Distributed Quantum Sensing

Distributed quantum sensing uses quantum states or quantum correlations shared among spatially separated sensor nodes to estimate a declared property of a field, signal, or set of local parameters. A network might estimate the average phase across several optical paths, a frequency difference between remote clocks, a magnetic-field gradient, or one spatial mode of a fluctuating field. The target is often a weighted functional

$$q := \mathbf w^{\mathsf T}\boldsymbol\theta = \sum_{m=1}^{M}w_m\theta_m,$$

where $\theta_m$ is the response accumulated at node $m$ and $\mathbf w$ defines the spatial mode of interest.

The phrase does not mean merely operating several quantum sensors. Independent quantum sensors followed by classical averaging already form a distributed sensor system. A specifically distributed quantum advantage requires a matched comparison showing that inter-node quantum resources improve the declared network task beyond the best allowed node-separable protocol. The comparison must include local nonclassical resources when those are allowed, not only coherent states or independent particles.

Entanglement is strongly task-dependent. It can directly concentrate information into a global functional such as an average. It need not improve, and can even obstruct, reconstruction of every local parameter. Loss at one node can erase a global correlation, and a shared oscillator, synchronized control, accepted-shot rule, or high-power local oscillator may dominate the experimental cost. The right question is therefore not whether a network is entangled, but whether its complete estimator gains information about the specified spacetime mode per matched resource and per unit wall time.

Required background. Quantum Measurement as Estimation supplies the likelihood and nuisance-parameter model; Classical and Quantum Fisher Information supplies local precision bounds; Standard Quantum Limit sets the independent-probe comparison; GHZ States explains the collective coherence used below.

Spatial field modes and local phase encoding

Local responses are filtered field samples

A sensor does not read an instantaneous field value at a mathematical point. Let $s(\mathbf x,t)$ be a physical field and let node $m$ have position $\mathbf x_m$, calibrated temporal response $h_m(t)$, and coupling coefficient $\kappa_m$. Its dimensionless encoded parameter can be written

$$\theta_m := \kappa_m \int_{-\infty}^{\infty} h_m(t) s(\mathbf x_m,t) \,\mathrm dt.$$

Finite sensor volume can be included by replacing the point sample with a spatial integral. Clock interrogation windows, interferometer sensitivity functions, dynamical-decoupling filters, and optical transit times all enter through $h_m$. A network therefore samples a collection of spacetime modes, not an abstract vector detached from hardware.

The target might be:

• a normalized average, $q=M^{-1}\sum_m\theta_m$;
• a difference, $q=(\theta_1-\theta_2)/2$;
• a finite-difference gradient or curvature;
• a matched filter for a predicted spatial profile;
• one coefficient in a basis expansion of a field; or
• the full vector $\boldsymbol\theta$.

These are different statistical problems. A state optimized for the average may deliberately contain little information about differences. It cannot then be credited with high-resolution imaging of all nodes.

The local encoding model

For $M$ nodes with locally acting generators $G_m$, a standard phase model is

$$U_{\boldsymbol\theta} = \bigotimes_{m=1}^{M} \exp\!\left(-i\theta_mG_m\right).$$

Operators at distinct nodes commute because they act on distinct tensor factors. The input state need not factorize: it may contain particle entanglement within each node, mode entanglement among nodes, or both. After the local interactions, measurements may be local with classical post-processing or genuinely joint. Which operations are allowed is part of the comparison.

For a parameter vector, a measurement produces a Fisher matrix $\mathbf F_{\theta}$. If all components are locally identifiable and an unbiased estimator is used in the regular asymptotic regime, then estimating $q=\mathbf w^{\mathsf T}\boldsymbol\theta$ through the vector model gives

$$\operatorname{Var}(\widehat q) \geq \frac{1}{\nu} \mathbf w^{\mathsf T} \mathbf F_{\theta}^{-1} \mathbf w,$$

for $\nu$ independent repetitions of the same phase vector. If the Fisher matrix is singular, the functional can still be locally estimable when $\mathbf w$ lies in its range: the corresponding regular bound uses $\mathbf w^{\mathsf T}\mathbf F_\theta^+\mathbf w$, with the Moore–Penrose inverse. A component of $\mathbf w$ along an unidentifiable direction prevents a locally unbiased finite-variance estimate. Replacing classical Fisher information by a quantum Fisher matrix gives a bound, whose attainability must still be checked for the allowed measurements.

This expression should not be confused with $\mathbf w^{\mathsf T}\mathbf F_{\theta}\mathbf w$. If the physics instead contains only one unknown $q$ through $\boldsymbol\theta=\mathbf vq$, then

$$F_q = \mathbf v^{\mathsf T} \mathbf F_{\theta} \mathbf v.$$

The first problem treats orthogonal field modes as unknown nuisance parameters. The second assumes they are absent or known. Switching between them can change a claimed quantum advantage.

For a meaningful comparison, fix the field functional, temporal windows, allowed local and joint operations, prior phase range, and resources per trial. In particular, distinguish a fixed probe capacity at each node from a total pool that can move between nodes. The distinction changes the separable bound even for ideal qubits, as the example below shows.

Covariance of an average or difference

Suppose calibrated local readouts obey

$$\mathbf y = \boldsymbol\theta + \boldsymbol\varepsilon, \qquad \mathbb E[\boldsymbol\varepsilon] = \mathbf0,$$

with noise covariance $\boldsymbol\Sigma$. The linear estimator $\widehat q=\mathbf w^{\mathsf T}\mathbf y$ has

$$\operatorname{Var}(\widehat q) = \mathbf w^{\mathsf T} \boldsymbol\Sigma \mathbf w.$$

Distributed quantum resources are useful when they reduce this quadratic form for the target mode. Negative cross-covariances can quiet an average; positive cross-covariances can quiet a difference. Reducing one spatial readout variance does not, by itself, require another spatial readout variance to increase: these local observables can commute. For optical squeezing the uncertainty constraint instead involves conjugate quadratures. For example, squeezing one collective $X$ mode can leave orthogonal $X$ modes at vacuum variance while increasing its conjugate $P$ variance.

The same formula also describes classical correlations. Synchronous interrogation with a common oscillator can reject oscillator noise in a difference, and subtracting a shared auxiliary sensor can create correlated residuals. Observing $\mathbf w^{\mathsf T}\boldsymbol\Sigma\mathbf w$ below the sum of individual variances is therefore not an entanglement witness. An inter-node quantum advantage requires a bound satisfied by every allowed node-separable protocol. Violating a valid metrological separability bound can itself certify the relevant entanglement under its stated measurement and calibration assumptions; separate state tomography is not always needed.

Nuisance-mode rejection

Suppose local phases follow a known linear field model

$$\boldsymbol\theta = \mathbf A\boldsymbol\beta,$$

where columns of $\mathbf A$ contain sampled spatial profiles and $\boldsymbol\beta$ contains their unknown amplitudes. To estimate coefficient $\beta_s$ without bias, weights must satisfy

$$\mathbf A^{\mathsf T}\mathbf w = \mathbf e_s.$$

Among linear unbiased estimators, the minimum-variance weights are

$$\mathbf w_\star = \boldsymbol\Sigma^{-1}\mathbf A \left( \mathbf A^{\mathsf T} \boldsymbol\Sigma^{-1} \mathbf A \right)^{-1} \mathbf e_s,$$

provided the indicated matrices have the required rank. Quantum correlations can reshape $\boldsymbol\Sigma$, but they cannot recover a field component that the sampled response matrix does not identify. Geometry and calibration come before squeezing.

Qubit benchmarks at fixed allocation

The cleanest hierarchy appears in a normalized qubit-phase model. Assume:

• $N$ two-level probes are distributed among $M$ nodes, with a fixed allocation $n_m$ used in every trial;
• each probe samples its local phase once through a generator with eigenvalue span one;
• all local phases not fixed by the estimand may be unknown;
• control, contrast, detection, and prior knowledge are ideal; and
• the same unknown phase vector is sampled independently $\nu$ times, without entanglement between trials, and the relevant local fringe branches are resolved.

Each protocol below may optimize its initial allocation subject to $\sum_m n_m=N$, but may not reassign that allocation between trials. The continuous optimizations presume nonzero weights; unused nodes need no probes. Integer allocations can raise the finite-$N$ minimum. These are local Cramér–Rao benchmarks, not globally unbiased single-shot error guarantees.

Independent product probes

Let node $m$ receive $n_m$ independent probes, with $\sum_m n_m=N$. Its local Fisher information is at most $n_m$ in this normalization. Independent local estimates give

$$\operatorname{Var}(\widehat q) \geq \frac{1}{\nu} \sum_{m=1}^{M} \frac{w_m^2}{n_m}.$$

Optimizing the allocation under the fixed total yields

$$n_m = N \frac{|w_m|}{\|\mathbf w\|_1}, \qquad \operatorname{Var}(\widehat q)_{\rm prod} \geq \frac{\|\mathbf w\|_1^2}{\nu N},$$

where

$$\|\mathbf w\|_1 := \sum_m|w_m|.$$

Equal allocation is optimal only for equal-magnitude weights. Sending the same number of probes to a nearly irrelevant node wastes the declared resource.

Node-separable states with local entanglement

A stronger comparator allows arbitrary entanglement among the $n_m$ probes inside each node but no entanglement between nodes. An ideal local GHZ-like state can provide Fisher information $n_m^2$ for $\theta_m$, giving

$$\operatorname{Var}(\widehat q) \geq \frac{1}{\nu} \sum_{m=1}^{M} \frac{w_m^2}{n_m^2}.$$

The optimized allocation is now

$$n_m = N \frac{|w_m|^{2/3}}{ \sum_j|w_j|^{2/3} },$$

and therefore

$$\operatorname{Var}(\widehat q)_{\rm node\text{-}sep} \geq \frac{ \left( \sum_m|w_m|^{2/3} \right)^3 }{ \nu N^2 }.$$

This is the optimized ideal benchmark for node-local entanglement with a fixed allocation in every simultaneous interrogation window. Beating the product-probe SQL while remaining above this bound demonstrates quantum enhancement, but not an advantage due specifically to entanglement between nodes.

Inter-node entanglement

A signed GHZ-type state can place the entire weighted functional into one collective phase. Allocate

$$n_m = N \frac{|w_m|}{\|\mathbf w\|_1}$$

and reverse the local coupling for negative $w_m$. The relative phase between the two GHZ branches is

$$\Phi = \sum_m \operatorname{sgn}(w_m)n_m\theta_m = \frac{N}{\|\mathbf w\|_1} q.$$

An ideal parity measurement then gives

$$\operatorname{Var}(\widehat q)_{\rm global} \geq \frac{\|\mathbf w\|_1^2}{\nu N^2}.$$

This construction is exact when the required $n_m$ are integers. Rounding them changes the sensed functional and generally introduces sensitivity to unknown orthogonal modes. Fractional interaction times or tunable local couplings can implement other real weights if those controls are allowed, but their maximum response must be recalculated. Averaging coefficients from different nonlinear parity fringes is not an automatic substitute: multiple settings require a joint likelihood and an identifiability check. The stated variance is an ideal lower bound whose finite-$N$ attainability depends on these conditions and on implementation of the coupling signs.

For an equal average, $w_m=1/M$, the hierarchy becomes

$$\begin{aligned} \operatorname{Var}(\widehat q)_{\rm prod} &\geq \frac{1}{\nu N}, \\ \operatorname{Var}(\widehat q)_{\rm node\text{-}sep} &\geq \frac{M}{\nu N^2}, \\ \operatorname{Var}(\widehat q)_{\rm global} &\geq \frac{1}{\nu N^2}. \end{aligned}$$

For equal allocation with $N/M$ an integer, the factor $M$ is the ideal variance advantage over separately entangled nodes constrained to keep that allocation. It is not a generic advantage for estimating all $M$ phases.

When moving the probes removes this advantage

Suppose instead that the phase vector is static and all $N$ probes can be assigned to one selected node in each trial. Prepare a local $N$-probe GHZ state at node $m$ for a fraction $f_m$ of the trials, keeping the node label. This protocol is separable between nodes. With resolved local fringes its total Fisher matrix is $\nu N^2\operatorname{diag}(f_m)$, so

$$\operatorname{Var}(\widehat q) \geq \frac{1}{\nu N^2}\sum_m\frac{w_m^2}{f_m}.$$

Choosing $f_m=|w_m|/\|\mathbf w\|_1$ reaches the same local bound as the global GHZ construction. No inter-node entanglement is required. Finite trial counts must approximate these fractions, and global phase ambiguities remain.

This time-sharing strategy is unavailable if each node has a fixed smaller capacity, redistribution is too slow, or the required phases change between windows and must be sampled simultaneously. Thus a total probe count alone does not establish an inter-node advantage. One must also specify the local capacities and the allowed allocation over time. At a fixed specified $\{n_m\}$, compare the corresponding bounds before optimizing over that set.

For imperfect devices, match interrogation time, state-preparation attempts, loss, readout, reference beams, and phase-branch resolution as well. Splitting one squeezed optical source among nodes can reduce the number of squeezers even without improving the optimum precision at fixed sample photons. That engineering benefit is distinct from a probe-count advantage.

Global Function Versus Full Field Map

A field map and a single weighted average have different optimal states. For independently encoded, commuting local parameters, additive local resource constraints, and a weighted sum of local estimation variances, broad results allow optimal sensor-separable states and local measurements (Proctor, Knott, and Dunningham). A cost coupling parameters, noncommuting local generators, or restrictions on local ancillas need a new analysis; these assumptions cannot be dropped merely because the sensors are spatially separated.

The situation changes when only one global function is required. A GHZ or collective squeezed mode can spend nearly all available sensitivity on $q=\mathbf w^{\mathsf T}\boldsymbol\theta$ while discarding information about orthogonal combinations. That concentration can improve precision under fixed local capacities. The time-sharing example above shows why it does not beat every conceivable strategy that first estimates local phases.

This yields a useful design rule:

Entangle across nodes only after deciding which spatial mode may be sacrificed and which one must be measured.

If the application later asks for a different weighting, the state, distribution network, local controls, or measurement basis may need to be reconfigured. One scalar network output cannot be retrospectively promoted to a complete field image.

Discrete-Variable Protocols

Signed GHZ sensing

For qubit or effective-spin nodes, the ideal global protocol prepares a coherence between two many-body branches. Local basis flips can reverse the sign with which selected probes accumulate phase. After encoding, the state has the form

$$|\psi(q)\rangle = \frac{1}{\sqrt2} \left( |\mathbf0\rangle + e^{-iK_qq} |\mathbf1\rangle \right), \qquad K_q = \frac{N}{\|\mathbf w\|_1}.$$

The strings $|\mathbf0\rangle$ and $|\mathbf1\rangle$ summarize the two signed branches; they need not be all-zero and all-one strings in the laboratory basis. An analysis rotation followed by local binary measurements produces a parity likelihood

$$p(\pi\mid q) = \frac12 \left[ 1 + \pi C \cos(K_qq+\phi_0) \right], \qquad \pi\in\{-1,+1\},$$

where $C$ is the final network contrast and $\phi_0$ is a controlled analysis phase. The classical Fisher information is

$$F_q = \frac{ C^2K_q^2 \sin^2(K_qq+\phi_0) }{ 1-C^2 \cos^2(K_qq+\phi_0) }.$$

At a quadrature working point,

$$F_q = C^2K_q^2.$$

Reduced projection noise alone is not enough: contrast determines the signal slope. A protocol with a narrow parity distribution and nearly vanishing contrast carries little phase information.

The parity can be assembled from local measurement records after classical communication. The sensor qubits need not be physically reunited after encoding. This does not enable faster-than-light signaling: no node can infer the global parity or the distant parameter from its local record alone.

Phase aliases and adaptive tiers

The same factor $K_q$ that increases local slope also shortens the fringe period:

$$q \sim q + \frac{2\pi}{K_q}.$$

An ideal $N$-probe global state therefore has local uncertainty of order $1/N$ but a fringe period of the same order. A single cosine readout also has a reflection ambiguity within a period. A broad prior requires a coarse product-state estimate, several interrogation strengths, adaptive analysis phases, or a Bayesian protocol that retains all branches. Quoting the local Fisher information while assuming the correct fringe index is known does not establish globally accurate sensing.

Split spin-squeezed ensembles

Atomic protocols can create correlations in one ensemble and then divide the atoms into spatial modes. Let $J_{z,m}$ be a local population-difference readout and let

$$\Gamma_{mn} := \frac12 \left\langle \Delta J_{z,m}\Delta J_{z,n} + \Delta J_{z,n}\Delta J_{z,m} \right\rangle.$$

For a locally linear, calibrated single-parameter response $\partial_q\langle\mathbf J_z\rangle=\mathbf r_q$, the variance of the normalized linear readout with coefficients $\mathbf a$ is

$$\Delta q^2 = \frac{ \mathbf a^{\mathsf T} \boldsymbol\Gamma \mathbf a }{ |\mathbf a^{\mathsf T}\mathbf r_q|^2 }.$$

If other field modes are unknown, the readout must also have zero response to those nuisance modes. Cross-covariances in $\boldsymbol\Gamma$ describe correlations, which may be classical; they are not alone an entanglement witness. Useful noise reduction must align with the desired response while retaining its slope. Particle entanglement within the original ensemble and mode entanglement among the separated outputs are related but distinct statements.

A network claim must compare the split state with independently optimized local squeezed ensembles when those are physically allowed. Comparing only with coherent-spin states establishes sub-projection-noise operation, not necessarily an inter-node advantage.

Continuous-Variable Protocols

Squeeze the target supermode

Optical and microwave networks often encode weak signals as quadrature displacements or small phase rotations. Choose normalized real mode coefficients $u_m$ with $\sum_m u_m^2=1$ and define

$$X_{\mathbf u} := \sum_{m=1}^{M}u_mX_m.$$

A passive splitter network can distribute one squeezed input so that this collective output quadrature inherits the squeezing:

$$\operatorname{Var}(X_{\mathbf u}) = \frac12e^{-2r}$$

in the convention where vacuum variance is $1/2$. Orthogonal combinations of the node quadratures contain vacuum noise. Thus each local marginal can look only weakly squeezed while the correctly weighted sum recovers the quiet input mode.

Suppose local transducers produce

$$\mathbb E[X_m] = g_m\theta_m.$$

The homodyne weights and splitter amplitudes must be chosen so that the collective response is proportional to $q=\sum_mw_m\theta_m$. Unequal transducer gains, link attenuation, or local oscillator phases rotate the realized mode away from $\mathbf w$. Calibration must therefore determine the response vector, not merely the optical splitting ratios.

The correct separable optical comparator

A coherent-state network supplies the familiar shot-noise baseline, but it is not always the strongest separable comparator. If local squeezers are allowed, each node can receive its own optimized squeezed state. The comparison must then fix:

• photons incident on each sample or transducer;
• coherent displacement and squeezed-vacuum photons;
• source-to-sample transmission;
• local-oscillator and reference power when operationally relevant;
• detector efficiency and bandwidth; and
• the number and quality of nonclassical sources.

Within specified displacement or small-phase Gaussian protocols, a shared squeezed mode can outperform independently squeezed inputs at fixed total mean photons; this does not establish an optimum over unrestricted optical states. For phase encoding, mean photon number alone does not even bound the QFI of arbitrary states: rare very-high-number components can give unbounded number variance. State restrictions or additional resource constraints matter (Zhang and Zhuang). Source count, sample photons, and wall time answer different practical questions.

Structured Noise and Protected Modes

Spatially correlated noise need not have the same profile as the signal. This creates an opportunity beyond simply reducing white projection noise. Let the Hamiltonian during interrogation be

$$\frac{H}{\hbar} = \beta_s \sum_m s_mG_m + \sum_{\mu=1}^{K} \beta_\mu \sum_m n_{\mu m}G_m,$$

where $G_m$ are dimensionless commuting generators, the profiles are dimensionless, and $\beta_s,\beta_\mu$ have units of angular frequency. For time-independent amplitudes, a coherence between two joint generator eigenstates whose eigenvalue vectors differ by $\mathbf k$ accumulates relative phase

$$\Phi = t \left[ \beta_s\mathbf k^{\mathsf T}\mathbf s + \sum_{\mu=1}^{K} \beta_\mu \mathbf k^{\mathsf T}\mathbf n_\mu \right].$$

If

$$\mathbf k^{\mathsf T}\mathbf n_\mu = 0 \quad \text{for every }\mu, \qquad \mathbf k^{\mathsf T}\mathbf s \ne 0,$$

then those nuisance generators give the same eigenvalue to both branches. Their unknown amplitudes, including time-dependent fluctuations with those same profiles, cannot randomize the relative phase; the signal still can. This protection applies to the specified commuting noise, not arbitrary local relaxation or additional noncommuting noise channels. Multilevel sensors can supply a richer set of branch differences $\mathbf k$ than two-level sensors.

This protection is model-based. It fails when the nuisance profile is misspecified, node positions drift, coupling gains change, or unmodeled noise has a component along $\mathbf k$. A serious experiment measures the residual transfer from every rejected mode and propagates profile uncertainty into the final estimate.

Loss, Decoherence, and State Age

Discrete-variable survival

A global GHZ coherence is fragile because every participating subsystem is part of the signal-bearing off-diagonal term. If losses are independent, probe $j$ survives with probability $\eta_j$, and any loss causes rejection, then

$$p_{\rm surv} = \prod_{j=1}^{N}\eta_j.$$

For independent local dephasing channels with coherence-amplitude magnitudes $c_j$, and after compensating known phases, the network contrast is

$$C_{\rm net} = C_0 \prod_{j=1}^{N}c_j.$$

Correlated phase noise need not obey this product law: a signed coherence can reject a common noise profile, as above. The conditional Fisher information may still scale as $C_{\rm net}^2N^2$, but the unconditional information per attempt is reduced by survival and preparation probability. Independent probes degrade gradually because lost probes can simply cease contributing; a monolithic GHZ state can lose its entire collective advantage after one erasure.

Heralded generation changes what is known, not what it costs. A failed link attempt may be excluded from the estimator only if its elapsed time, source uses, memory aging, and any consumed probes are included in the comparison.

Continuous-variable loss

For quadratures with vacuum covariance $I/2$, independent pure-loss channels transform a covariance matrix as

$$\boldsymbol V_{\rm out} = \boldsymbol L \boldsymbol V_{\rm in} \boldsymbol L + \frac12 \left( I-\boldsymbol L^2 \right),$$

where

$$\boldsymbol L = \operatorname{diag} \left( \sqrt{\eta_1},\ldots,\sqrt{\eta_M} \right)$$

for one quadrature per node. Means already encoded before this loss transform as $\boldsymbol\mu_{\rm out}=\boldsymbol L\boldsymbol\mu_{\rm in}$; a displacement added after the loss has a different response. Vacuum noise and unequal attenuation change the optimum splitter and electronic weights. For example, a single squeezed quadrature approaches variance $(1-\eta)/2$ as $r\to\infty$ at fixed $\eta<1$. Precision scaling must use this variance together with the actual encoding response; displacement and phase sensing do not have one interchangeable photon-scaling law.

Generation latency and memory decay

Let $A_m$ be the random time at which node $m$ obtains its required link. The network state becomes usable only when all required resources are available, so its age distribution depends on a maximum or a scheduling policy, not only on the mean single-link latency. Early memories wait while slow links retry. If coherence decays as $e^{-t/T_{2,m}}$, then the final contrast depends on the joint waiting-time distribution:

$$\mathbb E[C_{\rm net}] \sim C_0 \mathbb E\!\left[ \prod_m e^{-A_m^{\rm wait}/T_{2,m}} \right].$$

State-age tags also affect the information retained. If an observed waiting time $W$ sets a known contrast $C(W)$, the tagged parity record at quadrature has mean Fisher information $K_q^2\mathbb E[C(W)^2]$. Discarding the tag gives a mixture with contrast $\mathbb E[C(W)]$ and information $K_q^2\mathbb E[C(W)]^2$, which is no larger. This assumes the waiting-time distribution is independent of $q$ and the analysis phase remains calibrated.

Cutoffs can improve contrast by discarding old states but reduce delivery rate. The optimum policy depends on the sensing objective and field correlation time; it is not determined by entanglement fidelity alone.

Timing and Reference Frames

Distributed nodes must interrogate the same intended spacetime mode. In the frequency domain, one useful linear response is

$$q(\omega) = \sum_{m=1}^{M} w_m\kappa_m H_m(\omega) e^{-i\omega\tau_m} s(\mathbf x_m,\omega),$$

where $H_m(\omega)$ is the node transfer function and $\tau_m$ is its timing offset. A timing error rotates the weight by $e^{-i\omega\delta\tau_m}$ and can mix a nominally rejected field mode into the target. Synchronization must be specified relative to the signal bandwidth, not only as a timestamp resolution.

Phase sensing also needs a frame in which local analysis angles are defined. Unknown local-oscillator phases can be exactly degenerate with the physical phases:

$$\theta_m^{\rm obs} = \theta_m^{\rm sig} - \phi_m^{\rm ref}.$$

A common laser, stabilized fibre link, two-way transfer protocol, calibrated comb, pilot tone, or reference pulse may supply the frame. Its noise and power are resources. Entanglement does not create a shared phase convention by itself.

Unconditional Information Rate

Let $A$ denote an accepted network trial. The Fisher information in the full record decomposes as

$$F_q^{\rm trial} = F_q^{\rm flag} + p_A F_{q\mid A} + (1-p_A)F_{q\mid \bar A},$$

where

$$F_q^{\rm flag} = \frac{(\partial_qp_A)^2}{p_A(1-p_A)}.$$

When acceptance is independent of $q$ and rejected trials contain no useful information,

$$F_q^{\rm trial} = p_AF_{q\mid A}.$$

For independent trials of fixed duration, the information rate is

$$\mathcal I_q := \frac{F_q^{\rm trial}}{T_c},$$

with cycle time $T_c$ including state generation, herald communication, memory loading, synchronization, interrogation, readout, and reset. A distributed quantum advantage in averaging time requires

$$G_{\rm rate} := \frac{\mathcal I_q^{\rm ent}} {\mathcal I_q^{\rm sep}} > 1$$

under the same physical resource constraints. Conditional sensitivity can improve while information rate worsens if successful multipartite states are rare or slow.

For independent renewal cycles of random duration $T$, use $\mathcal I_q=\mathbb E[F_{q\mid\text{cycle tags}}]/\mathbb E[T]$ when the tag and duration distributions are parameter-independent. It is the ratio of accumulated information to accumulated time, not $\mathbb E[F_{q\mid\text{tags}}/T]$. Parameter-dependent timing carries its own information and belongs in the joint likelihood.

Worked comparison of precision and delivery rate

Consider $M=4$ equal-weight nodes and $N=400$ ideal qubit probes per accepted trial, with $100$ probes available at each node in every window. The local variance benchmarks per repetition are

$$\begin{array}{c|c|c} \text{protocol class} & \operatorname{Var}(\widehat q) & \Delta q \\ \hline \text{product probes} & 1/400 & 0.050 \\ \text{node-separable, locally entangled} & 4/400^2 & 0.0050 \\ \text{inter-node GHZ} & 1/400^2 & 0.0025 \end{array}$$

Now suppose the inter-node state is usable on only $12\%$ of attempts, has $C=0.80$, and takes twice the product-probe cycle time. Its information rate, in units of the product cycle time $T_0$, is

$$\mathcal I_q^{\rm ent} = \frac{ 0.12(0.80)^2(400)^2 }{2T_0} = \frac{6144}{T_0}.$$

The ideal product network has

$$\mathcal I_q^{\rm prod} = \frac{400}{T_0},$$

so the entangled protocol still has a factor $15.36$ information-rate gain over that baseline. But the ideal node-separable network has

$$\mathcal I_q^{\rm node\text{-}sep} = \frac{400^2/4}{T_0} = \frac{40000}{T_0}.$$

The same experiment would therefore beat independent particles while failing to beat an ideal protocol with local entanglement at every node. Both statements can be true. If all $400$ probes could instead be concentrated at different nodes on successive trials of a static field, the ideal time-shared local-GHZ rate would be $160000/T_0$. That stronger comparator is excluded by the fixed per-node capacity in this example, not by separability.

The global fringe also repeats after

$$\Delta q_{\rm wrap} = \frac{2\pi}{400} \simeq 1.57\times10^{-2} \mathrm{rad}.$$

A broader prior needs another sensing tier; the local variance alone does not resolve the branch.

Application Patterns

Remote clock comparison

Two remote clocks can estimate a frequency difference while rejecting a shared interrogation oscillator. If the local phases after Ramsey time $T$ are

$$\theta_m = (\omega_m-\omega_L)T,$$

then

$$\theta_1-\theta_2 = (\omega_1-\omega_2)T$$

is independent of the common oscillator frequency $\omega_L$ in the ideal synchronous model. Correlation spectroscopy can exploit this cancellation without inter-node entanglement. A remote Bell state can further improve the quantum statistical term of the differential parity readout.

Remote entanglement may improve a frequency-ratio measurement without improving either clock’s systematic shifts, uptime, time scale, or absolute realization of frequency. Optical Clocks explains the instrument and systematic shifts; Atomic Clocks for Quantum Estimation develops oscillator tracking and information per cycle. Common-oscillator cancellation requires matched sensitivity functions and the appropriate frequency-conversion factors, not merely simultaneous timestamps.

Gradients, curvature, and field modes

An array can choose weights orthogonal to low-order nuisance profiles. For nodes at positions $x_m$, the constraints

$$\sum_mw_m = 0, \qquad \sum_mw_mx_m = 0$$

reject a uniform offset and linear gradient. Curvature remains observable only if $\sum_mw_mx_m^2\ne0$; otherwise that mode is rejected too. Entangled multilevel sensors can encode the desired signed weights into one protected coherence. This is useful when the nuisance amplitudes are large but their spatial forms are accurately known.

The same logic applies to coherent-field searches, magnetic arrays, inertial gradiometers, and spectroscopy networks. It does not guarantee an application advantage: the signal must remain coherent across the baseline and interrogation window, while sensor positions, coupling tensors, and timing must be known well enough that nuisance leakage stays below the claimed gain.

Photonic and radio-frequency arrays

Continuous-variable optical networks can distribute one squeezed supermode to electro-optic, optomechanical, or interferometric transducers. Local homodyne currents are combined with weights matched to an average, edge difference, or beamforming mode. The architecture can be deterministic and high bandwidth, but its performance depends on transmission, phase-locked local oscillators, transducer gain, electronic covariance, and photons at the sample.

Discrete-variable photonic protocols can instead send path- or polarization-entangled photons through separated phase elements and recover a global phase from coincidence or parity statistics. Detection efficiency is decisive. A sub-shot-noise result conditioned on detecting every photon may disappear when launched photons and failed trials are included.

Privacy as a separate objective

Some global states store the sensed phase only in joint correlations. A single-node reduced state may then reveal no information about $q$, while authorized parties can reconstruct it from combined outcomes. This can support privacy against specified subsets of nodes.

Privacy is not automatic secrecy. A complete protocol must state which nodes, source, measurement station, and classical channels are trusted; what side information an adversary receives; and whether malicious devices can alter the encoding. Metrological privacy, cryptographic security, and device-independent certification are different claims.

For examples of these distinct mechanisms, Guo and colleagues study continuous-variable optical sensing, Nichol and colleagues use a remote clock Bell pair, and Bate and colleagues isolate a spatial signal from different noise profiles. These experiments use different resource constraints and cannot be ranked by a single quoted decibel gain. Distributed Quantum Sensing research discusses experimental milestones and open problems. Quantum Network Architectures develops the link, memory, and synchronization mechanisms; Network Verification explains how to test the delivered states under explicit trust assumptions.

Exercises

1. Optimize independent-probe allocation

For product probes, minimize

$$V = \sum_m\frac{w_m^2}{n_m}$$

subject to $n_m>0$ and $\sum_mn_m=N$, for nonzero weights. Treat $n_m$ as continuous and derive the optimum allocation and local variance bound.

Solution

Introduce a multiplier $\lambda$:

$$\mathcal L = \sum_m\frac{w_m^2}{n_m} + \lambda \left(\sum_mn_m-N\right).$$

Stationarity gives

$$-\frac{w_m^2}{n_m^2} + \lambda = 0,$$

so $n_m=|w_m|/\sqrt\lambda$. Normalization yields

$$n_m = N\frac{|w_m|}{\sum_j|w_j|}.$$

Substitution gives

$$V_{\min} = \frac{(\sum_m|w_m|)^2}{N} = \frac{\|\mathbf w\|_1^2}{N}.$$

For $\nu$ repetitions, divide this variance by $\nu$.

2. Optimize node-local Heisenberg resources

Allow ideal entanglement within each node, so the variance is

$$V = \sum_m\frac{w_m^2}{n_m^2}.$$

Find the optimal continuous $n_m$ at fixed $\sum_mn_m=N$ and fixed allocation between trials. Assume nonzero weights.

Solution

Stationarity of

$$\mathcal L = \sum_m\frac{w_m^2}{n_m^2} + \lambda \left(\sum_mn_m-N\right)$$

requires

$$-\frac{2w_m^2}{n_m^3} + \lambda = 0.$$

Hence $n_m\propto|w_m|^{2/3}$ and

$$n_m = N \frac{|w_m|^{2/3}}{\sum_j|w_j|^{2/3}}.$$

The minimum is

$$V_{\min} = \frac{(\sum_m|w_m|^{2/3})^3}{N^2}.$$

At integer allocations with more than one probe per relevant node, this locally entangled comparator is stronger than the product-probe SQL. The continuous optimum must be adjusted when counts would be fractional.

3. Reallocate a separable probe between trials

Three static phases have target $q=(\theta_1+\theta_2+\theta_3)/3$. There are $N=120$ probes per trial, and $\nu$ is a multiple of three. Compare using $40$-probe local GHZ states at every node in every trial with using all $120$ probes at one node for $\nu/3$ trials per node. Assume local fringe ambiguities are resolved and moving probes costs no time.

Solution

The fixed-allocation Fisher matrix is $\mathbf F_{\rm total}=\nu(40)^2I_3$. Thus $\operatorname{Var}(\widehat q)\geq3/(\nu120^2)$. For the time-shared strategy it is $\mathbf F_{\rm total}=(\nu/3)(120)^2I_3$, giving

$$\operatorname{Var}(\widehat q) \geq\frac{1}{\nu120^2}.$$

This matches the global-GHZ bound without entanglement between nodes. The second strategy requires a node to accommodate $120$ probes and the phases to remain unchanged while other nodes are sampled. A simultaneous snapshot with capacity $40$ per node does not permit this comparison.

4. Average and difference covariance modes

Two calibrated readouts have covariance

$$\boldsymbol\Sigma = \sigma^2 \begin{pmatrix} 1 & \rho\\ \rho & 1 \end{pmatrix}.$$

Find the variances of $q_+=(\theta_1+\theta_2)/2$ and $q_-=(\theta_1-\theta_2)/2$. Which sign of correlation helps each task?

Solution

Using $V=\mathbf w^{\mathsf T}\boldsymbol\Sigma\mathbf w$ gives

$$\operatorname{Var}(\widehat q_+) = \frac{\sigma^2}{2}(1+\rho),$$

and

$$\operatorname{Var}(\widehat q_-) = \frac{\sigma^2}{2}(1-\rho).$$

Negative covariance helps the average, whereas positive covariance helps the difference. “More correlation” is not a task-independent resource; its sign and orientation matter.

5. Reject offset and gradient

Three nodes lie at $x=-1,0,1$. Their phases follow

$$\theta(x) = a+bx+cx^2.$$

Find weights $\mathbf w$ such that $\widehat c=\sum_mw_m\theta(x_m)$ rejects $a$ and $b$ and has unit response to $c$. If independent readout noise at each node has variance $\sigma^2$, what is $\operatorname{Var}(\widehat c)$?

Solution

The constraints are

$$w_1+w_2+w_3 = 0,$$ $$-w_1+w_3 = 0, \qquad w_1+w_3 = 1.$$

Thus

$$\mathbf w = \left( \frac12,-1,\frac12 \right).$$

For independent equal noise,

$$\operatorname{Var}(\widehat c) = \sigma^2 \left( \frac14+1+\frac14 \right) = \frac32\sigma^2.$$

Quantum covariance engineering can reduce this target-mode variance, but the unbiasedness constraints still come from geometry and calibration.

6. Compare GHZ and product probes under the same loss

For a common phase, an $N$-probe GHZ state has conditional Fisher information $N^2$. Each probe survives independently with probability $\eta$, known loss flags carry no phase information, and any GHZ loss destroys its coherence. Product probes retain the information in every surviving particle. Derive the efficiency threshold for a GHZ advantage at equal cycle time, and evaluate it for $N=20$.

Solution

The all-survive probability is $\eta^N$, so $F_{\rm GHZ}=\eta^N N^2$. A product trial has mean $\eta N$ survivors, each contributing unit information: $F_{\rm prod}=\eta N$. For $\eta>0$ and $N>1$, the GHZ protocol wins when

$$\eta^{N-1}>\frac1N, \qquad \eta>N^{-1/(N-1)}.$$

For $N=20$ the threshold is about $0.854$. Comparing instead with a lossless product trial gives $20^{-1/20}\simeq0.861$, a different benchmark. Additional GHZ contrast or preparation-time penalties raise the matched-loss threshold; noise common to both protocols must be applied to both likelihoods.

7. Conditional gain but rate loss

An entangled protocol succeeds with probability $p_A=0.10$, has conditional Fisher information $F_{q\mid A}=1000$, and takes cycle time $5T_0$. Rejected trials carry no information. A separable protocol has $F_q=40$ every cycle and cycle time $T_0$. Which protocol has the larger information rate?

Solution

The entangled information per attempted trial is

$$F_q^{\rm trial} = p_AF_{q\mid A} = 100.$$

Its information rate is

$$\mathcal I_q^{\rm ent} = \frac{100}{5T_0} = \frac{20}{T_0}.$$

The separable rate is

$$\mathcal I_q^{\rm sep} = \frac{40}{T_0}.$$

The entangled protocol has much larger information in a successful record but only half the wall-time information rate.

8. Timing leakage into a differential channel

Two nodes estimate a difference with weights $(1/2,-1/2)$. A common nuisance is $s(t)=A\cos(\omega t)$, but the interrogation centers differ by $\delta\tau=\tau_1-\tau_2$. Show that the residual differential signal is at most approximately $A\omega|\delta\tau|/2$ for small timing mismatch.

Solution

The nuisance contribution is

$$q_{\rm leak} = \frac A2 \left[ \cos(\omega\tau_1) - \cos(\omega\tau_2) \right].$$

Writing $\bar\tau=(\tau_1+\tau_2)/2$ and using the cosine-difference identity gives

$$q_{\rm leak} = -A \sin(\omega\bar\tau) \sin\!\left( \frac{\omega\delta\tau}{2} \right).$$

For $|\omega\delta\tau|\ll1$,

$$|q_{\rm leak}| \lesssim \frac{A\omega|\delta\tau|}{2}.$$

The required synchronization accuracy therefore depends on nuisance amplitude and frequency as well as on the target uncertainty.

References

• J. Bate et al., “Experimental distributed quantum sensing in a noisy environment,” Physical Review Letters 135, 220801 (2025).

• J. Bringewatt, A. Ehrenberg, T. Goel, and A. V. Gorshkov, “Optimal function estimation with photonic quantum sensor networks,” Physical Review Research 6, 013246 (2024).

• Z. Eldredge, M. Foss-Feig, J. A. Gross, S. L. Rolston, and A. V. Gorshkov, “Optimal and secure measurement protocols for quantum sensor networks,” Physical Review A 97, 042337 (2018). See also the 2022 erratum, Physical Review A 105, 019903 (2022).

• W. Ge, K. Jacobs, Z. Eldredge, A. V. Gorshkov, and M. Foss-Feig, “Distributed quantum metrology with linear networks and separable inputs,” Physical Review Letters 121, 043604 (2018).

• X. Guo et al., “Distributed quantum sensing in a continuous-variable entangled network,” Nature Physics 16, 281–284 (2020).

• P. Kómár et al., “A quantum network of clocks,” Nature Physics 10, 582–587 (2014).

• B. C. Nichol et al., “An elementary quantum network of entangled optical atomic clocks,” Nature 609, 689–694 (2022).

• E. S. Polzik and J. Ye, “Entanglement and spin squeezing in a network of distant optical lattice clocks,” Physical Review A 93, 021404(R) (2016).

• T. J. Proctor, P. A. Knott, and J. A. Dunningham, “Multiparameter estimation in networked quantum sensors,” Physical Review Letters 120, 080501 (2018).

• T. Qian, J. Bringewatt, I. Boettcher, P. Bienias, and A. V. Gorshkov, “Optimal measurement of field properties with quantum sensor networks,” Physical Review A 103, L030601 (2021).

• Z. Zhang and Q. Zhuang, “Distributed quantum sensing,” Quantum Science and Technology 6, 043001 (2021).

• Q. Zhuang, Z. Zhang, and J. H. Shapiro, “Distributed quantum sensing using continuous-variable multipartite entanglement,” Physical Review A 97, 032329 (2018).