---
title: Choi-Shadow Estimators
url: https://www.emergentmind.com/topics/choi-shadow-estimators
type: topic
---

# Choi-Shadow Estimators

Searching arXiv for recent papers on Choi-shadow estimators and related Choi/shadow process tomography.
arxiv_search(query="Choi shadow estimators process tomography Choi shadows", max_results=10)
arxiv_search(query="shadow process tomography Choi isomorphism classical Choi shadows", max_results=10)
Choi-shadow estimators are estimator constructions that encode a quantum process into a Choi-state, Choi operator, branch Choi operator, or pseudo-Choi state, and then apply shadow-tomography machinery to estimate target functionals without full process reconstruction. In the canonical channel setting, a CPTP map \(E\) is represented by the Choi operator
\[
\eta = (\mathcal I_A\otimes E_B)\big[\ket{\omega}\bra{\omega}\big],
\]
with channel action recovered through
\[
E(\rho)=\operatorname{Tr}_A\!\big[(\rho^T\otimes \mathbb I_B)\,\eta\big].
\]
Consequently, any quantity of the form \(\operatorname{Tr}[E(\rho)O]\) becomes the linear functional \(\operatorname{Tr}[\eta(\rho^T\otimes O)]\), which can be estimated from shadow data rather than from a fully reconstructed process matrix [2110.03629]. Recent work extends this template to branch-resolved dynamic teleportation via classical Choi shadows [2604.28037], to generalized-measurement shadow process tomography with optimized POVMs [2506.23806], and to Hamiltonian learning via pseudo-Choi states that play a Choi-like role for generators rather than channels [2308.13020].

## 1. Core representation and estimator principle

The defining move is the Choi reduction: process inference is recast as state inference on a larger bipartite system. In one standard normalization, the Choi-state of a channel \(\mathcal N_A\) acting on a \(d\)-dimensional system is
\[
J_{\mathcal N_{AB}} = (\mathcal N_A \otimes \mathcal I_B)\!\left(\ketbra{\Phi^+}_{AB}\right),
\qquad
\ket{\Phi^+}_{AB}=\frac{1}{\sqrt d}\sum_{i=0}^{d-1}\ket{ii}_{AB},
\]
and the inverse relation is
\[
\mathcal N_A(\rho) = d\,\mathrm{tr}_B\!\left[(\mathbb I_A\otimes \rho^t)\,J_{\mathcal N_{AB}}\right].
\]
The Choi object therefore fully encodes the channel, and positivity plus the marginal condition
\[
J_{\mathcal N_{AB}} \ge 0,\qquad \mathrm{tr}_A[J_{\mathcal N_{AB}}]=\frac{\mathbb I_B}{d}
\]
characterize CPTP maps [2406.08360].

Taken together, the quantum-shadow papers suggest a common estimator architecture: encode the unknown process into a Choi-type object; measure that object through randomized or generalized measurements; construct unbiased or least-squares single-shot reconstructions; and evaluate linear observables on the reconstructed Choi samples. This shifts the computational burden from full tomography to observable-specific estimation, while preserving access to many channel properties from a common data stream [2110.03629].

| Framework | Encoded object | Distinguishing feature |
|---|---|---|
| Shadow process tomography [2110.03629] | Choi operator \(\eta\) | Direct channel-function estimation |
| Classical branch Choi shadows [2604.28037] | Branch Choi operator \(\mathcal C_m\) | Branch-resolved dynamic-circuit analysis |
| POVM-SPT [2506.23806] | Normalized Choi state \(\eta/d\) | Optimized POVMs minimize shadow norm |
| Pseudo-Choi shadows [2308.13020] | Pseudo-Choi state \(|\psi_c\rangle\) | Hamiltonian-coefficient estimation |

## 2. Standard Choi-shadow construction for channels

The standard shadow process tomography protocol samples the Choi operator indirectly by randomized preparation and measurement on the input and output sides. For each shot, one samples \(b_{\text{in}}\), prepares \(U_{\text{in}}\ket{b_{\text{in}}}\), applies the channel, applies \(U_{\text{out}}\), measures \(b_{\text{out}}\), and records
\[
z=\{b_{\text{in}},U_{\text{in}},U_{\text{out}},b_{\text{out}}\}.
\]
The corresponding Choi-shadow basis vector is
\[
\ket{z}=U_{\text{in}}^T\ket{b_{\text{in}}}\otimes U_{\text{out}}^\dagger \ket{b_{\text{out}}}.
\]
With \(\mathcal M_{U\otimes U}\) the induced shadow measurement map on the input-output system, the single-shot Choi shadow is
\[
\hat\zeta=\mathcal M_{U\otimes U}^{-1}\!\left(\ket{z}\bra{z}\right),
\]
and repeated sampling yields \(\zeta_m=\frac{1}{m}\sum_{j=1}^m \hat\zeta_j\) with \(\mathbb E[\zeta_m]=\eta\) [2110.03629].

Target functionals are then estimated linearly. For an input state \(\rho\) and output observable \(O\),
\[
o=\operatorname{Tr}[E(\rho)O]
      =\operatorname{Tr}\!\big[\eta\,(\rho^T\otimes O)\big],
\]
so one single shadow sample induces
\[
\hat o_j = \operatorname{Tr}\!\big[\hat\zeta_j(\rho^T\otimes O)\big].
\]
The recommended aggregation is median-of-means: split the \(m\) samples into \(K\) groups of size \(N\), average within each group, and take the median. In the local-Pauli specialization, the inverse map is explicit,
\[
\mathcal M_{U}^{-1}(A)=3A-\operatorname{Tr}(A)\mathbb I,
\]
and the single-qubit estimator building block becomes
\[
\tau_{b,\mu}=3\ket{b_\mu}\bra{b_\mu}-\mathbb I.
\]
A recurring practical point is that these estimators are not, in the first instance, reconstructed CPTP maps. They are shadows of the Choi operator from which channel properties are extracted linearly [2110.03629].

## 3. Major variants of the estimator family

A branch-resolved variant was introduced for dynamic-circuit teleportation on superconducting hardware. There the mid-circuit measurement defines a quantum instrument
\[
\rho \mapsto \mathcal{I}(\rho) = \sum_{m \in \Omega} \mathcal{E}_m(\rho)\otimes |m\rangle\langle m|,
\]
with post-correction branch channel
\[
\tilde{\mathcal{E}}_m = \zeta_m \circ \mathcal{E}_m, \quad \zeta_m(\rho) = U_m \rho U_m^\dagger.
\]
Each branch has a subnormalized branch Choi operator
\[
\mathcal{C}_m = (\mathbb{I}_R \otimes \tilde{\mathcal{E}}_m)\left[|\Phi_0^+\rangle\langle \Phi_0^+|\right],
\]
whose trace gives the branch probability \(p_m=\mathrm{Tr}(\mathcal C_m)\). Using a two-qubit classical shadow over the reference-output pair \((R,B)\), the branch-specific estimator is
\[
\hat{\mathcal{C}}_\ell = \frac{1}{N}\sum_{i=1}^N \mu_i\,\delta_{\ell_i,\ell}\,\tau_{b_i,s_i},
\]
with \(\mu_i=1\) for unmitigated physical correction and post-processing, and \(\mu_i=\alpha_{f_i}/p(f_i)\) for PROM mitigation [2604.28037].

A second major variant replaces random-unitary measurements by optimized generalized measurements. In POVM-SPT, a tensor-product POVM \(E_k^{\rm tot}\) is applied to the normalized Choi state \(\eta/d\), the least-squares reconstruction map is
\[
\chi_{\rm LS}=(\Phi_E^\dagger \Phi_E)^{-1}\Phi_E^\dagger,
\]
and single-shot Choi reconstructions are built as \( \hat{\rho}_k=C_E^{-1}(E_k)\), with the process estimator
\[
\hat{x}_k=\operatorname{Tr}\!\left[\hat{\eta}_k(\rho^T\otimes X)\right].
\]
The paper emphasizes that projection measurements are a special case of POVMs, so the optimized POVM cannot do worse than the best projective measurement in shadow-norm terms [2506.23806].

A third variant uses pseudo-Choi states for Hamiltonian learning. The pseudo-Choi state is defined on \(S,A,C\) by
\[
|{\psi_c}\rangle := \frac{({H} \otimes I_A)|{\Phi_d}\rangle_{SA}|0\rangle_C + |{\Phi_d}\rangle_{SA}|1\rangle_C}{\alpha},
\]
with \(\alpha=\sqrt{c_2^2+1}\). Coefficients are decoded through operators \(O_l\) and \(O_\alpha\) satisfying
\[
{\rm Tr}(\rho_c O_{l}) = \frac{c_l}{\alpha^2}, \qquad {\rm Tr}(\rho_c O_{\alpha}) = \frac{1}{\alpha^2},
\]
so that
\[
c_l = \frac{{\rm Tr}(\rho_c O_l)}{{\rm Tr}(\rho_c O_\alpha)}.
\]
This preserves the same estimator logic—encode, shadow, decode—but shifts the target from channels to Hamiltonian parameters [2308.13020].

## 4. Accuracy, shadow norms, and complexity

The finite-sample guarantees are organized around shadow norms. For channel shadows, if \(O_1,\dots,O_M\) are output operators and \(\rho_1,\dots,\rho_L\) are input states, then
\[
K=2\log(2ML/\delta),
\]
and
\[
N=\frac{34}{\epsilon^2}\,4^n\, \max_{1\le j\le M,\;1\le \ell\le L} f_{\mathcal U}(\rho_\ell)\,f_{\mathcal U}(O_j)
\]
suffices for simultaneous estimation at error \(\epsilon\) with probability at least \(1-\delta\). The paper attributes the \(4^n\) factor to the normalization convention for the unnormalized Choi operator, making it a process-specific overhead absent from ordinary normalized-state shadow estimation [2110.03629].

In POVM-SPT, variance is controlled by
\[
\|d\cdot \rho^T\otimes X\|_{E^{\rm tot}}^2
=
\lambda_{\max}\!\left(
\sum_k
\operatorname{Tr}[\hat{\eta}_k(\rho^T\otimes X)]^2\,E_k^{\rm tot}
\right),
\]
and the worst-case norm over a family \(\mathcal X\) is \(\kappa_{E^{\rm tot}}^2\). The median-of-means sample complexity is
\[
M=\mathcal{O}\!\left(\frac{\log(HG)\,\kappa^2_{E^{\rm tot}}}{\epsilon^2}\right),
\]
more explicitly with
\[
K=2\ln(2HG/\delta),\qquad \frac{M}{K}=\frac{34}{\epsilon^2}\max_{j,l}\|d\cdot \rho_l^T\otimes X_j\|_{E^{\rm tot}}^2.
\]
The reported numerical gains are an approximate 7-fold reduction in the squared shadow norm for single-qubit input states and a reported \(2^{180}\)-fold sample-complexity enhancement for 64-qubit input states relative to conventional SPT [2506.23806].

For pseudo-Choi Hamiltonian learning, the classical-shadow route yields query complexity
\[
\widetilde{\mathcal O}\!\left(\frac{M}{t^2\epsilon^2}\right),
\]
while replacing classical shadows with quantum mean estimation improves this to
\[
\widetilde{\mathcal O}\!\left(\frac{M}{t\epsilon}\right).
\]
The paper attributes the constant-observable behavior to bounded shadow norms for the decoding operators, stating \(O_i^2{}_{\text{shadow}}\le 6\) for Clifford shadows [2308.13020].

## 5. Operational uses and empirical behavior

The original channel-shadow framework was designed for estimating many targeted properties of large quantum channels rather than reconstructing the full process tensor. The paper gives explicit uses for transition probabilities, multitime correlation functions, channel concatenation, and the application of channel shadows to shadows of quantum states. It also identifies a sign problem in both concatenation and channel-on-state composition, because the induced coefficients become signed or quasiprobabilistic rather than strictly positive [2110.03629].

Branch-resolved classical Choi shadows provide a concrete experimental demonstration of the value of preserving Choi information at the branch level. In dynamic teleportation experiments, physical correction, post-processing adjustments, and PROM-mitigated physical application were compared against full tomography of the branch Choi operators. At a shadow budget of 73,728 shots on layout 1, the reported RMSE values for the perfect \(W_4\) resource were \(0.00694\) for physical correction and \(0.00779\) for post-processing on the primary observable set, and \(0.00797\) and \(0.00754\) on the full observable family. The feed-forward penalty
\[
\Delta^{\mathrm{FF}}=\sum_\ell \hat p_\ell^{\mathrm{post}}\Delta_\ell^{\mathrm{FF}}
\]
was \(0.0199\) and \(0.0259\) on layout 1, but \(0.0703\) and \(0.0698\) on layout 2 for the perfect and symmetric \(W_4\) resources, respectively. The paper reports a reversal in branch-quality ordering: PROM is best on the noisier-readout layout 1, whereas post-processing exceeds PROM for every branch on layout 2 [2604.28037].

Pseudo-Choi shadows also support robustness statements beyond ordinary tomography. When the true Hamiltonian contains additional orthogonal terms outside the modeled class, the known coefficients remain estimable and the residual normalization can reveal the missing component through
\[
\hat{\chi}^2 = (\hat{\gamma}^2 -1)\Delta^2 - \|\tilde c\|_2^2.
\]
For noisy resource states \(\tilde{\rho}_c = (1-\omega)\ket{\psi_c'}\bra{\psi_c'} + \omega \rho^\perp\), the paper states
\[
|{\rm Tr}(\tilde{\rho}_c O_l) - {\rm Tr}({\rho_c}' O_l)| \le 2\omega,
\]
so preparation noise adds an explicit additive error term [2308.13020].

## 6. Boundaries of the term and adjacent literatures

A common source of confusion is terminological. Not every result involving a Choi object and not every use of the word “shadow” belongs to the same estimator family. The operational interpretation of Choi rank, for example, is not an estimator construction but a channel property: if the encoded states in an entanglement-assisted exclusion task all have rank equal to the Choi rank \(r_c^{\mathcal N}\), then
\[
k \le \left\lfloor \frac{N\big(d^2-r_c^{\mathcal N}\big)}{d^2}\right\rfloor.
\]
This gives a universal exclusion bound, not a shadow estimator [2406.08360].

Likewise, infinite-dimensional analogues of Choi matrices generalize the representation theory of maps on von Neumann factors rather than shadow tomography. For suitable normal completely bounded maps \(\phi\), the paper defines two Choi-like objects,
\[
C_\phi := \Phi(E_0)\in B(H), \qquad D_\phi := \phi^*(E_0)\in T(H),
\]
and proves that \(C_\phi\ge 0\) or \(D_\phi\ge 0\) is equivalent to complete positivity of \(\phi\). It also shows that universal existence of these correspondences for all normal completely bounded maps holds if and only if the factor is of type I [2311.18240]. This suggests a broader “Choi-object as compressed witness” perspective, but not a direct shadow-tomography protocol.

Outside quantum information, the term “shadow” appears in unrelated estimator traditions. A semiparametric DID paper develops a shadow-variable-based estimator for the ATT under MNAR post-treatment outcome missingness, using a fully observed variable \(Z\), an odds-ratio model, and stacked estimating equations [2606.08474]. A separate model-uncertainty paper studies constrained M-estimation with Lagrangian shadow prices and individual shadow prices,
\[
\mathrm{ISP}_j(c) = 2\lambda(c)\,\mathrm{sign}(g_j)\,[\Sigma^{-1}g]_j,
\]
to quantify the empirical relevance of candidate restrictions [2604.15571]. These are methodologically distinct from Choi-shadow estimators in the quantum-process sense.

In the narrow technical sense established by the process-tomography literature, Choi-shadow estimators are therefore best understood as estimators that combine a Choi-type encoding with shadow-style measurement and reconstruction. Their common pattern is: represent the process by a state-like object, estimate linear functionals of that object efficiently, and trade full process recovery for scalable access to targeted observables and operational diagnostics [2110.03629].

Source: https://www.emergentmind.com/topics/choi-shadow-estimators