Why timing matters
A circuit may measure a qubit halfway through a computation and choose a later operation from the classical result. A final-measurement circuit can reproduce this behavior, but simply omitting the intermediate measurement is not enough. The outcome must be recorded in a way that preserves the relevant branches.
Definition 7.1.1
(Measurement followed by feed-forward)
Let be a finite-dimensional system containing a qubit . For let project onto the computational-basis outcome, with identity on the rest of . Start from a normalized pure state ; it may be entangled across and the other registers. The original process measures with projectors , records the classical bit , and then applies a unitary to . Its unnormalized branch is , and the outcome has probability .
When this probability is positive, dividing the branch by its norm gives the conditional output state. A zero-probability branch has no conditional state to normalize.
Replace the measurement by a coherent record
Add a fresh record qubit in . Apply controlled-NOT with as control and as target. In the order , the state becomes
Now apply the block-controlled unitary
It is unitary because the two record projectors are orthogonal and sum to the identity, and each is unitary. The joint state becomes
Theorem 7.1.2
(Deferred measurement)
In the circuit above, measuring in the computational basis at the end produces each with exactly the original probability and, for each positive-probability outcome, leaves in exactly the original conditional state. If the record is ignored instead, the state of is the same as after measuring, applying using the result, and discarding the result. The same holds for a mixed input.
Proof
(Comparing branches one by one)
The final state is . Reading gives with probability equal to the squared norm of the -th branch, which is because is unitary; the normalized post-measurement state of is the original conditional state. Therefore any later measurement of has the same joint statistics with . If the record is ignored, tracing it out gives , the same sum of unnormalized branch density operators as in the original process after discarding the classical result. By linearity the conclusion also holds for a mixed input; equivalently include a reference system in and let the circuit act trivially on it.
The record is a coherent control register until read. Controlled-NOT copies the computational-basis label; it does not clone an arbitrary unknown qubit. In particular, the record and measured system can become entangled.
Common mistake
(Copying a label is not cloning a state)
For , the record circuit produces on , an entangled state. A clone would be , which differs. The record knows the outcome label in each branch, not the amplitudes of the measured qubit, so reading never reveals 's state.
A concrete feed-forward circuit
Worked example 7.1.3
(Conditional X on a qubit in the plus state)
Take initially in , and let , on . In the original process, each intermediate measurement outcome occurs with probability ; the conditional maps the outcome-one state back to , so the final state is always .
In the deferred circuit, after the record controlled-NOT the joint state on is . After the controlled it is
Measuring gives 0 or 1 with probability , and is in either branch, exactly as before. This example also shows why a stored coherent record is not literally a classical register: before reading it, the record happens to be . Interfering it in another basis can expose coherence that was absent from an already-read classical record. Such interference changes the circuit and is outside the equivalence claim.
Worked example 7.1.4
(A zero-probability branch)
Let start in with the other registers in a state , so and . In the original process the outcome is with probability , and the outcome-one branch is the zero vector, for which no conditional state exists. In the deferred circuit the state after the record and is : the record is with certainty, and is in the outcome-zero state . Reading never produces , so the undefined conditional state is never needed.
Several measurements and an important failure mode
For a fixed finite sequence of specified computational-basis measurement locations and unitary feed-forward, allocate a fresh record for each measurement. Replace each measurement by its controlled-NOT record operation and make subsequent branch-dependent unitaries controlled by the earlier records. Use those records only as controls, never as targets of later gates or as inputs to coherent mixing.
Theorem 7.1.5
(Deferred measurement for a fixed sequence)
For a fixed finite sequence of computational-basis measurements, each followed by a unitary that may be controlled by earlier outcomes, the circuit with fresh records and a final read-out of all records gives the same joint distribution of all outcomes and the same conditional final state of the system as the original process.
Proof
(Induction over the number of stages)
After stages the state has the form , where ranges over the length- outcome histories and is the original unnormalized branch. Block-controlled unitaries act separately on each history; the next record operation splits each branch by its two projectors. Thus induction preserves the branch form and gives the same joint distribution of all recorded outcomes and the same conditional final system state. Read the records in that basis at the end.
Proof X-ray
Why records must be fresh controls and operations unitary
The induction invariant is the branch form , and it survives each stage only under three hypotheses. Records used only as controls act as projectors in the record basis, which commute with every later gate; this is why reading them at the end is the same as reading them at the time of measurement. A fresh record per measurement means a controlled-NOT writes onto and so copies the label; writing onto a used record would XOR the new label into the old one and destroy the history. Finally, unitary operations between measurements keep each branch a vector, so the squared norm of a branch is still its probability; an arbitrary noisy operation needs an explicit larger model or a separate argument.
Simply deleting a measurement can change observable behavior, as the next case shows.
Counterexample mode
Deleting an ignored measurement
Plausible claim: a measurement whose outcome is ignored can be deleted. Start with , measure it in the computational basis and ignore the outcome, then measure in the basis. The state after the first step is , so the second result is with probability . If the first measurement is omitted without making a record, the state remains and that probability is 1. In the coherent-record construction, tracing out the record leaves maximally mixed and the final probability is ; it is not equivalent to removing the measurement and its record altogether. The repair is to keep the record (or equivalently trace it out), not to delete it.
Worked example 7.1.6
(Two measurements with feed-forward and a perfectly correlated record)
Let and . Measure into record and apply to if the outcome is . Then measure into record . After the first record and the controlled (controlled by ), the state is . The second record operation is a controlled-NOT from to , giving
Reading both records gives and with probability each and never or . The original process agrees: the first outcome is uniform and sets equal to it, so the second outcome repeats the first. Every history is a pair of recorded bits, as in the proof.
Exercises
Quick check
(For the feed-forward example, compute the state immediately after the record controlled-NOT and before the controlled .)
Apply the controlled-NOT to .
Solution · Answer
, and the controlled-NOT with as control gives .
Quick check
(Suppose . What is the probability of record value 1? Why must one avoid dividing by its branch norm?)
Use the probability formula .
Solution · Answer
The probability is . The branch norm is , so dividing by it is undefined; a zero-probability outcome has no conditional state, and it never occurs in the record.
Quick check
(Explain why allowing a later Hadamard on the coherent record before reading it is not covered by the stated classical-feed-forward equivalence.)
Test it on the record of the concrete example.
Solution · Answer
The equivalence reads the record in the computational basis, so branches never interfere. A Hadamard on mixes the two branches. In the concrete example the state is , and a Hadamard on gives , so reading yields with probability instead of the original probability .