When more data become less informative: Finite-precision periodicization and collapse of forecast-error Lyapunov estimates
summary
The gist
Largest Lyapunov exponents (LLEs) quantify exponential sensitivity, but data-driven estimates are often obtained from finite-precision trajectories <ref:2608.16120#pg6> The paper shows that…
In short
The study investigated how increasing a single long record of chaotic data, computed with finite precision, affects a forecast-error Lyapunov exponent (LLE) estimate. Findings show that once the digital map enters an exact recurrence cycle due to finite state space limitations, the LLE estimate collapses toward zero. This means record length alone is not a reliable measure of information content for reduced-precision chaotic data.
Key concepts
- Largest Lyapunov Exponent (LLE)
- The LLE measures the exponential sensitivity of a system's future to small changes in its present state, quantifying chaos. In this study, it is estimated using forecast errors from a kNN predictor; when the estimate collapses to zero, it suggests the data is no longer providing new dynamical information.
- Finite-State Digital Map
- Deterministic digital maps (like those on embedded hardware) have a fixed, finite number of possible states. Because they are deterministic, once the system repeats a state exactly, its entire future trajectory must repeat. This causes exact digital recurrence after an initial transient phase.
- Exact-Neighbor Fraction (Fexact,K)
- This metric counts how often a test history in the data set has an exact copy among its K nearest neighbors. When this fraction rises to 100%, it signifies that the kNN predictor can perfectly replicate the future, causing the forecast error to drop and artificially suppress the LLE estimate.
Terminology used across episodes
This episode discusses
- When more data become less informative: Finite-precision periodicization and collapse of forecast-error Lyapunov estimates · Paper Radio
The paper
When more data become less informative: Finite-precision periodicization and collapse of forecast-error Lyapunov estimates · Read on arXiv
Institute of Physics and Technology, Petrozavodsk State University · Faculty of Electronics Technology, Industrial University of Ho Chi Minh City
DOI: 10.1016/j.chaos.2026.119363
Transcript
Introduction to the show: ident: Nonlinear Sciences. The day in nonlinear science: chaos, patterns, networks and adaptation.
Kai: I'm Kai.
Mira: And I'm Mira, here with Lev.
Kai: We're looking at the paper: "When more data become less informative".
Mira: Here's the gist of the paper. Largest Lyapunov exponents (LLEs) quantify exponential sensitivity, but data-driven estimates are often obtained from finite-precision trajectories <ref:2608.16120#pg6> The paper shows that increasing the length of a single reduced-precision chaotic record can eventually degrade a forecast-error LLE estimate <ref:2608.16120#pg10>.
Kai: The core finding concerns the saturation of predictive information in finite-precision systems.
Mira: The central mechanism is that deterministic digital maps have finite state spaces, leading to exact recurrence after a transient phase <ref:2608.16120#pg4>.
Kai: When this exact digital recurrence occurs, a kNN predictor can return an identical future, causing the forecast error to flatten and the LLE estimate to collapse toward zero <ref:2608.16120#pg5>.
Title and authors: Kai: We’re talking about this paper now, "When more data become less informative: Finite-precision periodicization and collapse of forecast-error Lyapunov estimates." The authors, Andrei Velichkoa and Viet-Thanh Phamb, are looking at the logistic map at r = four <ref:2608.16120#pg1,the logistic map at r = 4>.
Mira: That specific map is important because it’s a well-known example of deterministic chaos, but they test how the estimate behaves when you run it on hardware that uses finite precision arithmetic, like single precision versus double precision.
Lev: And they did this validation. They reproduced an ESP32 single-precision trajectory bit-for-bit with NumPy float32 across a million stored values. That means the software is acting as a perfect emulator for the tested hardware arithmetic for those initial sweeps.
Kai: But that exact match isn't perfect in terms of dynamics over the long run, which is where this paper gets interesting. They found that after ten thousand random float32 initial conditions, every trajectory reaches an exact recurrence before iteration seven thousand six hundred twelve on the ESP32 implementation <ref:2608.16120#pg1,10,000 random float32 initial conditions, every trajectory reaches an exact recurrence>.
Mira: That exact recurrence is key because it sets up a situation where the k-nearest neighbors predictor can return an identical future for those specific histories.
Lev: And that leads directly to the main point: when a long float32 record is used, the estimated LLE drops from zero point six eight five three at N = fifteen thousand down to zero point one eight two seven at N = twenty thousand and approximately zero by N = thirty thousand nine hundred ninety-nine as those exact train-test histories saturate <ref:2608.16120#pg1,to 0.1827 at N = 20,000 and approximately zero>.
Kai: So the number itself is what tells us that the estimate collapses because of this saturation of exact neighbors in the training set.
Mira: And they show that at a much larger budget, like N = one hundred thousand points, the long float32 record gives an LLE of zero point zero zero one six, while independently restarted short trajectories maintain a value near ln two <ref:2608.16120#pg1,the long float32 record gives>.
Lev: That comparison is telling because it shows that simply collecting more data doesn't keep the LLE stable if you have this kind of digital recurrence structure present in the map.
The paper's summary: Kai: To summarize what they found in "When more data become less informative: Finite-precision periodicization and collapse of forecast-error Lyapunov estimates," they are showing that data-driven LLE estimates can be fundamentally altered by finite precision.
Mira: They demonstrate that deterministic digital maps have a finite state space, which means once an exact machine state repeats, its complete future must also repeat. This forces the system into a cycle after a transient phase.
Lev: That forced periodicity is what causes the forecast error to flatten because the k-nearest neighbors predictor sees an identical future when it finds those exact copies in its training set.
Kai: The paper distinguishes this from other types of precision effects, like pseudo-orbit separation, which is used to estimate positive Lyapunov growth. This collapse is specific to how finite-state recurrence suppresses forecast error growth after a certain point.
Mira: They also found that the collapse scale isn't random; it correlates directly with the inherent digital period structure, specifically showing a link to Npred approximately equal to mu plus 3P divided by zero point seven.
Lev: That suggests the behavior isn't just about a single initial condition causing trouble; it’s tied to the underlying recurrence scale of the digital map itself across many seeds.
The paper's improvements: Kai: The authors suggest a few things they think are important improvements to how we look at this problem, focusing on acquisition protocols rather than changing the map itself.
Mira: They propose using an acquisition protocol called the restarted trajectory protocol. This involves distributing the total data budget among many independent short trajectories instead of one long continuous record.
Lev: That strategy is designed to preserve access to that pre-periodic chaotic transient, which is where the true instability lives, by preventing one single cycle from dominating the entire dataset.
Kai: They found that this protocol successfully delays the degradation of the LLE estimate compared to using a single long float32 record. The comparison at one hundred thousand points was key here <ref:2608.16120#pg1>.
Mira: So their suggestion is to shift focus away from just collecting more data and instead focus on *how* you collect that data—spreading it out.
Lev: And they found a limitation to this strategy too: the paper notes that finite-state collisions can still happen across those many independent trajectories if you go to very large total budgets.
Conclusion: Kai: So, wrapping up the "When more data become less informative: Finite-precision periodicization and collapse of forecast-error Lyapunov estimates," the main implication is that record length alone isn't a reliable measure of information content in reduced-precision chaotic data.
Mira: The practical conclusion is that when you're inferring the dynamics of an underlying real-valued system, long records from a finite-state implementation need to be checked for exact recurrence before you just assume they are supplying new dynamical information.
Lev: And for me, what we see is that the mechanism here is about recurrence and information saturation. A systematic decay of the estimated growth slope along with an increasing exact or near-exact neighbor fraction Fexact,K indicates the record is becoming dominated by recurrent patterns rather than supplying genuinely new dynamical information.
Kai: So for our audience, remember this: if you're working with reduced precision chaos, don't just collect more data; check for digital recurrence first.
Mira: And keep an eye on that exact neighbor fraction as a way to diagnose when your data is getting saturated by the system's own finite structure.
Lev: And for real hardware deployment, you need to be aware that long records might just be sampling duplicate futures from the digital cycle, which isn't necessarily the chaotic instability of the intended real-valued system.
Kai: That’s our take on this paper today regarding "When more data become less informative: Finite-precision periodicization and collapse of forecast-error Lyapunov estimates."
More episodes
- 2608.20276-Flip rate prediction in the double pendulum
- The day in conversation for 2026-11-18
- Daily Summary for 2026-11-18
- 2603.16051-Gradual emergence of temporal structures depending on the distance between neighboring callers in natural habitat of male treefrogs
- 2603.21801-Diffraction of deep-water solitons
- The day in conversation for 2026-11-17
- Daily Summary for 2026-11-17
- 2204.04335-Measuring and improving community resilience: a Fuzzy Logic approach
- 2604.11624-Prediction of chaotic dynamics from data: An introduction
- 2610.21575-Cooperation Dynamics with Environmental Feedback: Steady States and Transitions