Quantum-Inspired Tensor Network Autoencoders for Anomaly Detection: A MERA-Based Approach
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: High Energy Physics: Phenomenology. The day in particle phenomenology: models, signatures and constraints.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "Quantum-Inspired Tensor Network Autoencoders for Anomaly Detection".
Mira: Here's the gist of the paper. Quantum-Inspired Tensor Network Autoencoders for Anomaly Detection: A MERA-Based Approach investigates whether a multiscale tensor-network architecture can provide a useful inductive bias for reconstruction-based anomaly detection in collider jets, which is motivated by the natural organization of jet structure across angular and momentum scales The resulting picture is that the locality-preserving multiscale structure is well matched to jet data, and that the MERA disentanglers become beneficial precisely when the compression bottleneck is strongest Overall, the study supports locality-aware hierarchical compression as a useful inductive bias for jet anomaly detection.
Kai: Pruning redundant reasoning tokens The paper investigates whether a multiscale tensor-network architecture can provide a useful inductive bias for reconstruction-based anomaly detection in collider jets Jets are produced by a branching cascade, so their internal structure is naturally organised across angular and momentum scales This motivates an autoencoder that compresses information hierarchically and can reorganise short-range correlations before coarse-graining The paper compares this architecture to a dense autoencoder, the corresponding tree-tensornetwork limit, and standard classical baselines within a common background-only reconstruction framework.
Mira: How the training signal is built The anomaly detection framework involves learning what typical background events look like and then assigning a large anomaly score to events that are poorly described by that learned background model In reconstruction-based methods, a model is trained only on background events to compress and reconstruct its input The training objective for the MERA autoencoder is defined as arg min θ,ϑ 1/Btrain X x∈Btrain Lrec(x, xˆ) + λURU + λW RW Once trained on background jets only, the anomaly score is the reconstruction loss itself S(x) = Lrec(x, Dϑ⋆ (Eθ⋆ (x)).
Kai: Where the gains come from The study tests whether locality-aware hierarchical compression is genuinely supported by the data and whether MERA disentanglers contribute beyond a simpler tree hierarchy The nominal benchmark shows that MERA attains the largest test AUC at all three latent dimensions, reaching 0.7959, 0.7885, and 0.7963 at B = 8, 16, and 32 This performance is achieved with only about one third of the trainable parameters of the corresponding dense autoencoder.
Title and authors: Kai: We've covered what this paper is all about, focusing on how this MERA approach uses tensor networks to detect anomalies in collider jets. Now let's look at the title and the authors for a moment to see who’s behind this work.
Mira: The title is "Quantum-Inspired Tensor Network Autoencoders for Anomaly Detection: A MERA-Based Approach". It immediately tells us that we are dealing with quantum inspiration, tensor networks, and MERA.
Lev: I know the authors are Emre Gurkanli and Michael Spannowsky from Sinop University and KIT in Karlsruhe. They’re researchers deeply involved in quantum materials and theoretical physics.
Kai: Their background is very relevant because MERA itself is a concept that has deep roots in how we think about hierarchical structures, which is exactly what jet data possesses.
Mira: So, when they frame it as quantum-inspired, it suggests they are looking for some kind of structure in the data that might be related to quantum phenomena or at least a highly structured mathematical framework.
Lev: From an engineering standpoint, we have to ask what was actually built and cooled and measured. Is this just a theoretical construct, or is there an actual implementation on hardware?
Kai: Right now it seems like the focus is on the theoretical formulation and benchmarking against established baselines, but they are clearly aiming for something that could translate into practical AI systems.
Mira: The paper mentions that anomaly detection provides one such framework, and in collider physics, this program has grown rapidly recently because of community benchmarks clarifying both the promise and limitations of model agnostic searches.
Lev: That means the architecture needs to be flexible enough to respond to a wide class of signals while staying robust against ordinary fluctuations of the background. The structure matters a lot here.
Kai: So, as we look at this MERA-Based Approach, it’s about using that built-in structure to guide the learning process in a way that might be more efficient than just throwing a dense autoencoder at the problem.
The paper's summary: Mira: Now let's recap what they are summarizing in "Quantum-Inspired Tensor Network Autoencoders for Anomaly Detection: A MERA-Based Approach". They’re essentially saying they investigate whether this multiscale tensor-network architecture can provide a useful inductive bias for reconstruction-based anomaly detection in collider jets.
Kai: The key idea here is that because jets are produced by a branching cascade, their internal structure is naturally organized across angular and momentum scales. This naturally motivates an autoencoder that compresses information hierarchically and reorganizes those short-range correlations before coarse-graining.
Lev: So they formulate this MERA-inspired autoencoder acting directly on ordered jet constituents to exploit that natural organization. It’s not just a random compression scheme, it’s guided by the physics of the data generation process.
Mira: And they compare it against a dense autoencoder and the tree tensor network limit to see how much better this structure is for reconstruction-based anomaly detection in background-only scenarios.
Kai: The central picture emerging is that this locality-preserving multiscale structure is well matched to jet data, and that the MERA disentanglers become beneficial precisely when the compression bottleneck is strongest.
Lev: That means the model needs to be flexible enough to handle both fine details and broad features, but it gets extra help when it has to make big compromises in the middle of that process.
The paper's improvements: Mira: So what are the specific improvements they suggest? They are talking about how this approach can be made better than simpler models.
Kai: The paper suggests that the main improvement comes from explicitly focusing on local compression. The role of the disentangler is to seek a basis where those degrees of freedom that must be retained by the subsequent isometry are as informative as possible.
Lev: This means reorganizing short-distance angular correlations before forming a coarse description, which is exactly what we want when short distances need reorganization before a coarse description forms.
Mira: They showed that this locality-preserving ordering reduces the mean adjacent site separation to point zero one three one, compared to point zero three six one for pT ordering and point zero three nine seven plus or minus zero zero four for random permutations.
Kai: And they found that this same ordering yields the highest retained variance in the first two compression layers at point eight six nine and point nine two0, which is better than what you see with pT ordering or random control.
Lev: That means we’re getting a more robust representation of the local structure when we compress it down to those early layers. It’s about making sure the local details are preserved during that initial compression step.
Conclusion: Mira: So, to wrap up this paper, they really established that a MERA-based autoencoder is a viable and effective architecture for collider anomaly detection on ordered jet constituents. They also proved that locality-aware preprocessing isn't just some cosmetic choice but it’s substantive part of why the multiscale model works.
Kai: And crucially, they showed that in the regime where the hierarchy is forced to compress most aggressively, trainable disentanglers give a measurable improvement over the corresponding identity-disentangler limit.
Lev: This means we've established that this whole approach is sound for jet anomaly detection and that locality-aware hierarchical compression is a useful inductive bias.
Mira: It’s a concrete empirical case showing that the multiscale construction can outperform the standard dense autoencoder while doing so with a far smaller parameter budget.
Kai: We’re looking at how this paper suggests that AI can perform reconstruction-based anomaly detection in collider jets by utilizing a multiscale tensor-network architecture inspired by MERA, and it seems to be "stronger than the dense autoencoder in AUC" while using "about one third of the trainable parameters."
Mira: And when you look at the strongest-compression regime, incorporating trainable disentanglers into the MERA architecture improves performance over the corresponding identity-disentangler limit by reaching a mean test AUC gain of point zero zero eight one.
Kai: That's all we have on this paper for now. It really shows that even in complex physics, if you use a multiscale, locality-aware approach with the right architecture, you can get some real performance gains.
Emre Gurkanli, Michael Spannowsky
Department of Physics, Sinop University · Institute for Theoretical Physics, Campus Sud, Karlsruhe Institute of Technology (KIT) · Institute for Quantum Materials and Technologies, Karlsruhe Institute of Technology
hep-ph, cs.LG, quant-ph
Submitted: 2026-04-08
Updated: 2026-11-04
Comments: 29 pages, 5 figures
DOI: 10.1140/epjc/s10052-026-16489-z
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 80/100
The gist: Quantum-Inspired Tensor Network Autoencoders for Anomaly Detection: A MERA-Based Approach investigates whether a multiscale tensor-network architecture can provide a useful inductive bias for…
Terminology
Summary
Quantum-Inspired Tensor Network Autoencoders for Anomaly Detection: A MERA-Based Approach investigates whether a multiscale tensor-network architecture can provide a useful inductive bias for reconstruction-based anomaly detection in collider jets, which is motivated by the natural organization of jet structure across angular and momentum scales The resulting picture is that the locality-preserving multiscale structure is well matched to jet data, and that the MERA disentanglers become beneficial precisely when the compression bottleneck is strongest Overall, the study supports locality-aware hierarchical compression as a useful inductive bias for jet anomaly detection. p. 0 The resulting picture is that the
locality-preserving multiscale structure is well matched to jet data, and that the MERA
disentanglers become beneficial precisely when the compression bottleneck is strongest. Overall, the study supports locality-aware hierarchical compression as a useful inductive
bias for jet anomaly detection.Where the paper says it
Pruning redundant reasoning tokens
The paper investigates whether a multiscale tensor-network architecture can provide a useful inductive bias for reconstruction-based anomaly detection in collider jets Jets are produced by a branching cascade, so their internal structure is naturally organised across angular and momentum scales This motivates an autoencoder that compresses information hierarchically and can reorganise short-range correlations before coarse-graining The paper compares this architecture to a dense autoencoder, the corresponding tree-tensornetwork limit, and standard classical baselines within a common background-only reconstruction framework.
How the training signal is built
The anomaly detection framework involves learning what typical background events look like and then assigning a large anomaly score to events that are poorly described by that learned background model In reconstruction-based methods, a model is trained only on background events to compress and reconstruct its input The training objective for the MERA autoencoder is defined as arg min θ,ϑ 1/Btrain X x∈Btrain Lrec(x, xˆ) + λURU + λW RW Once trained on background jets only, the anomaly score is the reconstruction loss itself S(x) = Lrec(x, Dϑ⋆ (Eθ⋆ (x)).
Where the gains come from
The study tests whether locality-aware hierarchical compression is genuinely supported by the data and whether MERA disentanglers contribute beyond a simpler tree hierarchy The nominal benchmark shows that MERA attains the largest test AUC at all three latent dimensions, reaching 0.7959, 0.7885, and 0.7963 at B = 8, 16, and 32 This performance is achieved with only about one third of the trainable parameters of the corresponding dense autoencoder. p. 3 Thus, we turn that motivation into two concrete
tests: whether a locality-preserving constituent chain is measurably more compressible at
the local tensor-network level, and whether trainable MERA disentanglers improve on the
corresponding identity-disentangler limit. Our aim is therefore not only to test whether a MERA-inspired autoencoder can identify anomalous jets, but also to examine the physical argument behind its use: namely, that
an explicitly hierarchical architecture should offer a better inductive bias for jet data than
dense or purely linear baselines. Geometry and site ordering. MERA assumes a graph on which locality is defined,
whereas a jet is fundamentally an unordered set of constituents. 2 Jet Representation 11
3. 3 Locality-Preserving Ordering and Standardisation 12
3. To address these
questions, we combine benchmark comparisons with a training-free local-compressibility
diagnostic and a direct identity-disentangler ablation. The resulting picture is that the
locality-preserving multiscale structure is well matched to jet data, and that the MERA
disentanglers become beneficial precisely when the compression bottleneck is strongest. Quantum-Inspired Tensor Network Autoencoders for
Anomaly Detection: A MERA-Based Approach
Emre Gurkanli∗ a and Michael Spannowsky∗ b,c
aDepartment of Physics, Sinop University, Turkiye
b
Institute for Theoretical Physics, Campus S¨ud, Karlsruhe Institute of Technology (KIT), D-76128
Karlsruhe, Germany
c
Institute for Quantum Materials and Technologies, Karlsruhe Institute of Technology, Karlsruhe
76131, Germany
E-mail: egurkanli@sinop. tr, michael.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Testing the locality argument
The locality-preserving ordering is shown to yield the highest mean test ROCAUC at all three latent dimensions, with values of 0.780, 0.777, and 0.774 at B = 8, 16, and 32 This is consistent with the geometric and compressibility diagnostics reported in Table 2 where the locality-preserving ordering reduces the mean adjacent-site separation to ⟨∆R⟩ = 0.131 The disentangler study confirms that in the strongest-compression regime, trainable disentanglers improve performance over the corresponding identity-disentangler limit MERA(U = I). p. 3 The public Top Quark Tagging Reference Dataset, introduced as part of a broad
comparison of modern jet-tagging methods, provides a clean and widely used benchmark
for testing this idea in a setting [22]. Thus, we turn that motivation into two concrete
tests: whether a locality-preserving constituent chain is measurably more compressible at
the local tensor-network level, and whether trainable MERA disentanglers improve on the
corresponding identity-disentangler limit. If this ordering fails to place strongly correlated
constituents near each other, then the local disentanglers and isometries will act on the
wrong degrees of freedom and the MERA inductive bias is weakened. This is one of the
central modelling choices in collider applications, and it is why constituent ordering is not
a cosmetic preprocessing step but part of the architecture itself.Where the paper says it
Where the paper says it
Examining architectural advantages
The MERA architecture is motivated by the idea that local basis changes should occur before coarse-graining to retain relevant information The paper shows that the MERA gains are tied to the locality-aligned representation and hierarchical coarse-graining more generally The results support the geometric premise of the construction, showing that adjacent input blocks are most geometrically local and compressible with this ordering. This confirms that the locality-preserving chain aligns the ordered jet representation with the local compression steps of the MERA hierarchy. p. 5 , 48
Preprocessing
locality-preserving ordering, background standardisation, clipping
ordered constituent chain
nearest neighbours in the chain are geometrically nearby in the jet
Encoder layer 1
disentanglers U1 reorganise local correlations, then isometries
W1 coarse-grain 48 → 24
Encoder layer 2
U2 then W2: 24 → 12
Encoder layer 3
U3 then W3: 12 → 6
Encoder layer 4
U4 then W4: 6 → 3
Latent bottleneck
projection to z ∈ R
B
Decoder layer 4
W˜ 4 then U˜4: 3 → 6
Decoder layer 3
W˜ 3 then U˜3: 6 → 12
Decoder layer 2
W˜ 2 then U˜2: 12 → 24
Decoder layer 1
W˜ 1 then U˜1: 24 → 48
Reconstructed constituents
xˆi in the same ordered representation
Anomaly score
s(x) = 1
48
X
48
i=1
kxi − xˆik
2
Figure 1: Schematic overview of the MERA-based autoencoder used in this work. After
locality-preserving ordering and standardisation, the encoder alternates disentanglers Uℓ
and isometries Wℓ through the hierarchy 48 → 24 → 12 → 6 → 3, followed by a projection
to a latent vector z ∈ R
B. The public Top Quark Tagging Reference Dataset, introduced as part of a broad
comparison of modern jet-tagging methods, provides a clean and widely used benchmark
for testing this idea in a setting [22]. Thus, we turn that motivation into two concrete
tests: whether a locality-preserving constituent chain is measurably more compressible at
the local tensor-network level, and whether trainable MERA disentanglers improve on the
corresponding identity-disentangler limit.Where the paper says it
Where the paper says it
Mathematical challenges and constraints
Designing a MERA for anomaly detection is mathematically non-trivial due to several challenges These include specifying a map π: 1,..., N → 1,..., N that turns the set into an ordered sequence The irregular system size of jet representations requires a decision on padding or mixed-arity tensors near the top of the hierarchy Constrained optimization is difficult because parameter spaces are not Euclidean, requiring soft penalties or explicit retractions onto the constraint manifold Finally, once coarse-graining occurs, exact inversion is impossible because W⊤W!= Iχ2l.
Model comparisons and baselines
The study compares the MERA architecture against a dense autoencoder and classical reference baselines including Principal Component Analysis (PCA), Gaussian modelling, and Isolation Forest The nominal benchmark shows that MERA outperforms the dense autoencoder in AUC while using only about one third of the trainable parameters PCA becomes competitive only once the retained dimension is large enough to capture a sizeable fraction of the global background variance The Gaussian Background Model tests whether the standardized background can already be described adequately by a single ellipsoidal distribution in feature space Isolation Forest provides a conceptually different unsupervised baseline based on recursive random partitioning rather than reconstruction or density estimation. The training-free local-compressibility diagnostic is used to test the locality argument directly against pT ordering and random permutations. p. 0 To address these
questions, we combine benchmark comparisons with a training-free local-compressibility
diagnostic and a direct identity-disentangler ablation. The resulting picture is that the
locality-preserving multiscale structure is well matched to jet data, and that the MERA
disentanglers become beneficial precisely when the compression bottleneck is strongest.Where the paper says it
Conclusion
The study establishes that a MERA-based autoencoder is a viable and effective architecture for collider anomaly detection on ordered jet constituents Locality-aware preprocessing is not a cosmetic choice but a substantive part of why the multiscale model works Trainable disentanglers give a measurable improvement over the corresponding nodisentangler limit when the hierarchy is forced to compress most aggressively. Taken together, these results make our MERA-based tensor-network implementation a concrete empirical case that multiscale, locality-aware compression is a useful inductive bias for jet anomaly detection. p. 0 The resulting picture is that the
locality-preserving multiscale structure is well matched to jet data, and that the MERA
disentanglers become beneficial precisely when the compression bottleneck is strongest. Overall, the study supports locality-aware hierarchical compression as a useful inductive
bias for jet anomaly detection. We compare this architecture to a dense autoencoder, the corresponding tree-tensornetwork limit, and standard classical baselines within a common background-only reconstruction framework. The paper is organised around two main questions: whether localityaware hierarchical compression is genuinely supported by the data, and whether the disentangling layers of MERA contribute beyond a simpler tree hierarchy.Where the paper says it
Where the paper says it
References
[1] V. Belis, P. Odagiu, and T. K. Aarrestad, “Machine learning for anomaly detection in particle physics,” Rev. p. 4 In many-body physics they are used to represent states with constrained correlation structure; in machine learning the same idea has been adapted to
supervised classification, multiscale feature extraction, and generative modelling [16, 23–
29]. Collider applications of tensor-network-inspired methods have also begun to appear,
including matrix-product-state event reconstruction and quantum-probabilistic generative
or anomaly-detection frameworks [17–19]. ∗Corresponding author. arXiv:2604. Quantum-Inspired Tensor Network Autoencoders for
Anomaly Detection: A MERA-Based Approach
Emre Gurkanli∗ a and Michael Spannowsky∗ b,c
aDepartment of Physics, Sinop University, Turkiye
b
Institute for Theoretical Physics, Campus S¨ud, Karlsruhe Institute of Technology (KIT), D-76128
Karlsruhe, Germany
c
Institute for Quantum Materials and Technologies, Karlsruhe Institute of Technology, Karlsruhe
76131, Germany
E-mail: egurkanli@sinop. tr, michael.Where the paper says it
Where the paper says it
Where the paper says it
[2] G. Kasieczka et al., “The LHC Olympics 2020 a community challenge for anomaly detection in high energy physics,” Rept. p. 3 To the best of our knowledge, this is the first study to formulate collider anomaly
detection with an MERA-based autoencoder. Earlier anomaly-detection applications in
high-energy physics have considered quantum autoencoders and, more recently, tensornetwork models based on matrix product states acting in an externally learned latent space,
but not a MERA encoder–decoder operating directly on ordered jet constituents [20, 21]. ∗Corresponding author. arXiv:2604. arXiv:2604. 06541v1 [hep-ph] 8 Apr 2026Where the paper says it
Where the paper says it
Where the paper says it
[3] K. Fraser, S. Homiller, R. K. Mishra, B. Ostdiek, and M. D. Schwartz, “Challenges for unsupervised anomaly detection in particle physics,” JHEP 03 (2022) 066, [arXiv:2110. p. 1 3 Potential advantages of a MERA approach for anomaly detection and interpretability 8
2. 4 Mathematical Challenges 9
3 Benchmarking and Data Preprocessing 11
3. arXiv:2604. 06541v1 [hep-ph] 8 Apr 2026Where the paper says it
Where the paper says it
[4] T. Heimel, G. Kasieczka, T. Plehn, and J. M. Thompson, “QCD or What? p. 2 Hard, large-angle splittings
determine the coarse prong structure of the jet, while softer and more collinear radiation
refines that structure at progressively smaller scales. For anomaly detection, this is precisely
the type of organisation one would like a model to capture: signal jets may differ from QCD
not only through isolated features, but through how correlations persist or reorganise across
scales. ∗Corresponding author. arXiv:2604. arXiv:2604. 06541v1 [hep-ph] 8 Apr 2026Where the paper says it
Where the paper says it
Where the paper says it
[5] M. Farina, Y. Nakai, and D. Shih, “Searching for New Physics with Deep Autoencoders,” Phys. p. 0 Quantum-Inspired Tensor Network Autoencoders for
Anomaly Detection: A MERA-Based Approach
Emre Gurkanli∗ a and Michael Spannowsky∗ b,c
aDepartment of Physics, Sinop University, Turkiye
b
Institute for Theoretical Physics, Campus S¨ud, Karlsruhe Institute of Technology (KIT), D-76128
Karlsruhe, Germany
c
Institute for Quantum Materials and Technologies, Karlsruhe Institute of Technology, Karlsruhe
76131, Germany
E-mail: egurkanli@sinop. tr, michael. ∗Corresponding author. arXiv:2604. arXiv:2604. 06541v1 [hep-ph] 8 Apr 2026Where the paper says it
Where the paper says it
Where the paper says it
[6] J. H. Collins, P. Martin-Ramiro, B. Nachman, and D. Shih, “Comparing weak- and unsupervised methods for resonant anomaly detection,” Eur. p. 2 The reconstruction error is then used as the anomaly score, under the
expectation that atypical events will be reconstructed less accurately. Autoencoders have
therefore become a standard reference point in collider anomaly detection, including for
jet-based searches, resonant-search studies, and more specialized adversarial, normalized,
or graph-based constructions [4–11].Where the paper says it
[7] J. Barron, D. Curtin, G. Kasieczka, T. Plehn, and A. Spourdalakis, “Unsupervised hadronic SUEP at the LHC,” JHEP 12 (2021) 129
[8] A. Blance, M. Spannowsky, and P. p. 0 Quantum-Inspired Tensor Network Autoencoders for
Anomaly Detection: A MERA-Based Approach
Emre Gurkanli∗ a and Michael Spannowsky∗ b,c
aDepartment of Physics, Sinop University, Turkiye
b
Institute for Theoretical Physics, Campus S¨ud, Karlsruhe Institute of Technology (KIT), D-76128
Karlsruhe, Germany
c
Institute for Quantum Materials and Technologies, Karlsruhe Institute of Technology, Karlsruhe
76131, Germany
E-mail: egurkanli@sinop. tr, michael.Where the paper says it
[9] B. M. Dillon, L. Favaro, T. Plehn, P. Sorrenson, and M. Kr¨amer, “A normalized autoencoder for LHC triggers,” SciPost Phys. p. 1 3 Effect of the Disentanglers 21
6 Summary and Conclusions 23
1 Introduction
The Standard Model (SM) gives an excellent description of collider data, but it does not
explain several basic observations, including dark matter, neutrino masses, and the baryon
asymmetry of the Universe. The search for physics beyond the SM, therefore, remains a
central goal of the LHC programme.Where the paper says it
[10] O. Atkinson, A. Bhardwaj, C. Englert, P. Konar, V. S. Ngairangbam, and M. Spannowsky, “IRC-Safe Graph Autoencoder for Unsupervised Anomaly Detection,” Front. p. 0 Quantum-Inspired Tensor Network Autoencoders for
Anomaly Detection: A MERA-Based Approach
Emre Gurkanli∗ a and Michael Spannowsky∗ b,c
aDepartment of Physics, Sinop University, Turkiye
b
Institute for Theoretical Physics, Campus S¨ud, Karlsruhe Institute of Technology (KIT), D-76128
Karlsruhe, Germany
c
Institute for Quantum Materials and Technologies, Karlsruhe Institute of Technology, Karlsruhe
76131, Germany
E-mail: egurkanli@sinop. tr, michael.Where the paper says it
[11] T. Finke, M. Kr¨amer, A. Morandini, A. M¨uck, and I. Oleksiyuk, “Autoencoders for unsupervised anomaly detection in high energy physics,” JHEP 06 (2021) 161
[12] T. Finke, M. Hein, G. Kasieczka, M. Kr¨amer, A. M¨uck, P. Prangchaikul, T Quadfasel, D Shih, and M Sommerhalder, “Tree-based algorithms for weakly supervised anomaly detection,” Phys. p. 2 At the same time, recent studies have emphasized that
reconstruction quality alone is not enough: performance can depend strongly on the input
representation, the architecture, and the evaluation protocol [3, 11]. Related work has also
shown that non-neural or weakly supervised alternatives, such as tree-based strategies, can
be highly competitive in collider anomaly searches [12].Where the paper says it
[13] A. J. Larkoski, I. Moult, and B Nachman, “Jet substructure at the Large Hadron Collider: a review of recent advances in theory and machine learning,” Phys. p. 2 The basic idea is to learn what
typical background events look like and then assign a large anomaly score to events that
are poorly described by that learned background model [1]. In collider physics, this programme has grown rapidly in recent years, supported by community benchmarks and systematic comparisons that clarified both the promise and the limitations of model-agnostic
searches [2, 3].Where the paper says it
[14] G. Vidal, “Entanglement renormalization,” Phys. p. 2 Whether the additional
disentangling structure of MERA is genuinely useful beyond a simpler tree hierarchy is,
however, an empirical question rather than an assumption. We study a multiscale architecture inspired by the Multiscale Entanglement Renormalisation Ansatz (MERA) [14, 15].Where the paper says it
[15] G. Vidal, “Class of quantum many-body states that can be efficiently simulated,” Phys. p. 4 In many-body physics they are used to represent states with constrained correlation structure; in machine learning the same idea has been adapted to
supervised classification, multiscale feature extraction, and generative modelling [16, 23–
29]. Collider applications of tensor-network-inspired methods have also begun to appear,
including matrix-product-state event reconstruction and quantum-probabilistic generative
or anomaly-detection frameworks [17–19].Where the paper says it
[16] E. M. Stoudenmire and D. J. Schwab, “Supervised Learning with Tensor Networks,” Adv. p. 4 2 Constructing a MERA-Based Anomaly Detector
Tensor networks provide structured factorisations of high-order maps into products of
low-order tensors. In many-body physics they are used to represent states with constrained correlation structure; in machine learning the same idea has been adapted to
supervised classification, multiscale feature extraction, and generative modelling [16, 23–
29]. At the same time, recent studies have emphasized that
reconstruction quality alone is not enough: performance can depend strongly on the input
representation, the architecture, and the evaluation protocol [3, 11]. Related work has also
shown that non-neural or weakly supervised alternatives, such as tree-based strategies, can
be highly competitive in collider anomaly searches [12].Where the paper says it
Where the paper says it
[17] J. Y. Araz and M. Spannowsky, “Quantum-inspired event reconstruction with Tensor Networks: Matrix Product States,” JHEP 08 (2021) 112
[18] J. Y. Araz and M. Spannowsky, “Classical versus quantum: Comparing tensor-network-based quantum circuits on Large Hadron Collider data,” Phys. p. 3 Related tensor-network ideas have also been adapted to machine learning, where they offer
structured and often parameter-efficient representations of high-dimensional data [16]. In
collider physics, tensor-network-inspired approaches have already been explored for event
reconstruction, comparisons between classical and quantum tensor-network circuits, and
generative or anomaly-oriented modelling [17–19].Where the paper says it
[19] J. Y. Araz and M. Spannowsky, “Quantum-probabilistic Hamiltonian learning for generative modeling and anomaly detection,” Phys. p. 4 In many-body physics they are used to represent states with constrained correlation structure; in machine learning the same idea has been adapted to
supervised classification, multiscale feature extraction, and generative modelling [16, 23–
29]. Collider applications of tensor-network-inspired methods have also begun to appear,
including matrix-product-state event reconstruction and quantum-probabilistic generative
or anomaly-detection frameworks [17–19].Where the paper says it
[20] V. S. Ngairangbam, M. Spannowsky, and M Takeuchi, “Anomaly detection in high-energy physics using a quantum autoencoder,” Phys. p. 3 To the best of our knowledge, this is the first study to formulate collider anomaly
detection with an MERA-based autoencoder. Earlier anomaly-detection applications in
high-energy physics have considered quantum autoencoders and, more recently, tensornetwork models based on matrix product states acting in an externally learned latent space,
but not a MERA encoder–decoder operating directly on ordered jet constituents [20, 21].Where the paper says it
[21] E. Puljak, M. Pierini, and A Garcia-Saez, “Tensor Network for Anomaly Detection in the Latent Space of Proton Collision Events at the LHC,” Mach. p. 3 To the best of our knowledge, this is the first study to formulate collider anomaly
detection with an MERA-based autoencoder. Earlier anomaly-detection applications in
high-energy physics have considered quantum autoencoders and, more recently, tensornetwork models based on matrix product states acting in an externally learned latent space,
but not a MERA encoder–decoder operating directly on ordered jet constituents [20, 21].Where the paper says it
[22] G. Kasieczka et al., “The Machine Learning landscape of top taggers,” SciPost Phys. p. 3 physics as a tensor-network construction that alternates coarse-graining with disentangling
transformations, allowing relevant correlations to be retained across length scales. Related tensor-network ideas have also been adapted to machine learning, where they offer
structured and often parameter-efficient representations of high-dimensional data [16].Where the paper says it
[35] G. Kasieczka, T Plehn, J Thompson, and M Russel, “Top Quark Tagging Reference Dataset,” 3, 2019
[36] T. Sj¨ostrand et al., “An Introduction to PYTHIA 8. p. 1 Contents
1 Introduction 1
2 Constructing a MERA-Based Anomaly Detector 4
2. 1 From Ordered Jet Constituents to a MERA Encoder 4
2.Where the paper says it
[37] DELPHES 3 Collaboration, J de Favereau et al., “DELPHES 3, A modular framework for fast simulation of a generic collider experiment,” JHEP 02 (2014) 057
[38] M. Cacciari, G. P. Salam, and G Soyez, “The anti-kt jet clustering algorithm,” JHEP 04 (2008) 063
[39] G. E. Hinton and R. R. Salakhutdinov, “Reducing the Dimensionality of Data with Neural Networks,” Science 313 no. p. 2 At the same time, recent studies have emphasized that
reconstruction quality alone is not enough: performance can depend strongly on the input
representation, the architecture, and the evaluation protocol [3, 11]. Related work has also
shown that non-neural or weakly supervised alternatives, such as tree-based strategies, can
be highly competitive in collider anomaly searches [12].Where the paper says it
[40] K. Pearson, “LIII. On lines and planes of closest fit to systems of points in space,” Philos. p. 1 In practice, however, most searches are optimised
for specific benchmark signatures. This is a powerful strategy when the signal is known in
advance, but it can reduce sensitivity to unexpected phenomena that do not fit a standard
search template.Where the paper says it
[41] H. Hotelling, “Analysis of a complex of statistical variables into principal components,” J. p. 3 Our aim is therefore not only to test whether a MERA-inspired autoencoder can identify anomalous jets, but also to examine the physical argument behind its use: namely, that
an explicitly hierarchical architecture should offer a better inductive bias for jet data than
dense or purely linear baselines. To this end, we perform a comparison between a dense
autoencoder baseline, a MERA-based architecture, the corresponding tree tensor network
(TTN) limit, and classical unsupervised methods, including principal component analysis
(PCA), Gaussian modelling, and Isolation Forest.Where the paper says it
[42] P. C. Mahalanobis, “On the generalised distance in statistics,” Proc. p. 3 This is why we consider the MERA-type anomaly detection worth studying. If the
dominant background really lies on a multiscale manifold shaped by the QCD shower, then
a MERA-inspired autoencoder should be able to model that background economically: it
should retain the correlations that matter, discard redundant short-distance structure efficiently, and compress the event with fewer parameters than a dense network of comparable
reach.Where the paper says it
[43] F. T Liu, K M Ting, and Z.-H Zhou, “Isolation Forest,” in 2008 Eighth IEEE International Conference on Data Mining, pp. p. 3 Our aim is therefore not only to test whether a MERA-inspired autoencoder can identify anomalous jets, but also to examine the physical argument behind its use: namely, that
an explicitly hierarchical architecture should offer a better inductive bias for jet data than
dense or purely linear baselines. To this end, we perform a comparison between a dense
autoencoder baseline, a MERA-based architecture, the corresponding tree tensor network
(TTN) limit, and classical unsupervised methods, including principal component analysis
(PCA), Gaussian modelling, and Isolation Forest.Where the paper says it
[31] P.-A. Absil, R Mahony, and R Sepulchre, Optimization Algorithms on Matrix Manifolds. p. 3 To the best of our knowledge, this is the first study to formulate collider anomaly
detection with an MERA-based autoencoder. Earlier anomaly-detection applications in
high-energy physics have considered quantum autoencoders and, more recently, tensornetwork models based on matrix product states acting in an externally learned latent space,
but not a MERA encoder–decoder operating directly on ordered jet constituents [20, 21]. To the best of our knowledge, this is the first study to formulate collider anomaly
detection with an MERA-based autoencoder. Earlier anomaly-detection applications in
high-energy physics have considered quantum autoencoders and, more recently, tensornetwork models based on matrix product states acting in an externally learned latent space,
but not a MERA encoder–decoder operating directly on ordered jet constituents [20, 21].Where the paper says it
Where the paper says it
[32] L. De Lathauwer, B De Moor, and J Vandewalle, “A Multilinear Singular Value Decomposition,” SIAM J. Matrix Anal. p. 3 To the best of our knowledge, this is the first study to formulate collider anomaly
detection with an MERA-based autoencoder. Earlier anomaly-detection applications in
high-energy physics have considered quantum autoencoders and, more recently, tensornetwork models based on matrix product states acting in an externally learned latent space,
but not a MERA encoder–decoder operating directly on ordered jet constituents [20, 21]. To the best of our knowledge, this is the first study to formulate collider anomaly
detection with an MERA-based autoencoder. Earlier anomaly-detection applications in
high-energy physics have considered quantum autoencoders and, more recently, tensornetwork models based on matrix product states acting in an externally learned latent space,
but not a MERA encoder–decoder operating directly on ordered jet constituents [20, 21]. To the best of our knowledge, this is the first study to formulate collider anomaly
detection with an MERA-based autoencoder. Earlier anomaly-detection applications in
high-energy physics have considered quantum autoencoders and, more recently, tensornetwork models based on matrix product states acting in an externally learned latent space,
but not a MERA encoder–decoder operating directly on ordered jet constituents [20, 21].Where the paper says it
Where the paper says it
Where the paper says it
[33] P H Sch¨onemann, “A Generalized Solution of the Orthogonal Procrustes Problem,” Psychometrika 31 no. 1, (1966) 1–10
[34] K Batselier, A Cichocki, and N Wong, “MERACLE: Constructive Layer-Wise Conversion of a Tensor Train into a MERA,” Commun. Appl. Math. Comput. 3 no. 2, (2021) 257–279
[30] A Edelman, T A Arias, and S T Smith, “The Geometry of Algorithms with Orthogonality Constraints,” SIAM J. Matrix Anal. Appl. 20 no. 2, (1998) 303–353
[31] P.-A. Absil, R Mahony, and R Sepulchre, Optimization Algorithms on Matrix Manifolds. Princeton University Press, Princeton, 2008
[32] L. De Lathauwer, B De Moor, and J Vandewalle, “A Multilinear Singular Value Decomposition,” SIAM J. Matrix Anal. Appl. 21 no. 4, (2000) 1253–1278
[33] P H Sch¨onemann, “A Generalized Solution of the Orthogonal Procrustes Problem,” Psychometrika 31 no.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved AI system can perform reconstruction-based anomaly detection in collider jets by utilizing a multiscale tensor-network architecture inspired by MERA, which is shown to be
stronger than the dense autoencoder in AUC
while usingabout one third of the trainable parameters.
-
The improved system can incorporate locality-aware preprocessing, specifically reordering constituents based on geometric proximity in the angular and momentum plane, which is demonstrated to yield a higher retained variance fraction in local compression layers.
-
The system can incorporate trainable disentanglers within the MERA architecture, which
improve performance over the corresponding identity-disentangler limit
when operating in the strongest-compression regime, as evidenced by a mean test AUC gain of 0.0081 at latent dimension B=8.
Abstract
We investigate whether a multiscale tensor-network architecture provides a useful inductive bias for reconstruction-based anomaly detection in collider jets. Jets are produced by a branching cascade, so their internal structure is naturally organised across angular and momentum scales. This motivates an autoencoder that compresses information hierarchically and can reorganise short-range correlations before coarse-graining. Guided by this picture, we formulate a Multiscale Entanglement Renormalization Ansatz (MERA)-inspired autoencoder acting on ordered jet constituents. To our knowledge, a MERA-inspired autoencoder has not previously been proposed, and this architecture has not been explored in collider anomaly detection. We compare this architecture to a dense autoencoder, the corresponding tree-tensor-network limit, and classical baselines within a background-only reconstruction framework. The paper is organised around two main questions: whether locality-aware hierarchical compression is genuinely supported by the data, and whether the disentangling layers of MERA contribute beyond a simpler tree hierarchy. To address these questions, we combine benchmark comparisons with a training-free local-compressibility diagnostic and a direct identity-disentangler ablation. In the nominal benchmark, the MERA-inspired model attains a higher mean AUC than the dense autoencoder at all three latent dimensions while using roughly one-third of its trainable parameters. The locality-preserving multiscale structure is well matched to jet data, and, at the strongest compression bottleneck, the trained disentanglers move substantially away from the identity and are accompanied by a positive mean performance shift. Overall, the study supports locality-aware hierarchical compression as a useful and parameter-efficient inductive bias for reconstruction-based jet anomaly detection in this benchmark setting.
Sources
- Machine Learning for Anomaly Detection in Particle Physics
- The LHC Olympics 2020: A Community Challenge for Anomaly Detection in High Energy Physics
- Challenges for Unsupervised Anomaly Detection in Particle Physics
- QCD or What?
- Searching for New Physics with Deep Autoencoders
- Autoencoders for unsupervised anomaly detection in high energy physics
- Jet Substructure at the Large Hadron Collider: A Review of Recent Advances in Theory and Machine Learning
- Supervised Learning with Quantum-Inspired Tensor Networks
- Quantum-inspired event reconstruction with Tensor Networks: Matrix Product States
- Classical versus Quantum: comparing Tensor Network-based Quantum Circuits on LHC data
- Quantum-probabilistic Hamiltonian learning for generative modelling & anomaly detection
- Tensor Network for Anomaly Detection in the Latent Space of Proton Collision Events at the LHC
- The Machine Learning Landscape of Top Taggers
- A Practical Introduction to Tensor Networks: Matrix Product States and Projected Entangled Pair States
- Machine Learning by Unitary Tensor Network of Hierarchical Tree Structure
- Tree Tensor Networks for Generative Modeling
- An Introduction to PYTHIA 8.2
- DELPHES 3, A modular framework for fast simulation of a generic collider experiment
- The anti-k_t jet clustering algorithm
Related papers
- MadNIS at NLO
- The NANOGrav 15-year data set: A Limit on Continuous Spin Gravitons
- DIS-secting the universe with high-energy neutrinos
- Crowdsourcing Gravitational Waves from Superradiant Axions
- Classification of g-modes for neutron stars with a strong transition: Novel universal relation including slow stable hybrid stars
- Scalar axion field of toroidal electromagnetic pulses