Local Measurement & Classical Feedforward

• Local measurement and classical feedforward is an operational paradigm that integrates local quantum measurements with adaptive classical controls based on recorded outcomes.
• It underpins protocols in quantum networking, LOCC, and measurement-based quantum computation by converting local outcomes into deterministic gates and state-preparation maps.
• It employs adaptive strategies, scheduling algorithms, and continuous-feedback techniques to overcome causal ambiguities and enhance performance in distributed quantum systems.

Local measurement and classical feedforward is an operational paradigm in which measurements are performed on local degrees of freedom—sites, modes, qubits, or network nodes—while subsequent bases, corrections, controls, or decision rules are chosen as functions of recorded outcomes. In distributed quantum protocols it mediates basis propagation, outcome feedback, and Pauli-frame updates; in LOCC it defines adaptive POVMs with bounded classical messages; in measurement-based quantum computation it converts local measurements on an entangled resource into deterministic gates or state-preparation maps; and in continuous-measurement theory it appears as a classical signal derived from a local record and fed back into the quantum dynamics (Søndergaard et al., 12 Feb 2026, Dutra et al., 10 Oct 2025, Tantivasadakarn et al., 2022, Tilloy, 2024).

1. Operational definitions and formal models

A standard fixed-direction one-round LOCC measurement from $A\to B$ has the form

$$M^\lambda = \sum_{a=1}^m A^a \otimes B^{\lambda|a}, \qquad \sum_a A^a = I_A, \qquad \sum_\lambda B^{\lambda|a} = I_B \text{ for each } a,$$

with $|\{a\}| \le 2^{m_{\mathrm{bits}}}$ enforcing the message-size bound. In this representation, Alice measures first, communicates a classical symbol $a$, and Bob chooses a conditional POVM. The same framework distinguishes adaptive from non-adaptive strategies: non-adaptive measurements admit a parent POVM on the second party plus classical post-processing, whereas adaptive measurements allow the second party’s POVM to depend on the first outcome (Dutra et al., 10 Oct 2025).

In measurement-based state preparation, a “shot” is one simultaneous layer of single-site measurements followed by classical feedforward and a finite-depth unitary correction. The minimal number of such measurement layers defines a hierarchy of phases under finite-depth unitaries, measurements, and feedforward. In this formulation, multiple rounds count only when outcomes are used adaptively (Tantivasadakarn et al., 2022).

In continuous hybrid quantum–classical dynamics, local measurement is specified by observables $L_k$ and a normalized current

$$dY_{k,t} = \langle L_k + L_k^\dagger\rangle_t\, dt + dW_{k,t},$$

with innovation noise $dW_{k,t}$. The corresponding conditional quantum evolution is

$$d\rho_t = -i[H(t),\rho_t]dt + \sum_k D[L_k]\rho_t\, dt + \sum_k \sqrt{\eta_k}\, H[L_k]\rho_t\, dW_{k,t},$$

while classical controller variables evolve by stochastic differential equations driven by the same record. In this setting, the quantum-to-classical channel is the measurement current, and the classical-to-quantum channel is a causal feedback Hamiltonian or control law (Tilloy, 2024).

2. Causal ordering in distributed quantum networks

In measurement-based quantum networking, local measurements on a pre-shared graph state generate outcomes that determine local Clifford correction operations on other nodes. Classical feedforward communicates measurement bases and outcomes hop-by-hop so that later measurements and end-node corrections are consistent with the actual measurement sequence. The central difficulty is that heterogeneous measurement durations and variable local processing can make a missing outcome ambiguous: a node cannot tell whether the cause is slow local measurement or delayed classical propagation. The ambiguity persists even in simple line networks with only Pauli measurements, and Lamport clocks do not remove it because concurrent local measurements exchange no direct causal messages (Søndergaard et al., 12 Feb 2026).

The temporal framework introduced for this setting imposes a time-division architecture with quantum slots of length $T_q$ and classical slots of length $T_c=1$. A measuring node $$M^\lambda = \sum_{a=1}^m A^a \otimes B^{\lambda|a}, \qquad \sum_a A^a = I_A, \qquad \sum_\lambda B^{\lambda|a} = I_B \text{ for each } a,$$0 receives a slot-and-basis assignment $$M^\lambda = \sum_{a=1}^m A^a \otimes B^{\lambda|a}, \qquad \sum_a A^a = I_A, \qquad \sum_\lambda B^{\lambda|a} = I_B \text{ for each } a,$$1, where $$M^\lambda = \sum_{a=1}^m A^a \otimes B^{\lambda|a}, \qquad \sum_a A^a = I_A, \qquad \sum_\lambda B^{\lambda|a} = I_B \text{ for each } a,$$2 and $$M^\lambda = \sum_{a=1}^m A^a \otimes B^{\lambda|a}, \qquad \sum_a A^a = I_A, \qquad \sum_\lambda B^{\lambda|a} = I_B \text{ for each } a,$$3. Causality preservation is characterized by two constraints. The feedforward constraint requires

$$M^\lambda = \sum_{a=1}^m A^a \otimes B^{\lambda|a}, \qquad \sum_a A^a = I_A, \qquad \sum_\lambda B^{\lambda|a} = I_B \text{ for each } a,$$4

so that a node measures only after it knows its logical basis. The adjacency constraint requires

$$M^\lambda = \sum_{a=1}^m A^a \otimes B^{\lambda|a}, \qquad \sum_a A^a = I_A, \qquad \sum_\lambda B^{\lambda|a} = I_B \text{ for each } a,$$5

so that adjacent nodes never measure concurrently when $$M^\lambda = \sum_{a=1}^m A^a \otimes B^{\lambda|a}, \qquad \sum_a A^a = I_A, \qquad \sum_\lambda B^{\lambda|a} = I_B \text{ for each } a,$$6 measurements are used. With only $$M^\lambda = \sum_{a=1}^m A^a \otimes B^{\lambda|a}, \qquad \sum_a A^a = I_A, \qquad \sum_\lambda B^{\lambda|a} = I_B \text{ for each } a,$$7 measurements, the correction operations that land on a common neighbor belong to $$M^\lambda = \sum_{a=1}^m A^a \otimes B^{\lambda|a}, \qquad \sum_a A^a = I_A, \qquad \sum_\lambda B^{\lambda|a} = I_B \text{ for each } a,$$8 and commute, which makes concurrent non-adjacent measurements unambiguous. The slot index therefore induces a partial order: all measurements in slot $$M^\lambda = \sum_{a=1}^m A^a \otimes B^{\lambda|a}, \qquad \sum_a A^a = I_A, \qquad \sum_\lambda B^{\lambda|a} = I_B \text{ for each } a,$$9 must be causally interpreted before slot $|\{a\}| \le 2^{m_{\mathrm{bits}}}$0 (Søndergaard et al., 12 Feb 2026).

The resource-state model is a simple graph $|\{a\}| \le 2^{m_{\mathrm{bits}}}$1 with graph state

$|\{a\}| \le 2^{m_{\mathrm{bits}}}$2

and single-qubit Pauli measurements $|\{a\}| \le 2^{m_{\mathrm{bits}}}$3 with $|\{a\}| \le 2^{m_{\mathrm{bits}}}$4. The analysis concentrates on $|\{a\}| \le 2^{m_{\mathrm{bits}}}$5 measurements because the sufficiency of the adjacency rule depends on the commutation structure of their correction operations. A stated caveat is that if coherence times are too short to delay measurements into prescribed slots, or if $|\{a\}| \le 2^{m_{\mathrm{bits}}}$6-basis measurements are included, stronger constraints or additional signaling may be required (Søndergaard et al., 12 Feb 2026).

3. Scheduling algorithms, local-network coordination, and constant-time feedforward

For line networks, the scheduling problem becomes constructive. If the path from source $|\{a\}| \le 2^{m_{\mathrm{bits}}}$7 to receiver $|\{a\}| \le 2^{m_{\mathrm{bits}}}$8 has length $|\{a\}| \le 2^{m_{\mathrm{bits}}}$9, the sequential regime $a$0 yields one measurement per slot and $a$1. In the fully parallel regime $a$2, many inner nodes become feedforward-eligible early and the slot count can drop substantially; the summary bounds are

$a$3

The proposed line-network procedure defines

$a$4

then, in each round, chooses the first element of $a$5 and assigns slot $a$6 to that node and every odd-indexed element in $a$7. In a 1D cluster, each round assigns $a$8 nodes while maintaining both feedforward and adjacency (Søndergaard et al., 12 Feb 2026).

A related networking construction replaces global routing and global Pauli-frame tracking by strictly local measurement and feedforward. In graph-state generation with relay nodes, the key identity for a star graph with center $a$9 and leaves $L_k$0 is

$L_k$1

Under the independent-neighborhood condition, the recovery after an $L_k$2-basis measurement collapses to a single-qubit correction on a designated neighbor: $L_k$3 This supports protocols for star-subgraph generation, center transformation, and star fusion in which each measurement outcome is a single bit broadcast only to immediate neighbors, and each correction is a constant-depth local Clifford. The stated complexity is amortized $L_k$4 local decision complexity, with entanglement cost per delivered terminal qubit

$L_k$5

Within the dual-species trapped-ion timing model, the total execution time remains far below the quoted memory coherence time $L_k$6, and the main bottleneck is readout fidelity rather than classical delay (Zheng et al., 15 Jun 2026).

This conjunction of results isolates two distinct coordination layers. One is a slot-based temporal layer that guarantees a unique causal interpretation of outcomes. The other is a locality-preserving graph-transformation layer in which byproducts are engineered to collapse to $L_k$7 feedforward rules. A plausible implication is that large-scale quantum networking requires both kinds of structure: temporal disambiguation and locality of corrective action.

4. Adaptive discrimination, LOCC structure, and network nonclassicality

In one-round LOCC, the convex hull of $L_k$8 with at most $L_k$9 first-stage outcomes admits an exact characterization in the limit $$dY_{k,t} = \langle L_k + L_k^\dagger\rangle_t\, dt + dW_{k,t},$$0 via constrained symmetric extensions, and the corresponding SDP hierarchy yields convergent upper bounds for tasks such as minimum-error state discrimination. The same work gives an analogous hierarchy for the convex hull of non-adaptive one-round protocols, thereby certifying gaps between adaptive and non-adaptive strategies. Explicit examples show directional asymmetry and message-size dependence: for an iso-entangled two-qubit basis with one bit of communication, the optimal success probabilities are

$$dY_{k,t} = \langle L_k + L_k^\dagger\rangle_t\, dt + dW_{k,t},$$1

with $$dY_{k,t} = \langle L_k + L_k^\dagger\rangle_t\, dt + dW_{k,t},$$2 for all $$dY_{k,t} = \langle L_k + L_k^\dagger\rangle_t\, dt + dW_{k,t},$$3. For the Double Trine ensemble, adaptive one-round LOCC strictly outperforms all non-adaptive strategies, and non-projective POVMs outperform projective ones (Dutra et al., 10 Oct 2025).

A multishot channel-discrimination variant uses local probe preparation, per-shot Helstrom measurements, and classical feedforward. For two qubit amplitude-damping channels, the Bayesian strategy updates the full posterior after each outcome, whereas the Markovian strategy conditions only on the immediately preceding outcome. The reported conclusion is that Bayesian feedforward is only slightly advantageous, and only in a limited region centered around $$dY_{k,t} = \langle L_k + L_k^\dagger\rangle_t\, dt + dW_{k,t},$$4; away from that region, Markovian and Bayesian performance essentially coincide and match the global strategy shown in the numerical plots (Rexiti et al., 2 Jun 2025).

Measurement adaptivity also reduces measurement error. For the imperfect $$dY_{k,t} = \langle L_k + L_k^\dagger\rangle_t\, dt + dW_{k,t},$$5 POVM

$$dY_{k,t} = \langle L_k + L_k^\dagger\rangle_t\, dt + dW_{k,t},$$6

the minimal total error is $$dY_{k,t} = \langle L_k + L_k^\dagger\rangle_t\, dt + dW_{k,t},$$7. For $$dY_{k,t} = \langle L_k + L_k^\dagger\rangle_t\, dt + dW_{k,t},$$8 uses, adaptive circuits strictly outperform all non-adaptive circuits in the region $$dY_{k,t} = \langle L_k + L_k^\dagger\rangle_t\, dt + dW_{k,t},$$9; for $dW_{k,t}$0, adaptive advantage holds on almost the entire parameter space except the boundaries $dW_{k,t}$1 and $dW_{k,t}$2; and across the class the relative advantage $dW_{k,t}$3 can be unbounded as $dW_{k,t}$4 increases (Byrne et al., 19 Jun 2026).

Classical feedforward can even replace entangled measurements in network nonlocality. In the bilocal scenario, separable measurements at the central node augmented with bidirectional feedforward achieve full network nonlocality and minimal network nonclassicality. The adaptive central-node measurement is

$dW_{k,t}$5

with final output $dW_{k,t}$6. For two singlets, the achieved witness values are $dW_{k,t}$7, and under identical Werner-state noise the separable-feedforward strategy certifies full network nonlocality for $dW_{k,t}$8 (Polino et al., 13 Apr 2026).

5. Measurement-based computation, finite-shot preparation, and nonlinear feedforward

In cluster-state MBQC, local projective measurements drive computation and classical feedforward updates later measurement angles. For a perfect 1D cluster implementing $dW_{k,t}$9, the adaptive angles satisfy

$$d\rho_t = -i[H(t),\rho_t]dt + \sum_k D[L_k]\rho_t\, dt + \sum_k \sqrt{\eta_k}\, H[L_k]\rho_t\, dW_{k,t},$$0

On SLOCC-transformed cluster resources, two strategies were isolated. N-type transformations preserve unitarity of the teleported gate but can induce non-Pauli byproducts, repaired quasi-deterministically by alternating with unitary sites. B-type transformations preserve a Pauli-$$d\rho_t = -i[H(t),\rho_t]dt + \sum_k D[L_k]\rho_t\, dt + \sum_k \sqrt{\eta_k}\, H[L_k]\rho_t\, dW_{k,t},$$1 byproduct but yield non-unitary complex-angle rotations, repaired probabilistically in 1D and promoted to universality in higher-dimensional constructions by percolation (D'Souza et al., 2011).

The finite-shot hierarchy of measurement and feedforward makes this structure algebraic. Single-site measurement layers together with feedforward can prepare all Abelian quantum doubles and all nilpotency-class-two non-Abelian quantum doubles in one shot; general metabelian groups require two shots; solvable groups require exactly $$d\rho_t = -i[H(t),\rho_t]dt + \sum_k D[L_k]\rho_t\, dt + \sum_k \sqrt{\eta_k}\, H[L_k]\rho_t\, dW_{k,t},$$2 shots along the derived series; and non-solvable quantum doubles together with Fibonacci and double Fibonacci are conjectured not to be preparable by any finite number of shots. In the nil-2 construction, outcomes label Abelian anyons from a Lagrangian subgroup and are removed by single-layer string feedforward (Tantivasadakarn et al., 2022).

Optical continuous-variable MBQC introduces nonlinear feedforward. In the demonstrated nonlinear quadrature measurement, the adaptive homodyne angle and rescaling are

$$d\rho_t = -i[H(t),\rho_t]dt + \sum_k D[L_k]\rho_t\, dt + \sum_k \sqrt{\eta_k}\, H[L_k]\rho_t\, dW_{k,t},$$3

and the measured observable is

$$d\rho_t = -i[H(t),\rho_t]dt + \sum_k D[L_k]\rho_t\, dt + \sum_k \sqrt{\eta_k}\, H[L_k]\rho_t\, dW_{k,t},$$4

With a non-Gaussian ancilla, the experiment reported a $$d\rho_t = -i[H(t),\rho_t]dt + \sum_k D[L_k]\rho_t\, dt + \sum_k \sqrt{\eta_k}\, H[L_k]\rho_t\, dW_{k,t},$$5 reduction of measurement excess noise relative to vacuum ancilla, thereby realizing the adaptive primitive required for deterministic measurement-induced non-Gaussian processing and for nonlinear classical post-processing in GKP-style fault-tolerant protocols (Sakaguchi et al., 2022).

A complementary structural result shows that a broad class of linear-depth sequential unitary circuits and constant-depth measurement-feedback circuits are dual under a spacetime rotation. Applied to GHZ preparation, the rotation maps a non-invertible Kramers–Wannier duality circuit to a constant-depth gauging circuit in which local $$d\rho_t = -i[H(t),\rho_t]dt + \sum_k D[L_k]\rho_t\, dt + \sum_k \sqrt{\eta_k}\, H[L_k]\rho_t\, dW_{k,t},$$6-measurements and feedforward deterministically prepare the GHZ state. The same logic extends to toric-code and fractal-order constructions, and also yields protocols that use only a constant number of qubits to measure disorder operators and measurement-induced long-range order (Lu et al., 16 Jul 2025).

6. Continuous-time feedback, complexity boundaries, and unsettled questions

For Gaussian bosonic circuits with adaptive local measurements and Gaussian feedforward, the decisive parameter for the quantum mean-value problem is the total number $$d\rho_t = -i[H(t),\rho_t]dt + \sum_k D[L_k]\rho_t\, dt + \sum_k \sqrt{\eta_k}\, H[L_k]\rho_t\, dW_{k,t},$$7 of adaptively measured modes, or equivalently the number $$d\rho_t = -i[H(t),\rho_t]dt + \sum_k D[L_k]\rho_t\, dt + \sum_k \sqrt{\eta_k}\, H[L_k]\rho_t\, dW_{k,t},$$8 of adaptive measurement-and-feedforward steps. When $$d\rho_t = -i[H(t),\rho_t]dt + \sum_k D[L_k]\rho_t\, dt + \sum_k \sqrt{\eta_k}\, H[L_k]\rho_t\, dW_{k,t},$$9, mean-value estimation for product observables admits efficient classical randomized algorithms even for highly non-Gaussian product inputs; with sufficiently many adaptive steps, the same architectures become universal and mean-value estimation becomes BQP-complete. This contrasts with sampling, where non-Gaussianity alone often induces hardness. The classical algorithms rely on median-of-means estimators, chain-rule sampling, and a generalization of Gurvits’s second algorithm to arbitrary product inputs and Gaussian circuits (Oh et al., 31 Aug 2025).

In continuous measurement-feedback dynamics, locality and causality are imposed directly at the stochastic-equation level. Local currents $T_q$0 are the only quantum-to-classical signals, while feedback Hamiltonians are causal functionals of the past record. Markovian feedback proportional to $T_q$1 produces an averaged Lindbladian

$T_q$2

and non-Markovian feedback is realized by filtering the record through classical variables obeying linear stochastic dynamics. The measurement-based formulation guarantees complete positivity without additional positivity constraints (Tilloy, 2024).

Several recurring misconceptions are explicitly contradicted by the literature. Asynchronous timestamps or buffering do not reconstruct the causal order of independent local measurements in heterogeneous quantum networks (Søndergaard et al., 12 Feb 2026). Separable measurements do not preclude full network nonlocality when bidirectional local feedforward is allowed (Polino et al., 13 Apr 2026). Adaptivity does not always help: for the symmetric imperfect $T_q$3 measurement with $T_q$4, for $T_q$5, and for the trine POVM under the stated criterion, adaptive and non-adaptive performance coincide (Byrne et al., 19 Jun 2026). Non-Gaussian ingredients do not by themselves determine hardness of expectation-value estimation in adaptive Gaussian circuits; bounded adaptivity can preserve efficient classical simulability (Oh et al., 31 Aug 2025).

Open problems are correspondingly heterogeneous. General graphs make maximum-independent-set selection per round NP-hard in causality-preserving network scheduling, and practical deployments require explicit robustness to loss, jitter, and guard times (Søndergaard et al., 12 Feb 2026). Multi-round LOCC with per-round message budgets remains open beyond the one-round hierarchies (Dutra et al., 10 Oct 2025). In finite-shot state preparation, non-preparability conjectures for non-solvable quantum doubles and Fibonacci order remain unresolved (Tantivasadakarn et al., 2022). Taken together, these results suggest that local measurement and classical feedforward is best understood not as a single protocol family, but as a unifying control principle whose power depends on how locality, timing, adaptivity, and classical side information are formalized in each physical setting.