An official website of the United States government
Eddy covariance serves as one the most effective techniques for long-term monitoring of ecosystem fluxes, however long-term data integrations rely on complete timeseries, meaning that any gaps due to missing data must be reliably filled. To date, many gap-filling approaches have been proposed and extensively evaluated for mature and/or less actively managed ecosystems. Random forest regression (RFR) has been shown to be stable and perform better in these systems than alternative approaches, particularly when filling longer gaps. However, the performance of RFR gap filling remains less certain in more challenging ecosystems, e.g., actively managed agri-ecosystems and following recent land-use change due to management disturbances, ecosystems with relatively low fluxes due to low signal to noise ratios, or for trace gases other than carbon dioxide (e.g., methane). In an extension to earlier work on gap filling global carbon dioxide, water, and energy fluxes, we assess the RFR approach for gap filling methane fluxes globally. We then investigate a range of gap-filling methodologies for carbon dioxide, water, energy, and methane fluxes in challenging ecosystems, including European managed pastures, Southeast Asian converted peatlands, and North American drylands. Our findings indicate that RFR is a competent alternative to existing research standard gap-filling algorithms. The marginal distribution sampling (MDS) is still suggested for filling short (< 12 days) gaps in carbon dioxide fluxes, but RFR is better for filling longer (> 30 days) gaps in carbon dioxide fluxes and also for gap filling other fluxes (e.g. sensible heat, latent energy and methane). In addition, using RFR with globally available reanalysis environmental drivers is effective when measured drivers are unavailable. Crucially, RFR was able to reliably fill cumulative fluxes for gaps > 3 moths and, unlike other common approaches, key environment-flux responses were preserved in the gap-filled data.
more »
« less
This content will become publicly available on August 22, 2026
Gap labels and asymptotic gap opening for full shifts
We discuss gap labeling for operators generated by the full shift over a compact subset of the real line. The set of Johnson–Schwartzman gap labels is the algebra generated by weights of clopen subsets of the support of the single-site distribution. Due to the presence of a dense set of periodic orbits, it is impossible to find a sampling function for which all gaps allowed by the gap labeling theorem open simultaneously. Nevertheless, for a suitable choice of the single-site distribution, we show that for generic sampling functions, each spectral gap opens in the large-coupling limit. Furthermore, we show that for other choices of weights there are gaps that cannot open for purely diagonal operators.
more »
« less
- PAR ID:
- 10647438
- Publisher / Repository:
- EMS Press
- Date Published:
- Journal Name:
- Journal of Spectral Theory
- Volume:
- 15
- Issue:
- 3
- ISSN:
- 1664-039X
- Page Range / eLocation ID:
- 1383 to 1407
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
From the general inverse theory of periodic Jacobi matrices, it is known that a periodic Jacobi matrix of minimal period p≥2 may have at most p−2 closed spectral gaps. We discuss the maximal number of closed gaps for one-dimensional periodic discrete Schrödinger operators of period p. We prove nontrivial upper and lower bounds on this quantity for large p and compute it exactly for p≤6. Among our results, we show that a discrete Schrödinger operator of period four or five may have at most a single closed gap, and we characterize exactly which potentials may exhibit a closed gap. For period six, we show that at most two gaps may close. In all cases in which the maximal number of closed gaps is computed, it is seen to be strictly smaller than p−2, the bound guaranteed by the inverse theory. We also discuss similar results for purely off-diagonal Jacobi matrices.more » « less
-
Machine learning models often perform poorly under subpopulation shifts in the data distribution. Developing methods that allow machine learning models to better generalize to such shifts is crucial for safe deployment in real-world settings. In this paper, we develop a family of group-aware prior (GAP) distributions over neural network parameters that explicitly favor models that generalize well under subpopulation shifts. We design a simple group-aware prior that only requires access to a small set of data with group information and demonstrate that training with this prior yields state-of-the-art performance -- even when only retraining the final layer of a previously trained non-robust model. Group aware-priors are conceptually simple, complementary to existing approaches, such as attribute pseudo labeling and data reweighting, and open up promising new avenues for harnessing Bayesian inference to enable robustness to subpopulation shifts.more » « less
-
Machine learning models often perform poorly under subpopulation shifts in the data distribution. Developing methods that allow machine learning models to better generalize to such shifts is crucial for safe deployment in real-world settings. In this paper, we develop a family of group-aware prior (GAP) distributions over neural network parameters that explicitly favor models that generalize well under subpopulation shifts. We design a simple group-aware prior that only requires access to a small set of data with group information and demonstrate that training with this prior yields state-of-the-art performance -- even when only retraining the final layer of a previously trained non-robust model. Group aware-priors are conceptually simple, complementary to existing approaches, such as attribute pseudo labeling and data reweighting, and open up promising new avenues for harnessing Bayesian inference to enable robustness to subpopulation shifts.more » « less
-
Abstract We consider discrete one-dimensional Schrödinger operators whose potentials are generated by sampling along the orbits of a general hyperbolic transformation. Specifically, we show that if the sampling function is a non-constant Hölder continuous function defined on a subshift of finite type with a fully supported ergodic measure admitting a local product structure and a fixed point, then the Lyapunov exponent is positive away from a discrete set of energies. Moreover, for sampling functions in a residual subset of the space of Hölder continuous functions, the Lyapunov exponent is positive everywhere. If we consider locally constant or globally fiber bunched sampling functions, then the Lyapuonv exponent is positive away from a finite set. Moreover, for sampling functions in an open and dense subset of the space in question, the Lyapunov exponent is uniformly positive. Our results can be applied to any subshift of finite type with ergodic measures that are equilibrium states of Hölder continuous potentials. In particular, we apply our results to Schrödinger operators defined over expanding maps on the unit circle, hyperbolic automorphisms of a finite-dimensional torus, and Markov chains.more » « less
