Entanglement Witnesses and Local Tomography Limits

Local field measurements can certify entanglement without reconstructing an entire continuum state. The defensible claim is fixed by the calibrated observables and the model class: a statistically significant witness violation is a one-sided proof of entanglement, whereas a finite covariance matrix determines a state only after a Gaussian assumption. This page makes both statements quantitative for two regulated wavepacket modes and supplies an explicit non-Gaussian look-alike that defeats covariance tomography.

Required background. Local measurement instruments supplies the effects, updates, and calibration data behind measured correlators.

Helpful background. Choosing an entanglement measure separates the operational question from a convenient numerical diagnostic.

Chapter map. Use the route from a local coupling to a field instrument, the chapter claim-validity table, and the three independent validity questions to place this construction within the full measurement protocol.

A witness licenses one conclusion

For a declared bipartition, an entanglement witness is a Hermitian operator $W$ obeying

$$\operatorname{tr}(W\sigma)\ge0$$

for every separable state $\sigma$, while at least one entangled state has $\operatorname{tr}(W\rho)\lt0$. Therefore a confidence interval for $w=\operatorname{tr}(W\rho)$ that lies entirely below zero certifies entanglement under the calibration, sampling, and support assumptions. The reverse implication is false: $w\ge0$ means only that this witness did not detect the state.

For the two regulated modes used below, “separable” means that the state can be written as a convex mixture

$$\sigma=\sum_\lambda p_\lambda\, \sigma_A^{(\lambda)}\otimes\sigma_B^{(\lambda)}, \qquad p_\lambda\ge0, \qquad \sum_\lambda p_\lambda=1.$$

Classical correlations are allowed in such a mixture; what is excluded is quantum entanglement across the declared $A|B$ split. This definition belongs to the regulated type-I mode model. For continuum local algebras, the operational split must instead be stated at the algebraic level or through an explicitly controlled split approximation.

The EPR test below is a variance inequality rather than the expectation value of one fixed, state-independent Hermitian operator: empirical means enter when the records are centered. It nevertheless has the same one-sided certifying role. Local displacements used to center the modes do not create entanglement, and the separable bound applies to arbitrary first moments provided the variances are computed about those measured means.

There are three distinct claim levels:

1. Witness violation: the measured state is not separable across the stated modes or algebras.
2. Model-based quantification: a resource value is inferred within a class such as two-mode Gaussian states.
3. Tomography: the measured effects are informationally complete for the admitted state space.

Each step adds assumptions. In QFT, one must also specify the local algebras or normalized wavepacket functions, control the domains of unbounded quadratures, and show that the measurement instruments have the advertised supports.

Two-wavepacket covariance benchmark

Take two regulated wavepacket modes with

$$R=(q_A,p_A,q_B,p_B)^T, \qquad [R_j,R_k]=i\Omega_{jk},$$

and covariance

$$V_{jk}=\frac12\langle R_jR_k+R_kR_j\rangle -\langle R_j\rangle\langle R_k\rangle.$$

The quadratures are dimensionless. In this convention the symplectic form is $\Omega=J\oplus J$, with $J_{12}=1$, and a vacuum mode has $V_{qq}=V_{pp}=1/2$. These choices fix both the separable bound $2$ and the partial-transpose threshold $1/2$; importing formulas written for $[q,p]=2i$ would introduce a factor-of-two error.

Prepare the exact two-mode squeezed vacuum with $r=1/2$:

$$V(r)=\frac12 \begin{pmatrix} \cosh 2r&0&\sinh 2r&0\\ 0&\cosh 2r&0&-\sinh 2r\\ \sinh 2r&0&\cosh 2r&0\\ 0&-\sinh 2r&0&\cosh 2r \end{pmatrix}.$$

Define the EPR witness

$$\mathcal D =\operatorname{Var}(q_A-q_B) +\operatorname{Var}(p_A+p_B).$$

For $[q_j,p_j]=i$, every separable state obeys $\mathcal D\ge2$. With the signs used here, this is the $a=-1$ case of Duan et al. 2000, Theorem 1, Eqs. (2)–(3); a local phase reversal on mode $B$ converts it to the equivalent $a=1$ convention. The exact squeezed state gives

$$\mathcal D=2e^{-2r}=2e^{-1}=0.7357588823\lt2.$$

The same conclusion follows from partial transposition. Reversing $p_B$ gives the smallest symplectic eigenvalue

$$\widetilde\nu_-=\frac12e^{-2r} =\frac{e^{-1}}2 =0.1839397206\lt\frac12.$$

The phase-space reflection and covariance uncertainty test are derived in Simon 2000, Eqs. (5)–(11), with Gaussian necessity and sufficiency after Eq. (19).

Now model each quadrature readout as adding independent calibrated noise of variance $\delta=0.04$. Then $V_{\mathrm{obs}}=V+\delta\mathbf1_4$ and

$$\mathcal D_{\mathrm{obs}} =\mathcal D+4\delta =0.8957588823,$$ $$\widetilde\nu_{-,\mathrm{obs}} =\widetilde\nu_-+\delta =0.2239397206.$$

The complete numerical matrix is useful as a reproducible benchmark:

$$V_{\mathrm{obs}}= \begin{pmatrix} 0.8115403174&0&0.5876005968&0\\ 0&0.8115403174&0&-0.5876005968\\ 0.5876005968&0&0.8115403174&0\\ 0&-0.5876005968&0&0.8115403174 \end{pmatrix}.$$

Here $0.8115403174=\cosh(1)/2+0.04$ and $0.5876005968=\sinh(1)/2$. Substituting these rounded entries gives each measured EPR variance as $0.4478794412$ and reproduces the table below to the stated precision.

Both raw noisy statistics still violate the separable thresholds. If calibrated noise is subtracted, its calibration uncertainty must be propagated rather than treated as an exact correction.

Check	Exact regulated state	With $\delta=0.04$ readout noise	Separable threshold
EPR variance $\mathcal D$	$0.7357588823$	$0.8957588823$	$2$
Partial-transpose eigenvalue $\widetilde\nu_-$	$0.1839397206$	$0.2239397206$	$1/2$

The logarithmic-negativity convention on this page is $E_N(\rho)=\log_2\lVert\rho^{T_B}\rVert_1$, as defined by Vidal and Werner 2002, Eqs. (1)–(2). For a two-mode Gaussian state it becomes $\max\{0,-\log_2(2\widetilde\nu_-)\}$. Thus the exact benchmark has

$$E_N=-\log_2(2\widetilde\nu_-) =\frac{1}{\ln2} =1.4426950409\ \text{bits}.$$

That quantitative value is licensed by the Gaussian model and the exact benchmark state. The witness violation itself needs no Gaussian assumption.

A reproducible witness and covariance protocol

The shortest experiment that evaluates $\mathcal D$ uses two settings on independently prepared copies. In the first setting, jointly read the commuting observables $q_A$ and $q_B$ and retain the shot-by-shot differences $z_q^{(k)}=q_A^{(k)}-q_B^{(k)}$. In the second, jointly read $p_A$ and $p_B$ and retain $z_p^{(k)}=p_A^{(k)}+p_B^{(k)}$. With the unbiased sample variance

$$s_x^2=\frac{1}{N_x-1}\sum_{k=1}^{N_x} \left(z_x^{(k)}-\overline z_x\right)^2, \qquad \widehat{\mathcal D}=s_q^2+s_p^2,$$

the decision rule is an upper-confidence test: certify entanglement only when a simultaneous upper bound $U_{\mathcal D}$ is smaller than $2$. Reporting only the point estimate is insufficient.

Here is a fully numerical finite-sample benchmark. Take $N_q=N_p=100$ independent Gaussian records and sample summaries $s_q^2=s_p^2=0.4478794412$, matching the noisy covariance above. For a family error probability $\alpha=0.05$, give each variance a one-sided error probability $\alpha/2$. The exact chi-squared bounds are

$$U_x=\frac{(N_x-1)s_x^2}{\chi^2_{\alpha/2,N_x-1}}, \qquad \chi^2_{0.025,99}=73.36108019,$$

so the union bound gives simultaneous coverage of at least $95\%$ and

$$U_{\mathcal D}=U_q+U_p =2(0.6044085578) =1.2088171156\lt2.$$

This synthetic summary therefore passes the predeclared witness rule without subtracting the readout noise. The calculation assumes independent, identically distributed Gaussian preparations within each setting and no adaptive discarding of shots. If those assumptions are not justified, the experiment needs a confidence construction valid for its actual dependence and tail behavior; the threshold itself must not be relaxed.

The same two settings are not full covariance tomography. A real symmetric $4\times4$ covariance has ten independent entries. On repeatable preparations, one concrete phase-setting schedule is

$$(q_A,q_B),\quad(q_A,p_B),\quad(p_A,q_B),\quad(p_A,p_B), \quad(x_{A,\pi/4},x_{B,\pi/4}),$$

where $x_{j,\pi/4}=(q_j+p_j)/\sqrt2$. The first four settings determine the four single-mode variances and the four cross-mode covariances. The last determines the two local symmetrized covariances through

$$V_{q_jp_j}=\operatorname{Var}(x_{j,\pi/4}) -\frac12\left(V_{q_jq_j}+V_{p_jp_j}\right).$$

Means are estimated from the same setting records. The reconstructed matrix must first pass the physicality check $V+i\Omega/2\ge0$ within uncertainty. Partial transposition and its symplectic spectrum may then be evaluated over the entire confidence region, not merely at the best-fit matrix. These ten second moments and four means determine a two-mode Gaussian state. They remain only a projection of an unrestricted non-Gaussian state.

Failure controls and countermodels

Three controls should accompany the positive benchmark.

Vacuum normalization control. Set $r=0$ before adding readout noise. The exact output must be $\mathcal D=2$ and $\widetilde\nu_-=1/2$; with $\delta=0.04$, it must be $\mathcal D_{\mathrm{obs}}=2.16$ and $\widetilde\nu_{-,\mathrm{obs}}=0.54$. A vacuum result below either separable threshold falsifies the claimed normalization, uncertainty propagation, or calibration pipeline.

Phase-sign control. Apply a local phase reversal to mode $B$ but deliberately keep the old combinations $q_A-q_B$ and $p_A+p_B$. At $r=1/2$ the statistic becomes $2e=5.4365636569$, although the state is still entangled. Replacing the combinations by $q_A+q_B$ and $p_A-p_B$ restores the violation. This check tests phase labels and demonstrates concretely why a non-violation cannot establish separability.

Noise-model control. Record detector output with the signal blocked and retain the full noise covariance. For zero-mean additive noise independent of the field,

$$\mathcal D_{\mathrm{obs}}=\mathcal D +\operatorname{Var}(\eta_{q_A}-\eta_{q_B}) +\operatorname{Var}(\eta_{p_A}+\eta_{p_B}).$$

The shortcut $\mathcal D+4\delta$ follows only for four independent noises of equal variance. Repeating the analysis with allowed calibration covariances and with every retained shot is an adversarial stress test against an underestimated error budget or selection-dependent result. The covariance look-alike in the next section is a separate countermodel to the stronger tomography claim.

A non-Gaussian covariance look-alike

Second moments do not identify a general state. Consider

$$\rho_{\mathrm{sep}}=\tau_1\otimes\tau_1, \qquad \tau_1=\sum_{n=0}^{\infty}\frac{1}{2^{n+1}}|n\rangle\langle n|,$$

where each thermal state has mean occupation one, and compare it with

$$\rho_{\mathrm{ent}}=|\psi_2\rangle\langle\psi_2|, \qquad |\psi_2\rangle=\frac{|00\rangle+|22\rangle}{\sqrt2}.$$

Both have zero first moments. In both states,

$$\langle q_A^2\rangle =\langle p_A^2\rangle =\langle q_B^2\rangle =\langle p_B^2\rangle =\frac32,$$

and all cross second moments vanish, so

$$V_{\mathrm{sep}}=V_{\mathrm{ent}} =\frac32\mathbf1_4.$$

Yet $\rho_{\mathrm{sep}}$ is a product state, whereas $|\psi_2\rangle$ has Schmidt coefficients $(1/\sqrt2,1/\sqrt2)$, one bit of entanglement entropy, and negativity $1/2$. Covariance tomography would reconstruct the separable Gaussian product thermal state from either data set and would miss the entanglement of $|\psi_2\rangle$.

The twins can be separated by a higher-order observable. The geometric thermal distribution has $\langle n_A^2\rangle=3$, whereas the reduced state $(|0\rangle\langle0|+|2\rangle\langle2|)/2$ of $|\psi_2\rangle$ has $\langle n_A^2\rangle=2$. Since $n=(q^2+p^2-1)/2$, this is fourth order in quadratures. Measuring it rejects the particular covariance countermodel, but one additional moment still does not make the measurement set informationally complete.

This does not invalidate a negative covariance witness: a genuine violation remains sufficient for entanglement. It invalidates the stronger claims that non-violation proves separability or that second moments determine an arbitrary state. Higher-order moment tests form a hierarchy rather than a single covariance condition; see Shchukin and Vogel 2005, Eqs. (10)–(21).

From detector data to a field claim

Let $\mathcal S=\{E_1,\ldots,E_n\}$ be the calibrated effects actually measured. Two states satisfying

$$\operatorname{tr}(E_j\rho)=\operatorname{tr}(E_j\sigma) \quad\text{for every }j$$

are indistinguishable in this experiment. A reconstruction algorithm can choose one representative, but uniqueness follows only if $\mathcal S$ is informationally complete on the declared state class.

A defensible certification should report the mode functions or local algebras; switching, smearing, and detector resolution; every measured setting; sample-independence and stationarity assumptions; the estimator and confidence construction; calibration and multiple-testing corrections; and regulator convergence. Compare the confidence region with the entire separable model, not only with one nominal covariance matrix. If both separable and entangled states remain compatible with the data, report a bound or an inconclusive result.

It helps to separate the inference into three records. The target record identifies the wavepacket modes and the bipartition. The instrument record maps raw probe outcomes to calibrated effects, including loss, additive noise, cross-talk, and setting labels. The statistical record states which trials were included and turns those outcomes into a confidence region. Only after these agree may a mode-level witness be interpreted as a field-theoretic statement. Agreement of a reconstruction algorithm with its own assumed model is not an independent validation.

Limitations

The benchmark concerns two type-I wavepacket modes under a regulator. It does not turn a pair of continuum type-III local algebras into tensor factors, and it does not establish that arbitrary quadrature effects have compactly supported detector realizations. A continuum claim requires an operationally specified subalgebra or split approximation and convergence under resolution and cutoff changes. The Gaussian chi-squared interval also addresses sampling variation only; it does not cover drift, misspecified detector response, mode overlap, or cutoff bias unless those effects are separately bounded.

Common pitfalls

Treating non-violation as separability. A witness is one-sided. The non-Gaussian state $|\psi_2\rangle$ passes every second-moment separability threshold in this example and is nevertheless entangled.

Subtracting detector noise without uncertainty. Calibration is data, not an identity. Propagate its confidence region into the witness statistic.

Calling a covariance fit tomography. A covariance matrix fixes a Gaussian state, not a general state. State the Gaussian assumption explicitly or weaken the claim.

Exercises

Derive $\mathcal D=2e^{-2r}$ and $\widetilde\nu_-=e^{-2r}/2$ for the covariance matrix above.

Solution

The $q$ block gives

$$\operatorname{Var}(q_A-q_B)=\cosh2r-\sinh2r=e^{-2r}.$$

The negative $p_Ap_B$ covariance gives the same result for $p_A+p_B$, hence $\mathcal D=2e^{-2r}$. Partial transposition flips the sign of $p_B$. After reordering to $(q_A,q_B,p_A,p_B)$, the covariance consists of two identical blocks

$$\frac12 \begin{pmatrix} \cosh 2r&\sinh 2r\\ \sinh 2r&\cosh 2r \end{pmatrix}.$$

Because the two blocks coincide, their smaller ordinary eigenvalue is also the smaller symplectic eigenvalue:

$$\widetilde\nu_-=\frac{\cosh2r-\sinh2r}{2} =\frac{e^{-2r}}2.$$

Independent readout noise of variance $\delta$ is added to each quadrature. Show why $\mathcal D$ increases by $4\delta$.

Solution

If the two $q$ noises are independent, the variance of their difference gains $\delta+\delta=2\delta$. The independent $p$ noises add another $2\delta$ to the variance of their sum. Therefore $\mathcal D_{\mathrm{obs}}=\mathcal D+4\delta$. For $\delta=0.04$, the increase is $0.16$.

Verify that $\rho_{\mathrm{sep}}$ and $\rho_{\mathrm{ent}}$ have the same covariance but different entanglement.

Solution

A thermal mode with mean occupation $\bar n=1$ has $\langle q^2\rangle=\langle p^2\rangle=\bar n+1/2=3/2$, and the product has no cross moments. In $|\psi_2\rangle$, either local reduced state is $(|0\rangle\langle0|+|2\rangle\langle2|)/2$, also with mean occupation one and vanishing $\langle a^2\rangle$. Operators quadratic across the modes change the occupation of each mode by at most one, so they cannot connect $|00\rangle$ with $|22\rangle$; all cross second moments vanish. The first state is explicitly a product. The second has two equal nonzero Schmidt coefficients, so its entanglement entropy is one bit and its negativity is $1/2$.

For the thermal distribution, $\operatorname{Var}(n)=\bar n(\bar n+1)=2$, so $\langle n^2\rangle=\operatorname{Var}(n)+\bar n^2=3$. The local reduction of $|\psi_2\rangle$ instead gives $\langle n^2\rangle=(0^2+2^2)/2=2$. Thus a fourth-order measurement distinguishes this pair even though every first and second moment agrees.

Reproduce the simultaneous finite-sample decision in the benchmark and state what would make the conclusion inconclusive.

Solution

For each setting, $N-1=99$, $s_x^2=0.4478794412$, and the lower $2.5\%$ chi-squared quantile is $73.36108019$. Hence

$$U_x=\frac{99(0.4478794412)}{73.36108019} =0.6044085578.$$

Giving the two one-sided bounds error probabilities $0.025$ each makes their joint failure probability no larger than $0.05$. Their sum is $U_{\mathcal D}=1.2088171156\lt2$, so the predeclared rule certifies entanglement under independent Gaussian sampling and the stated calibration. If the upper sum were $2$ or larger, the result would be inconclusive, not evidence of separability. Dependence between preparations, non-Gaussian sampling without a valid replacement interval, postselection, or unbounded calibration error would also prevent this particular confidence statement.

References

• Duan, L.-M., Giedke, G., Cirac, J. I., and Zoller, P. (2000). “Inseparability Criterion for Continuous Variable Systems.” Physical Review Letters 84, 2722–2725. DOI. Open PDF.
• Shchukin, E., and Vogel, W. (2005). “Inseparability Criteria for Continuous Bipartite Quantum States.” Physical Review Letters 95, 230502; errata 95, 249904 (2005) and 96, 129902 (2006). DOI. Open PDF.
• Simon, R. (2000). “Peres–Horodecki Separability Criterion for Continuous Variable Systems.” Physical Review Letters 84, 2726–2729. DOI. Open PDF.
• Vidal, G., and Werner, R. F. (2002). “Computable Measure of Entanglement.” Physical Review A 65, 032314. DOI. Open PDF.