Fixed point compositionality via low-rank gluing rules in inhibition-dominated threshold-linear networks

arXiv:2606.07336 · q-bio.NC · Submitted 2026-06-05 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: I'm Ines, and with me are Marcus and Yuki, guest researcher.

Marcus: Today's paper: "Fixed point compositionality via low-rank gluing rules in inhibition-dominated threshold-linear networks".

Ines: Brains routinely generate highly flexible and complex behaviors on a relatively stable structure and limited resources, and this ability is rooted in compositionality, which allows for efficient task decomposition.

Marcus: First, who's behind it and why it matters.

Title and authors: Ines: So we're looking at this paper by Juliana Londono Alvarez and Robert J. and Nancy D. Carney titled "Fixed point compositionality via low-rank gluing rules in inhibition-dominated threshold-linear networks." It seems like they're trying to mathematically nail down how modular structure connects to the ability of these biological systems to do complex things efficiently.

Marcus: I think the title itself points directly at the core focus, which is that structural organization, specifically modularity, drives functional compositionality in these types of networks. For us in genomics data science, it’s interesting because we often look at how different parts of a dataset interact and whether that interaction reveals underlying biological processes.

Yuki: From a population genetics standpoint, I'm curious if this mathematical framework can help us understand why certain behavioral motifs appear consistently across different species, suggesting an evolutionary constraint on how these networks are built.

Ines: Exactly, Yuki; the paper is trying to move beyond just saying that modularity and compositionality are linked and provide a rigorous mathematical characterization for inhibition-dominated threshold-linear networks—TLNs. They introduce a new coupling mechanism called low-rank gluings to investigate this relationship more deeply.

Marcus: The authors are focusing on how the global fixed points of these networks are constrained, specifically showing that they have to be combinations of the local fixed points from their component modules. That's a specific constraint on what the network can actually achieve dynamically.

Yuki: That idea of constraints on global states based on local ones is significant because it suggests that complexity isn't just random, but follows a predictable structural recipe, which ties nicely into how we model evolutionary changes in gene regulatory networks.

Ines: Right, and the paper then introduces low-rank gluings as this novel way to connect these modules, and they go further by defining rank-one gluings which give a complete characterization of which combinations of local fixed points actually result in global ones.

Marcus: That move to rank-one gluings is what makes the analysis more tractable mathematically, moving it from a general idea to a concrete set of rules for assembly. From a data perspective, this means we can start predicting network behavior based on how the components are linked rather than just observing every possible connection.

Title and authors: Yuki: It’s interesting how they link this structural rule—the gluing rule—to specific combinatorial structures like cyclic unions and clique unions in graph-based networks, which is where it gets really applied.

Ines: They extend the theory from general TLNs to graph-based ones, specifically generalized combinatorial threshold-linear networks, or gCTLNs, proving that these structural rules hold up even in more complex graph structures defined by a directed graph G.

Marcus: When they apply these rules to clique unions and disjoint unions in those gCTLNs, it shows that the emergent dynamic attractors of the glued networks still behave as compositions of the attractors from the subnetworks. That’s a strong statement about preserving functional compositionality even when you build up more complex graph topologies.

Yuki: If this holds true across different graph types, it suggests that the fundamental principle of modular compositionality is quite robust and isn't just an artifact of one specific network topology. This has implications for how we think about gene regulatory networks evolving over time.

Ines: The paper then demonstrates that these gluing rules give a recipe for engineering compositional dynamics, allowing us to build networks with a large repertoire of predictable attractors, from discrete fixed points up to compositional limit cycles.

Marcus: From an engineering standpoint, that is really valuable because it gives us a structured way to design computational systems where we know exactly what kinds of behaviors we can expect and how they arise from the components. It’s about controlling the output by controlling the input structure.

Yuki: That control aspect is what interests me; if we can predict these attractors, it might help us identify which structural configurations are biologically plausible or favored in evolution based on their dynamic stability.

Ines: The main results center around four theorems: Theorem one defines fixed point compositionality via low-rank gluings, Theorem two gives the complete characterization for rank-one gluings by determining if local fixed points "die" or "survive," and then Theorems three and four characterize the dynamics for cyclic unions, clique unions, and disjoint unions in gCTLNs.

Marcus: The distinction between whether the component fixed points survive or die in Theorem two is a crucial statistical element; it’s essentially determining which local motifs contribute to the final global state, which sounds like a way to filter out noisy or irrelevant components from a complex system.

Title and authors: Yuki: I wonder how this survival/death mechanism translates into population dynamics; does the stability of these attractors relate to the fitness landscape of an organism?

Ines: The overall implication is that we have established a principled recipe for building architectures that generate many predictable attractors, which are compositions ranging from simple fixed points to more complex limit cycles. This suggests a principle by which biological networks might create intricate behaviors from limited building blocks.

Marcus: So, the big implication is that the core computational function seems heavily encoded in the graph structure itself rather than just being determined by every single synaptic weight value. That shifts our focus from purely parameter tuning to architectural design for functional outcomes.

Yuki: That’s a big shift for evolutionary biology; it implies that selection might favor networks with certain structural gluing rules because they provide a richer repertoire of stable dynamic states, which could be linked to adaptive evolution.

Ines: We're wrapping up the discussion on this paper, "Fixed point compositionality via low-rank gluing rules in inhibition-dominated threshold-linear networks." The work establishes a blueprint for creating architectures with many predictable attractors by using low-rank couplings as the assembly mechanism.

Marcus: It’s really about taking that abstract concept of compositionality and giving us concrete, mathematical rules—the rank-one gluings—to actually build things with, which is a very practical step for anyone working with large network models.

Yuki: For the history of species, it suggests that the underlying connectivity structure is a primary driver in generating the variety of behaviors we observe across different organisms.

Ines: Indeed, and we should keep an eye on how these structural rules apply to other types of biological networks beyond just TLNs, as that's where the real biological insight lies.

Marcus: We definitely need to keep track of this paper because it provides a way to predict the emergent behavior from a structured assembly process rather than just running simulations and hoping for the right outcome.

Yuki: I think we'll be watching how these mathematical constraints interface with empirical data from genomic studies in the coming months, trying to find those population-level correlations.

The paper's summary: Ines: So, this paper boils down to saying that the way these neural networks are put together—their structure—dictates exactly what kinds of complex behaviors they can actually produce dynamically, and that structure is governed by these specific low-rank coupling rules.

Marcus: Exactly; it’s not just about having a bunch of connections, but how those connections are mathematically constrained, which means the network's final stable states have to be built out of the stable states of its smaller parts.

Yuki: From a population genetics view, that constraint is fascinating because it suggests that the evolutionary path might be biased toward specific structural arrangements because only those arrangements allow for a wide variety of possible dynamic outcomes.

Ines: Right, and they introduce these low-rank gluings as the mechanism to formally study this relationship, showing that global fixed points are necessarily combinations of local ones based on their supports.

Marcus: The real statistical meat there is how they define those rank-one gluings, which gives a complete way to predict whether a combination of local parts will actually result in a viable global state or if it will just die out.

Yuki: That survival versus death criterion is what I find most compelling for understanding how biological systems maintain stability over evolutionary time, as we can hypothesize that only certain structural combinations persist in a population.

Ines: It gives us this recipe for engineering dynamics; if we know the local motifs and the rules for gluing them, we can predict the entire repertoire of behaviors—from simple fixed points to limit cycles—that this architecture supports.

Marcus: That means instead of just tuning every single synapse value and hoping for a stable pattern, we design the architecture based on these structural rules to guarantee certain functional outcomes.

Yuki: And when you extend this theory to graph-based networks, like those in gCTLNs, they show that this compositional principle holds up even when the overall connectivity structure gets much more complicated with cyclic or clique unions.

Ines: That extension is important because it proves this isn't just some abstract model for simple networks; it works for the more intricate, realistic architectures we see in biology.

Marcus: I think the implication for genomics data science is that when we look at complex gene regulatory circuits, this framework gives us a way to filter out noise by identifying which structural components are essential contributors to the overall dynamic behavior.

Yuki: It suggests that the core computational function might be less about the precise molecular details of every synapse and more about the underlying graph topology itself, which has huge implications for how we model deep evolutionary history.

The paper's improvements: Ines: So, we're looking at how this research suggests refining the original work by moving beyond just fixed points to look at more complex dynamics, and what that means for understanding biological systems that operate over longer time scales.

Marcus: Right, it seems they aren't just stopping at finding stable states; they’re pushing into compositional limit cycles, which is a much richer dynamic space for modeling real-world biological processes than just static fixed points.

Yuki: That complexity is key because in evolution, sustained oscillations or rhythms are often what drive adaptation and fitness across different environments, so exploring these limit cycle attractors gives us a better look at how species might maintain stable, adaptive behaviors.

Ines: They’re suggesting that the low-rank gluing rules aren't just for static points; they provide a "recipe" for building networks that support these compositional limit cycles, meaning we can engineer sustained rhythms from simpler components.

Marcus: From a statistical standpoint, this means our metrics for batch effects or cohort differences need to account for not just the average state, but the entire dynamical landscape of attractors that the network is capable of entering.

Yuki: I wonder if we can use these compositional rules to predict which structural modifications in gene regulatory networks would lead to a stable, adaptive oscillation suitable for a specific environmental stressor.

Ines: Precisely; this moves us from describing what the network *is* in terms of static states to predicting what it *does* over time based on its assembly structure.

Marcus: It’s about moving the statistical analysis from just looking at correlations between inputs and outputs to looking at how structural rules constrain the entire possible dynamic trajectory space.

Yuki: That connects back to population genetics because if certain structural rules favor a specific set of attractors, those structures might become strongly selected for in an environment that demands that particular kind of sustained activity.

Ines: So, the improvement is essentially gaining a dynamical toolkit: using the same low-rank gluing concept to characterize more complex, time-dependent behaviors rather than just steady states.

Marcus: It gives us a way to structure our analysis so we aren't missing potential dynamic states because our statistical tests only looked for equilibrium points.

Yuki: And when you look at the wider history of species, this suggests that the structural constraints on how information flows—the gluing rules—are as important as the sequence of genes themselves in determining the long-term evolutionary trajectory.

Ines: It really highlights how fundamental architectural principles, like compositionality, underpin complex biological function across vastly different scales of time and structure.

Conclusion: Ines: So, to wrap up our discussion on "Fixed point compositionality via low-rank gluing rules in inhibition-dominated threshold-linear networks," this paper essentially provides a rigorous mathematical blueprint showing that network structure dictates functional capability through these low-rank coupling mechanisms.

Marcus: It's really about taking the abstract idea of modularity and giving us concrete, rank-one rules for assembly, which means we can actually predict the resulting dynamics instead of just running simulations hoping for a certain outcome.

Yuki: From my perspective in population genetics, this suggests that the underlying connectivity architecture is a primary driver in generating the variety of behaviors we observe across different organisms.

Ines: Exactly; if you can predict the attractors based on structure, you get a better idea of what structural configurations are evolutionarily viable because those are the ones that allow for a wide repertoire of predictable states.

Marcus: It shifts our focus away from just tuning individual synapse values and toward designing robust architectures where the functional output is encoded in the graph structure itself, which is super valuable when we're dealing with large-scale genomic data sets.

Yuki: I think this framework offers a way to hypothesize about selection pressures that favor specific gluing rules because those rules might provide a richer set of stable dynamic states suitable for survival in certain conditions.

Ines: That’s the core biological question: what kind of structural recipe allows limited building blocks to generate such complex, predictable dynamics?

Marcus: It gives us a statistical tool to filter out irrelevant components by looking at how local motifs survive or die when they get coupled into the larger network.

Yuki: And that survival mechanism has huge implications for how we model long-term stability in biological systems because it links structural persistence directly to dynamic viability.

Ines: We've seen how this principle extends into graph networks, and the ability to predict attractors ranging from fixed points to limit cycles really opens up new ways for us to design computational models that mimic biological flexibility.

Marcus: It means we can build systems where the functional outcome is more predictable based on the schematic of how those components are linked rather than just observing every possible synaptic weight setting in a cohort.

Yuki: And this principle, established by "Fixed point compositionality via low-rank gluing rules in inhibition-dominated threshold-linear networks," provides a strong foundation for thinking about the evolutionary constraints on network architecture itself.

Ines: It’s definitely a piece of work that gives us a principled recipe for building architectures with many predictable behaviors from limited building blocks.

Marcus: We'll be watching how these structural rules interface with empirical data from genomic studies in the coming months to see if we can find those population-level correlations you mentioned, Yuki.

Yuki: I think that's exactly where the next big connection will be made as we try to link these abstract mathematical constraints back to observable traits in diverse species.

Juliana Londono Alvarez, Robert J. Carney, Nancy D. Carney

Brown University

q-bio.NC

Submitted: 2026-06-05

Updated: 2026-09-28

Comments: 41 pages, 17 figures. Version submitted to SIAM Journal on Life Sciences (SIALS)

Code: https://github.com/juliana-londono/low-rank-gluings

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 83/100

The gist: Brains routinely generate highly flexible and complex behaviors on a relatively stable structure and limited resources, and this ability is rooted in compositionality, which allows for efficient task

Key concepts

Fixed Point Compositionality: This is the property where the overall stable states (fixed points) of a large network can be precisely determined by combining the stable states of its smaller, individual parts. The paper shows this happens when networks are connected using specific low-rank coupling methods.
Low-Rank Gluings: This is a novel method for connecting different subnetworks within a larger network. Instead of simple connections, these involve 'low-rank couplings,' which are specific mathematical ways to link components that restrict the resulting global fixed points to be combinations of the local ones.
Inhibition-Dominated TLNs: These are specific types of dynamical systems used in the study, characterized by recurrent connections and external input where inhibition plays a dominant role. The network's behavior is studied by looking at its fixed points, which represent stable states or steady behaviors over time.

Terminology

Modularity — 18× in this paper · explained in 9 episodes
Limit Cycles — 14× in this paper · explained in 2 episodes
Core — 7× in this paper · explained in 2 episodes
Structural Constraints — 4× in this paper · explained in 2 episodes
Adaptation — 2× in this paper · explained in 1 episode
Computational Biology — 2× in this paper · explained in 2 episodes
Degeneracy — 2× in this paper · explained in 3 episodes
Robustness — 2× in this paper · explained in 5 episodes
Scaffolding — 2× in this paper · explained in 3 episodes
Trace — 2× in this paper · explained in 2 episodes
Adam — 1× in this paper · explained in 2 episodes
Adjacency Matrix — 1× in this paper · explained in 1 episode
Entrainment — 1× in this paper · explained in 2 episodes
Generalization — 1× in this paper · explained in 13 episodes
Graph Construction — 1× in this paper · explained in 3 episodes
Inverse Problem — 1× in this paper · explained in 2 episodes
Latent Dynamics — 1× in this paper · explained in 2 episodes
Mean-Field Dynamics — 1× in this paper · explained in 2 episodes
Metastability — 1× in this paper · explained in 3 episodes
Ordinary Differential Equations (ODEs) — 1× in this paper · explained in 2 episodes
Task Decomposition — 1× in this paper · explained in 2 episodes

Summary

Brains routinely generate highly flexible and complex behaviors on a relatively stable structure and limited resources, and this ability is rooted in compositionality, which allows for efficient task decomposition. This work formally investigates how structural modularity supports functional compositionality in inhibition-dominated threshold-linear networks (TLNs) by introducing low-rank gluings as a novel coupling mechanism.

The gist

The global fixed points of these networks are constrained to be combinations of the local fixed points of their constituent modules.

Fixed Point Compositionality in TLNs

The study focuses on inhibition-dominated threshold-linear networks (TLNs), defined by ODEs where firing rates are governed by recurrent connections and a constant external input. Fixed point compositionality is defined as the property where global fixed points are obtained as combinations of local fixed points, specifically concerning their supports. The paper introduces low-rank gluings, where component subnetworks with arbitrary internal connectivity are connected via specific low-rank couplings. The central finding is that for these networks, the global fixed points of these networks are constrained to be combinations of the local fixed points of their constituent modules.

Low-Rank Gluings and Rank-1 Gluings

The paper introduces a novel class of modular network assembly called low-rank gluings, where component subnetworks are connected via specific low-rank couplings. A more structured subclass, called rank-1 gluings, provides a complete characterization that determines which combinations of local fixed points yield global ones. The mathematical tools used to prove these results include Lemma 1 (a determinant factorization for matrices with a rank-1 block) and Theorem 5 (sign conditions), which are applied to factorize the quantities denoted as sσi.

Application to Graph-Based Networks (gCTLNs)

The theory is extended from general TLNs to graph-based networks, specifically generalized combinatorial threshold-linear networks (gCTLNs). The paper proves that these structural rules are more robust than initially posited by applying the results to graph structures defined by a directed graph G. This involves proving Theorems 3 and 4, which establish gluing rules for cyclic unions (Theorem 3), clique unions (Theorem 4a), and disjoint unions (Theorem 4b). These theorems show that the emergent dynamic attractors of glued networks also behave as compositions of the attractors of the subnetworks.

Attractor Dynamics and Compositional Recipes

The structural rules provide a mathematically tractable recipe for engineering compositional dynamics, enabling the construction of networks with a combinatorially large repertoire of predictable attractors that can be understood from simpler component motifs. The paper demonstrates that these gluing rules allow for the construction of attractors ranging from compositions of discrete fixed points to compositional limit cycles. For example, in the mollusk locomotion example, a cyclic union architecture supports eight attractors corresponding precisely to the eight octants in a cube.

Key Results Summary

The main results consist of four theorems:

Theorem 1 (fixed point compositionality): Low-rank gluings restrict global fixed point supports to be unions of local fixed point supports, at most one per component.
Theorem 2 (rank-1 gluings): Provides a complete characterization for rank-1 gluings, determining which combinations of local fixed points yield global ones based on whether they die or survive.
Theorem 3 (generalized cyclic union): Characterizes fixed points for cyclic unions in gCTLNs.
Theorem 4 (generalized clique and disjoint union): Characterizes fixed points for clique and disjoint unions in gCTLNs, distinguishing between cases where component fixed points survive or die.

Conclusion and Implications

The work establishes a principled recipe to build architectures with combinatorially many attractors that range from compositions of discrete fixed points to compositional limit cycles, suggesting a principle by which biological networks might generate complex behaviors from limited building blocks. The findings show that functional compositionality is robust, as the rules extend successfully from CTLNs to the more flexible gCTLN family, implying that the core computational function is primarily encoded in the graph structure rather than in the precise values of the synapses.

Roadmap

The paper proceeds by developing low-rank gluing theory (Theorem 1), extending it to rank-1 gluings (Theorem 2), applying these to graph-based networks (Theorems 3 and 4), and concluding with an example application to mollusk locomotion. The discussion highlights the link between structural modularity and functional compositionality across continuous and discrete dynamical systems.

Limitations

Open questions remain regarding the formal proof that observed trajectories are attractors, the exact boundaries of structural constraints, and extending survival rules for general TLNs beyond gCTLNs. However, the framework provides a "bare-bones scaffolding or structural blueprint on which the richer complexities of biological brains might rest.

Improvements for AI systems

Based on the research presented in this paper, here are specific improvements that could be made to AI systems, along with what those improved systems could achieve:

)The Improved System: Compositionally Robust Neural Architectures (CRNA)

The core improvement is moving from monolithic or arbitrarily connected neural networks to architectures explicitly designed around structured modularity and low-rank inter-module coupling, governed by mathematically tractable gluing rules.

Specific improvements include:

Moving from arbitrary connectivity to constrained low-rank inter-component couplings (low-rank gluings).
Enforcing structural compositionality through rank-1 gluing rules where applicable.
Implementing dynamic control mechanisms based on fixed point decomposition (survival vs. death of local motifs).

)Specific Capabilities and Applications:

The CRNA system, enabled by these mathematical principles, can perform the following specific functions:

Robust Task Decomposition and Recomposition: The system can decompose complex, high-dimensional tasks into a vast repertoire of simpler, reusable primitives (local fixed points/motifs). This allows the AI to efficiently handle novel tasks by combining known building blocks in new ways, mirroring biological compositionality.

Combinatorially Large Attractor Diversity: Instead of being limited to a few fixed points or simple limit cycles, the system can be engineered to possess a combinatorially large number of predictable attractors (ranging from discrete fixed points to compositional limit cycles). This allows it to maintain multiple distinct modes or strategies for solving the same high-level problem simultaneously.
Predictive Behavioral Switching (Attractor Transitions): The system can be designed with precise, predictable switching mechanisms between these distinct modes. By tuning external inputs (like transient pulses), the AI can reliably transition from one compositional strategy to another, enabling flexible and context-aware behavior without needing a complete re-training cycle for every new scenario.
Feature Binding and Coherent Representation: The system excels at dynamically binding distinct features or sensory inputs into a single, coherent representation (fusion attractors). This is crucial for complex perception tasks where multiple interacting signals must be integrated into one meaningful output (e.g., understanding a complex scene).
Model-Based Control for Complex Locomotion/Motor Skills: For robotic agents or bio-inspired systems, the CRNA can generate sequential control patterns (like gait) by combining component motifs in specific cyclic or clique union structures, allowing for flexible, multi-modal locomotion that is robust to parameter variations (heterogeneity) while maintaining a clear functional structure.

In summary, the improved AI system moves beyond simple pattern recognition toward a structured intelligence capable of efficient modular task execution and robust, context-dependent behavioral switching across a vast repertoire of predictable dynamical states.

Abstract

Brains routinely generate highly flexible and complex behaviors on a relatively stable structure and limited resources. A key mechanism underlying this ability is compositionality, which allows the brain to efficiently decompose complex tasks into simpler, reusable primitives. While network modularity has often been linked to compositionality in biological and artificial networks, a rigorous mathematical characterization of this relationship in nonlinear networks is still lacking. In this work, we formally investigate how structural modularity supports functional compositionality in inhibition-dominated threshold-linear networks (TLNs). We introduce a novel class of modular network assembly called low-rank gluings, where component subnetworks with arbitrary internal connectivity are connected via specific low-rank couplings. We prove that the global fixed points of these networks are constrained to be combinations of the local fixed points of their constituent modules. For a more structured subclass, called rank-1 gluings, we provide a complete characterization that determines which combinations of local fixed points yield global ones. We apply these results to graph-based networks, extending fixed point decomposition rules from combinatorial threshold-linear networks (CTLNs) to the more flexible family of generalized CTLNs (gCTLNs), thereby proving that these structural rules are more robust than initially posited. Finally, we demonstrate that these gluing rules provide a mathematically tractable recipe for engineering compositional dynamics, enabling the construction of networks with a combinatorially large repertoire of predictable attractors that can be understood from simpler component motifs, ranging from compositions of fixed points to compositional limit cycles.