Quantum Supremacy Paradox: Why Simulated Annealing Outperforms Quantum Annealing in Detecting Bitcoin Mixers

01.11.2026

In 2024, researchers from National Taiwan University published a paper that, at first glance, promised to confirm the power of quantum computing in blockchain analysis. However, the results were unexpected: the classical Simulated Annealing (SA) algorithm not only outperformed quantum annealing in classifying Bitcoin mixers, but also did so approximately 1,200 times faster at the feature selection stage. This paper attempts to understand this paradox, which has important implications for the entire field of quantum technologies. We demonstrate that “quantum-inspired” is not synonymous with “quantum-superior,” and that honest comparison with classical methods remains the primary tool of scientific hygiene.

Keywords:  quantum annealing, simulated annealing, QUBO, Bitcoin mixers, feature selection, quantum-inspired algorithms, fair comparison, blockchain data classification.

The questions this article answers (selected from the research questions framework):

  1. Why did quantum annealing lose to classical simulated annealing in detecting Bitcoin mixers?
  2. What testable hypotheses explain this result?
  3. What does this mean for assessing quantum supremacy in real-world applications?
  4. What experiments need to be done to understand when quantum methods really win?

2. Introduction: A Paradox Worth Exploring

⚙️ Detailed research option

Imagine you’ve built a quantum computer capable of solving optimization problems that classical machines can’t. You apply it to a real-world problem—identifying Bitcoin addresses used for money laundering. And you discover that the trusty simulated annealing algorithm, invented in 1983, performs better, faster, and more accurately.

This is precisely what happened in the work of Ming-Fong Hsieh and colleagues from National Taiwan University [1]. Their study classifies Bitcoin addresses into six categories: exchanges, faucets, gambling, trading platforms, mixers, and mining pools. They particularly focus on mixers, services that combine the cryptocurrency of different users to make transactions more difficult to trace. Mixers are of interest to regulators, law enforcement agencies, and blockchain researchers, as they can be used for money laundering [2, 3].

The task is challenging: over 900 million Bitcoin transactions, 69 features per address, class imbalance (289,006 mixers versus 14,720 faucets), high dimensionality, and dynamic data evolution. The authors proposed using quantum-inspired feature selection (QIFS) based on the QUBO model to reduce the training time of machine learning models while maintaining accuracy.

They compared three methods:  Simulated Annealing (SA)  , a classical probabilistic algorithm;  Quantum Annealing (QA)  , a quantum annealing algorithm on real D-Wave hardware; and  QA BQM  , a hybrid D-Wave solver for binary quadratic model problems.

A paradox in one paragraph

⚙️ Detailed research option

Quantum annealing, which is supposed to exploit quantum effects to more efficiently find the global minimum, performed worse than the classical algorithm in terms of solution quality and was approximately 1,200 times slower at the feature selection stage. The quantum method selected fewer features, which sped up subsequent training, but at the cost of information loss and accuracy.

This paradox is not a curiosity. It raises fundamental questions about when quantum and quantum-inspired methods truly provide an advantage, and when classical algorithms remain unbeatable. In this article, we will examine the mechanism of the QUBO model, analyze experimental results, propose five hypotheses explaining the victory of SA, and formulate testable questions for future research.

3. What are Bitcoin mixers and why is it important to classify them?

⚙️ Detailed research option

Bitcoin, launched in 2009 by Satoshi Nakamoto [4], is a decentralized ledger that serves as a medium of exchange and a store of value [5]. However, the blockchain’s transparency creates a privacy issue: all transactions are publicly visible. Mixers solve this problem by combining the cryptocurrency of multiple users into target wallets, making it difficult to trace the source of funds [6].

On the one hand, mixers are a privacy tool. On the other, they can be used for money laundering, concealing illicit income [7]. For regulators and blockchain researchers, the ability to automatically identify mixers is an important task.

The authors of [1] used WalletExplorer.com to obtain a dataset of 694,676 labeled Bitcoin addresses associated with 3,459,773 transactions. The distribution by class is presented in Table 1.

Table 1. Distribution of addresses and transactions by category

CategoryAddressesInteracting addressesInteracting transactions
Exchange (stock exchanges)163,537675 6942,582,337
Faucet14,72015,35987 245
Gambling78,698107,661170,815
Market (trading platforms)102 12957 198181,337
Mixer (mixers)289,006166,536234,449
Pool (mining pools)46,58626 127203,590
Total694 6761,048,5753,459,773

The class imbalance is clear: mixers make up 41.6% of the dataset, while faucets make up only 2.1%. This creates challenges for training models, which we’ll discuss in Section 9.

4. Mechanism: How quantum-inspired feature selection works

⚙️ Detailed research option

4.1. QUBO model: balance of influence and dependence

⚙️ Detailed research option

The feature selection problem is formulated as an unconstrained quadratic binary optimization (QUBO) problem. The idea, adapted from Milne et al. [10], is to select a subset of K features that maximize their impact on the target variable (the “mixer” class) and minimize the interdependence (correlation) between the features.

Let x j  ∈ {0,1} be a binary variable indicating whether feature j is selected. The objective function is:

f(x) = −α Σ j=1 n  x j  |ρ oj | + (1−α) Σ j=1 n  Σ k≠j  x j x k  |ρ jk |

where ρ oj  is the Spearman correlation coefficient between feature j and the target variable, ρ jk  is the correlation between features j and k, and α ∈ [0,1] is the weighting coefficient.

The first term, with a minus sign, maximizes the influence of features on the class. The second term minimizes the dependence between the selected features. The parameter α controls the trade-off between these two objectives.

The task comes down to finding:

x* = arg min X  [−X T  QX]

where Q is a matrix encoding the factors of influence and independence.

Why is this a “quantum-inspired” approach?  The QUBO model naturally maps onto the Ising model—a system of interacting spins. The energy minimization problem for the Ising model can be solved using both classical (Simulated Annealing) and quantum (Quantum Annealing) methods. The term “quantum-inspired” means that the algorithm uses ideas from quantum mechanics (tunneling, superposition) but can be executed on classical hardware.

4.2. Three optimization methods

⚙️ Detailed research option

Simulated Annealing (SA)  is a classical probabilistic method proposed by Kirkpatrick and colleagues in 1983 [11]. It simulates the process of slow cooling of a material: at high temperatures, the system can transition to states with higher energy, which allows avoiding local minima; with a gradual decrease in temperature, the probability of such transitions decreases. SA is implemented in the library  dwave-neal.

Quantum Annealing (QA)  is a method that utilizes D-Wave quantum hardware to directly solve QUBO problems. Quantum annealing utilizes tunneling to penetrate energy barriers, which theoretically can lead to faster global minimum detection. The experiment utilized a D-Wave Advantage 6.4 system with 5,612 working qubits.

QA BQM  is a hybrid D-Wave solver (version 2.2) for general binary quadratic model problems. It combines quantum and classical components, decomposing the problem into parts that can be solved by quantum hardware and parts that can be processed classically.

5. Experimental setup

⚙️ Detailed research option

5.1. Data and features

⚙️ Detailed research option

The authors deployed a full Bitcoin node with 64 GB of RAM, a 12th-generation Intel Core i9-12900K processor, a 6 GB GeForce GTX 1060 GPU, and a 4 TB SSD. Obtaining the full transaction history took four months and covered 544,283 blocks and 345,882,038 transactions. Due to computational limitations, the first 1,000 transactions for each address were analyzed.

For each address, 69 features were calculated, grouped into five categories: basic statistics, advanced statistics, distribution moments, and transaction patterns. Examples of features include address lifetime, the ratio of received to sent transactions, average and standard deviations of BTC and USD balances, and the frequency of certain transaction types.

5.2. Metrics and Validation

⚙️ Detailed research option

Ten-fold cross-validation was used. Precision, recall, F1-score, accuracy, AUC, and training time were evaluated. Seven classifiers were compared: logistic regression, AdaBoost-SAMME, Random Forest, XGBoost, LightGBM, SVM, and a neural network.

6. Results: A Fair Comparison

⚙️ Detailed research option

6.1. Feature selection time

⚙️ Detailed research option

Table 2. Feature selection time for the “mixer” class

MethodNumber of selected featuresTime
Simulated Annealing (SA)230.0055 s
Quantum Annealing (QA)96.644 s
QA BQM76.0041 s

SA turned out to be approximately 1200 times faster than QA and 1090 times faster than BQM. Moreover, SA selected more features, which provided the model with more information for classification.

Key result:  SA selected 23 features in 0.0055 seconds—faster than the time required for quantum hardware to initialize and calibrate. Quantum methods selected fewer features (9 and 7), which reduced subsequent training time, but at the cost of information loss.

6.2. Quality of classification

⚙️ Detailed research option

Table 3. Results of Random Forest and its variants with feature selection for the “mixer” class

MethodPrecisionRecallF1AccuracyAUCTraining time
Random Forest (all 69 features)0.870.970.920.950.993305 min
SA Random Forest (23 features)0.850.970.910.940.992212 min
QA Random Forest (9 features)0.580.620.600.860.88125 min
QA BQM Random Forest (7 features)0.650.760.750.900.93183 min

The main conclusion:  SA Random Forest reduced training time by 30.3% (from 305 to 212 minutes) compared to the full feature set, while maintaining the F1 score at 91% (versus 92% for the full feature set). QA Random Forest and QA BQM Random Forest performed significantly worse in both F1 score and AUC.

6.3. Detailed analysis by classes

⚙️ Detailed research option

Random Forest with a full feature set showed strong results for the Exchange (F1 = 0.92), Market (F1 = 0.85), and Mixer (F1 = 0.91) classes. The Faucet (F1 = 0.26) and Gambling (F1 = 0.52) classes remained problematic due to class imbalance and heterogeneity in address behavior.

SA Random Forest maintained high performance for Exchange, Market, and Mixer, while slightly reducing recall for Pool (from 0.69 to 0.63). QA Random Forest showed significantly lower results for all classes, especially for Faucet (precision = 0.08, recall = 0.08) and Gambling (F1 = 0.31).

6.4 Importance of Features

⚙️ Detailed research option

An information gain analysis showed that Random Forest with a full feature set relies primarily on balance sheet characteristics (USD and BTC), transaction volumes, and distribution moments. SA Random Forest retained its emphasis on balance sheet characteristics but added the lifetime feature (the length of time an address is active), which proved important for understanding user behavior.

QA Random Forest and QA BQM Random Forest, by selecting fewer features, missed some important information, which affected the classification quality.

7. Why SA won: five hypotheses

⚙️ Detailed research option

The victory of classical SA over quantum QA requires an explanation. We propose five hypotheses that could be tested in future experiments.

Hypothesis 1: The problem size is too small for a quantum advantage

⚙️ Detailed research option

The QUBO problem contains 69 binary variables. For modern quantum annealers, this is a relatively small problem. Quantum advantage typically manifests itself on problems with thousands or millions of variables, where classical methods encounter combinatorial explosion. On small problems, classical heuristics are very effective.

Question to be tested:  At what size of a QUBO task does QA begin to outperform SA in terms of solution quality and time?

Hypothesis 2: Noise and errors in quantum hardware

⚙️ Detailed research option

Modern quantum annealers are susceptible to noise, decoherence, and calibration errors. The D-Wave Advantage 6.4 with 5,612 qubits is an impressive system, but its qubits are not perfect. These errors can cause the quantum annealer to find a suboptimal rather than an optimal solution.

Question to be tested:  How do the error rates of quantum hardware and the quality of the QUBO problem solution relate to different noise levels?

Hypothesis 3: The QUBO formulation is not optimal for QA

⚙️ Detailed research option

The QUBO model adapted from Milne et al. may not be perfectly balanced for quantum annealing. The parameter α, the correlation weights, and the structure of the Q matrix may create an energy landscape that classical SA explores more efficiently than quantum annealing.

Question to be tested:  How do different formulations of QUBO (different α, different correlation measures) affect the relative advantage of SA and QA?

Hypothesis 4: Quantum Access Overhead

⚙️ Detailed research option

The QA runtime (6.644 s) includes not only the quantum processor’s execution time but also overhead costs: task compilation, hardware access queuing, calibration, and result reading. SA runs locally and does not incur these costs.

Question to be tested:  What proportion of QA time is spent on quantum operations, and what proportion is spent on classical overhead?

Hypothesis 5: SA is better at exploring landscapes with multiple local minima.

⚙️ Detailed research option

A feature selection problem with a correlation matrix can have a complex landscape with multiple local minima. SA with gradual cooling and probabilistic transitions can explore this landscape more efficiently than quantum annealing, which in practice often operates as a stochastic classical algorithm due to hardware limitations.

Question to be tested:  How do the distributions of the found SA and QA solutions compare across multiple runs?

8. What does this mean for the development of quantum technologies?

⚙️ Detailed research option

8.1. Quantum-inspired ≠ quantum-transcending

⚙️ Detailed research option

The term “quantum-inspired” is often perceived as a promise of superiority. The work of Sie et al. shows that this is not the case. Quantum-inspired algorithms can be useful, but their advantage is not guaranteed and depends on the problem, its size, hardware, and formulation.

A warning for researchers:  Don’t confuse the use of quantum ideas (the Ising model, QUBO, tunneling) with quantum advantage. Classical quantum annealing, inspired by quantum mechanics, can perform better than real quantum annealing on modern hardware.

8.2. Honest comparison is the main tool

⚙️ Detailed research option

The work by Sie et al. is valuable precisely because the authors conducted a fair comparison of SA, QA, and QA BQM on a single problem with the same metrics. Many studies of quantum algorithms are limited to theoretical estimates or simulations. This study, however, used real D-Wave quantum hardware, and the results reveal the weaknesses of the quantum approach.

8.3. Practical value of SA for blockchain analysis

⚙️ Detailed research option

For practical Bitcoin address classification tasks, SA Random Forest with 23 features provides the best balance between quality and training time. This is an important result for the industry: it’s not always necessary to wait for quantum supremacy to obtain an effective solution.

8.4 Relationship with cryptography and security

⚙️ Detailed research option

Although the work focuses on classification rather than cryptanalysis, it demonstrates a general principle: quantum technologies develop unevenly. While quantum factorization (Shor) and discrete logarithm algorithms pose a long-term threat to cryptography, quantum optimization methods on modern hardware are often inferior to classical ones. This is important to consider when assessing quantum threats and planning the migration to post-quantum cryptography.

9. Limitations and Uncertainty

⚙️ Detailed research option

9.1 Class imbalance

⚙️ Detailed research option

The dataset is highly imbalanced: 289,006 mixers versus 14,720 faucets. This results in low performance for minority classes (Faucet F1 = 0.26). The authors plan to use SMOTE (Synthetic Minority Oversampling Technique) [15] to improve the recall of minority classes.

9.2. Limit of 1,000 transactions per address

⚙️ Detailed research option

Due to computational limitations, only the first 1,000 transactions for each address were analyzed. This may miss important information about the behavior of addresses with a long history.

9.3. Feature selection only for the “mixer” class

⚙️ Detailed research option

QIFS was optimized specifically for the “mixer” class. Applying the same features to other classes may not be optimal. Each class may require its own set of features.

9.4 Limitations of quantum equipment

⚙️ Detailed research option

The D-Wave Advantage 6.4 is a modern, but not perfect, system. QA results could improve with the introduction of more advanced quantum annealers with lower noise and better connectivity.

10. Open questions and next experiments

⚙️ Detailed research option

Based on the analysis of the work of Sie et al. and the framework of the research questions, we formulate the following test questions:

  1. Problem size:  At what number of features (69, 200, 1000, 5000) does QA start to outperform SA in F1-measure with a fixed selection time?
  2. Noise Level:  How does the quality of a QA solution depend on the noise level of quantum equipment, and is there a threshold below which QA becomes competitive?
  3. QUBO Formulation:  How do different values ​​of α and correlation measures (Spearman, Pearson, mutual information) affect the relative advantage of SA and QA?
  4. Hybrid approaches:  Can a combination of SA and QA (e.g. SA for pre-screening, QA for refinement) produce better quality than either method alone?
  5. Transfer to other classes:  Does the advantage of SA over QA remain when selecting features for other classes (Exchange, Gambling, Market)?
  6. Time Series:  How to account for the dynamic evolution of the blockchain—new transaction types, changing address behavior—in the feature selection process?
  7. Post-quantum perspective:  If quantum annealers become powerful enough to solve QUBO problems with thousands of variables, will the balance of power between SA and QA change?

11. Conclusion: A lesson for researchers

⚙️ Detailed research option

The work by Sie et al. is an important contribution to the field of quantum-inspired methods and blockchain analysis. Its main lesson is not that “quantum is better” or “classical is better,” but that honest comparison and understanding of the limits of applicability are essential for scientific progress.

SA Random Forest with 23 features achieves a 91% F1 score for mixers with a 30% reduction in training time. This is a practical result that can be used today. QA and BQM QA are inferior on modern hardware, but their potential is not exhausted.

For readers interested in quantum technologies, this paper serves as a reminder: quantum supremacy isn’t universal. It depends on the problem, its size, hardware, and its formulation. And sometimes, good old Simulated Annealing turns out to be the best quantum algorithm—because it’s inspired by quantum mechanics but isn’t constrained by its hardware limitations.

Further reading:  We recommend reading the original work by Sie et al. [1], as well as the research question framework used to structure this analysis. For understanding the QUBO model, the work by Milne et al. [10] is useful, and for the context of Bitcoin address classification, the work by Toyoda et al. [13] and Lin et al. [14] is useful.

12. Sources

⚙️ Detailed research option

  1. Sie, M.-F., Chang, Y.-J., Lin, C.-L., Chang, C.-R., Liao, S.-W. Efficient Bitcoin Address Classification Using Quantum-Inspired Feature Selection. National Taiwan University, 2024.
  2. Wu, L., Hu, Y., Zhou, Y., Wang, H., Luo, X., Wang, Z., Zhang, F., Ren, K. Towards understanding and demystifying bitcoin mixing services. Proceedings of the Web Conference 2021, pp. 33–44 (2021).
  3. Rathore, M.M., Chaurasia, S., Shukla, D. Mixers detection in bitcoin network: a step towards detecting money laundering in crypto-currencies. IEEE International Conference on Big Data (Big Data), pp. 5775–5782 (2022).
  4. Nakamoto, S. Bitcoin: A peer-to-peer electronic cash system. Decentralized business review (2008).
  5. Baur, DG, Dimpfl, T. The volatility of bitcoin and its role as a medium of exchange and a store of value. Empirical Economics 61(5), 2663–2683 (2021).
  6. Liu, M., Chen, H., Yan, J. Detecting roles of money laundering in bitcoin mixing transactions: A goal modeling and mining framework. Frontiers in Physics 9, 665399 (2021).
  7. Ashfaq, T., Khalid, R., Yahaya, AS, Aslam, S., Azar, AT, Alsafari, S., Hameed, IA A machine learning and blockchain based efficient fraud detection mechanism. Sensors 22(19), 7162 (2022).
  8. Chang, Y.-J., Sie, M.-F., Liao, S.-W., Chang, C.-R. The prospects of quantum computing for quantitative finance and beyond. IEEE Nanotechnology Magazine 17(2), 31–37 (2023).
  9. Chang, Y.-J., Nien, C.-F., Huang, K.-P., Zhang, Y.-T., Cho, C.-H., Chang, C.-R. Quantum computing for optimization with ising machine. IEEE Nanotechnology Magazine 18(3), 15–22 (2024).
  10. Milne, A., Rounds, M., Goddard, P. Optimal Feature Selection in Credit Scoring and Classification Using a Quantum Annealer. 1QB Information Technologies (2017).
  11. Kirkpatrick, S., Gelatt Jr, CD, Vecchi, MP Optimization by simulated annealing. Science 220(4598), 671–680 (1983).
  12. Breiman, L. Random forests. Machine Learning 45, 5–32 (2001).
  13. Toyoda, K., Ohtsuki, T., Mathiopoulos, PT Multi-class bitcoin-enabled service identification based on transaction history summarization. IEEE iThings/GreenCom/CPSCom/SmartData, pp. 1153–1160 (2018).
  14. Lin, Y.-J., Wu, P.-W., Hsu, C.-H., Tu, I.-P., Liao, S.-W. An evaluation of bitcoin address classification based on transaction history summarization. IEEE ICBC, pp. 302–310 (2019).
  15. Chawla, N.V., Bowyer, K.W., Hall, L.O., Kegelmeyer, W.P. SMOTE: synthetic minority over-sampling technique. Journal of Artificial Intelligence Research 16, 321–357 (2002).