In classical dynamical systems theory, we often assume the world is smooth. We use the Jacobian matrix—a collection of derivatives—to understand how a system responds to small changes. If the eigenvalues of that matrix cross a certain threshold, the system might transition from a steady state to an oscillation. This is the essence of a standard bifurcation.
However, many real-world systems are not smooth. They involve "switches," "thresholds," or "clipping." A climate model might have a threshold for ice-sheet collapse; a neural network uses ReLU functions that have a sharp corner at zero; a mechanical sensor might tap against a surface. In these cases, the traditional tools of calculus break down because the derivative is undefined or discontinuous at the boundary.
Researchers are instead turning to the theory of non-smooth dynamical systems to explain why these systems exhibit sudden, "jumping" behaviors that smooth models cannot predict.
The Geometry of the Border Collision
The central mechanism in these systems is the border-collision bifurcation. In a smooth system, a new behavior (like a vibration) usually starts small and grows as you turn a parameter dial. In a piecewise-smooth system, a behavior can emerge at full scale the moment the system's state "collides" with a boundary where the governing physics changes.
Daniel Durstewitz and Zahra Monfared have shown that in piecewise-linear recurrent neural networks (PLRNNs), these collisions are the primary drivers of complexity. When the trajectory of a network's hidden states hits the "corner" of a ReLU activation function, the system can instantly jump from a stable fixed point to a chaotic attractor. Their work in Existence of n-cycles and border-collision bifurcations in piecewise-linear continuous maps provides a roadmap for predicting these transitions by mapping exactly where orbits intersect these switching manifolds.
Why Systems "Jump"
One of the most recognizable features of non-smooth dynamics is the jump phenomenon. This occurs when a marginal change in an external force causes a sudden, dramatic spike or drop in the system's response.
In mechanical engineering, Kiichi Satō and Koichi OTOMI observed this in vertical beams under periodic loading. As the frequency of the load changes, the beam doesn't just vibrate more; it reaches a "hard-spring" threshold where the amplitude suddenly leaps to a much higher state. Similarly, Donatello Materassi and Michele Basso found that Atomic Force Microscopes exhibit these jumps because of hysteresis in the tip-sample interaction.
Mathematically, these jumps are often caused by the existence of multiple stable states for the same input frequency. As Jianwei Wu notes in the context of power systems, this bistability means that a tiny fluctuation can trigger a disproportionately large increase in oscillation magnitude, posing a risk to grid stability.
Modeling the "Saw-Tooth" Climate
The stakes of understanding non-smooth bifurcations are perhaps highest in climate science. Earth’s history is full of abrupt shifts—like the sudden onset of ice ages—that appear "saw-tooth" in the geological record: a slow buildup followed by a rapid crash.
Traditional smooth models struggle to replicate this because they predict that periodic orbits should emerge with zero amplitude. Kgomotso S. Morupisi argues in An analysis of the periodically forced PP04 climate model that we must treat the climate as a Filippov system. By modeling discontinuities—such as the sudden release of carbon dioxide once a threshold is hit—as surfaces where the physics changes instantaneously, we can explain how large-scale glacial cycles manifest instantly.
Furthermore, Rachel Kuske has demonstrated that these non-smooth features can lead to earlier-than-expected tipping points. When seasonal cycles resonate with these internal thresholds, the system can collapse far sooner than a smooth model would suggest.
Beyond the Jacobian
Because traditional calculus fails at the "border," researchers have developed new mathematical workarounds. Andrew M. Stuart explored singular limits, where a reaction zone is shrunk to a single point to make the math tractable. In the realm of discrete-state systems, Diego Fasoli developed semi-analytical frameworks for binary neural networks. Instead of looking for derivatives, these methods analyze the transition probability matrices between states to identify qualitative shifts in behavior.
By moving away from the requirement of smoothness, this research direction provides a more accurate language for a world that is often defined by its hard edges and sudden changes.