Tensor Autoregression: Compressing Climate into a Few Numbers Without Losing Meaning ✛
Context
Climatologists have tracked El Niño for decades, but the exact location of its "basin" in the equatorial Pacific remains a matter of debate. Ocean temperature data form a tensor (three-dimensional array) of size 51×41×7, where each slice corresponds to longitude, latitude, and depth. Direct autoregression would require estimating 14637² parameters, which is impossible with 503 observations. Existing approaches based on Tucker decomposition provide interpretation but lose efficiency at high tensor order, while CP decomposition is efficient but lacks a clear statistical interpretation. The work sits at the intersection of dimensionality reduction and time series analysis, alongside studies of the large-scale structure of the universe and machine learning, but differs in its focus on tensor autoregression with supervised factor modeling.
Methods
The authors use CP decomposition (canonical polyadic) to extract features from the response tensor and the covariate tensor separately, then build a linear regression between the resulting features. The autoregressive coefficient tensor is represented as a product of loading matrices, giving a low-rank structure. A sparse tensor is added to account for heterogeneous signals, allowing local anomalies to be captured. Parameter estimation is performed via least squares with an alternating algorithm (ALS), where each step solves a standard linear regression. Hyperparameters are selected using a hold-out sample and rolling forecast.
Results
The model with low-rank and sparse components achieved the lowest prediction errors on ocean temperature data: mean squared error of 0.271 versus 0.361 for the purely low-rank model and 0.373–0.440 for CP, Tucker, and Lasso approaches. Spatial loading maps revealed regions corresponding to the El Niño basin, confirming physical interpretability. Theoretically, non-asymptotic error bounds are established, growing as the square root of model complexity divided by sample size. Simulations confirmed the convergence of estimates and the influence of the non-identifiability radius on separating low-rank and sparse parts.
Implications
This work is the first to combine the interpretability of factor modeling with the efficiency of CP decomposition in tensor autoregression, overcoming limitations of previous methods. Unlike neighboring works using Tucker decomposition or pure sparsity, the proposed approach maintains linear growth in the number of parameters with tensor order while also identifying significant regions. This enables analysis of high-dimensional time series without losing physical meaning, which is critical for climatology and other data sciences.
Future development
Possible directions include adding sparsity to loading matrices to further reduce sample size requirements, constructing confidence intervals for coefficients, and using tensor ring decomposition to improve representation accuracy. Also promising is separating periods before and during El Niño to improve interpretation of covariate features.
Impact
The method is applicable in climatology, neurobiology, economics, and finance, where data are naturally represented as tensors and exhibit temporal dependence.
Next steps
Test the model on other climate datasets and investigate statistical properties of the estimates to draw conclusions about coefficient significance.
Key open problems
In time series statistics, the question of balancing interpretability and efficiency of dimensionality reduction for tensor data remains open. This work directly addresses this problem by proposing a structure that preserves both properties, paving the way for more complex models with dynamic factors.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors build an autoregressive model for tensor time series in which the coefficients have a CP low-rank structure, while a sparse component captures local anomalies. Parameter estimation is performed via least squares with an alternating algorithm, and hyperparameters are selected on a hold-out sample. On ocean temperature data, the model outperforms its counterparts in forecast accuracy and highlights regions corresponding to the El Niño basin.
Where its strength is
Combining CP decomposition with a sparse component keeps the number of parameters growing linearly with the tensor order while still singling out significant regions, which yields interpretability without sacrificing efficiency.
Where it could go next
- It might be interesting to see how the proposed tensor autoregression combines with drought forecasting methods based on reverse reading of weather data, where interpretability and handling multivariate series are also important.based on: How a neural network, reading weather backwards, warns of drought months in advance arXiv:2607.00331
- Nearby lies an approach to reconstructing the neutron star equation of state using AI, which employs decomposition and interpretable features; perhaps one could transfer their techniques for extracting physically meaningful components to climate tensors.based on: Listening to the Stars: AI Decoder Instantly Reconstructs the Neutron Star Equation of State arXiv:2607.12722
- The archive also contains work on cosmic web detection (JADE) that applies dimensionality reduction methods to large astrophysical datasets; one could consider adapting their ideas for analyzing spatiotemporal climate structures.based on: JADE: The AI Detective That Reads the Cosmic Web in Light Trails arXiv:2606.31988
Worth trying
- Use a rolling forecast with a variable horizon to assess how stable the identified sparse anomalies are over time.
- Apply the method to drought data from paper 2607.00331 and compare the interpretability of the identified regions with their approaches.
- Check how adding the sparse component affects the stability of CP factors when data contain missing values.
Why this may matter
The proposed method could prove useful for analyzing any high-dimensional tensor time series where one needs to balance forecast accuracy against physical interpretability, especially in climatology and neuroscience.
Nearby works in our archive
- How to Track Heat Across an Entire Room with Just One Thermometersimilarity 0.49
- Listening to the Stars: AI Decoder Instantly Reconstructs the Neutron Star Equation of Statesimilarity 0.48
- How a neural network, reading weather backwards, warns of drought months in advancesimilarity 0.48
- Neural Network by Ear: How to Capture a Particle's Energysimilarity 0.48
- Cloud Spices: How a Neural Network Improves Climatesimilarity 0.48
- Thermometers in Plasma Lie in Unisonsimilarity 0.47
- JADE: The AI Detective That Reads the Cosmic Web in Light Trailssimilarity 0.46
The archive is sparse around this work: the closest review is only 0.49, against a usual 0.64. It looks like few have drilled nearby.
The ocean temperature tensor contains 14637 elements, but the proposed model uses only about 300 parameters for prediction.
Key numbers
- tensor size: 51×41×7
- number of parameters in full model: 14637²
- number of observations: 503
- mean squared prediction error: 0.271