-controlled gates. The classical bit b_0 that controlled the X on Bob's qubit is now the quantum state of e_A; wire a CNOT from e_A to e_B. The classical bit b_1 that controlled the Z on Bob's qubit is now the quantum state of q_A; wire a controlled-Z from q_A to e_B. Only then, after the two controlled gates, do Alice's qubits get measured — and the measurements are now at the end of the circuit, where they no longer affect anything.

The two ways to draw teleportation. In the mid-measurement form (top), Alice measures at the cut, classical bits fly across, Bob conditions. In the deferred form (bottom), everything is quantum until the very end — the Bell measurement becomes a CNOT, a Hadamard, and two deferred measurements, while the classical conditions become quantum CNOT and CZ gates.

The deferred circuit is the one that lives in proofs. When someone writes "teleportation is (\text{H} \otimes I \otimes I) \cdot (\text{CNOT} \otimes I) \cdot (I \otimes \text{CNOT}) \cdot (I \otimes I \otimes \text{CZ}) applied to |\psi\rangle \otimes |\Phi^+\rangle," they are writing the deferred form. It is one clean unitary followed by measurements. The mid-measurement form has four cases to analyse; the deferred form has one.

The mid-measurement form is what the hardware actually runs when Alice and Bob are genuinely separated — there is real classical communication and Bob cannot apply a gate until the bits arrive. The deferred form is what theorists reason about. The deferred-measurement principle says the two forms compute the same function, so you can prove correctness once, in the form that is cleanest, and trust both implementations.

A toy syndrome-correction example

A second, smaller example drives the same point home. Take a single data qubit and a single syndrome qubit. Some noise process might have flipped the data qubit; the syndrome qubit is supposed to tell you whether it did, and if so, you apply an X to the data qubit to undo the flip.

In the mid-measurement form: measure the syndrome qubit in the computational basis. If the outcome is 1, apply X to the data qubit. If the outcome is 0, do nothing. Measure the data qubit at the end.

In the deferred form: do a CNOT from syndrome to data. That is it. Measure both qubits at the end.

Why this works. A CNOT with the syndrome as control and the data as target applies X to the data iff the syndrome is |1\rangle, and does nothing iff the syndrome is |0\rangle. That is exactly the action the classical conditional was going to perform — but done quantumly, with the syndrome qubit itself (rather than its measurement outcome) driving the gate. The classical "if syndrome=1 then flip data" becomes the quantum "CNOT from syndrome to data," and the syndrome measurement moves to the end of the circuit where it no longer controls anything.

Why it matters — three concrete uses

1. Theoretical simplification

A circuit with mid-circuit measurements is a hybrid classical-quantum object: its description mixes unitary evolution, random classical outcomes, and classically conditioned subsequent unitaries. Correctness proofs have to branch on outcomes, and the algebra becomes a case analysis. Deferring every measurement to the end gives you a pure unitary circuit with terminal measurements — one object to reason about. Every modern quantum-algorithms textbook uses this as the default presentation for exactly this reason.

2. Hardware and simulator adaptation

Most classical simulators of quantum circuits are written to evolve a pure state vector under a sequence of unitary gates and then measure. Mid-circuit measurement requires either branching the simulation (re-running the circuit for each outcome) or tracking a probabilistic mixture of post-measurement states. Both are expensive. The deferred form needs neither — it is one unitary followed by one measurement, which is the simulator's native mode. For this reason, every quantum-algorithm paper you read presents its circuits in deferred form, even when the "real" hardware would run them with mid-circuit measurements.

Some hardware also lacks the infrastructure to do mid-circuit measurement at all. Until recently, most superconducting-qubit machines could only measure at the end of a pulse sequence. Any algorithm described with mid-circuit measurements had to be compiled into its deferred equivalent before it could run. The deferred measurement principle is what guarantees the compilation is correct.

3. Dynamic-circuit support is actually a big deal

The interesting twist: on modern hardware, running the deferred form of some circuits is more expensive than the mid-measurement form. A controlled-U gate takes several CNOTs and single-qubit rotations to synthesise; a classically conditioned gate takes just one U plus a classical bit fed back. IBM's "dynamic circuits" feature (launched 2022 on their superconducting stack), Quantinuum's native mid-circuit measurement and feed-forward on their trapped-ion machines, and the measurement-based-quantum-computation model (Raussendorf-Briegel, 2001) all take advantage of real mid-circuit measurement to save gate count and circuit depth.

So the deferred-measurement principle is a mathematical equivalence, but the two sides are not equally cheap to run on any given architecture. On older or simulation-bound stacks, defer everything. On a platform with fast mid-circuit measurement and feed-forward, keep the measurement mid-circuit. The principle tells you the transformation is safe; the hardware tells you which direction to transform.

Worked examples

Example 1 — Bell-state preparation with "deferred reset"

Setup. You want to prepare a Bell state |\Phi^+\rangle = \tfrac{1}{\sqrt{2}}(|00\rangle + |11\rangle) starting from two qubits in an unknown state. One way to do this robustly is: measure the first qubit, and if the outcome is 1, apply X to send it to |0\rangle; then apply X to the second qubit conditionally similarly; then do the usual \text{CNOT}\cdot(H \otimes I) preparation. Rewrite the reset step in deferred form.

Step 1 — the mid-measurement circuit. Qubit q_0 is in some unknown state |\psi\rangle. Measure it: with some probability you get 0 (state is already |0\rangle, do nothing) or 1 (state is now |1\rangle; apply X to send it to |0\rangle). Why the conditional X works: after measurement, the qubit is definitely in a computational-basis state. If that state is |1\rangle, applying X turns it into |0\rangle. This is a classical reset built from quantum primitives — possible precisely because measurement forces the qubit out of superposition.

Step 2 — apply deferred measurement. Replace the classically-conditioned X with a quantum-controlled X — but conditioned on what? The original classical bit was the measurement outcome of q_0. So the deferred version uses q_0 itself as the control: apply a CNOT with q_0 as the control and q_0 as the target... but a qubit cannot be its own control.

Step 3 — the right rewrite. You need an auxiliary qubit. Introduce an ancilla a in |0\rangle. Do a CNOT from q_0 to a (this copies the computational-basis value of q_0 onto a, in a way that is reversible because both end up correlated). Now use a as the quantum control for the X on q_0: if a is |1\rangle, flip q_0. Measure a at the end. Why the ancilla is needed: deferred measurement replaces a classical wire with a quantum control line, but the control and target of a quantum gate must be different qubits. Copying the would-be-classical bit onto an ancilla gives you a legal control qubit.

Step 4 — check the net effect. Start with |\psi\rangle_{q_0} |0\rangle_a = (\alpha|0\rangle + \beta|1\rangle)_{q_0} |0\rangle_a. After CNOT from q_0 to a: \alpha|00\rangle + \beta|11\rangle. After CNOT from a to q_0: \alpha|00\rangle + \beta|01\rangle = |0\rangle_{q_0}(\alpha|0\rangle + \beta|1\rangle)_a.

Result. Qubit q_0 is now |0\rangle with certainty; the original quantum amplitudes of q_0 have been transferred to the ancilla a, and measuring a in the computational basis reproduces the same statistics as measuring q_0 before the reset. The deferred-measurement circuit did exactly what the mid-measurement circuit did — it reset q_0 to |0\rangle — but without ever performing a measurement until the end.

The reset-to-$|0\rangle$ operation in its mid-measurement form and in its deferred form. The deferred form uses one extra qubit (an ancilla) but never performs a measurement until the very end.

Example 2 — syndrome correction, 2 qubits

Setup. A data qubit is in some state |\chi\rangle. A bit-flip error may or may not have happened; the syndrome qubit holds the bit 1 if an error happened and 0 otherwise. The correction is: if the syndrome is 1, apply X to the data; if the syndrome is 0, do nothing. Write this first as a mid-measurement circuit, then as a deferred circuit.

Step 1 — the mid-measurement form. Measure the syndrome qubit in the computational basis. Outcome s \in \{0, 1\}. If s = 1, apply X to the data. Otherwise, do nothing. Leave the data qubit in whatever state it ends up in.

Step 2 — write the unitary in each branch. When s = 0, the operation on the data is I. When s = 1, the operation on the data is X. So we need a controlled operation with syndrome as the control and with I on |0\rangle-branch and X on |1\rangle-branch — that is exactly a CNOT with syndrome as control and data as target. Why the CNOT does it: the CNOT is by definition "do nothing if control is |0\rangle, do X if control is |1\rangle." That matches the classical conditional on the syndrome exactly, but done as a quantum gate.

Step 3 — the deferred form. Apply a CNOT from the syndrome qubit to the data qubit. Then, at the end of the circuit, measure both qubits.

Step 4 — verify on a specific case. Suppose the syndrome qubit is in the superposition \tfrac{1}{\sqrt{2}}(|0\rangle + |1\rangle) (which a realistic syndrome-extraction circuit might produce) and the data qubit is in state |\chi\rangle. Joint state before: \tfrac{1}{\sqrt{2}}(|0\rangle + |1\rangle)_s |\chi\rangle_d = \tfrac{1}{\sqrt{2}}(|0\rangle_s |\chi\rangle_d + |1\rangle_s |\chi\rangle_d). After CNOT from syndrome to data: \tfrac{1}{\sqrt{2}}(|0\rangle_s |\chi\rangle_d + |1\rangle_s X|\chi\rangle_d). Now measure syndrome. If outcome is 0: data is |\chi\rangle (no correction needed). If outcome is 1: data is X|\chi\rangle (the correction has been applied). Both outcomes occur with probability 1/2, and each produces the correct post-correction data state.

Result. The deferred circuit — one CNOT plus terminal measurements — produces the same joint output distribution as the mid-measurement circuit with the classical conditional X. In fact the deferred form is strictly more transparent here: it makes it obvious that both branches of the syndrome superposition get the correct correction applied, whereas the mid-measurement form required you to analyse the two outcomes separately.

What this shows. Syndrome-and-correct structures in error correction are almost always presented in the deferred form when proving correctness, exactly because the CNOT captures the conditional cleanly. They are almost always executed in mid-measurement form on real hardware, because the CNOT is cheaper than the full end-of-circuit analysis would be when each syndrome extraction runs thousands of times per computation.

Common confusions

"Deferring changes the outcome of the mid-circuit measurement." It does not. In the deferred form, the final measurement of a produces exactly the same classical bit b with the same probability as the mid-circuit measurement did. The distribution over (b, \text{everything else}) is identical. The deferred form does not erase the measurement — it just delays when you read the output.
"Deferring changes whether the circuit is 'really' measured." In quantum mechanics, every measurement-like interaction changes the joint state; what makes something "a measurement" is whether a classical record is produced. In the deferred form you postpone producing that record, but the coherence of qubit a with the rest of the register is still present until you read out. This is a real difference in intermediate states (the deferred form keeps the whole thing pure), but not in the final output distribution.
"The deferred form is always better." No. Mathematically the two forms are equivalent, but on a physical device the deferred form generally requires a controlled-U gate where the mid-measurement form used a classically conditional gate. On a trapped-ion machine with \mus feed-forward latency and a native X, the mid-measurement form can easily be 2-4x cheaper in gate count. On a superconducting device without fast readout and feed-forward, only the deferred form is possible. Use the right form for the right hardware.
"Deferring works for any measurement, not just computational-basis." Yes, provided you preface the deferred measurement with the same basis-rotation unitary you would have used for the mid-circuit measurement. Measuring in the X basis is (\text{H on } a) \cdot (\text{measure } a \text{ in Z}). In the deferred form, keep the Hadamard mid-circuit, but replace the mid-circuit Z-basis measurement with a quantum control off a, and only measure a in Z at the end. The rotation unitary is not the measurement; it is part of the gate sequence and defers just fine.
"Deferring makes debugging easier." The opposite is often true. In the mid-measurement form, you can print the classical bit after the measurement and see what happened. In the deferred form, the corresponding information is locked inside the quantum state until the final measurement, and you cannot read it out without running a separate circuit. Real hardware engineers often prefer mid-circuit measurement precisely because it is observable during execution.
"The deferred-measurement principle means mid-circuit measurement is useless." It would be, if all you cared about were input-output statistics. In practice, mid-circuit measurement is essential for error correction (you must measure syndromes fast and correct before more errors accumulate), measurement-based quantum computation (the whole computation is a sequence of mid-circuit measurements on a cluster state), and dynamic quantum-classical algorithms (where the classical bit chooses what circuit to run next, and you cannot know that circuit until you see the bit). The principle tells you when mid-circuit and terminal measurement are interchangeable; it does not tell you they always are.

Going deeper

Everything you need to use the deferred-measurement principle in reading and writing quantum algorithms is in the sections above: the statement, the one-line proof, the teleportation rewrite, the syndrome-correction toy. The rest of this section goes further: the formal proof stated against the four postulates, the precise relationship between deferred measurement and the sibling principle of implicit measurement (the next chapter), the measurement-based quantum computation programme that lives on the other side of the trade, and the hardware-feedback landscape that determines which form is cheaper on which device.

Formal proof against postulate 3 plus linearity

The Kraus operators for a projective measurement in the computational basis on qubit a are K_0 = |0\rangle\langle 0|_a and K_1 = |1\rangle\langle 1|_a. The operation "measure a, get outcome b, then apply U_b to the rest" is the CPTP (completely positive trace preserving) map

\mathcal{E}_{\text{mid}}(\rho) \;=\; \sum_b (I_{\text{else}} \otimes U_b) \cdot (K_b \otimes I_{\text{else}}) \cdot \rho \cdot (K_b^\dagger \otimes I_{\text{else}}) \cdot (I_{\text{else}} \otimes U_b)^\dagger.

The deferred version is the unitary C_a(U_0, U_1) = |0\rangle\langle 0|_a \otimes U_0 + |1\rangle\langle 1|_a \otimes U_1 applied as a conjugation, then a terminal measurement in the Z basis on a. Unravelling the terminal measurement yields the same sum of Kraus terms as \mathcal{E}_{\text{mid}}, provided the classical outcome is recorded (so we are comparing the same bit against the same bit). Straightforward algebra (expand C_a(U_0,U_1) \rho C_a(U_0,U_1)^\dagger and trace against |b\rangle\langle b|_a \otimes I to get the joint (b, \text{outcome}) distribution) reproduces the same expression. Linearity over the decomposition of \rho in the computational basis of a is what lets the algebra go through term by term.

This is the formal statement that tools like Stinespring dilation systematise: every non-unitary map (including measurement plus conditional evolution) can be re-expressed as a unitary on a larger system followed by a terminal measurement or a partial trace. The deferred-measurement principle is a special case where the "larger system" is just the original qubit a itself.

Implicit versus deferred — related, not the same

The next chapter (implicit measurement) discusses the principle that any qubit left unmeasured at the end of a circuit is statistically equivalent to one that was measured and had its outcome discarded. It is tempting to conflate the two principles. They are distinct:

Deferred measurement is about moving a measurement from mid-circuit to the end of the circuit, keeping the classical bit it produces. Same information extracted, different point in time.
Implicit measurement is about ignoring a qubit at the end of the circuit. The information is not extracted; the partial trace takes its place.

Combine them and you get a clean slogan: every mid-circuit measurement whose outcome you ultimately discard can be replaced by "do nothing" on that qubit — the deferred-measurement principle first moves the measurement to the end, and then the implicit-measurement principle replaces the terminal measurement by discarding the qubit. The slogan is the working version of what happens to ancilla qubits at the end of many quantum algorithms.

Measurement-based quantum computation — the other side

There is a whole computational model (measurement-based quantum computation, MBQC, due to Raussendorf and Briegel in 2001) in which the only operations are mid-circuit single-qubit measurements on a large entangled "cluster state." The classical outcomes of earlier measurements determine the bases in which later measurements are performed; this adaptive structure is what drives the computation forward.

The deferred-measurement principle applies here, too: any MBQC computation can in principle be rewritten as a unitary circuit with terminal measurements, where every mid-circuit measurement is replaced by a quantum-controlled gate. But doing so throws away the whole point of MBQC. The model derives its practical advantages — in photonic hardware, it is easier to prepare a big entangled state once than to do gates on demand — from keeping the measurements mid-circuit. The deferred-measurement principle describes an equivalence of models, not a reason to prefer one.

Hardware feedback latency

On a real device, a classically-conditioned gate (the mid-measurement form) is a pulse sequence that depends on the measurement outcome becoming available during the circuit. This "feed-forward" capability has a latency — the time between the measurement pulse ending and the next gate pulse that can use its outcome. On superconducting hardware in 2024, this latency is in the hundreds of nanoseconds to a few microseconds. On trapped-ion hardware, it can be shorter. On optical MBQC hardware with fast switches, sub-microsecond.

If the feed-forward latency is longer than the coherence time of the qubits that are idling while waiting, the mid-measurement form fails — the data qubits decohere before the correction can be applied. In that regime, the deferred form is mandatory. If feed-forward is fast enough, the mid-measurement form saves the gate overhead of the quantum-controlled operation, and it is preferable.

Quantinuum's H1 and H2 trapped-ion machines, which have been the hardware flagship for mid-circuit-measurement demonstrations since 2020, exploit this advantage to run a wide range of error-correction and dynamic-circuit protocols in their native, mid-measurement form. IBM's dynamic-circuit programme on superconducting hardware, launched publicly in 2022, is a similar bet — pushing feed-forward latency down so that mid-circuit measurement becomes cheap. The deferred-measurement principle provides the safety net: whatever the hardware runs, the theoretical analysis lives in the deferred form, and translating back and forth is always correct.

Where this leads next

Implicit measurement — the sibling principle about unmeasured qubits at the end of a circuit.
Classical control — the notation for classically conditioned gates in circuit diagrams, and its quantum-control counterpart.
Quantum teleportation — the canonical example, presented fully in both mid-measurement and deferred forms.
Measurement-based quantum computation — the model that deliberately lives on the mid-circuit side.
Controlled-U gates — the primitive that makes the deferred form possible.
Quantum error correction — introduction — where mid-circuit measurement is non-negotiable and the deferred form becomes a proof device rather than an implementation guide.

References

Nielsen and Chuang, Quantum Computation and Quantum Information (2010), §4.4 — the standard statement of the deferred-measurement principle. Cambridge University Press.
John Preskill, Lecture Notes on Quantum Computation, Ch. 4 — measurement, deferred measurement, and MBQC. theory.caltech.edu/~preskill/ph229.
Wikipedia, Deferred Measurement Principle — encyclopaedic version with the generic circuit diagram.
IBM Quantum, Dynamic circuits documentation — the engineering perspective on mid-circuit measurement and feed-forward latency.
Qiskit Textbook, Measurement and classical control — practical examples of mid-measurement vs deferred circuits in simulator and hardware.
Raussendorf and Briegel, A one-way quantum computer (2001) — the foundational paper on measurement-based computation. arXiv:quant-ph/0010033.