Rodeo Algorithm: A Survey Across Domains

Updated 7 February 2026

Rodeo Algorithm is a family of adaptive methods spanning quantum state preparation, robust deep learning outlier detection, nonparametric estimation, and probabilistic ODE solving.
In quantum computing, it uses controlled evolution with ancilla measurements to exponentially suppress unwanted spectral components and isolate desired eigenstates.
RODEO in deep learning and statistics leverages adaptive sampling, text-guided diffusion, and coordinatewise derivative tests to enhance detection, estimation, and efficiency.

The term "Rodeo Algorithm" denotes multiple independently developed, field-specific algorithms, notably in quantum computing for eigenstate preparation and spectrum estimation, robust outlier detection in adversarial deep learning, probabilistic ODE solvers, nonparametric statistics, and online object detection. Each reflects distinct methodological traditions and technical architectures. This article first details the Rodeo Algorithm in quantum computing, then covers the data-centric adversarial outlier detection method RODEO, and finally provides a technical survey of other important Rodeo-style algorithms.

1. The Rodeo Algorithm for Quantum Eigenstate Preparation and Spectrum Estimation

Motivation and Core Framework

The Rodeo Algorithm (RA) is a quantum state-preparation and spectral estimation protocol that projects an arbitrary initial state onto a chosen energy eigenstate of a (generally many-body) Hamiltonian via repeated controlled evolutions and projective ancilla measurements. It achieves exponential suppression of undesired spectral components with asymptotically optimal circuit depth (Choi et al., 2020, Qian et al., 2021, Bonitati, 2024, Patkowski et al., 5 Feb 2026, Patkowski et al., 21 Oct 2025, Bee-Lindgren et al., 2022).

Formally, given a Hamiltonian $H$ with eigenstates $|E_j\rangle$ and an initial state $|\psi_I\rangle = \sum_j c_j |E_j\rangle$ , RA uses an ancilla qubit to control the application of time-evolution unitaries $U_n = \exp(-i H t_n)$ and phase rotations $P(E t_n) = e^{iEt_n}$ . The ancilla is measured after each cycle; only runs in which the measurement sequence produces the desired outcome (frequently “0” or “1,” depending on convention) are kept.

Key mathematical primitives:

Single-cycle amplitude update: after one ancilla-controlled cycle with time $t$ , amplitudes are multiplied by $f_j(t)=\tfrac{1}{2} + \tfrac{1}{2} e^{-i(E_j-E)t}$ .
Repeating $N$ cycles, the filter applied on the energy eigenstate $|E_j\rangle$ becomes $\prod_{n=1}^N \cos\bigl[(E_j-E)t_n/2\bigr]$ .
For states with $E_j\neq E$ , the amplitude is exponentially suppressed with $N$ ; for $E_j=E$ the amplitude survives.

Success probability for all $N$ ancilla cycles is:

$P_\mathrm{success}(E) = \sum_j |c_j|^2 \prod_{n=1}^N \cos^2\left(\frac{(E_j - E) t_n}{2}\right)$

By scanning over $E$ , this produces sharp peaks in $P_\mathrm{success}$ at eigenvalues of $H$ (Choi et al., 2020, Bonitati, 2024, Qian et al., 2021).

Circuit Realization, Scaling, and Practical Protocols

Ancilla qubits: Only one is needed with mid-circuit measurement/reset; otherwise, $N$ .
Gate decomposition: Each cycle entails a Hadamard on the ancilla, controlled- $e^{-iHt_n}$ on the system, a phase gate, and a final Hadamard plus measurement.
For multi-qubit $H$ , $e^{-iHt_n}$ is synthesized via Trotter-Suzuki or LCU decomposition; controlled variants require Toffoli-style gate decompositions, scaling as $O(\mathrm{poly}(n))$ in number of system qubits (Bee-Lindgren et al., 2022, Bazavov et al., 2024).

Resource scaling:

Projective filtering error $\delta$ → number of cycles $N = O(\log(1/\delta))$ .
Total time budget to error $\epsilon$ for eigenvalue estimation: $O((\log\epsilon)^2/(p \epsilon))$ where $p=|c_j|^2$ is initial overlap (Choi et al., 2020, Qian et al., 2021).
State preparation to error $\Delta$ : $O(\log\Delta/p)$ .
With random time sampling (e.g., $t_n$ from $\mathcal{N}(0,\sigma^2)$ ), suppression of off-resonant eigencomponents is only average-case optimal; geometric/deterministic sampling (see below) achieves provable worst-case bounds (Cohen et al., 2023, Patkowski et al., 5 Feb 2026).

Geometric and Deterministic Time Sampling for Optimal Suppression

Random (e.g., Gaussian) $t_n$ selection yields exponentially large fluctuations in the suppression of unwanted eigenstates: the suppression is log-normally distributed, so single-run performance exhibits large variance (Cohen et al., 2023). By selecting $t_n$ as a geometric sequence— $t_n = t_0 r^{n-1}$ for $0 $r$

For a gap $\Delta_\text{min}$ and total allowed evolution time $T$ , setting $r$ near $0.5$ for $T\lesssim \pi/\Delta_\text{min}$ and flattening $r\rightarrow1$ for larger $T$ transitions the convergence from power-law to exponential. This approach outperforms the original random protocol by several orders of magnitude in suppression rate for the same computational effort (Patkowski et al., 5 Feb 2026, Cohen et al., 2023).

Limitations and Hybridization Strategies

The primary limitation is the dependence on initial overlap $|c_j|^2$ with the target eigenstate. In large many-body systems, this can scale unfavorably with system size (Orthogonality Catastrophe). Preconditioning with adiabatic ramps, QAOA, or a variational Rodeo Algorithm (VRA) can boost overlap. The fusion method incrementally builds larger systems from exactly solved subsystems, alternately using adiabatic couplings and local RA filtering, offering polynomial scaling in system size and precision for high-fidelity preparation in large-scale systems (Patkowski et al., 21 Oct 2025, Bonitati, 2024).

Experimental Validation and Extensions

Sub-percent eigenvalue estimation and eigenstate preparation have been achieved on superconducting and ion-trap hardware for systems up to $n=2$ qubits, with robustness to CNOT and readout error (Qian et al., 2021, Bee-Lindgren et al., 2022).
Applications demonstrated include Hamiltonian spectrum extraction, ground/excited state preparation, computation of quantum expectation values (Hellmann–Feynman theorem), and thermodynamic state counting (Rocha et al., 2023, Wang et al., 2024, Bazavov et al., 2024).
The protocol is adaptable to yield the degeneracy function $\Omega(E)$ by cycling over computational basis input states and statistically averaging ancilla measurement outcomes (Rocha et al., 2023).

2. RODEO for Robust Outlier Detection in Deep Learning

Data-Centric Synthesis of “Near-Distribution” Outliers

RODEO (Robust Outlier Detection via Exposing Adaptive Out-of-Distribution Samples) is a framework for adversarially robust outlier detection in image classification. The central idea is to synthesize “near-distribution” and diverse out-of-distribution (OOD) samples by text-guided diffusion-based editing of inlier data (Mirzaei et al., 28 Jan 2025).

For each inlier image $x_0$ , RODEO generates prompts using Word2Vec and negative-adjective augmentations (e.g., “broken screw” for a “screw” class).
A pretrained diffusion model (GLIDE with CLIP guidance) is initialized at an intermediate noise level and denoised to maximize CLIP-similarity to the OOD prompt, with stochastic noise level $t_0\sim\mathcal{U}(0.3T, 0.6T)$ conferring diversity.
Generated samples are filtered with a CLIP-image/text similarity threshold to ensure proximity to the inlier manifold but preserve outlier semantics.

The synthesized OOD set is used to augment standard outlier exposure and adversarial training:

$L_\mathrm{RODEO}(\theta) = L_\mathrm{in}(\theta) + \lambda_1\,L_\mathrm{OE}(\theta) + \lambda_2\,L_\mathrm{adv}(\theta)$

where $L_\mathrm{OE}$ is over the adaptively generated OOD set, $L_\mathrm{adv}$ enforces robustness to $\ell_\infty$ -norm adversarial perturbations, and the train set includes both inlier and generated OOD data (Mirzaei et al., 28 Jan 2025).

Performance Benchmarks

Substantial improvement in adversarial ( $\mathrm{PGD}$ -1000) AUROC: for one-class novelty detection, clean-AUROC $\sim84\%$ ; adversarial AUROC $\sim64\%$ vs near-zero for prior methods.
On open-set recognition, robustness is doubled over baselines like PLP and ATD.
For out-of-distribution detection, clean-AUROC is $\sim93\%$ ; under strong attack, AUROC is $\sim69\%$ vs $5\%$ for EXOE.
RODEO demonstrates consistently improved robustness on standard (CIFAR-10, CIFAR-100) and high-resolution medical image benchmarks (Mirzaei et al., 28 Jan 2025).

Methodological Insights

Diversity and proximity to the inlier manifold, quantified by FID, density, and coverage metrics, are crucial for effective adversarial robustness.
Synthetic or fixed OOD datasets that are too far from the decision boundary yield weak improvements.
Text-guided diffusion schemes for OOD generation can leverage information from both vision and language domains to modulate semantic/pixel-level diversity.
The main limitation is reliance on semantic class labels for inlier data; large-scale generation takes $\sim1$ hour for $10^4$ low-resolution samples on a single GPU.

3. Rodeo Algorithm and Extensions in Nonparametric Statistics

Sparsity-Adaptive Kernel Conditional Density Estimation

The statistical Rodeo (Regularization of Derivative Expectation Operator) algorithm, originally by Wasserman and Lafferty, has engendered several nonparametric methods for adaptive bandwidth selection, most notably in conditional density estimation with high-dimensional data [$2106.14669$, $1801.06477$].

The key idea is to exploit the partial derivative of the empirical estimator with respect to each bandwidth parameter as a test statistic for relevance. If, for a coordinate $j$ , $\left|\frac{\partial}{\partial h_j} \hat f_{h}(w)\right|$ falls below a data-dependent threshold, $h_j$ is “grown” (made less selective), else it is shrunk.
These coordinatewise procedures adaptively select low-dimensional “active” bandwidths, scaling computational cost as $O(d n \log n)$ while achieving minimax convergence rates of $(\log n/n)^{s/(2s+r)}$ where $r$ is the number of relevant variables (Nguyen et al., 2021).
Empirical studies confirm that as $d$ increases (with $r$ small), estimation error remains nearly stable up to $d\sim \log n$ , provided sparsity holds.

Algorithmic Structure

Initialization: all bandwidths set to a moderate constant value.
Alternate “Reverse” and “Direct” coordinatewise passes: in Reverse, small derivative directions are grown; in Direct, large derivative directions are shrunk. Each pass is governed by coordinatewise empirical thresholds.
Active bandwidths are iterated until either no significant directions remain or the overall bandwidth product falls below a lower limit.

4. Probabilistic ODE Solvers: rodeo Library

The rodeo Python library implements a linear-scaling family of probabilistic solvers for ODE parameter inference, based on state-space modelization and Kalman filtering (Wu et al., 26 Jun 2025). Key ideas:

The ODE's solution is modeled as a $q$ -times integrated Brownian motion with transitions $X_{n+1}|X_n \sim \mathcal{N}(A X_n, Q)$ , and the ODE is enforced via "model interrogations" formulated as noisy linear constraints.
Each solver step applies Kalman filtering and (if desired) smoothing to obtain the posterior mean and covariance for the latent state trajectory.
The core algorithm and all likelihood approximations (e.g., Fenrir, DALTON, MAGI) are implemented in JAX, supporting AD and JIT compilation for both simulation and gradient-based inference.
For a system of $d$ ODEs and $N$ time steps, computational complexity is $O(d N)$ , achieved by block-diagonalizing the state-update for independent variables.

5. Additional Domains: Online Object Detection and Autoencoding

RODEO also denotes (1) a replay-based online object detector for streaming class-incremental learning, and (2) a robust de-aliasing autoencoder for MRI/CT reconstruction.

In streaming detection, RODEO maintains performance under “catastrophic forgetting” by replaying compressed mid-network features using product quantization, achieving state-of-the-art performance on PASCAL VOC 2007 and MS COCO (Acharya et al., 2020).
In real-time imaging, RODEO trains a single-layer autoencoder with robust $\ell_1$ reconstruction via Split-Bregman splitting, yielding $5$– $10\times$ speed-ups for MRI and CT image reconstruction versus compressed sensing (Mehta et al., 2019).

6. Comparative Table of Rodeo Algorithms

Domain	Core Principle	Notable Reference(s)
Quantum Computing (Eigenstate)	Controlled quantum evolution + ancilla filtering	(Choi et al., 2020, Bonitati, 2024, Patkowski et al., 5 Feb 2026, Rocha et al., 2023, Qian et al., 2021)
Adversarial Outlier Detection	Adaptive text-guided diffusion OOD synthesis with adversarial training	(Mirzaei et al., 28 Jan 2025)
Nonparametric Statistics	Coordinatewise derivative tests for adaptive bandwidth	(Nguyen et al., 2021, Nguyen, 2018)
Probabilistic ODE Solvers	Kalman-filtered Gaussian process state-space modeling	(Wu et al., 26 Jun 2025)
Online Object Detection	Feature replay via product quantization buffer	(Acharya et al., 2020)
Medical Image Reconstruction	Single-layer autoencoder with robust loss and Bregman splitting	(Mehta et al., 2019)

7. Outlook and Limitations

Across all domains, the “Rodeo” algorithm family distinguishes itself by seeking sharp adaptivity and resource-efficient suppression or selection—whether it is in projection onto eigensolutions, robust boundary definition, or sparsity-adaptive estimation.

Key constraints include:

In quantum applications, performance hinges on initial state overlap and efficacy of system-specific controlled-evolution decompositions.
In adversarial outlier detection, computational bottlenecks arise from large-scale synthetic sample generation, and reliance on class semantics may restrict generality.
In statistical settings, theoretical guarantees depend on local sparsity and smoothness—if these break down, minimax rates may not be attainable.

Further advances are expected in integration of hybrid algorithms (e.g., fusion methods in quantum RA), hardware-oriented stochastic/deterministic protocol optimizations, nontrivial high-dimensional sparsity detection, and more general data-centric adversarial paradigms. For comprehensive details and proofs, see (Choi et al., 2020, Bonitati, 2024, Rocha et al., 2023, Patkowski et al., 5 Feb 2026, Mirzaei et al., 28 Jan 2025).