Classically-Controlled Quantum Instruments

• Classically-controlled quantum instruments are operations where a measurement produces a classical outcome and an outcome-dependent quantum state for later processing.
• They enable mid-circuit measurements and feed-forward control, linking measurement outcomes with conditional evolution in adaptive and error-correcting circuits.
• Recent advances focus on rigorous error modeling, branchwise representations, and compositional techniques that bridge classical control with quantum programming semantics.

Searching arXiv for recent and foundational papers on quantum instruments, classically controlled operations, composition, and error modeling. Search query 1: "all:quantum instruments mid-circuit measurement classically controlled feed-forward" Classically-controlled quantum instruments are quantum operations that produce a joint quantum-classical output: a classical outcome is recorded explicitly, and an outcome-dependent post-measurement quantum state is passed to later stages of a computation. In the standard formulation, an instrument is a family of completely positive, trace-non-increasing maps whose sum is trace-preserving, or equivalently a map of the form

$I:\rho \mapsto \sum_i E_i(\rho)\otimes \ketbra{i},$

with outcome probability $p(i|\rho)=\operatorname{Tr}[E_i(\rho)]$ and conditional post-measurement state $\rho_i=E_i(\rho)/\operatorname{Tr}[E_i(\rho)]$. This is the natural formalism for mid-circuit measurements, adaptive circuits, and quantum error correction, because a POVM specifies only outcome statistics and a single CPTP channel suppresses the classical branch label, whereas an instrument retains both (Hashim, 20 Apr 2026).

1. Formal definition and conceptual scope

A quantum instrument may be written either as a family $\mathcal I=\{E_i\}_i$ of completely positive maps with $\sum_i E_i$ trace-preserving, or as a hybrid map

$$I:\mathcal L(\mathcal H_A)\mapsto \mathcal L(\mathcal H_B)\otimes \mathcal L(\mathcal H_K),$$

where $\mathcal H_K$ is the Hilbert space of the classical register and the output is block-diagonal in the basis $\{\ket{i}\}$ of classical outcomes (Hashim, 20 Apr 2026). In this form, the classical register is not treated as an ordinary coherent subsystem; it is represented by orthogonal labels $\ketbra{i}$, and the quantum state correlated with outcome $i$ lives in a different output space. This separation is what makes later classical feed-forward meaningful.

For non-destructive measurement, especially in fault-tolerant architectures, the relevant instrument often acts only on a subsystem. An ideal projective instrument may be written

$p(i|\rho)=\operatorname{Tr}[E_i(\rho)]$0

with orthogonal projectors $p(i|\rho)=\operatorname{Tr}[E_i(\rho)]$1 satisfying $p(i|\rho)=\operatorname{Tr}[E_i(\rho)]$2. In the subsystem-measurement case discussed for mid-circuit measurement, $p(i|\rho)=\operatorname{Tr}[E_i(\rho)]$3, so one subsystem is measured while the complementary subsystem $p(i|\rho)=\operatorname{Tr}[E_i(\rho)]$4 remains available for subsequent computation (McLaren et al., 2023). This setting is the canonical interface between measurement and later classically controlled correction.

This suggests that “classically-controlled quantum instrument” is best understood operationally rather than as a single universally fixed formal subclass. Across the recent literature, it denotes the primitive that sits at a branch point of a hybrid computation: a classical record is emitted, and the branch-resolved quantum state is preserved for later use (Hashim, 20 Apr 2026).

2. Mid-circuit measurement, feed-forward, and adaptive computation

The immediate use-case for classically-controlled instruments is mid-circuit measurement. These are “required to model mid-circuit measurements (MCMs) and the dependence of the post-measurement state on the measurement outcome,” and this role is “essential for applications using MCMs, such as adaptive circuits and quantum error correction” (Hashim, 20 Apr 2026). In this regime, the branch label is not a terminal report. It is a control signal that selects later operations, while the associated branch state is the actual quantum input to the next stage.

This architecture is explicit in quantum error correction and quantum non-demolition protocols. In a trapped-ion loss-detection example, the instrument is realized by coupling data and ancilla, measuring the ancilla, and obtaining a hybrid output

$p(i|\rho)=\operatorname{Tr}[E_i(\rho)]$5

after which the classical outcome determines whether a replacement qubit is inserted and how later syndrome processing proceeds (Stricker et al., 2021). The same logic underlies ancilla readout and syndrome extraction in surface-code-style settings: the instrument precedes the corrective map, but its correctness already determines whether the subsequent feed-forward can be trusted (McLaren et al., 2023).

A natural inference, consistent with the instrument formalism and explicitly noted in the literature, is that if a conditional continuation $p(i|\rho)=\operatorname{Tr}[E_i(\rho)]$6 is applied after the instrument, the adaptive step has the form

$p(i|\rho)=\operatorname{Tr}[E_i(\rho)]$7

That inference is not presented as a primary equation in the note on interpretation, but it is exactly the operational meaning of the repeated claim that MCMs are reused in further computations (Hashim, 20 Apr 2026).

Dynamic-circuit compilation studies treat the same structure from the programming side. A recent framework supports mid-circuit computational-basis measurements, resets, and classically controlled operations with Boolean guards, and rewrites part of the runtime feed-forward into probabilistic compile-time structure via “probabilistic controls” $p(i|\rho)=\operatorname{Tr}[E_i(\rho)]$8. On randomly generated dynamic circuits it reports about $p(i|\rho)=\operatorname{Tr}[E_i(\rho)]$9 classical feedforward reduction on average, with $\rho_i=E_i(\rho)/\operatorname{Tr}[E_i(\rho)]$0 removal in favorable settings and $\rho_i=E_i(\rho)/\operatorname{Tr}[E_i(\rho)]$1 removal at fixed width with increasing depth (Fulginiti et al., 27 May 2026). The underlying semantic object is still measurement-plus-conditional continuation, i.e. an instrument followed by classically controlled evolution.

3. Representations and interpretability

A central issue is how to represent an instrument so that both its classical statistics and its quantum back-action remain visible. One route is branchwise superoperator or transfer-matrix representation. For an ordinary process $\rho_i=E_i(\rho)/\operatorname{Tr}[E_i(\rho)]$2, the Pauli transfer matrix is

$\rho_i=E_i(\rho)/\operatorname{Tr}[E_i(\rho)]$3

and for an instrument one assigns one $\rho_i=E_i(\rho)/\operatorname{Tr}[E_i(\rho)]$4 superoperator to each outcome branch $\rho_i=E_i(\rho)/\operatorname{Tr}[E_i(\rho)]$5 (Hashim, 20 Apr 2026). The interpretive subtlety is that each branch is CP but generally not TP, so a branch PTM cannot be read as if it were an independent deterministic gate. Only the sum of the branches satisfies the TP constraint; in the two-outcome case,

$\rho_i=E_i(\rho)/\operatorname{Tr}[E_i(\rho)]$6

This branchwise view has a precise operational meaning. The top row of the branch PTM encodes outcome probabilities and the effective POVM effect,

$\rho_i=E_i(\rho)/\operatorname{Tr}[E_i(\rho)]$7

while the rest of the matrix captures back-action and conditional state update (Hashim, 20 Apr 2026). The left column describes what state is prepared on that branch from a maximally mixed input; the unital block shows how coherences are transmitted or destroyed. This is why a confusion matrix or assignment-fidelity metric is only a partial description: it captures computational-basis discrimination, but not generally the conditional state propagated forward in an adaptive protocol.

A second, more structural representation is the Choi picture. For branch maps $\rho_i=E_i(\rho)/\operatorname{Tr}[E_i(\rho)]$8, one may reconstruct each branch from its Choi operator and compare instruments at the level of the full block-diagonal Choi state. A useful identity is that for block-diagonal instruments the square-root fidelity decomposes over outcome blocks: $\rho_i=E_i(\rho)/\operatorname{Tr}[E_i(\rho)]$9 so the generalized instrument fidelity combines both correct outcome statistics and correct branch-conditioned state updates (McLaren et al., 2023). This is exactly the quantity that matters when the classical output is later consumed by control logic.

4. Noise models, distances, and certification

Generic imperfect instruments are difficult to analyze directly. For subsystem measurements, an imperfect implementation may be written

$\mathcal I=\{E_i\}_i$0

where each $\mathcal I=\{E_i\}_i$1 is completely positive and $\mathcal I=\{E_i\}_i$2 is trace-preserving (McLaren et al., 2023). In this form, experimentally relevant error modes include reporting the wrong outcome, flipping the measured subsystem after measurement, and applying stochastic noise to the unmeasured subsystem.

A major recent development is the definition of stochastic and uniform stochastic instruments as the instrument analogue of stochastic noise models for gates. For arbitrary implementations there are efficiently computable bounds on the diamond distance from the ideal instrument. Writing $\mathcal I=\{E_i\}_i$3, one has

$\mathcal I=\{E_i\}_i$4

for suitable $\mathcal I=\{E_i\}_i$5, and

$\mathcal I=\{E_i\}_i$6

which reduces certification to branchwise Choi data (McLaren et al., 2023). In practice, this makes instrument-level validation compatible with Choi- or Kraus-based reconstruction workflows.

For the special class of uniform stochastic instruments, the analysis simplifies sharply. The key equality is

$\mathcal I=\{E_i\}_i$7

where $\mathcal I=\{E_i\}_i$8 is the probability that no error at all occurs during the instrument (McLaren et al., 2023). In that class, diamond distance, generalized process infidelity, and error probability coincide. The caveat is essential: the same equivalence fails for nonuniform stochastic instruments, because diamond distance is controlled by the worst branch whereas fidelity averages over branches.

A common misconception is therefore that a single readout metric suffices. Recent guidance on branchwise superoperators argues the opposite: assignment error, measurement-axis misalignment, dephasing, coherent rotation during readout, and non-unital effects can all coexist inside one instrument, and a good confusion matrix does not imply a good branch-conditioned post-measurement state (Hashim, 20 Apr 2026).

5. Characterization, simulability, and resource limitations

Experimental characterization of instruments cannot simply recycle standard channel tomography. Since each branch map is CP but not generally TP, trace preservation must not be imposed branchwise. A general recipe is to reconstruct each $\mathcal I=\{E_i\}_i$9 separately, using conditional data and positivity constraints but omitting the usual TP constraint. In Choi form, one may use

$\sum_i E_i$0

or a positivity-constrained weighted estimator, subject to $\sum_i E_i$1 but not $\sum_i E_i$2 (Stricker et al., 2021). Applied to a trapped-ion loss-detection instrument, this exposed false positives and false negatives that were invisible in qubit-only or unconditional descriptions, and subsequent logical-level simulations showed that false negatives are much more harmful than false positives (Stricker et al., 2021).

A separate line of work studies which instruments can be realized by a restricted classically controlled architecture: classical randomness, projective measurements, and outcome-conditioned quantum channels. Such “projective instruments” have the form

$\sum_i E_i$3

and their simulability is tied to Schmidt-number constraints on the branch Choi operators (Khandelwal et al., 2 Mar 2025). For qubit input, the resulting condition is both necessary and sufficient; for higher dimension, an SDP relaxation gives an efficient necessary criterion (Khandelwal et al., 2 Mar 2025). A key consequence is that projective simulability at the POVM level does not imply simulability at the instrument level: the outcome statistics may be reproducible even when the post-measurement state-update structure is not.

Resource-theoretic work sharpens this point further by defining “non-interactive instruments”

$\sum_i E_i$4

which discard the input and generate classical outcome and output state independently of it (Hsieh et al., 29 Mar 2026). In that formal sense, classically controlled behavior is identified with internal classical randomness plus conditional state preparation, and everything outside that free set is “interactive.” The associated robustness

$\sum_i E_i$5

admits three operational meanings, including entanglement preservation after local measurement and recovery of classical information generated from measuring half of a maximally entangled state (Hsieh et al., 29 Mar 2026). This suggests that “classically controlled instrument” can denote either a broad operational category or a much narrower free class, depending on context.

6. Composition, programming, and physical realization

For discrete adaptive circuits, composition is often written informally as “measure, record the branch, then apply the next branch-conditioned map.” For general measurable outcome spaces, however, composition requires an integral of channel-valued functions against an instrument. A recent Heisenberg-picture theory defines instruments as CP-subunital-map-valued measures $\sum_i E_i$6 and composes them by

$\sum_i E_i$7

using the Okamura–Ozawa normal extension to a von Neumann tensor product (Booth et al., 26 Jun 2026). The resulting multiplication defines a parameterised monad, and the category of quantum Markov kernels becomes its Kleisli category. In finite-outcome settings this reduces to the expected sum over branches; the contribution is the continuous-outcome generalization.

Programming-language semantics has moved in a parallel direction. One recent typed language combines classical control—control flow based on classical information, potentially resulting from measurement—with quantum control—control based on superposition—by introducing a modality $\sum_i E_i$8 that embeds pure quantum types into a mixed-state calculus. Measurement appears as

$\sum_i E_i$9

and denotational soundness takes the branch-summing form

$$I:\mathcal L(\mathcal H_A)\mapsto \mathcal L(\mathcal H_B)\otimes \mathcal L(\mathcal H_K),$$0

(Dave et al., 27 Nov 2025). Although this is not packaged as a first-class “instrument type,” it is equivalent in content to outcome-indexed CP branches plus classical continuation.

At the systems level, classically controlled instruments depend on a nontrivial classical control plane. Hardware-aware control theory models the classical signal chain as a distortion map $$I:\mathcal L(\mathcal H_A)\mapsto \mathcal L(\mathcal H_B)\otimes \mathcal L(\mathcal H_K),$$1, so that one optimizes $$I:\mathcal L(\mathcal H_A)\mapsto \mathcal L(\mathcal H_B)\otimes \mathcal L(\mathcal H_K),$$2 rather than an idealized pulse directly (Hincks et al., 2014). Trapped-ion implementations similarly rely on FPGA-based PI/PID loops, DDS frequency synthesis, DAC-based electrode control, shared atomic-standard-derived clocks, and explicit feedforward updates such as

$$I:\mathcal L(\mathcal H_A)\mapsto \mathcal L(\mathcal H_B)\otimes \mathcal L(\mathcal H_K),$$3

to preserve Raman resonance under repetition-rate drift (Mount et al., 2015). In that sense, the abstract instrument interface—classical record plus branch-resolved quantum action—rests on a substantial classical infrastructure that stabilizes, times, and routes the control information on which adaptive quantum computation depends.