Principle of Implicit Measurement

A quantum circuit almost always uses more qubits than it reports. Ancillas hold scratch data, reversible arithmetic leaves garbage registers, a fan-out gate copies a control — and at the end only a chosen subset of wires is measured to read off the answer. The principle of implicit measurement governs the fate of everything left over: any qubit discarded at the end of a circuit — never measured, never touched by a later gate — may, without loss of generality, be treated as though it had been measured in the computational basis. Nothing you can observe on the qubits you keep can tell the difference. The reason is that the only observable consequence of a quantum state is the statistics of measurements on the registers you look at, encoded in the reduced density operator got by partial-tracing away the rest. The pivotal fact is that this reduced operator is identical whether or not you first perform a complete projective measurement on the discarded subsystem and throw the outcome away. Measure-and-forget and simply-forget give the same density matrix on the remainder, because the partial trace is blind to exactly the coherences a measurement would destroy. That invisibility makes the principle free to use and powerful to reason with. After an implicit measurement the discarded register carries a definite but unknown classical value, so the global state collapses from an entangled superposition into a probabilistic mixture of product branches, and conditioning on that hidden value turns a coherent, entangled analysis into a sum over classical cases — a controlled gate becomes a classically controlled one. The principle is the exact dual of deferred measurement: one promotes classical wires to quantum control, the other demotes unused quantum wires to classical bits, and together they let the designer slide the quantum–classical boundary at will.

Key takeaways

The principle of implicit measurement says any qubit left unobserved at the end of a circuit may, with no loss of generality, be assumed to have been measured in the computational basis.
It holds because a measurement on a subsystem that is never read and never reused cannot change the statistics of any other measurement — information travels with what you do with an outcome, and here you do nothing.
Mathematically it is the identity that tracing out a register equals first completely dephasing it and then tracing out: $\mathrm{Tr}_B(\rho_{AB})=\mathrm{Tr}_B\big(\Delta_B[\rho_{AB}]\big)$ , because off-diagonal coherences of $B$ never survive the partial trace.
It lets you replace coherent ancillas by classical random variables, collapsing an entangled analysis into a sum over classical branches — the standard argument that uncomputed garbage does no harm.
It is the exact dual of deferred measurement: one pushes measurements to the end and promotes classical control to quantum, the other lets unused terminal qubits be treated as already measured.
Key idea: tracing out a subsystem is identical to first completely dephasing it in any orthonormal basis and then tracing out, so an unobserved terminal register may be treated as measured because $\mathrm{Tr}_B(\rho_{AB})=\sum_k (\mathbb{1}_A\otimes\langle k|_B)\,\rho_{AB}\,(\mathbb{1}_A\otimes|k\rangle_B)$ .

The core insight

Tracing out a subsystem is identical to first completely dephasing it in any orthonormal basis and then tracing out: an unobserved register at the end of a circuit may be treated as measured because $\mathrm{Tr}_B(\rho_{AB})=\sum_k (\mathbb{1}_A\otimes\langle k|_B)\,\rho_{AB}\,(\mathbb{1}_A\otimes|k\rangle_B)$ .

Where this lives today

The principle is invisible infrastructure in nearly every rigorous argument about real quantum hardware. In fault-tolerant quantum error correction, the formal proofs that a logical circuit behaves correctly routinely invoke implicit measurement to treat leftover, unread ancillas as classical bits — collapsing an intractable coherent analysis of thousands of physical qubits into a sum over classical error branches. In channel theory it is why a noisy gate found by process tomography can be described equivalently as decoherence into an environment or as a random application of Kraus operators — the two sides of equation (3). It underwrites the "garbage does no harm" argument that lets reversible-arithmetic circuits — the modular exponentiation inside Shor's algorithm, for instance — leave uncomputed scratch registers behind without corrupting the measured output, provided those registers are genuinely terminal. And in one-way quantum computation, where the whole computation is a cascade of measurements on a cluster state, it is the bookkeeping that says any cluster qubit never fed forward can be dropped unread. Every time a circuit diagram simply ends a wire, the principle is silently at work.

The physical picture

Picture a two-qubit entangled state in which qubit $B$ is about to be thrown away. Whatever happens to $B$ behind the closed lid of the trash can — whether a janitor measures it or leaves it alone — cannot reach back and change the odds of a measurement you make on qubit $A$ . This is the no-signalling backbone of quantum mechanics: information is transmitted not by the act of measuring a subsystem but by what you subsequently do with the outcome, and here you do nothing. The first figure makes it concrete: a CNOT builds a Bell pair, and whether you discard the control directly or measure it first and then discard it, the target is left in exactly the same state.

Figure

A CNOT prepares a Bell pair from control plus and target zero; discarding the control directly (left) or measuring it first and then discarding it (right) leaves the target in the same maximally mixed state ½I with p₀ = p₁ = ½

Why does the measurement leave no fingerprint? Because discarding a subsystem is the partial trace, and the partial trace keeps only the diagonal of the traced-out register in whatever basis you sum over. A complete projective measurement is precisely the operation that erases that register's off-diagonal coherences — the completely dephasing channel $\Delta_B$ . Since the trace was going to throw those coherences away anyway, inserting $\Delta_B$ first changes nothing downstream. The second figure shows the block picture — the joint state as a grid of operator blocks indexed by the discarded register's basis — and its collapse under dephasing, with the Bell pair as a numeric anchor.

Figure

The joint state written as a block grid over the discarded register B: dephasing Δ_B deletes the off-diagonal coherence blocks in red, and the partial trace keeps only the blue diagonal blocks, so Tr_B and Tr_B∘Δ_B give the identical reduced state — for the Bell pair each diagonal block equals ½, giving ρ_A = ½I

A one-line slogan carries the whole idea: a quantum wire that is never read and never reused is already classical. The circuit cannot tell whether you painted it classical at the very end or left it quantum, because nothing downstream ever queries the difference. And that is exactly the time-reverse of deferred measurement — deferred measurement promotes a classical wire to a coherent control provided you measure it later; implicit measurement demotes a quantum wire heading nowhere to a classical bit that no one will read.

The core equations

The starting point is the definition of the partial trace written as a sum over an orthonormal basis of the discarded register. It is this form — a sum of "sandwich" terms — that makes the connection to measurement transparent, because each term looks like a projected component of the state.

\mathrm{Tr}_B(\rho_{AB})=\sum_k (\mathbb{1}_A\otimes\langle k|_B)\,\rho_{AB}\,(\mathbb{1}_A\otimes|k\rangle_B) \qquad (1)

where: $\rho_{AB}$ — joint density operator of registers $A$ (kept) and $B$ (discarded); $\{|k\rangle_B\}$ — any orthonormal basis of $B$ 's Hilbert space; $\mathrm{Tr}_B$ — partial trace over $B$ , returning the reduced state on $A$ ; $\mathbb{1}_A$ — identity on $A$ . Reading it: the reduced state is the sum over $B$ 's basis of the $A$ -operators got by clamping $B$ into each basis state on both sides. An unrecorded computational-basis measurement of $B$ is modelled by the completely dephasing channel, which conjugates the state by every basis projector and sums. It keeps the diagonal blocks of $\rho$ in the chosen basis and annihilates the off-diagonal coherences.

\Delta_B[\rho_{AB}]=\sum_k (\mathbb{1}_A\otimes P_k)\,\rho_{AB}\,(\mathbb{1}_A\otimes P_k),\qquad P_k=|k\rangle\langle k|_B \qquad (2)

where: $\Delta_B$ — completely dephasing (decohering) channel on $B$ , the model of an unread projective measurement; $P_k$ — projector onto $|k\rangle_B$ , with $\sum_k P_k=\mathbb{1}_B$ . Reading it: $\Delta_B$ is what an unrecorded measurement does — it forgets which $k$ occurred but records that some definite $k$ did. The principle itself is the operator identity equating the two: tracing out $B$ gives the same reduced state as first dephasing $B$ and then tracing out. This single equation is the whole theorem.

\mathrm{Tr}_B(\rho_{AB})=\mathrm{Tr}_B\big(\Delta_B[\rho_{AB}]\big) \qquad (3)

where: the equality sign carries the content — discarding $\equiv$ measure-then-discard, exactly, for any $\rho_{AB}$ . Reading it: nothing observable on $A$ is sensitive to whether $B$ was silently measured before being thrown away. The mechanism is visible if you expand $\rho_{AB}$ into blocks labelled by $B$ 's basis. Dephasing keeps the diagonal blocks and drops the rest, and those are precisely the blocks the trace uses.

\rho_{AB}=\sum_{j,k}\rho^{(jk)}_A\otimes|j\rangle\langle k|_B \;\xrightarrow{\;\Delta_B\;}\; \sum_{k}\rho^{(kk)}_A\otimes|k\rangle\langle k|_B \qquad (4)

where: $\rho^{(jk)}_A$ — the operator on $A$ sitting in the $|j\rangle\langle k|_B$ block of $\rho_{AB}$ . Reading it: the coherence blocks $j\neq k$ are exactly what a measurement erases and what the trace ignores, so their loss is invisible on $A$ . Because the surviving diagonal blocks each carry a weight, the reduced state is literally a post-measurement mixture — a convex combination of normalized conditional states, one per hidden outcome.

p_k=\mathrm{Tr}\big[(\mathbb{1}_A\otimes P_k)\,\rho_{AB}\big],\qquad \rho_A=\sum_k p_k\,\rho_A^{(k)} \qquad (5)

where: $p_k$ — probability the implicit measurement of $B$ yields outcome $k$ ; $\rho_A^{(k)}=\rho^{(kk)}_A/p_k$ — normalized conditional state of $A$ given that outcome. Reading it: after discarding $B$ , register $A$ is genuinely in one branch $\rho_A^{(k)}$ with probability $p_k$ — you just do not know which. The canonical example is a CNOT that entangles a control into a Bell pair and is then discarded; the reduced target is maximally mixed.

\mathrm{CNOT}_{c\to t}\,(|+\rangle_c\otimes|0\rangle_t)=\tfrac{1}{\sqrt2}(|00\rangle+|11\rangle)\;\xrightarrow{\;\mathrm{Tr}_c\;}\;\tfrac12\mathbb{1} \qquad (6)

where: $c,t$ — control and target of the CNOT; $|+\rangle=\tfrac{1}{\sqrt2}(|0\rangle+|1\rangle)$ ; $\mathrm{Tr}_c$ — discard of the control. Reading it: the target's $\tfrac12\mathbb{1}$ is exactly what measuring the control and forgetting the result gives — the Bell pair's single-qubit marginal. Finally, the principle underwrites the Stinespring picture: every quantum channel is a unitary on the system plus an ancilla, followed by discarding the ancilla — and that discard is an implicit measurement.

\mathcal{E}(\rho)=\mathrm{Tr}_B\big(U(\rho\otimes|0\rangle\langle0|_B)U^\dagger\big) \qquad (7)

where: $U$ — unitary on system-plus-ancilla; $B$ — ancilla initialized to $|0\rangle$ and discarded; $\mathcal{E}$ — the resulting channel on the system. Reading it: because the discard may be read as a measurement, the channel admits both a Kraus operator-sum form and an "environment was measured" story.

Derivation ADiscarding a register equals dephasing then discarding

The goal is to prove the operator identity at the heart of the principle — that for any joint state, tracing out the terminal register $B$ gives exactly the same reduced state as first passing $B$ through the dephasing channel $\Delta_B$ and only then tracing it out. We do it twice for confidence: once by pushing the projectors through the trace using idempotence and completeness, once by expanding into basis blocks so the surviving pieces are manifestly a post-measurement mixture. Both routes must land on $\sum_k\rho^{(kk)}_A$ .

Step 1. Start from the dephased state and trace it, using linearity to break the sum apart: $\mathrm{Tr}_B(\Delta_B[\rho_{AB}])=\sum_k\mathrm{Tr}_B\big((\mathbb{1}\otimes P_k)\rho_{AB}(\mathbb{1}\otimes P_k)\big)$ . Why: the trace is linear, so a channel written as a sum of conjugations can be traced term by term.

Step 2. Cycle one projector around the partial trace over $B$ — legitimate because $P_k$ acts only on $B$ — to gather the two copies: $=\sum_k\mathrm{Tr}_B\big((\mathbb{1}\otimes P_k^2)\rho_{AB}\big)$ . Why: cyclicity of the trace holds within the traced-out factor, letting the right-hand $P_k$ move over to meet the left-hand one.

Step 3. Collapse using idempotence and completeness of the projectors: since $P_k^2=P_k$ and $\sum_k P_k=\mathbb{1}_B$ , the sum telescopes to $\sum_k\mathrm{Tr}_B\big((\mathbb{1}\otimes P_k)\rho_{AB}\big)=\mathrm{Tr}_B(\rho_{AB})$ . Why: a complete set of orthogonal projectors resolves the identity, so summing the projected pieces reconstructs the full partial trace — dephasing has vanished.

Step 4. Confirm by the block route. Write the most general bipartite operator as $\rho_{AB}=\sum_{j,k}\rho^{(jk)}_A\otimes|j\rangle\langle k|_B$ and take the direct trace: $\mathrm{Tr}_B(\rho_{AB})=\sum_{j,k}\rho^{(jk)}_A\,\langle k|j\rangle=\sum_k\rho^{(kk)}_A$ , since $\langle k|j\rangle=\delta_{jk}$ . Why: only the diagonal blocks $j=k$ have nonzero trace over $B$ , so the coherence blocks never enter the reduced state to begin with.

Step 5. Apply dephasing to the same block form: $\Delta_B[\rho_{AB}]=\sum_k\rho^{(kk)}_A\otimes|k\rangle\langle k|_B$ , whose trace is again $\sum_k\rho^{(kk)}_A$ . Identifying $p_k=\mathrm{Tr}(\rho^{(kk)}_A)$ and $\rho_A^{(k)}=\rho^{(kk)}_A/p_k$ exhibits it as $\sum_k p_k\rho_A^{(k)}$ — a genuine post-measurement mixture. Why: dephasing deletes exactly the off-diagonal blocks the trace was already ignoring, so the two operations commute past the trace and the name "implicit measurement" is earned.

Go deeper · Derivation B · 1 minImplicit measurement of a CNOT control wire

Derivation BImplicit measurement of a CNOT control wire

The goal is to show, with a fully general control state, that a CNOT whose control is discarded afterward acts on the target exactly as a classically controlled bit-flip — a coin flip that applies $X$ with some probability. This is the workhorse instance of the principle: it converts a coherent two-qubit entangler into a one-qubit mixture and makes visible how the control's off-diagonal coherences are annihilated not by fiat but by the trace itself. We conjugate an arbitrary control and target by CNOT, then trace out the control.

Step 1. Expand the control in the computational basis, $\rho_c=\sum_{a,b\in\{0,1\}}c_{ab}\,|a\rangle\langle b|_c$ , with the diagonal entries fixing the outcome probabilities $p_0=c_{00}$ and $p_1=c_{11}$ . Why: writing the control block-by-block lets us track precisely which entries survive the trace and which do not.

Step 2. Use the control-target form of the gate, $\mathrm{CNOT}=|0\rangle\langle0|_c\otimes\mathbb{1}_t+|1\rangle\langle1|_c\otimes X_t$ , equivalently $\mathrm{CNOT}=\sum_a|a\rangle\langle a|_c\otimes X_t^{a}$ . Why: in this diagonal-in-the-control form each control basis state simply selects a power of $X$ on the target.

Step 3. Conjugate a single control block: $\mathrm{CNOT}\,(|a\rangle\langle b|_c\otimes\rho_t)\,\mathrm{CNOT}^\dagger=|a\rangle\langle b|_c\otimes X^a\rho_t X^b$ . Why: the left gate factor picks $X^a$ from the ket side and the right gate factor picks $X^b$ from the bra side, so the labels $a,b$ independently set the two powers of $X$ .

Step 4. Trace out the control, which enforces $\mathrm{Tr}_c(|a\rangle\langle b|_c)=\delta_{ab}$ so only $a=b$ terms survive: $\mathrm{Tr}_c(\cdots)=\sum_a c_{aa}\,X^a\rho_t X^a$ . Why: this Kronecker delta is the coherence-killing step — the off-diagonal control amplitudes $c_{01},c_{10}$ are wiped out exactly as an implicit measurement of the control would wipe them.

Step 5. Collect the two surviving terms to reach the boxed result, $\mathrm{Tr}_c\big(\mathrm{CNOT}\,(\rho_c\otimes\rho_t)\,\mathrm{CNOT}^\dagger\big)=c_{00}\,\rho_t+c_{11}\,X\rho_t X=p_0\,\rho_t+p_1\,X\rho_t X$ . Why: measuring the control first would have set $c_{01}=c_{10}=0$ and left $c_{00},c_{11}$ untouched, giving the identical operator — so the control may be implicitly measured, reproducing Nielsen and Chuang Sec. 4.4.

Go deeper · Derivation C · 1 minEvery quantum channel hides an implicit measurement

Derivation CEvery quantum channel hides an implicit measurement

The goal is to connect the principle to open-system dynamics: to show that the Stinespring dilation — unitary on system-plus-ancilla, then discard the ancilla — is an implicit measurement of the ancilla, and that reading it as such recovers the Kraus operator-sum form of the channel with the correct operators. This is why a channel has a dual life: an "environment decohered us" story and an "operators $E_k$ act at random" story are the two sides of equation (3).

Step 1. Write the channel as a dilation, $\mathcal{E}(\rho)=\mathrm{Tr}_B\big(U(\rho\otimes|0\rangle\langle0|_B)U^\dagger\big)$ , with the ancilla $B$ prepared in $|0\rangle$ and discarded at the end. Why: Stinespring's theorem guarantees every completely positive trace-preserving map arises this way, so a terminal discard is the universal source of irreversibility.

Step 2. Because $B$ is terminal, apply the principle: insert $\Delta_B$ before the trace at no cost, $\mathcal{E}(\rho)=\mathrm{Tr}_B\big(\Delta_B[\,U(\rho\otimes|0\rangle\langle0|_B)U^\dagger]\big)$ . Why: equation (3) lets us imagine the environment being measured in its basis $\{|k\rangle_B\}$ just before it is discarded, converting a coherent trace into a classical sum over outcomes.

Step 3. Evaluate the trace against that basis and define the Kraus operators $E_k=\langle k|_B\,U\,|0\rangle_B$ , acting on the system alone: $\mathcal{E}(\rho)=\sum_k \langle k|_B U|0\rangle_B\,\rho\,\langle0|_B U^\dagger|k\rangle_B=\sum_k E_k\,\rho\,E_k^\dagger$ . Why: each implicit-measurement outcome $k$ selects a system operator $E_k$ ; summing over the unread outcomes is exactly the operator-sum representation.

Step 4. Check normalization: $\sum_k E_k^\dagger E_k=\langle0|_B U^\dagger\big(\sum_k|k\rangle\langle k|_B\big)U|0\rangle_B=\langle0|_B U^\dagger U|0\rangle_B=\mathbb{1}$ , using completeness of $\{|k\rangle_B\}$ and unitarity of $U$ . Why: the Kraus operators satisfy $\sum_k E_k^\dagger E_k=\mathbb{1}$ precisely because the implicit measurement is complete and $U$ preserves probability — the channel is trace-preserving.

Limiting cases

The principle is exact, not an approximation — but whether you are allowed to invoke it depends entirely on the discarded qubit's causal future inside the circuit.

Case	Result	Meaning
Truly discarded (terminal)	$\rho_A=\mathrm{Tr}_B(\rho_{AB})=\mathrm{Tr}_B(\Delta_B[\rho_{AB}])$	$B$ has no future gate and is never read; dephasing it leaves $\rho_A$ exactly unchanged. Implicit measurement is legal, exact, and free.
Spectator still coupled	$U_{AB}\,\Delta_B[\rho_{AB}]\,U_{AB}^\dagger \neq \Delta_B\big[U_{AB}\,\rho_{AB}\,U_{AB}^\dagger\big]$	If a later gate $U_{AB}$ still touches $B$ , dephasing it first generally changes the outcome — the principle is forbidden until after $B$ 's last interaction.
Outcome used classically	$\sum_k p_k\,\rho_A^{(k)}\otimes\|k\rangle\langle k\|_B$	If measuring $B$ controls a later classical operation, you must keep the labelled, correlated mixture — you may still call it a measurement, but you cannot forget the label.
Any measurement basis	$\mathrm{Tr}_B=\mathrm{Tr}_B\circ\Delta_B^{(\mathcal{U})}$ for any orthonormal $\mathcal{U}$	The dephasing may be in any basis, not just the computational one; the trace is blind to coherences in every basis, so the "implicit" measurement can be imagined in whichever basis is convenient.

Common trapThe principle applies to any leftover qubit, even one still entangled downstream

The single most common misuse is treating a qubit as implicitly measured while it is still coherently coupled to the rest of the computation. The principle is licensed only for qubits that are truly terminal — neither measured nor used by any later gate. A "spectator" ancilla that a downstream gate will still touch is not discardable: dephasing it now would erase interference that the later gate depends on, changing the final statistics. The test is causal, not spatial — it does not matter how the qubit was entangled or what gates hit it earlier; what matters is that its causal future inside the circuit is empty. Equivalently, implicit measurement is valid exactly at and after a qubit's last interaction with everything else.

Worked examples

1 · A biased control discarded after a CNOT (fully worked)

Step 1. Read off the control populations: $p_0=c_{00}=0.7$ and $p_1=c_{11}=0.3$ , with no off-diagonal coherences ( $c_{01}=c_{10}=0$ ) since $\rho_c$ is already diagonal. The derivation's boxed result applies directly.

Step 3. Interpret it. The final target is the classical mixture $0.7\,|0\rangle\langle0|+0.3\,|1\rangle\langle1|$ — diagonal, hence carrying no coherence: a bit that is $0$ with probability $0.7$ and $1$ with probability $0.3$ . Because $\rho_c$ had no coherences to begin with, measuring the control first would have changed nothing, and the principle says the same is true even had it been coherent — the trace kills $c_{01},c_{10}$ regardless. The CNOT-plus-discard acts as a classical channel that copies the control's bit distribution onto the target.

2 · Discarding one qubit of a GHZ state

Three qubits are prepared in the GHZ state $|\mathrm{GHZ}\rangle=\tfrac{1}{\sqrt2}(|000\rangle+|111\rangle)$ . Qubit $C$ is discarded. What is the state of the remaining pair $AB$ , and what does implicit measurement predict?

Step 1. Form $\rho_{ABC}=|\mathrm{GHZ}\rangle\langle\mathrm{GHZ}|=\tfrac12(|000\rangle+|111\rangle)(\langle000|+\langle111|)$ and trace out $C$ using $\mathrm{Tr}_C(|x\rangle\langle y|_C)=\delta_{xy}$ . Only the $C$ -diagonal terms $|000\rangle\langle000|$ and $|111\rangle\langle111|$ survive.

Step 2. The surviving terms give $\rho_{AB}=\tfrac12\big(|00\rangle\langle00|+|11\rangle\langle11|\big)$ — the cross terms $|000\rangle\langle111|$ and its conjugate vanish because $\langle1|0\rangle_C=0$ .

Step 3. Read the physics. The pair is left in the classically correlated mixture $\tfrac12(|00\rangle\langle00|+|11\rangle\langle11|)$ , not an entangled superposition: with probability $\tfrac12$ both are $0$ , with probability $\tfrac12$ both are $1$ . This is precisely what implicit measurement predicts — discarding $C$ is equivalent to measuring it in the computational basis, and each outcome $c\in\{0,1\}$ collapses $AB$ into $|cc\rangle$ with probability $p_c=\tfrac12$ . GHZ entanglement is fragile to loss: dropping one qubit leaves only classical correlation.

Physics of the portrait

The principle is a distilled corollary of a century of work on how a subsystem inherits its state from a larger whole — the density matrix, the partial trace, and the measurement postulate — finally packaged as a circuit-designer's tool in the language of quantum computation.

Origins · 1 minprinciple of implicit measurement and how this was discovered

principle of implicit measurement

the circuit theorem Tr_B = Tr_B∘Δ_B, that an unread terminal register may be treated as measured · the density matrix, the partial trace, and the quantum circuit model

The principle has no single discoverer; it is the crystallization of ideas that took decades to assemble. Its raw material is the density matrix, introduced independently by Lev Landau and John von Neumann in 1927 precisely to describe a subsystem of a larger, possibly entangled whole — Landau's point was that a component of a composite system generally has no wavefunction of its own, only a statistical operator got by what we now call the partial trace. Von Neumann's 1932 axiomatization supplied the other half: the projection postulate, and the idea that a measurement whose result is discarded is a decohering, diagonal-forming operation. The bridge to open-system dynamics came with Stinespring's 1955 dilation theorem — every channel is a unitary on a larger space followed by a partial trace — and Kraus's 1971 operator-sum representation, which made the "measure the environment" reading explicit. What Nielsen and Chuang did around 2000 was to name the resulting circuit rule — terminal, unread wires may be treated as measured — and pair it with its dual, deferred measurement, giving circuit designers a licence to slide the quantum–classical boundary without touching a single output probability.

History

1927

— Landau and von Neumann introduce the density matrix to describe a subsystem of a larger whole

1932

— von Neumann axiomatizes quantum measurement and the projection postulate

1955

— Stinespring proves every channel dilates to a unitary followed by a discarded ancilla

1971

— Kraus casts open-system dynamics in the operator-sum form

2000

— Nielsen and Chuang state the principles of deferred and implicit measurement as circuit rules

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information (10th anniv. ed.), Sec. 4.4 (principles of deferred and implicit measurement) · Preskill, Lecture Notes on Quantum Computation, Ch. 3 (measurement, density operators, partial trace) · Watrous, The Theory of Quantum Information, Ch. 2 (partial trace and channels) · Wilde, Quantum Information Theory (2nd ed.), Ch. 4 (the completely dephasing channel) · Kitaev, Shen & Vyalyi, Classical and Quantum Computation, Sec. 7 · Mermin, Quantum Computer Science, Ch. 1–2. The key results in this article are: the operator identity $\mathrm{Tr}_B(\rho_{AB})=\mathrm{Tr}_B(\Delta_B[\rho_{AB}])$ , established via both the projector-collapse route ( $P_k^2=P_k$ , $\sum_k P_k=\mathbb{1}$ ) and the diagonal-block route; the CNOT-control reduction $p_0\rho_t+p_1 X\rho_t X$ , illustrated with the biased example $0.7|0\rangle\langle0|+0.3|1\rangle\langle1|$ ; the Bell marginal $\tfrac12\mathbb{1}$ and the GHZ reduction $\tfrac12(|00\rangle\langle00|+|11\rangle\langle11|)$ ; and the Kraus form $E_k=\langle k|U|0\rangle$ with $\sum_k E_k^\dagger E_k=\mathbb{1}$ . All numerical values ( $p_k$ , $\tfrac12$ , $0.7$ , $0.3$ ) are standard textbook quantities.

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

Ask the tutor: Ask me to walk through why discarding one qubit of your favourite entangled state is the same as measuring it — and where in a real circuit you are and are not allowed to make that swap.

What is the difference between implicit measurement and deferred measurement?

They are exact duals that let you move the quantum–classical boundary in opposite directions. Deferred measurement says an intermediate measurement can be postponed to the end of the circuit by replacing its classical control with coherent quantum control — it turns a classical wire into a quantum one and pushes the measurement later. Implicit measurement runs the other way: it says a terminal quantum wire that is never read and never reused may be treated as already measured, turning an unused quantum wire into a classical bit. Deferred measurement is about promoting classical to quantum; implicit measurement is about demoting unused quantum to classical. Used together, they let a designer prove that a coherent "all-unitary" version of a protocol and its "measure-and-correct" version produce identical output statistics.

If the discarded qubit is never measured, in what sense was it measured?

In the sense that its reduced effect on everything you keep is identical to what an unread measurement would have produced. Discarding a subsystem is the partial trace, which retains only the diagonal of that subsystem in any basis; an unrecorded projective measurement, the dephasing channel $\Delta_B$ , deletes exactly the off-diagonal coherences — and those coherences never survive the trace anyway. So the density operator on the rest of the world is bit-for-bit the same whether or not the measurement happened. Physically, once a qubit is entangled with a huge environment and lost, it has effectively been measured by that environment; the principle just formalizes that a terminal wire in a circuit is in the same situation, with the circuit itself playing the role of the environment.

Does the choice of measurement basis matter?

No — and that is the surprising part. The partial trace is basis-independent, so equation (1) gives the same reduced state summed over any orthonormal basis of the discarded register. You may therefore imagine the implicit measurement in whichever basis is convenient; the computational basis is the usual default, but any complete orthonormal set gives the same reduced state on the qubits you keep. What is not free is the finer, labelled information: if you actually recorded and used an outcome, the basis would matter, because different bases give different post-measurement branches. The principle only promises invariance of the reduced state you get when the outcome is thrown away.

Why can I not use the principle on a qubit that a later gate still touches?

Because dephasing a qubit destroys interference, and a downstream gate may depend on exactly that interference. Concretely, $\Delta_B$ does not commute past a subsequent entangling gate $U_{AB}$ : measuring $B$ first collapses its coherences, so when $U_{AB}$ later acts it sees a classical mixture instead of a superposition, and the final statistics differ. The principle is licensed only once the qubit's causal future inside the circuit is empty — at or after its very last interaction. A useful mental check: could any information still flow out of this qubit into a measured wire? If yes, it is not discardable; if no, it is terminal and you may treat it as measured.

How does this relate to a channel having both a Kraus and an environment picture?

Every quantum channel can be built as a unitary on the system plus a fresh ancilla, followed by discarding the ancilla — the Stinespring dilation. That terminal discard is an implicit measurement, so you are free to imagine the ancilla being measured in its basis just before it is thrown away. Doing so turns the coherent partial trace into a classical sum over ancilla outcomes $k$ , and the system operator selected by outcome $k$ is exactly the Kraus operator $E_k=\langle k|U|0\rangle$ . That is why the same noise process can be described either as "the environment decohered the system" or as "one of the operators $E_k$ was applied at random with $\sum_k E_k^\dagger E_k=\mathbb{1}$ ." The two descriptions are the two readings of the identity $\mathrm{Tr}_B=\mathrm{Tr}_B\circ\Delta_B$ .

Does discarding a qubit destroy entanglement, and is that a loss of information?

It destroys the entanglement between the discarded qubit and the rest, and yes, that is an irreversible loss. The GHZ example is the clean illustration: dropping one qubit turns a genuinely entangled three-party state into a merely classically correlated two-party mixture. The lost information is the coherence — the relative phase between the branches — carried away with the discarded qubit and unrecoverable from the remainder. This is the microscopic origin of decoherence, and it is why reversible computation carefully uncomputes scratch registers back to $|0\rangle$ before discarding them: to avoid paying this price on data that still matters.