Principle of Deferred Measurement

A circuit that measures a qubit mid-computation and uses that bit to decide whether to fire a later gate looks, at first, like an irreducibly two-phase object — quantum, then classical, then quantum again. The principle of deferred measurement says it is not. Every such measure-and-feed-forward step can be rewritten so the measurement slides all the way to the end, with the classically-controlled gate replaced by a fully coherent quantum-controlled gate driven by the still-unmeasured qubit. The output statistics on every register are identical, exactly, not approximately. Two facts make this work. First, a measurement in the computational basis is diagonalising: it keeps the populations $|0\rangle\langle0|$ and $|1\rangle\langle1|$ and destroys the coherences between them. Second, a controlled gate whose control is destined to be measured in that same basis only couples to those surviving populations — the off-diagonal terms it creates are annihilated by the terminal measurement before they can matter. So "measure then act classically" and "act coherently then measure" compute the same channel; this article is the proof and a tour of what it buys. The payoff runs two ways. Conceptually, deferred measurement — with its companion, the principle of implicit measurement — proves that intermediate measurements add no computational power: any circuit equals one unitary block followed by a single terminal layer of measurements, the normal form behind the definition of BQP. Practically, it is the licence under which teleportation, syndrome extraction, gate teleportation, and measurement-based computation are all understood as one coherent circuit wearing a classical disguise.

Key takeaways

Any mid-circuit measurement whose result only conditions a later gate can be postponed to the end, replacing the classically-controlled gate by a coherent controlled- $U$ on the still-live qubit.
It works because a computational-basis measurement is diagonalising: the coherences a coherent control creates are exactly the off-diagonal terms the terminal measurement erases, so early and late measurement give identical statistics.
The proof is one projector identity, $(M_m\otimes I)\,\Lambda_q(U)=M_m\otimes U^{m}$ , applied branch by branch, giving equal probabilities and post-states.
Deferring costs a qubit: the control must survive to the end, so a measured-and-reused qubit needs a fresh ancilla to carry its bit.
With the principle of implicit measurement it puts every circuit in a normal form — one unitary, then a terminal measurement layer — the basis for treating intermediate measurement as free in complexity theory.
Key idea: Measure-then-classically-control equals coherently-control-then-measure — replace each classical wire feeding a gate by a quantum control and slide the measurement to the end, because a computational-basis measurement kills exactly the coherences the coherent control would otherwise expose.

The core insight

Measure-then-classically-control equals coherently-control-then-measure: replace each classical wire feeding a gate by a quantum control and slide the measurement to the end, since a computational-basis measurement is diagonalising and annihilates precisely the coherences the coherent control creates.

Where this lives today

The principle is now a daily engineering decision. Real superconducting and trapped-ion machines have gained mid-circuit measurement with real-time feed-forward — IBM's "dynamic circuits" and Quantinuum's H-series measure, decode, and condition a later gate within a coherence time — and the principle tells a compiler when it may trade feed-forward for a coherent controlled gate: defer to avoid slow classical loops, or measure early and reset to save qubits. The choice is sharp in error correction: Google's 2023 below-threshold surface-code runs feed syndromes forward with real-time decoders, precisely where deferring everything would cost too many ancillas, while fault-tolerant gate constructions descend from Gottesman–Chuang gate teleportation, deferred by design. Photonic measurement-based platforms at PsiQuantum and Xanadu take it to the limit, computing almost entirely by adaptive measurement on cluster states — and deferred measurement is the guarantee that this equals a coherent circuit.

The physical picture

Picture a control qubit and the two ways its "decision" reaches a target. In the early-measurement circuit a meter collapses the control to $|0\rangle$ or $|1\rangle$ , and a classical wire carries the bit forward to switch a gate $U$ on the target on or off. In the deferred circuit nothing is measured yet: the control stays coherent, a quantum wire runs into a controlled- $U$ , and only at the very end does a meter read it out. The principle claims these two circuits are the same operation — the first figure draws that as two interchangeable diagrams.

Figure

The deferred-measurement rewrite Mz → Λc(U) ≡ Λq(U) → Mz measuring the control then applying U classically equals a coherent controlled-U followed by a terminal measurement with identical output on every register

Why is this allowed? A computational-basis measurement does one thing to the control's density matrix: it keeps the two diagonal populations and deletes the off-diagonal coherences. A coherent controlled- $U$ entangles control and target, but look only at the target and it is already a classical mixture over the control's values — the entanglement decohered it exactly as an early measurement would, and the terminal read merely reveals the branch. The second figure makes this quantitative: prepare the control so outcome $1$ has probability $0.70$ , and the gate leaves the target in a $0.30 / 0.70$ mixture whether or not the control has been measured.

Figure

With the control prepared so that outcome 1 has probability 0.70 a coherent controlled-U leaves the target in the mixture 0.30 ρχ + 0.70 U ρχ U† and the control off-diagonal coherences are erased by the terminal measurement so the statistics match early measurement exactly

The core equations

The whole principle is one equivalence between two ways of wiring a decision. Write $M_z$ for a computational-basis measurement of the control, $\Lambda_c(U)$ for a classically-controlled $U$ (fire $U$ iff the bit is $1$ ), and $\Lambda_q(U)$ for the coherent quantum-controlled gate. The identity swaps their order.

M_z \to \Lambda_c(U) \;\;\equiv\;\; \Lambda_q(U) \to M_z \qquad (1)

where: $M_z$ — computational-basis ( $Z$ ) measurement of the control; $\Lambda_c(U)$ — classically-controlled $U$ ; $\Lambda_q(U)$ — coherent controlled- $U$ ; $\equiv$ — equal output state and statistics on every register. Reading: an arrow is "then"; left measures first, right measures last, and the two are the same map. The coherent controlled- $U$ replaces the classical wire: it applies $U$ to the target on the $|1\rangle$ branch and nothing on the $|0\rangle$ branch, keeping both in superposition.

\Lambda_q(U) = |0\rangle\langle 0| \otimes I + |1\rangle\langle 1| \otimes U \qquad (2)

where: $|0\rangle\langle 0|,\,|1\rangle\langle 1|$ — projectors onto the control's computational basis; $I$ — identity on the target; $U$ — fired when the control is $|1\rangle$ . Reading: it is block-diagonal in the control, never mixing the two control values — the exact property the terminal measurement needs. The measurement is a pair of orthogonal projectors resolving the identity; their orthogonality $M_m M_n = \delta_{mn} M_m$ is the engine of the proof.

M_0 = |0\rangle\langle 0|, \quad M_1 = |1\rangle\langle 1|, \quad M_0+M_1 = I \qquad (3)

where: $M_0, M_1$ — projectors for outcomes $0$ and $1$ ; completeness $M_0+M_1=I$ makes the outcomes exhaustive. Reading: because $M_0 M_1 = 0$ , any density-matrix term connecting the two control values is annihilated the instant a projector lands. The early-measurement circuit implements a channel on $\rho$ : for each outcome $m$ , project the control with $M_m$ , conditionally apply $U^m$ ( $U^0=I$ ), and sum over the unrecorded outcome — which, the trailing projectors being idempotent, collapses to a clean form.

\rho' = \sum_{m\in\{0,1\}} (M_m \otimes U^{m})\,\rho\,(M_m \otimes U^{m})^{\dagger} \qquad (4)

where: $m$ — the summed-over outcome; $U^m$ — $I$ for $m=0$ , $U$ for $m=1$ ; $\dagger$ — Hermitian conjugate. Reading: the measure-first branch as a sum of two Kraus-like terms — what the deferred circuit must reproduce. Everything hinges on one line of algebra: pushing a terminal projector back through the coherent gate turns it into exactly the conditional operator of the classical branch.

(M_m\otimes I)\,\Lambda_q(U) = M_m \otimes U^{m} \qquad (5)

where: the left side measures the control after the coherent gate; the right is the measure-first conditional operation. Reading: because $M_m\,|j\rangle\langle j| = \delta_{mj}\,M_m$ , only the control-value- $m$ block survives, leaving $U^m$ on the target — the two circuits equal in one step. The flagship application is teleportation: Bob's classically-controlled Pauli correction $Z^{m_1}X^{m_2}$ becomes two coherent controls driven by Alice's still-unmeasured qubits, her measurements sliding to the end.

Z^{m_1} X^{m_2} \;\longrightarrow\; \Lambda_{A_1}(Z)\,\Lambda_{A_2}(X) \qquad (6)

where: $m_1,m_2$ — Alice's two measured bits; $X,Z$ — Pauli corrections; $\Lambda_{A_1}(Z)$ — coherent controlled- $Z$ from $A_1$ , $\Lambda_{A_2}(X)$ — controlled- $X$ from $A_2$ . Reading: feed-forward becomes coherent control, and the protocol collapses to one unitary block then two terminal reads.

Derivation AExactness of deferral for computational-basis control

The goal is to prove, with no approximation, that the two circuits of equation (1) implement the identical channel on an arbitrary joint control–target input $\rho$ : measuring the control then classically applying $U$ gives the same outcome probabilities and post-measurement states, outcome by outcome, as applying the coherent controlled- $U$ first and measuring at the end. We compare the two branch operators directly; since the outcomes are exhaustive, matching them proves the channels equal.

Step 1. Write the classical branch. Measuring the control gives outcome $m$ with post-state $\sigma_m = (M_m\otimes U^{m})\,\rho\,(M_m\otimes U^{m})^{\dagger}$ , the projector $M_m$ on the control and conditional $U^m$ on the target. Why: this is equation (4) read term by term — measure, then fire $U$ only on the $m=1$ branch.

Step 2. Write the coherent branch. Apply $\Lambda_q(U)$ first, then measure the control, giving outcome $m$ with post-state $\tau_m = (M_m\otimes I)\,\Lambda_q(U)\,\rho\,\Lambda_q(U)^{\dagger}\,(M_m\otimes I)$ . Why: the gate is fully coherent and the only measurement is terminal, so the projector sits on the outside.

Step 3. Push the terminal projector through the gate. Using $\Lambda_q(U)=|0\rangle\langle0|\otimes I+|1\rangle\langle1|\otimes U$ and $M_m|j\rangle\langle j|=\delta_{mj}M_m$ , only the control-value- $m$ block survives: $(M_m\otimes I)\,\Lambda_q(U)=M_m\otimes U^{m}$ , equation (5). Why: the projector selects one control value and the target operator attached to it — the coherent gate cannot leave a coherence once a computational-basis projector lands.

Step 4. Conclude equality. Substituting Step 3 into Step 2 gives $\tau_m = (M_m\otimes U^{m})\,\rho\,(M_m\otimes U^{m})^{\dagger} = \sigma_m$ for every $m$ , so $\mathrm{tr}\,\tau_m=\mathrm{tr}\,\sigma_m$ and the post-states match; summing over $m$ gives $\mathcal{E}_{\text{defer}}(\rho)=\mathcal{E}_{\text{classical}}(\rho)$ . Why: matching every branch operator is a proof of channel equality on all inputs, including entangled ones, not a spot check.

Go deeper · Derivation B · 1 minWhy the coherences are harmless

Derivation BWhy the coherences are harmless

The goal is to see explicitly that the off-diagonal terms a coherent control creates never reach the output, by computing the target's reduced state after $\Lambda_q(U)$ and showing it is already the same classical mixture an early measurement would produce — before the control is measured at all. This is the physics behind the previous derivation: entanglement with a to-be-measured control is, for the target, indistinguishable from decoherence.

Step 2. Trace out the control. The reduced target is $\rho_t = |a|^2\,|\chi\rangle\langle\chi| + |b|^2\,U|\chi\rangle\langle\chi|U^{\dagger} + a b^{*}\,U|\chi\rangle\langle\chi|\,\langle0|1\rangle + \text{c.c.}$ , and the cross terms carry $\langle0|1\rangle=0$ . Why: the control basis states are orthogonal, so the interference terms vanish — the target is a mixture even though nothing has been measured yet.

Step 3. Read off the mixture. What remains is $\rho_t = |a|^2\,|\chi\rangle\langle\chi| + |b|^2\,U|\chi\rangle\langle\chi|U^{\dagger}$ , exactly the state early measurement produces ( $|\chi\rangle$ with probability $|a|^2$ , $U|\chi\rangle$ with $|b|^2$ ). Why: the entanglement itself did the decohering; measuring the control later only records which branch, without changing the target's statistics.

Step 4. Confirm by measuring the control. A terminal read gives outcome $1$ with $\Pr(m{=}1)=|b|^2$ and target $U|\chi\rangle\langle\chi|U^{\dagger}$ — identical to measure-first. Why: with $a=\sqrt{0.3},\,b=\sqrt{0.7}$ this is $\Pr(1)=0.70$ and the $0.30/0.70$ mixture of the second figure, whichever way you order the measurement.

Go deeper · Derivation C · 1 minDeferring teleportation's Pauli corrections

Derivation CDeferring teleportation's Pauli corrections

The goal is to convert the standard measure-and-feed-forward teleportation protocol into a fully coherent circuit followed by two terminal measurements, showing how deferred measurement removes the real-time classical communication the textbook description seems to require. In ordinary teleportation Alice measures two qubits, sends bits $m_1,m_2$ to Bob, and Bob applies $Z^{m_1}X^{m_2}$ ; we replace that feed-forward by coherent control and slide Alice's measurements to the end.

Step 1. Identify the classical controls. After Alice's Bell-basis rotation (a CNOT then a Hadamard), Bob's qubit holds $X^{m_2}Z^{m_1}|\psi\rangle$ , and his correction $Z^{m_1}X^{m_2}$ is classically controlled by her two bits. Why: these conditional Paulis are the only place the classical outcome is used, so they are exactly the gates the principle makes coherent.

Step 2. Replace with coherent control. Use Alice's still-unmeasured qubits $A_1$ (from $m_1$ ) and $A_2$ (from $m_2$ ) as quantum controls: $Z^{m_1}X^{m_2}\to\Lambda_{A_1}(Z)\,\Lambda_{A_2}(X)$ , a controlled- $Z$ and controlled- $X$ on Bob's qubit. Why: equation (5) guarantees each coherent control reproduces its classical counterpart once $A_1,A_2$ are measured at the end.

Step 3. Defer the measurements. By identity (1), slide Alice's two measurements past the coherent corrections to the end. Bob's output is now correct with no feed-forward: a unitary block (Bell prep, disentangling, coherent corrections) followed by two terminal reads. Why: nothing between the corrections and the end depends on the outcomes, so the measurements commute to the end exactly.

Step 4. Check a branch numerically. For $|\psi\rangle=0.6|0\rangle+0.8|1\rangle$ and outcome $m_1{=}1,m_2{=}0$ , Bob's uncorrected qubit is $Z|\psi\rangle=0.6|0\rangle-0.8|1\rangle$ , and $Z^{1}X^{0}=Z$ restores $|\psi\rangle$ . Why: the coherent controlled- $Z$ from $A_1$ gives the same conditional map on this branch, so deferring changes only when the reads happen, not the recovered state.

Limiting cases

The identity is exact, but whether a rewrite is legal depends on how the measured qubit and its bit are used downstream.

Case	Result	Meaning
Result only conditions a later gate	$M \to U^{m}\;\equiv\;\Lambda(U)\to M$	Deferral is exact — replace by the coherent controlled- $U$ and measure after.
Qubit reset or reused after measurement	$M \to \text{reset}\to(\cdots)$	Cannot defer directly; CNOT the bit onto a fresh ancilla first, then defer, or the identity fails.
Outcome drives classical post-processing	$m \to f(m)\to \text{output}$	The quantum part defers unchanged; any classical function of $m$ is computed from the terminal read instead.
Non-computational-basis measurement	$\{P_k\}$ general	Conjugate the measurement to the $Z$ basis by a unitary $W$ , defer there, then undo $W$ — the diagonal-control argument applies to the rotated outcomes.
Adaptive gate depends on the outcome	$U(m)$ from $m$	Deferrable only if the whole family $\{U(m)\}$ folds into one coherent controlled operation on the ancilla.

Common trapDeferral is not free — a reused qubit needs a fresh ancilla

The principle is often quoted as "you can always push measurements to the end", and the missing fine print is at the cost of keeping the control alive. Deferral is exact only if the measured qubit is not destroyed, reset, or re-measured before the point you defer to. A circuit that measures a qubit then resets it to $|0\rangle$ for reuse cannot simply slide that measurement to the end — the qubit's post-measurement role has nowhere coherent to live. The fix is to copy the bit onto a fresh ancilla with a CNOT (which copies a computational-basis bit without violating no-cloning) and defer that ancilla's measurement. So each deferred measurement can cost an extra qubit and a two-qubit gate — and where fast measurement is plentiful but qubits scarce, the trade runs the other way, which is why real machines still use feed-forward.

Worked examples

1 · Early versus late measurement give the same numbers (confidence builder)

A control is prepared in $|\phi\rangle=\sqrt{0.3}\,|0\rangle+\sqrt{0.7}\,|1\rangle$ , a target in $|\chi\rangle$ , and $U$ is applied to the target conditioned on the control. Compute the probabilities and post-measurement target state (i) measuring first then acting classically, and (ii) coherent controlled- $U$ then measuring last. Show they agree.

Step 1. Measure-first. The Born rule gives $\Pr(0)=\mathbf{0.30}$ and $\Pr(1)=\mathbf{0.70}$ . On outcome $0$ the target is untouched, $|\chi\rangle$ ; on outcome $1$ it becomes $U|\chi\rangle$ . The unconditioned target is $0.30\,|\chi\rangle\langle\chi|+0.70\,U|\chi\rangle\langle\chi|U^{\dagger}$ .

Step 2. Coherent-then-measure. $\Lambda_q(U)$ gives $\sqrt{0.3}\,|0\rangle|\chi\rangle+\sqrt{0.7}\,|1\rangle U|\chi\rangle$ ; tracing out the control kills the cross term via $\langle0|1\rangle=0$ , leaving the reduced target $0.30\,|\chi\rangle\langle\chi|+0.70\,U|\chi\rangle\langle\chi|U^{\dagger}$ — identical to Step 1, before the control is measured at all.

Step 3. Now read the control. By equation (5) the terminal projector gives the branch $(M_m\otimes U^m)|\phi\rangle|\chi\rangle$ : $\Pr(0)=0.30$ with target $|\chi\rangle$ , $\Pr(1)=0.70$ with target $U|\chi\rangle$ . Both probabilities and post-states match Step 1 exactly — the coherent controlled- $U$ does the classical control's job with the measurement at the end. (Matrix multiplication with $U=X$ confirms every number.)

2 · Deferring teleportation on a concrete state

Alice teleports $|\psi\rangle=0.6\,|0\rangle+0.8\,|1\rangle$ to Bob. For outcome $m_1{=}1,\ m_2{=}0$ , show that classical feed-forward and deferred coherent correction leave Bob with the same qubit.

Step 1. Feed-forward path. After Alice's measurement gives $(m_1,m_2)=(1,0)$ , Bob's qubit is $X^{m_2}Z^{m_1}|\psi\rangle=Z|\psi\rangle=0.6\,|0\rangle-\mathbf{0.8}\,|1\rangle$ . Bob applies the classical correction $Z^{m_1}X^{m_2}=Z$ , recovering $\mathbf{0.6\,|0\rangle+0.8\,|1\rangle}=|\psi\rangle$ .

Step 2. Deferred path. Replace the correction by coherent controls (equation 6): a controlled- $Z$ from $A_1$ and a controlled- $X$ from $A_2$ , applied while both are still unmeasured. On the branch $A_1=1,A_2=0$ the controlled- $Z$ fires and the controlled- $X$ does not, applying $Z$ to Bob's qubit — the same operator as Step 1.

Step 3. Read at the end. Measuring $A_1,A_2$ selects branch $(1,0)$ with probability $\tfrac14$ , and by equation (5) the coherent controlled- $Z$ has already delivered exactly $Z$ on Bob's line. Bob holds $\mathbf{0.6\,|0\rangle+0.8\,|1\rangle}$ with no classical message ever sent — the feed-forward is absorbed into a unitary run ahead of time, and the deferred circuit reproduces all four outcomes identically.

3 · When you may not defer, and the ancilla that rescues it

A circuit measures qubit $q$ mid-computation, uses the bit to control a gate $U$ on another register, then resets $q$ to $|0\rangle$ and reuses it. Can the measurement be deferred? If not, fix it.

Step 1. Spot the obstruction. To defer, $q$ must survive coherently to the end to serve as the control of $\Lambda_q(U)$ . But the circuit resets $q$ right after measuring it, overwriting the very qubit that would carry the deferred control. Naive deferral is illegal here.

Step 2. Copy the bit first. Before the reset, CNOT $q$ (control) onto a fresh ancilla $a$ in $|0\rangle$ : $|b\rangle_q|0\rangle_a\mapsto|b\rangle_q|b\rangle_a$ copies the classical bit without cloning a superposition, since these are basis states.

Step 3. Defer on the ancilla. Let $a$ — never reset — be the coherent control of $\Lambda_q(U)$ and defer $a$ 's measurement to the end; $q$ is free to reset and reuse. The rewrite is legal at the cost of one ancilla and one CNOT, the price the warning names.

Physics of the portrait

Deferred measurement is not a lone trick but the load-bearing beam under a cluster of results — teleportation, gate teleportation, syndrome extraction, and measurement-based computation are all one coherent circuit seen through a classical veil. The portrait is of the principle itself and the people who, across a decade, turned "measure and feed forward" into "one unitary, then read out".

Origins · 1 minthe deferred measurement principle and how this was discovered

the deferred measurement principle

coherent control ≡ classical control up to a terminal read · circuit normal forms, teleportation, and measurement-based computing

The deferred measurement principle has no single discoverer; it crystallised as the quantum circuit model matured in the 1990s, and its authority comes from the uses it licenses. The template was set in 1993 when Charles Bennett and co-workers cast teleportation as a measurement followed by a classically-controlled Pauli correction — the archetype of measure-and-feed-forward, and the first place people saw the correction could be made coherent. Peter Shor's 1996 fault-tolerant constructions raised the same question at scale: does conditioning recovery on measured syndromes need real-time classical control, or only its coherent shadow? Daniel Gottesman and Isaac Chuang answered in 1999 with gate teleportation, showing that deferring the corrections turns teleportation into a universal computational primitive — the basis of fault-tolerant gate constructions still used today. Nielsen and Chuang's 2000 text codified the two companion statements as named principles and proved the normal form. In 2001 Robert Raussendorf and Hans Briegel built the one-way quantum computer entirely from adaptive single-qubit measurements on a cluster state, where deferred measurement is what guarantees the adaptive machine and a coherent circuit compute the same unitary — a bookkeeping convenience become the reason an entire alternative model of computation is known to equal the circuit model.

History

1993

— Bennett and co-workers give teleportation with a measure-and-feed-forward Pauli correction

1996

— Shor's fault-tolerant scheme conditions recovery on measured error syndromes

1999

— Gottesman and Chuang defer the corrections, making teleportation a universal gate primitive

2000

— Nielsen and Chuang codify the deferred and implicit measurement principles and the circuit normal form

2001

— Raussendorf and Briegel build the one-way computer from adaptive measurements

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §4.4 (measurement, principles of deferred and implicit measurement) and §1.3.7 (teleportation) · Preskill, Lecture Notes for Physics 219, Ch. 5 · Watrous, The Theory of Quantum Information, Ch. 2 · Wilde, Quantum Information Theory, 2nd ed., Ch. 4–5 · Kitaev, Shen & Vyalyi, Classical and Quantum Computation, §8 · Mermin, Quantum Computer Science, Ch. 1–2 · Gottesman & Chuang, "Quantum Teleportation is a Universal Computational Primitive" (1999) · Raussendorf & Briegel, "A One-Way Quantum Computer" (2001).

Provenance. A verifier should confirm the projector identity $M_m|j\rangle\langle j|=\delta_{mj}M_m$ , hence $(M_m\otimes I)\Lambda_q(U)=M_m\otimes U^{m}$ (Derivation 1); the coherence cancellation via partial trace and $\langle0|1\rangle=0$ , reproducing the $0.30/0.70$ mixture for $a=\sqrt{0.3},b=\sqrt{0.7}$ (Derivation 2); and the teleportation branch $Z|\psi\rangle=0.6|0\rangle-0.8|1\rangle\to|\psi\rangle$ under the $Z$ correction (Derivation 3). All gates and the teleportation circuit follow Nielsen & Chuang §4.4 and §1.3.7; the deferred and implicit measurement statements are their named principles.

Prerequisites. Quantum Circuit Model · CNOT Gate · Controlled-U Gates · Circuit Identities and Equivalences. Leads to. Controlled-U Gates · Circuit Identities and Equivalences.

Ask the tutor: Show me any circuit with a mid-computation measurement and I will walk you through deferring it to the end — or show you why it needs an ancilla first.

Does deferring a measurement change the answer at all?

No — the equality is exact, not an approximation. Deferred measurement is a theorem: for every input, including entangled ones, the two circuits' outcome probabilities and post-measurement states agree to the last digit, via the projector identity $(M_m\otimes I)\Lambda_q(U)=M_m\otimes U^m$ applied branch by branch. What can change is the resource cost — deferring may need an extra ancilla to keep a measured-and-reused qubit's bit alive — never the computed statistics, which is what makes it safe as a rewrite rule in compilers and proofs.

Why doesn't the coherence created by the controlled gate leak into the result?

Because the control is measured in the same basis the gate is diagonal-blocked in. A coherent controlled- $U$ does create control–target entanglement, and off-diagonal coherences appear in the joint state. But trace out the control and those cross terms carry a factor $\langle0|1\rangle=0$ and vanish — the target is already a classical mixture, and the terminal measurement just records which branch occurred. So the coherence is real while the control is live but invisible to any measurement ending in a computational-basis read of it — measure the control in a different basis and deferral no longer holds.

What is the principle of implicit measurement, and how is it different?

It is the companion rule. Deferred measurement moves a mid-circuit measurement to the end; implicit measurement says any qubit left unmeasured at the end may be treated as though measured in the computational basis, without changing the statistics of the qubits you do measure — for the same reason, a qubit never observed again might as well have decohered. Together the two let a designer slide measurement layers freely to the end and assume a clean read-out there: the "one unitary, then a terminal measurement layer" normal form.

If it's all equivalent, why do real machines still use feed-forward?

Because the equivalence is about statistics, not cost. Deferring keeps the control coherent to the end and, when the qubit is reused, spends an extra ancilla — so a fully deferred error-correction cycle can demand far more qubits than a device has. Feed-forward instead measures early, frees the qubit by resetting it, and pays with a fast classical loop. On today's machines qubits are scarce and fast measurement is increasingly available, so feed-forward is often cheaper — which is why "dynamic circuits" exist. The principle lets a compiler choose between the two knowing the answer is identical.

How does deferred measurement make teleportation a computational primitive?

Ordinary teleportation moves a state with a Bell pair, a Bell measurement, and a classically-controlled Pauli correction. Gottesman and Chuang noticed that if you defer the correction — make it coherent — and choose the entangled resource cleverly, the "correction" can be an arbitrary gate applied to the teleported state. Teleportation stops being mere state transfer and becomes a way to apply a gate baked into an offline-prepared resource. This "gate teleportation" is how hard gates like $T$ are done fault-tolerantly by consuming magic states — a cornerstone of scalable quantum computing.

Does copying the bit onto an ancilla violate the no-cloning theorem?

No, and the distinction is the point. No-cloning forbids copying an unknown superposition $\alpha|0\rangle+\beta|1\rangle$ with one unitary. But CNOTing a computational-basis state $|b\rangle$ ( $b$ a definite $0$ or $1$ ) onto an ancilla copies a classical bit: $|b\rangle|0\rangle\mapsto|b\rangle|b\rangle$ is perfectly unitary and legal. The measured qubit's information is classical the instant it is read, so duplicating it to carry a deferred control is allowed; the CNOT would clash with no-cloning only on a superposition, which deferral never requires.