Adaptive Non-local Observables in Quantum Circuits

• Adaptive Non-local Observables are trainable multi-qubit Hermitian measurement operators that replace fixed readouts in variational quantum circuits and quantum neural networks.
• They leverage k-local measurements to probe entangled qubit correlations, expanding the effective function space without the need for deeper circuits.
• Diagonal reformulations like DANO reduce parameter complexity while maintaining performance improvements in tasks such as classification, reinforcement learning, and super-resolution.

Adaptive Non-local Observables (ANOs) are trainable, multi-qubit Hermitian measurement operators used as the readout layer of variational quantum circuits (VQCs) and quantum neural networks (QNNs). In the ANO formulation, the model no longer learns only a parameterized unitary; it also learns the observable used to extract outputs from the prepared quantum state. This replaces the conventional fixed-measurement regime—typically built from local Pauli observables—with a jointly optimized, potentially $k$-local Hermitian operator. Across the ANO literature, the adaptive aspect refers to training the observable parameters together with circuit parameters, while the non-local aspect refers to allowing the observable to act jointly on multiple qubits rather than being restricted to single-qubit or strictly local Pauli readouts (Lin et al., 18 Apr 2025, Lin et al., 25 Jul 2025, Lin et al., 20 Jan 2026).

1. Definition and conceptual basis

In standard VQCs, prediction is formed from an encoding unitary and a variational unitary followed by a fixed Hermitian observable. One representative expression is

$$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$

with training loss

$$L(\theta;\mathcal D) = \frac{1}{|\mathcal D|} \sum_{j=1}^{|\mathcal D|} \left\| \langle H\rangle(x^{(j)};\theta) - y^{(j)} \right\|^2.$$

Within this regime, $H$ is fixed and only $\theta$ is learned (Lin et al., 18 Apr 2025).

ANOs replace that fixed $H$ by a trainable observable $H(\phi)$. In reinforcement learning and super-resolution formulations, the core output becomes

$$f_{\theta,\phi}(x) = \langle \psi_0 | W^\dagger(x)\,U^\dagger(\theta)\,H(\phi)\,U(\theta)\,W(x) |\psi_0\rangle,$$

or an equivalent expression with $V(x)$ in place of $W(x)$ (Lin et al., 25 Jul 2025, Lin et al., 20 Jan 2026). The observable is therefore part of the hypothesis class, not merely a fixed measurement appended to a trained circuit.

The non-local qualifier refers to $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$0-local observables acting on a $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$1-qubit subsystem of an $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$2-qubit register, with $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$3. These observables can couple multiple qubits jointly and thereby probe correlations across entangled subsystems. The adaptive qualifier refers to end-to-end optimization of the observable parameters $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$4 together with circuit parameters $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$5 (Lin et al., 20 Jan 2026).

A central conceptual motivation is the Heisenberg-picture reinterpretation of variational models. Since

$$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$6

optimizing the circuit can be viewed as moving the observable through Hermitian-operator space via conjugation. The ANO framework extends this viewpoint by allowing the observable itself to vary across equivalence classes rather than remaining confined to the orbit generated from one fixed $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$7 (Lin et al., 18 Apr 2025).

2. Observable-space formulation and canonical parameterizations

A general $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$8-local Hermitian observable is parameterized as a $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$9 Hermitian matrix with $$L(\theta;\mathcal D) = \frac{1}{|\mathcal D|} \sum_{j=1}^{|\mathcal D|} \left\| \langle H\rangle(x^{(j)};\theta) - y^{(j)} \right\|^2.$$0,

$$L(\theta;\mathcal D) = \frac{1}{|\mathcal D|} \sum_{j=1}^{|\mathcal D|} \left\| \langle H\rangle(x^{(j)};\theta) - y^{(j)} \right\|^2.$$1

with real parameters

$$L(\theta;\mathcal D) = \frac{1}{|\mathcal D|} \sum_{j=1}^{|\mathcal D|} \left\| \langle H\rangle(x^{(j)};\theta) - y^{(j)} \right\|^2.$$2

and lower-triangular entries determined by Hermitian symmetry so that $$L(\theta;\mathcal D) = \frac{1}{|\mathcal D|} \sum_{j=1}^{|\mathcal D|} \left\| \langle H\rangle(x^{(j)};\theta) - y^{(j)} \right\|^2.$$3 (Lin et al., 18 Apr 2025, Lin et al., 25 Jul 2025, Lin et al., 20 Jan 2026). Embedded into an $$L(\theta;\mathcal D) = \frac{1}{|\mathcal D|} \sum_{j=1}^{|\mathcal D|} \left\| \langle H\rangle(x^{(j)};\theta) - y^{(j)} \right\|^2.$$4-qubit system, such an operator has $$L(\theta;\mathcal D) = \frac{1}{|\mathcal D|} \sum_{j=1}^{|\mathcal D|} \left\| \langle H\rangle(x^{(j)};\theta) - y^{(j)} \right\|^2.$$5 real degrees of freedom (Tseng et al., 14 May 2026).

The observable-space perspective is formalized through unitary similarity. Two Hermitian operators are taken to be equivalent when

$$L(\theta;\mathcal D) = \frac{1}{|\mathcal D|} \sum_{j=1}^{|\mathcal D|} \left\| \langle H\rangle(x^{(j)};\theta) - y^{(j)} \right\|^2.$$6

Because Hermitian matrices are diagonalizable, the quotient space is summarized in the ANO literature as

$$L(\theta;\mathcal D) = \frac{1}{|\mathcal D|} \sum_{j=1}^{|\mathcal D|} \left\| \langle H\rangle(x^{(j)};\theta) - y^{(j)} \right\|^2.$$7

In this language, a conventional VQC with fixed $$L(\theta;\mathcal D) = \frac{1}{|\mathcal D|} \sum_{j=1}^{|\mathcal D|} \left\| \langle H\rangle(x^{(j)};\theta) - y^{(j)} \right\|^2.$$8 explores only the orbit $$L(\theta;\mathcal D) = \frac{1}{|\mathcal D|} \sum_{j=1}^{|\mathcal D|} \left\| \langle H\rangle(x^{(j)};\theta) - y^{(j)} \right\|^2.$$9, whereas ANOs allow the model to move across observable classes (Lin et al., 18 Apr 2025).

This observable-space formulation leads directly to the diagonal reformulation introduced in Diagonal ANO (DANO). DANO restricts the learned observable to

$H$0

using the fact that every Hermitian operator admits

$H$1

The diagonal matrix $H$2 is the canonical representative of the equivalence class under unitary similarity (Tseng et al., 14 May 2026).

The corresponding density result is stated as follows: if $H$3 and $H$4, then

$H$5

is dense in $H$6. The approximation argument uses the generalized Solovay–Kitaev theorem to approximate the diagonalizing unitary, with observable error bounded by

$H$7

The resulting claim is that DANO is not a restriction in principle, but a reparameterization through canonical diagonal representatives, provided the circuit family can approximate the relevant eigenbasis (Tseng et al., 14 May 2026).

3. Circuit architectures, measurement schemes, and learning rules

The ANO pipeline consists of three components: an encoding layer, a variational unitary, and an adaptive measurement layer. In reinforcement learning, the circuit is described as an encoding layer $H$8, a variational layer $H$9, and a trainable Hermitian observable $\theta$0, with joint optimization of $\theta$1 and $\theta$2 (Lin et al., 25 Jul 2025). In super-resolution, the same structure appears with an encoding unitary $\theta$3, layered entangling unitaries $\theta$4, and an ANO measurement producing the reconstructed high-resolution output (Lin et al., 20 Jan 2026).

For super-resolution, the paper gives a representative encoding

$\theta$5

and a layered variational unitary

$\theta$6

where $\theta$7 denotes entangling gates such as CNOT layers (Lin et al., 20 Jan 2026). The reconstruction objective is

$\theta$8

with $\theta$9 (Lin et al., 20 Jan 2026).

In quantum reinforcement learning, the ANO-VQC is used as the function approximator in both DQN and A3C. For DQN, the Bellman regression objective is

$H$0

where both circuit and observable parameters are updated jointly (Lin et al., 25 Jul 2025). For A3C, the actor uses

$H$1

while the critic is another ANO-based quantum approximator, trained with the stated policy, value, and entropy losses (Lin et al., 25 Jul 2025).

The original ANO QNN paper also introduces two explicit measurement constructions. The first is a sliding $H$2-local ANO, which measures overlapping $H$3-qubit windows cyclically across the register; for $H$4 and $H$5, the measurement groups are

$H$6

The second is pairwise combinatorial measurement, which chooses a subset $H$7 of qubits and measures every pair in $H$8, producing

$H$9

outputs (Lin et al., 18 Apr 2025). The pairwise scheme is designed to be parameter-efficient while still capturing rich pairwise interactions.

4. Expressivity, locality, and computational trade-offs

The primary claim attached to ANOs is that adaptive measurement expands the effective function class of a fixed-depth quantum circuit without increasing circuit depth. In the reinforcement-learning formulation, the ablation results are summarized by the statement that adaptive measurements enhance the function space without increasing circuit depth (Lin et al., 25 Jul 2025). In the super-resolution formulation, fixed Pauli readouts are described as a measurement bottleneck, and ANOs are introduced to enlarge the measurable function class by making the observable itself trainable (Lin et al., 20 Jan 2026).

The DANO paper makes this function-space argument explicit through a partial-Fourier viewpoint: $H(\phi)$0 where the encoding determines the accessible frequencies $H(\phi)$1, while the circuit plus measurement determine the coefficients $H(\phi)$2. Enlarging the observable class therefore increases flexibility in fitting the target function (Tseng et al., 14 May 2026).

Locality $H(\phi)$3 acts as a direct expressivity parameter. Higher $H(\phi)$4-local observables couple more qubits jointly and capture correlations that single-qubit or strictly local measurements can miss. The original ANO paper summarizes the complementarity between locality and rotations through the inclusion

$H(\phi)$5

arguing that variational rotations can compensate partially for low locality by mixing features before measurement, although genuinely higher-local observables remain more expressive (Lin et al., 18 Apr 2025).

This expressivity gain carries a steep parameter cost in full ANO. For $H(\phi)$6-local observables with $H(\phi)$7, a general Hermitian operator requires $H(\phi)$8 real parameters, whereas DANO trains only the diagonal eigenvalues and therefore requires $H(\phi)$9 parameters (Tseng et al., 14 May 2026). In the reported 16-qubit MNIST experiment with $$f_{\theta,\phi}(x) = \langle \psi_0 | W^\dagger(x)\,U^\dagger(\theta)\,H(\phi)\,U(\theta)\,W(x) |\psi_0\rangle,$$0, full ANO needs $$f_{\theta,\phi}(x) = \langle \psi_0 | W^\dagger(x)\,U^\dagger(\theta)\,H(\phi)\,U(\theta)\,W(x) |\psi_0\rangle,$$1 real parameters across sliding observables, while DANO needs only $$f_{\theta,\phi}(x) = \langle \psi_0 | W^\dagger(x)\,U^\dagger(\theta)\,H(\phi)\,U(\theta)\,W(x) |\psi_0\rangle,$$2, a $$f_{\theta,\phi}(x) = \langle \psi_0 | W^\dagger(x)\,U^\dagger(\theta)\,H(\phi)\,U(\theta)\,W(x) |\psi_0\rangle,$$3 reduction (Tseng et al., 14 May 2026). For statevector simulation, the same paper states that measurement-side classical computation scales as $$f_{\theta,\phi}(x) = \langle \psi_0 | W^\dagger(x)\,U^\dagger(\theta)\,H(\phi)\,U(\theta)\,W(x) |\psi_0\rangle,$$4 for ANO and $$f_{\theta,\phi}(x) = \langle \psi_0 | W^\dagger(x)\,U^\dagger(\theta)\,H(\phi)\,U(\theta)\,W(x) |\psi_0\rangle,$$5 for DANO, removing the extra $$f_{\theta,\phi}(x) = \langle \psi_0 | W^\dagger(x)\,U^\dagger(\theta)\,H(\phi)\,U(\theta)\,W(x) |\psi_0\rangle,$$6 factor from the measurement-side classical cost (Tseng et al., 14 May 2026).

The observable spectrum also changes the numerical range of the model output. The ANO literature notes that Pauli observables effectively restrict outputs to $$f_{\theta,\phi}(x) = \langle \psi_0 | W^\dagger(x)\,U^\dagger(\theta)\,H(\phi)\,U(\theta)\,W(x) |\psi_0\rangle,$$7, whereas a general Hermitian observable satisfies

$$f_{\theta,\phi}(x) = \langle \psi_0 | W^\dagger(x)\,U^\dagger(\theta)\,H(\phi)\,U(\theta)\,W(x) |\psi_0\rangle,$$8

A trainable spectrum therefore broadens the accessible output range when the task requires it (Lin et al., 18 Apr 2025, Lin et al., 25 Jul 2025).

5. Reported applications and empirical findings

The ANO framework has been applied to supervised classification, quantum reinforcement learning, and quantum super-resolution, with DANO introduced as a parameter-efficient reformulation. The reported results are summarized below.

Domain	Setting	Reported finding
Classification	Banknote Authentication	2-local w/ rotations: $$f_{\theta,\phi}(x) = \langle \psi_0 | W^\dagger(x)\,U^\dagger(\theta)\,H(\phi)\,U(\theta)\,W(x) |\psi_0\rangle,$$9
Classification	MNIST, sliding $V(x)$0-local	Accuracy rises from $V(x)$1 at 1-local to $V(x)$2 at 5-local
Classification	MNIST, pairwise combinatorial	16-qubit pairwise: $V(x)$3 with 3130 parameters
Reinforcement learning	DQN/A3C benchmarks	ANO-VQC agents outperform baseline VQCs
Super-resolution	MNIST, $V(x)$4	3-local ANOs outperform 2-local ANOs in MSE, PSNR, and SSIM
Diagonal reformulation	Reduced MNIST and Yale B	DANO improves test accuracy from $V(x)$5 to $V(x)$6; Yale B reaches $V(x)$7 at 10-local

In classification, the original ANO paper reports that all ANO variants outperform the fixed Pauli baseline on Banknote Authentication. The reported test accuracies are $V(x)$8 for the Pauli baseline, $V(x)$9 for 1-local with rotations, $W(x)$0 for 2-local with rotations, and $W(x)$1 for 3-local with rotations. On MNIST resized to $W(x)$2, sliding $W(x)$3-local ANOs improve from $W(x)$4 at 1-local to $W(x)$5 at 5-local, while the pairwise combinatorial scheme reaches $W(x)$6 with 3130 parameters, outperforming the 5-local sliding model at $W(x)$7 with 10,304 parameters (Lin et al., 18 Apr 2025).

In quantum reinforcement learning, the ANO-VQC architecture is used within both DQN and A3C. On CartPole under DQN, “3-local w/ R.” learns fastest and reaches the 500-step cap earlier and more stably; “Only R.” converges more slowly and to a lower average reward; and “Only measurement” performs similarly to or slightly better than the classical VQC baseline, but worse than the full ANO plus rotation model. On MountainCar, increasing locality from 3-local to 6-local improves learning speed and final reward, while at 6-local the versions with and without rotation gates perform very similarly. In A3C, “3-local w/ R.” crosses average reward 400 around episode 12,000 on CartPole, and on MiniGrid SimpleCrossing S9N1 it surpasses about 0.8 success by episode 6000, compared with about 0.4 for “Only R.” and about 0.3 for “Only measurement” (Lin et al., 25 Jul 2025).

In super-resolution, the 2026 study presents the first study to investigate quantum circuits for SR and uses ANO-VQCs on MNIST digits downsampled from $W(x)$8 to $W(x)$9, then reconstructed to $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$00, $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$01, and $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$02 for $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$03, $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$04, and $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$05, respectively. For 2-local ANOs, the reported metrics are: $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$06: MSE $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$07, LPIPS $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$08, PSNR $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$09, SSIM $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$10; $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$11: MSE $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$12, LPIPS $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$13, PSNR $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$14, SSIM $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$15; $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$16: MSE $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$17, LPIPS $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$18, PSNR $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$19, SSIM $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$20. For 3-local ANOs, the metrics are: $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$21: MSE $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$22, LPIPS $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$23, PSNR $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$24, SSIM $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$25; $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$26: MSE $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$27, LPIPS $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$28, PSNR $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$29, SSIM $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$30; $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$31: MSE $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$32, LPIPS $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$33, PSNR $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$34, SSIM $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$35. The main reported pattern is that 3-local ANOs outperform 2-local ANOs in MSE, PSNR, and SSIM, with a slight LPIPS increase (Lin et al., 20 Jan 2026).

The DANO reformulation preserves the ANO principle while reducing measurement-side parameterization cost. In the reported experiments, increasing locality from pure VQC to DANO improves test accuracy from $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$36 to $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$37 on reduced MNIST; on Yale B face recognition, accuracy rises from $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$38 to $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$39 at 10-local DANO; and a “rescue” experiment shows that freezing a weak VQC and training only the diagonal observable can raise performance to $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$40 (Tseng et al., 14 May 2026).

6. Conceptual boundaries, related precedents, and common misconceptions

A recurring source of confusion is the scope of the term “adaptive.” In the ANO literature, adaptivity means that the observable parameters are optimized during training together with circuit parameters. It does not denote a sequential measurement protocol in which later observables are selected on the basis of earlier measurement outcomes (Lin et al., 25 Jul 2025, Lin et al., 20 Jan 2026).

This distinction is particularly important when comparing ANOs with earlier work on non-local observables. “On the informational completeness of local observables” proves a static reconstruction theorem rather than introducing ANOs as a measurement class. For tripartite states $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$41 and $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$42 on $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$43 with identical marginals on $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$44 and $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$45, the paper establishes

$$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$46

where

$$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$47

Its multipartite generalization uses a Markov entropy decomposition

$$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$48

and yields the locally checkable bound

$$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$49

Since trace distance controls expectation-value differences for arbitrary observables, the paper shows that local reduced density matrices can determine expectations of many nonlocal observables under small conditional-mutual-information conditions. The relation to ANOs is therefore structural rather than formal: both treat some nonlocal observable expectations as determined by local information, but the 2014 paper does so via local entropy bounds and Markov shields rather than by defining trainable adaptive observables (Kim, 2014).

An even broader distinction arises in holography. “Quantum corrections to dynamical holographic thermalization: entanglement entropy and other non-local observables” studies geodesic lengths, Wilson loops, and entanglement entropy as non-local probes of nonequilibrium dynamics in planar $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$50 SYM. The paper analyzes how $$\langle H \rangle(x;\theta) = \bra{\psi_0} V^\dagger(x)\, U^\dagger(\theta)\, H\, U(\theta)\, V(x)\ket{\psi_0},$$51 and leading non-planar corrections modify the thermalization behavior of these observables, finding a UV/IR split in which shorter scales thermalize faster and larger scales more slowly. These are non-local observables in the field-theoretic and holographic sense, but they are not ANOs: the paper does not propose a trainable observable class, a Heisenberg-picture QNN formulation, or a joint optimization rule for measurement operators (Baron et al., 2013).

The resulting taxonomy is therefore precise. ANOs are a quantum-machine-learning framework in which measurement is elevated to a trainable component of the model. Earlier work on non-local observables in quantum information and holography is relevant because it demonstrates how non-local expectation values can encode global structure or dynamical information, but such work does not by itself instantiate the ANO formalism (Kim, 2014, Baron et al., 2013).