Controlled Operations

A controlled operation is one joint transformation on a control register and a target register. It is not shorthand for measuring the control and then deciding which gate to run. When the control is coherent, the branch labels remain quantum amplitudes, branch phases can interfere, and the operation can entangle the control with the target.

The central construction is a block-diagonal unitary: each orthogonal control sector selects one target unitary. This makes its action and inverse simple to calculate, but physically implementing the control may require more than access to the target gate alone.

Helpful background. Multi-qubit gates supplies tensor order and CNOT examples; single-qubit gates explains why two matrices differing by a global phase describe the same isolated gate channel.

Controlled Operations as Coherent Conditionals

Let $C$ be a control system and $T$ a target system. A coherent conditional applies different target operations in orthogonal control sectors while preserving superpositions among those sectors. In the simplest qubit case,

$$C_1(U) = |0\rangle\langle0|_C\otimes I_T +|1\rangle\langle1|_C\otimes U_T.$$

With control first and target second, this is the block matrix

$$C_1(U) = \begin{pmatrix} I&0\\ 0&U \end{pmatrix}.$$

If the input control is $a|0\rangle+b|1\rangle$, then

$$C_1(U) \bigl(a|0\rangle+b|1\rangle\bigr)|\psi\rangle = a|0\rangle|\psi\rangle +b|1\rangle U|\psi\rangle.$$

No measurement has occurred. If $U|\psi\rangle$ is not proportional to $|\psi\rangle$ and both amplitudes are nonzero, the result is entangled. If a measurement record instead selects a later operation, the overall process is an instrument or channel with different coherence properties.

A controlled gate must specify the control basis, tensor order and exact target matrix. In particular, a phase that was irrelevant for the isolated target can become measurable relative to the inactive branch.

Projector and Block-Diagonal Semantics

Let $\{P_j\}$ be a complete orthogonal resolution of the control space:

$$P_jP_k=\delta_{jk}P_j, \qquad \sum_jP_j=I_C.$$

Let every $U_j$ be unitary on the same target space. The general block-controlled operator is

$$W=\sum_jP_j\otimes U_j.$$

Orthogonality removes cross terms:

$$\begin{aligned} W^\dagger W &= \sum_{j,k}P_jP_k\otimes U_j^\dagger U_k\\ &= \sum_jP_j\otimes I_T =I_C\otimes I_T. \end{aligned}$$

Its inverse is branchwise,

$$W^{-1}=W^\dagger = \sum_jP_j\otimes U_j^\dagger.$$

If the projectors overlap, fail to span the declared control space, or select blocks of different target dimension, this formula does not yet define the advertised complete unitary. A promised action on only part of the space needs either a full extension or a statement that the action outside the promised subspace is unspecified.

Register order is part of the definition. In control–target order, $C_1(U)=\operatorname{diag}(I,U)$. In target–control order the matrix is related by the SWAP permutation, not by silently reinterpreting indices. Operators on composite systems develops the tensor-product and projector algebra.

Open, Negative, and Multiple Controls

An open control activates on computational-basis value zero:

$$C_0(U) = P_0\otimes U+P_1\otimes I = (X\otimes I)C_1(U)(X\otimes I),$$

where $P_b=|b\rangle\langle b|$. The word “negative” refers to this declared basis convention; it is not a basis-independent physical property.

For a Boolean predicate $g(c)$ on a multi-bit control string,

$$P_g = \sum_{c:g(c)=1}|c\rangle\langle c|,$$

and

$$C_g(U) = (I-P_g)\otimes I+P_g\otimes U.$$

A Toffoli gate is the special case $g(c_1,c_0)=c_1c_0$ with $U=X$. More complicated predicates can be computed into an ancilla, used as a control, and then uncomputed.

Mixed-polarity controls are obtained by conjugating the appropriate control wires with $X$. Their resource cost depends on the declared alphabet: a diagram may count one mixed-polarity primitive, while a decomposition into positive controls counts the surrounding flips. State which model is being used.

Multiplexed and SELECT Operations

A selector register can choose among many coherent branches. For $k$ selector qubits and target unitaries $U_j$,

$$\operatorname{SELECT}(U) = \sum_{j=0}^{2^k-1} |j\rangle\langle j|\otimes U_j.$$

This is a unitary exactly when the control projectors form a complete orthogonal resolution and every branch is unitary on the common target space. Its inverse is

$$\operatorname{SELECT}(U)^{-1} = \sum_j|j\rangle\langle j|\otimes U_j^\dagger.$$

Branch phases are operational. Replacing $U_j$ by $e^{i\gamma_j}U_j$ multiplies the SELECT operator by a diagonal phase operation on the selector register. Equal isolated channels therefore do not imply equal multiplexed gates.

One SELECT call is not automatically one elementary gate or one oracle query. Its implementation may need branch logic, data access, ancillas and routing. Gate decomposition constructs multiplexor networks; algorithmic primitives uses SELECT within larger algorithms.

Coherent Control versus Classical Feedforward

A controlled unitary preserves amplitudes between control sectors. For a control state $a|0\rangle+b|1\rangle$ and target state $|\psi\rangle$,

$$C_1(U) \bigl(a|0\rangle+b|1\rangle\bigr)|\psi\rangle = a|0\rangle|\psi\rangle +b|1\rangle U|\psi\rangle.$$

The two branches can therefore interfere later. A circuit control dot denotes this coherent block action; it does not secretly measure the control.

Measurement followed by classical feedforward is a different process. If the computational-basis outcome is forgotten, its channel is

$$\mathcal E(\rho) = \sum_{b=0}^{1} (P_b\otimes U_b)\rho(P_b\otimes U_b)^\dagger, \qquad U_0=I, \quad U_1=U.$$

This sum deletes the off-diagonal control blocks $P_0\rho P_1$ and $P_1\rho P_0$. Retaining the measurement record gives a classical–quantum record rather than restoring the lost branch coherence. Coherent control and feedforward may agree on computational-basis truth tables while giving different phase-sensitive predictions.

Measurement in circuits explains these measured branches, and mid-circuit measurement and feedforward follows the causal use of their records. Their full mathematical description uses quantum instruments.

Global Phases Become Relative

Two target operators $U$ and $e^{i\gamma}U$ induce the same isolated unitary channel. Adding a control promotes their phase difference to a control-relative operation:

$$C_1(e^{i\gamma}U) = \bigl(P(\gamma)\otimes I\bigr)C_1(U), \qquad P(\gamma)=P_0+e^{i\gamma}P_1.$$

Only a common phase multiplying the entire joint operator is globally irrelevant. For example,

$$C_1(-X)=(Z_C\otimes I)C_1(X),$$

so the two controlled gates can give orthogonal control states even though $X$ and $-X$ define the same isolated channel. Likewise, independent SELECT branch phases $e^{i\gamma_j}$ implement the selector-register phase

$$\sum_j e^{i\gamma_j}|j\rangle\langle j|.$$

Thus exact operator equality, equality up to a global phase, and equality of isolated channels are different requirements once a control is added. For an example where many such relative phases combine, see the quantum Fourier transform.

Eigenstate Kickback and Target Entanglement

For input $|+\rangle|\psi\rangle$, define

$$z=\langle\psi|U|\psi\rangle.$$

After the controlled operation, the reduced control state is

$$\rho_C = \frac12 \begin{pmatrix} 1&z^*\\ z&1 \end{pmatrix}, \qquad \operatorname{Tr}\rho_C^2 = \frac{1+|z|^2}{2}.$$

More generally, for control amplitudes $a,b$, the determinant is

$$\det\rho_C = |ab|^2\left(1-|\langle\psi|U|\psi\rangle|^2\right).$$

When both control amplitudes are nonzero, the output factors exactly when $U|\psi\rangle$ is proportional to $|\psi\rangle$. Otherwise the distinct target branches entangle control and target. Failure to entangle one chosen input does not prove that the gate is local or non-entangling on all inputs.

If $U|u\rangle=e^{i\theta}|u\rangle$, then

$$C_1(U)|+\rangle|u\rangle = \frac{|0\rangle+e^{i\theta}|1\rangle}{\sqrt2} \otimes|u\rangle.$$

The eigenphase appears on the control while the target remains unchanged. Phase kickback develops this mechanism for Boolean and modular oracles, and phase estimation combines several controlled powers to estimate an eigenphase.

For a nonnegative integer $r$,

$$C_1(U)^r=C_1(U^r),$$

but this algebraic identity says nothing about whether $U^r$ costs one powered query, $r$ controlled-base queries explicitly supplied by the access model, depth $r$, or physical evolution time $rt$. Those choices describe different supplied operations and different costs.

When can a target gate be controlled?

If a phase-fixed known circuit is declared as $U=G_m\cdots G_1$, controlling every factor gives

$$C_1(U)=C_1(G_m)\cdots C_1(G_1).$$

A decomposition valid only up to a target-global phase needs the corresponding control phase restored. Gate-by-gate control also carries the decomposition’s ancilla, depth, connectivity, and approximation assumptions; these additional controls generally require additional gates.

Completely unknown black-box access is more restrictive. A channel oracle cannot distinguish $U$ from $e^{i\gamma}U$, yet their controlled operations differ. Standard circuit access to an arbitrary unknown unitary therefore does not automatically supply $C_1(U)$. This is an access-model obstruction, not a claim that the block-diagonal matrix fails to exist.

Araújo and coauthors formulate this impossibility for unknown operations in the standard circuit model. Practical schemes can work with a richer device. For example, Zhou and coauthors use an extended Hilbert space containing a known identity sector large enough to hold the target state. Coherent transfers can place the inactive branch in that sector, apply the device, and undo the transfers. The relative action of the active and reference sectors is now part of the supplied physical operation. A path bypass with a controlled relative phase can play the same role. These constructions do not contradict the channel-only limitation.

The quantum-oracle page distinguishes base, inverse, controlled and powered access. An experimental or numerical test should also include a superposition: a computational-basis truth table can miss branch phases entirely.

With identical control projectors and a fixed relative-phase convention,

$$\|C_1(U)-C_1(V)\|=\|U-V\|$$

in operator norm. Independently rephasing a target block changes the controlled operation and invalidates that comparison.

Worked example: a phase-sensitive controlled rotation

Take control first and target second, and suppose the available target gate is exactly

$$U=e^{i\pi/5}R_z(2\pi/3) =\operatorname{diag}\bigl(e^{-2\pi i/15},e^{8\pi i/15}\bigr).$$

Its controlled matrix is $\operatorname{diag}(1,1,e^{-2\pi i/15},e^{8\pi i/15})$. On $|+\rangle_c|1\rangle_t$, the target is an eigenstate, so

$$C_1(U)|+\rangle_c|1\rangle_t =\frac{|0\rangle+e^{8\pi i/15}|1\rangle}{\sqrt2}\otimes|1\rangle.$$

The control’s two equatorial components are therefore

$$\begin{aligned} \langle X_c\rangle&=\cos(8\pi/15)\simeq-0.104528,\\ \langle Y_c\rangle&=\sin(8\pi/15)\simeq0.994522. \end{aligned}$$

Discarding the target-global factor $e^{i\pi/5}$ before adding the control would instead predict $\langle X_c\rangle=1/2$ and $\langle Y_c\rangle=\sqrt3/2$. The two circuits have the same basis-label action, but different interference. This is why the exact target phase matters in a controlled construction.

Worked example: four coherent branches

Order two selector qubits $(c_1,c_0)$ before target $t$. Choose

$$U_{00}=I,\qquad U_{01}=X,\qquad U_{10}=Z,\qquad U_{11}=H,$$

and prepare $|++\rangle_C|0\rangle_T$. The SELECT output is

$$|\Psi\rangle=\frac12\bigl( |00\rangle|0\rangle+|01\rangle|1\rangle +|10\rangle|0\rangle+|11\rangle|+\rangle\bigr).$$

The selector labels are orthogonal, so tracing them out removes cross terms between different labels:

$$\begin{aligned} \rho_T&=\frac14\bigl(2|0\rangle\langle0|+|1\rangle\langle1|+|+\rangle\langle+|\bigr)\\ &=\frac18\begin{pmatrix}5&1\\1&3\end{pmatrix}. \end{aligned}$$

Its Bloch vector is $(1/4,0,1/4)$, of length $1/\sqrt8$. Hence

$$\lambda_\pm=\frac{1\pm1/\sqrt8}{2}, \qquad \operatorname{Tr}\rho_T^2=\frac9{16}.$$

The mixed target is entangled with the selector because the full state is pure. Each branch is a Hermitian involution, so $W^\dagger=W$ and $W^2=I$: applying the same SELECT again restores the original product state. In contrast, classically choosing and recording one branch can produce the same reduced $\rho_T$ without retaining these coherent selector–target correlations. The reduced target alone does not distinguish those preparations.

Exercises

Exercise 1: Projector Proof and Inverse

Let $W=\sum_jP_j\otimes U_j$. Prove that $W$ is unitary and derive its inverse when the $P_j$ form a complete orthogonal resolution and every $U_j$ is unitary. Identify what fails if projectors overlap or one branch is nonunitary.

Solution

Using $P_j^\dagger=P_j$ and $P_jP_k=\delta_{jk}P_j$,

$$W^\dagger W = \sum_{j,k}P_jP_k\otimes U_j^\dagger U_k = \sum_jP_j\otimes I = I.$$

The same calculation gives $WW^\dagger=I$, so

$$W^{-1}=W^\dagger = \sum_jP_j\otimes U_j^\dagger.$$

If projectors overlap, cross terms $P_jP_k\otimes U_j^\dagger U_k$ survive and the branches are not an orthogonal block decomposition. If a branch is nonunitary, its diagonal block contributes $P_j\otimes U_j^\dagger U_j\ne P_j\otimes I$. If the projectors are orthogonal but incomplete, the sum acts as zero on the omitted control subspace rather than as a complete unitary.

Exercise 2: Control Coherence and Entanglement

For input $|+\rangle|\psi\rangle$, derive the reduced control state after $C_1(U)$, its purity, and the exact factorization condition. Explain why a controlled gate need not entangle every input.

Solution

The joint output is

$$|\Psi\rangle = \frac{|0\rangle|\psi\rangle+|1\rangle U|\psi\rangle}{\sqrt2}.$$

Writing $z=\langle\psi|U|\psi\rangle$ and tracing out the target gives

$$\rho_C = \frac12 \begin{pmatrix} 1&z^*\\ z&1 \end{pmatrix}, \qquad \operatorname{Tr}\rho_C^2 = \frac{1+|z|^2}{2}.$$

The output is pure on the control, hence factorized, exactly when $|z|=1$. Equality in Cauchy–Schwarz for normalized $|\psi\rangle$ and $U|\psi\rangle$ then requires $U|\psi\rangle=e^{i\theta}|\psi\rangle$. Thus an eigenstate produces phase kickback without entanglement, while a generic target state can entangle. One non-entangling input does not classify the whole gate as local.

Exercise 3: A Mixed-Polarity 10 Control

Controls are ordered $(c_1,c_0)$ before target $t$. Activate $X_t$ only for $c_1c_0=10$. Give the full truth table, a decomposition using positive-control $CCX$, its abstract resources, and the output on $|++0\rangle$.

Solution

With $P=|10\rangle\langle10|$,

$$W=(I-P)\otimes I+P\otimes X.$$

It fixes $000,001,010,011,110,111$ and swaps

$$100\longleftrightarrow101.$$

Using right-to-left action,

$$W = (I\otimes X\otimes I) CCX (I\otimes X\otimes I).$$

The two $X$ gates open and close the $c_0$ control. In the abstract alphabet $\{X,CCX\}$ this has width three, size three, depth three, and no ancilla. A native mixed-polarity primitive may be counted as size and depth one only if that alphabet is explicitly declared.

On the stated input,

$$W|++0\rangle = \frac{|000\rangle+|010\rangle+|101\rangle+|110\rangle}{2}.$$

Tracing out the controls leaves target probabilities $3/4$ and $1/4$, so

$$\operatorname{Tr}\rho_t^2 = \left(\frac34\right)^2 +\left(\frac14\right)^2 =\frac58,$$

and

$$S(\rho_t) = h_2(1/4) =0.811278\ \text{bits}.$$

Exercise 4: Distinguishing X from Minus X

Explain why $X$ and $-X$ define the same isolated unitary channel but different controlled channels. Give a control-superposition and target-eigenstate test that distinguishes them exactly.

Solution

The isolated channels agree because

$$(-X)\rho(-X)^\dagger=X\rho X.$$

After control is added,

$$C_1(-X)=(Z_C\otimes I)C_1(X).$$

Choose control $|+\rangle$ and target $|+\rangle$, for which $X|+\rangle=|+\rangle$. Then

$$C_1(X)|+\rangle|+\rangle =|+\rangle|+\rangle,$$

whereas

$$C_1(-X)|+\rangle|+\rangle =|-\rangle|+\rangle.$$

The outputs are orthogonal. An $X$-basis measurement of the control distinguishes them with certainty, demonstrating that branch-relative phase is observable.

Exercise 5: Changing a SELECT branch

Replace the $H$ branch in the four-branch example by $S=\operatorname{diag}(1,i)$ and prepare $|++\rangle_C|+\rangle_T$. Find the output, reduced target state, spectrum, purity, and inverse. Is the new SELECT operator an involution?

Solution

Define $|s\rangle=(|0\rangle+i|1\rangle)/\sqrt2$. Since $I|+\rangle=X|+\rangle=|+\rangle$, $Z|+\rangle=|-\rangle$, and $S|+\rangle=|s\rangle$, the output is

$$|\Psi\rangle = \frac12\left( |00\rangle|+\rangle +|01\rangle|+\rangle +|10\rangle|-\rangle +|11\rangle|s\rangle \right).$$

Tracing out the selector gives

$$\rho_T = \frac18 \begin{pmatrix} 4&1-i\\ 1+i&4 \end{pmatrix}.$$

Its eigenvalues and purity are

$$\lambda_\pm = \frac{1\pm1/\sqrt8}{2}, \qquad \operatorname{Tr}\rho_T^2=\frac9{16}.$$

Trace one and positive eigenvalues verify a valid reduced state. The exact inverse replaces every branch by its adjoint:

$$W^{-1} = |00\rangle\langle00|\otimes I +|01\rangle\langle01|\otimes X +|10\rangle\langle10|\otimes Z +|11\rangle\langle11|\otimes S^\dagger.$$

Because $S^2=Z\ne I$, the new $W$ is not an involution. This block description does not yet specify an elementary-gate decomposition.

Exercise 6: Controlled powers and their costs

Suppose $U|u\rangle=e^{2\pi i(3/16)}|u\rangle$. For $r=4$, find the kicked-back phase under controlled $U^r$. Then compare a powered-query interface, repetition of a known base circuit, and controlled physical time evolution.

Solution

The powered eigenphase is

$$\left(e^{2\pi i(3/16)}\right)^4 =e^{2\pi i(3/4)} =-i.$$

Thus $C_1(U^4)|+\rangle|u\rangle$ places the relative phase $-i$ on the active control branch. The algebra alone does not fix cost:

• a declared powered oracle may count one query to $U^4$;
• a model that explicitly supplies controlled-$U$ as its base oracle counts four sequential controlled-$U$ queries; plain unknown-$U$ access does not supply this interface;
• a known circuit may repeat its controlled decomposition four times, with four times its size and usually its depth unless parallel structure is proved;
• if $U=e^{-iHt/\hbar}$, direct controlled evolution for $U^4=e^{-iH(4t)/\hbar}$ uses physical evolution time $4t$ unless another simulation interface is supplied.

Query count, gate count, depth, and evolution time must therefore be reported separately.

Exercise 7: Diagnosing Unknown-Black-Box Control

Use $U$ and $e^{i\gamma}U$ to explain why channel-only black-box access does not specify one controlled operation. Give one additional physical resource that can make a practical controlization scheme possible.

Solution

The two black-box channels are identical:

$$U\rho U^\dagger = (e^{i\gamma}U)\rho(e^{i\gamma}U)^\dagger.$$

Their controlled operators differ by

$$C_1(e^{i\gamma}U) = \bigl(P(\gamma)\otimes I\bigr)C_1(U),$$

which is observable on a control superposition. A channel-only interface therefore contains too little information to choose a unique branch phase, so it cannot universally define controlled $U$ in the standard circuit model.

One extra resource is an extended target space with an identity sector of sufficient dimension, plus coherent transfers into and out of that sector. The inactive branch can then bypass the unknown action without revealing which branch occurred. This supplies a known relative reference action absent from the channel-only oracle; it does not establish that an arbitrary unmodified black box can be controlled.

Exercise 8: An energy offset in controlled evolution

Let $H=(\hbar\omega/2)Z$, with $\omega t=\pi/2$. An ideal device applies $U(t)=e^{-iHt/\hbar}$ only on control value one. Find the output on $|+\rangle_c|0\rangle_t$ and the control’s $X,Y$ expectations. What changes if the active-branch Hamiltonian becomes $H+cI$?

Solution

Since $|0\rangle$ has energy $\hbar\omega/2$, the output is

$$\frac{|0\rangle+e^{-i\pi/4}|1\rangle}{\sqrt2}\otimes|0\rangle.$$

Thus $\langle X_c\rangle=1/\sqrt2$ and $\langle Y_c\rangle=-1/\sqrt2$. Adding $cI$ to the active branch only multiplies its amplitude by $e^{-ict/\hbar}$, changing the interference phase to $-\pi/4-ct/\hbar$. The controlled joint Hamiltonian has changed by $c|1\rangle\langle1|_c\otimes I_t$, not by a global multiple of the full identity. A common energy shift of the entire joint Hamiltonian would remain unobservable. This distinction reconciles the controlled phase with the freedom to choose a common energy zero.

References

• M. Araújo, A. Feix, F. Costa, and Č. Brukner, “Quantum circuits cannot control unknown operations,” New Journal of Physics 16, 093026 (2014), doi:10.1088/1367-2630/16/9/093026.
• A. Barenco, C. H. Bennett, R. Cleve, D. P. DiVincenzo, N. Margolus, P. Shor, T. Sleator, J. A. Smolin, and H. Weinfurter, “Elementary gates for quantum computation,” Physical Review A 52, 3457–3467, 1995, doi:10.1103/PhysRevA.52.3457.
• V. Bergholm, J. J. Vartiainen, M. Möttönen, and M. M. Salomaa, “Quantum circuits with uniformly controlled one-qubit gates,” Physical Review A 71, 052330 (2005), doi:10.1103/PhysRevA.71.052330.
• D. W. Berry, A. M. Childs, R. Cleve, R. Kothari, and R. D. Somma, “Simulating Hamiltonian dynamics with a truncated Taylor series,” Physical Review Letters 114, 090502 (2015), doi:10.1103/PhysRevLett.114.090502.
• R. Cleve, A. Ekert, C. Macchiavello, and M. Mosca, “Quantum algorithms revisited,” Proceedings of the Royal Society A 454, 339–354 (1998), doi:10.1098/rspa.1998.0164.
• M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information, 10th Anniversary Edition, Cambridge University Press (2010), doi:10.1017/CBO9780511976667.
• X.-Q. Zhou, T. C. Ralph, P. Kalasuwan, M. Zhang, A. Peruzzo, B. P. Lanyon, and J. L. O’Brien, “Adding control to arbitrary unknown quantum operations,” Nature Communications 2, 413 (2011), doi:10.1038/ncomms1392.