CTMRG as a tensor-network message-passing algorithm
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "CTMRG as a tensor-network message-passing algorithm".
Mira: Here's the gist of the paper. The paper introduces Corner Transfer Matrix Renormalization Group (CTMRG), formulated as a set of local equations on general planar graphs, which are then used to build a message-passing algorithm similar to belief propagation for constructing Matrix Product States (MPS) environments for any connected subgraph.
Kai: These local equations depend only on the small neighborhoods of the full planar graph and are consistently glued together.
Mira: The general theory involves fixing partial gauges, rewriting corner matrices in terms of simplest ones (C k), and solving an iterative cycle: first, solving eigenproblems for every graph face and imposing a gauge choice to fix corner matrices; second, using these projectors and corner matrices to update edge messages (A e) via message equations while fixing residual gauge by requiring elementwise closeness between successive message tensors; finally, normalizing the tensors and monitoring convergence via factors kappa.
Kai: The CTMRG approach is shown to reconcile two research branches: it can be formulated as a message-update algorithm on general planar graphs, mirroring belief propagation.
Paper summary: Kai: So we're looking at this paper called "CTMRG as a tensor-network message-passing algorithm". It's about how you can use corner transfer matrix renormalization group equations to build something that helps you calculate environments for any connected piece of a larger graph.
Mira: Exactly. The core idea is formulating these corner transfer matrix equations as local rules on general planar graphs, and then using those rules to build message tensors that describe the environment around any subgraph, sort of like how belief propagation works for smaller pieces <ref:2610.00631#pg3>.
Lev: That sounds like a way to approximate those hard tensor network contractions without the exponential cost scaling we usually see <ref:2610.00631#pg2>.
Kai: Right, and what they're claiming is that this approach systematically builds these components from local building blocks, allowing you to apply it to any graph, whether it's finite or even an infinite periodic one <ref:2610.00631#pg1>.
Mira: It’s interesting because the equations only depend on small neighborhoods of the whole graph and they glue together consistently, which is what makes them work for any connected subgraph <ref:2610.00631#pg1>.
Lev: That gluing part is key for practical use, because it means we can construct an MPS environment on the external legs of a piece, not just on the whole thing <ref:2610.00631#pg3>.
Kai: They test this idea on a few different lattices, including ferromagnetic Ising models on square and hyperbolic lattices like (four five), (five four), and (eight three) <ref:2610.00631#pg2>.
Mira: And they show that their prediction for the partition function matches exact contractions for finite open square lattices <ref:2610.00631#pg2>.
Lev: For the hyperbolic lattices, they determine where the magnetization phase transition happens by fitting squared magnetization data <ref:2610.00631#pg2>.
Kai: They also have to deal with some numerical issues, like instabilities when inverting corner matrices and problems with biorthogonal projectors, so they limit the bond dimension on faces with ill-conditioned corners <ref:2610.00631#pg2>.
Mira: And they monitor how well it's converging using these factors called kappa <ref:2610.00631#pg2>.
Lev: It sounds like a necessary step to make the algorithm run on real hardware, because you can't just assume everything is perfectly conditioned <ref:2610.00631#pg2>.
Kai: Beyond the simulations, they looked at some theoretical stuff too, showing that inverse corner matrices behave in certain ways that suggest they are equivalent to other corner matrices we've seen before <ref:2610.00631#pg4>.
Mira: But they also point out something important about the external field h, where if the field is non-zero, the projectors and tensors are related in a way we expect from biorthogonal relations <ref:2610.00631#pg4>.
Lev: So when you set the external field to zero, which is often what you want for equilibrium calculations, they find that the system becomes integrable <ref:2610.00631#pg4>, which leaves a residual gauge freedom that isn't fixed by this procedure <ref:2610.00631#pg4>.
Kai: That residual gauge freedom is something we have to be careful about when building things for actual quantum hardware, because you can't just leave it as is <ref:2610.00631#pg4>.
Mira: They also clarify a specific condition: the relation between the message tensors A, the corner matrices C, and the projector P only holds if their gauge choices match on that final index <ref:2610.00631#pg4>.
Lev: It sounds like a very precise set of conditions you have to meet for this approximation to be reliable <ref:2610.00631#pg4>.
Kai: So, what does this actually mean when we look at the title, "CTMRG as a tensor-network message-passing algorithm"?
Mira: It means they reconciled two different ways of looking at tensor networks—one based on renormalization group methods and another based on message passing <ref:2610.00631#pg3>.
Lev: For someone building hardware, it suggests a path to construct the environment description using local updates rather than needing a full, massive contraction calculation upfront <ref:2610.00631#pg2>.
Kai: If we think about what this means for the world, it's that we might be able to tackle these complex 2D statistical mechanics problems on planar graphs with controlled error scaling instead of just relying on methods that scale exponentially <ref:2610.00631#pg2>.
Mira: It suggests a more structured way to approximate the physics, linking local rules to global structure through these message passing steps <ref:2610.00631#pg3>.
Lev: The systematic construction from building blocks is what makes it applicable across different graph types, which is good for testing on varied physical systems like those hyperbolic lattices <ref:2610.00631#pg2>.
Kai: And the fact that they can build an MPS environment on the external legs of any connected subgraph means we are not stuck analyzing just the whole system at once <ref:2610.00631#pg3>.
Mira: It moves us closer to applying these tools to more complex 2D statistical mechanics problems that don't have simple exact solutions <ref:2610.00631#pg2>.
Lev: The limitation they state is that the method is formulated on planar graphs, and the future work will look at generalizing this approach to three-dimensional graphs <ref:2610.00631#pg4>.
Kai: And they also mentioned looking into Variational Uniform Matrix Product States, or VUMPS, as a direction for future research <ref:2610.00631#pg4>.
Mira: It sounds like the next step is to see how this message passing framework fits with other variational methods on tensor networks <ref:2610.00631#pg4>.
Lev: For hardware, it means we need to figure out how to make these local equations run efficiently when we scale up the physical system size <ref:2610.00631#pg2>.
Conclusion: Kai: So we’re wrapping up this look at Corner Transfer Matrix Renormalization Group, or CTMRG, which they frame as a tensor network message passing algorithm.
Mira: It basically takes these complex corner transfer matrix equations and turns them into local rules on general planar graphs to build up these message tensors that describe the environment around any connected piece of a system.
Lev: So it’s about taking something really complicated in renormalization group theory and turning it into a message passing algorithm similar to belief propagation for constructing these matrix product states.
Kai: The authors are showing how you can systematically build these components from local building blocks, meaning you can apply this technique to almost any graph, whether it’s finite or even an infinite one with periodic boundaries.
Mira: It really bridges two different worlds here—the way renormalization group works and the way message passing algorithms operate. It gives us a structured way to approximate those big tensor network contractions without having to deal with the full exponential scaling upfront.
Lev: For hardware, that means we might be able to construct the environment description on just the external legs of a subgraph, which is a lot cleaner than trying to contract everything at once on real quantum hardware.
Kai: And they did test this on several lattices, including ferromagnetic Ising models on square and hyperbolic shapes like four by five or eight by three. The results matched exact calculations for finite open square lattices, and they even found where the magnetization phase transition occurs on those hyperbolic grids.
Mira: But the caveats are important—they had to limit how large the bond dimension gets on faces with tricky corners because inverting those corner matrices can be numerically unstable or ill-conditioned.
Lev: That makes sense, because if you're running this on real hardware, you can't just assume everything is perfectly conditioned; you have to manage that error scaling carefully.
Kai: Theoretically, they found that the inverse corner matrices behave in a way that suggests they might be equivalent to other corner matrices we’ve seen before in the literature.
Mira: And when there’s an external field applied, like a non-zero field h, the projectors and tensors relate exactly as expected from biorthogonal relations, but when you set the field to zero, which is what you usually want for equilibrium calculations, they find the system becomes integrable.
Lev: Integrable means something special for error correction; it leaves a residual gauge freedom that isn't fixed by this procedure, so we have to be careful about how we interpret those results.
Kai: This paper suggests a powerful way to systematically build up these complex descriptions of quantum many-body environments using local message passing rules on planar graphs.
Mira: It points toward a new methodology for approximating the physics of 2D systems by linking local geometric structures directly to the construction of tensor network states <ref:2610.00631#pg1>.
Lev: The next big step, as they suggest, is generalizing this whole idea from planar graphs in two dimensions up to three dimensions and exploring things like variational uniform matrix product states.
Kai: It gives us a better toolkit for tackling these types of 2D statistical mechanics problems on lattices that are too complex for simple exact solutions <ref:2610.00631#pg1>.
National Science Center "Kharkiv Institute of Physics and Technology" · Haiqu, Inc. · V.N. Karazin Kharkiv National University
cond-mat.stat-mech
Submitted: 2026-09-30
Updated: 2026-11-15
Comments: Comments welcome; 56 pages, 38 figures
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 87/100
The gist: The paper introduces Corner Transfer Matrix Renormalization Group (CTMRG), formulated as a set of local equations on general planar graphs…
Key concepts
- Corner Transfer Matrix Renormalization Group (CTMRG)
- A framework formulated as local equations on planar graphs used to build MPS environments. It involves solving iterative cycles: first fixing corner matrices by solving eigenproblems on faces, then updating edge messages, and finally normalizing the tensors to monitor convergence.
- Message-Passing Algorithm
- The CTMRG approach constructs MPS environments using a message-passing style algorithm, similar to belief propagation. This allows the method to be applied systematically to arbitrary planar graphs by gluing local equations together based on small neighborhoods.
- Matrix Product States (MPS) Environment
- CTMRG tensors describe the environment of an arbitrary connected subgraph as a generalized MPS on its external legs. This means they can represent complex quantum states or correlation functions associated with that subgraph in a structured, tensor-network format.
- Gauge Fixing and Corner Matrices ($C_k$)
- The general theory requires fixing partial gauges by rewriting corner matrices in terms of simpler ones ($C_k$). This process is crucial for solving the iterative cycle, ensuring consistency between different parts of the graph structure.
Terminology
Summary
The paper introduces Corner Transfer Matrix Renormalization Group (CTMRG), formulated as a set of local equations on general planar graphs, which are then used to build a message-passing algorithm similar to belief propagation for constructing Matrix Product States (MPS) environments for any connected subgraph. p. 2 These local equations depend only on the small neighborhoods of the full planar graph and are consistently glued together. p. 2 The general theory involves fixing partial gauges, rewriting corner matrices in terms of simplest ones (C k), and solving an iterative cycle: first, solving eigenproblems for every graph face and imposing a gauge choice to fix corner matrices; second, using these projectors and corner matrices to update edge messages (A e) via message equations while fixing residual gauge by requiring elementwise closeness between successive message tensors; finally, normalizing the tensors and monitoring convergence via factors kappa.
The CTMRG approach is shown to reconcile two research branches: it can be formulated as a message-update algorithm on general planar graphs, mirroring belief propagation. p. 2 Furthermore, the CTMRG tensors can describe the environment of an arbitrary connected subgraph as a generalized MPS on its external legs. p. 2 The process systematically constructs CTMRG from building blocks, allowing application to arbitrary graphs (finite and periodic infinite). p. 2
The method has been tested on various lattices, including ferromagnetic Ising models on square and hyperbolic lattices like (4, 5), (5, 4), and (8, 3). p. 0 The CTMRG prediction for the partition function agrees with exact contractions for finite open square lattices. p. 1 For hyperbolic lattices, magnetization phase transition positions are determined by fitting squared magnetization data. p. 0 Numerical instabilities related to corner-matrix inverses and ill-conditioning of biorthogonal projectors are addressed by limiting the maximal bond dimension on faces with ill-conditioned corners. p. 9 The convergence is monitored through the behavior of kappa. p. 2
The paper also explores theoretical aspects, showing that inverse corner matrices (C-1) approximately obey certain relations, suggesting their equivalence to corner matrices defined in prior literature. p. 1 For a nonzero external field h, the projectors and tensors A are related as expected from the biorthogonal relation; however, at h=0, the system is integrable, leaving a residual gauge freedom not fixed by the procedure in Appendix A. p. 4 The relation between A, C, and projector (P) holds only if the gauge on the final index of P matches that of the initial index. p. 5 Future research directions include generalizing CTMRG to 3D graphs, exploring Variational Uniform Matrix Product States (VUMPS), using QR decomposition for projectors, and dualizing to CTMRG on the interaction-round-a-face (IRF) model. p. 1
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
-
Develop a message-passing algorithm for general planar graphs by formulating corner transfer matrix equations as local equations on subgraphs, allowing construction of an MPS environment for any connected subgraph.
-
Implement a systematic method to derive CTMRG projectors from the CTM update equations by using the auxiliary eigenproblem derived from enlarged corner matrices, which can be solved via truncation of eigenvalues for practical implementation.
-
Create a robust framework to handle numerical instabilities in CTMRG by dynamically choosing bond dimensions based on the condition number of corner matrices or by employing special gauges that equalize corner matrices around a face.
-
Implement an algorithm to compute the partition function for ferromagnetic Ising models on various planar and hyperbolic lattices, achieving near-exact results compared to exact calculations obtained via the Kac Ward determinant.
-
Extend CTMRG to general 3d graphs or cell complexes by generalizing the message-passing formulation from planar graphs, or by applying a variational formulation of CTM equations.
Sources
- Tensor networks for complex quantum systems
- Matrix Product States and Projected Entangled Pair States: Concepts, Symmetries, and Theorems
- Tensor Network Algorithms: a Route Map
- Developments in the Tensor Network -- from Statistical Mechanics to Quantum Entanglement
- Lecture Notes of Tensor Network Contractions
- Second Renormalization of Tensor-Network States
- Coarse-graining renormalization by higher-order singular value decomposition
- Tensor-Entanglement-Filtering Renormalization Approach and Symmetry Protected Topological Order
- Tensor Network Renormalization
- Renormalization of tensor networks using graph independent local truncations
- Gauging tensor networks with belief propagation
- Tensor Networks contraction and the Belief Propagation algorithm
- Simulating quantum dynamics in two-dimensional lattices with tensor network influence functional belief propagation
- Efficient tensor network simulation of IBM's Eagle kicked Ising experiment
- Confinement in the Transverse Field Ising model on the Heavy Hex lattice
- Dynamics of disordered quantum systems with two- and three-dimensional tensor networks
- Quantum Advantage: a Tensor Network Perspective
- Resolving Structure in Prethermal Floquet Dynamics with Precision Quantum Computation
- Large-scale quantum annealing simulation with tensor networks and belief propagation
- Efficient tensor network simulation of IBM's largest quantum processors
Related papers
- Emergent long-tail dynamics in driven magnets with dynamical frustration
- Statistical Mechanics of Dynamical System Identification
- Second law of thermodynamics in closed quantum many-body systems
- Chemical potentials from structure factors: II. Charged multi-component mixtures
- Genuine quantum scars in Floquet chaotic many-body systems
- Stochastic Thermodynamics for Autoregressive Generative Models: A Non-Markovian Perspective