Autocorrelation Reintroduces Spectral Bias in KANs for Time Series Forecasting
A new study reveals that Kolmogorov-Arnold Networks (KANs), previously thought to overcome spectral bias under independent inputs, suffer from reintroduced spectral bias when applied to time series forecasting due to strong temporal autocorrelation in lagged observations. Theoretical analysis and empirical validation show that the bias intensifies with higher autocorrelation, challenging the applicability of standard KANs for such tasks. To mitigate this, the authors propose using the Discrete Cosine Transform (DCT) to decorrelate inputs, which experimentally reduces the low-frequency preference in time series forecasting. The findings are detailed in arXiv:2604.23518.
Key facts
- KANs were thought to overcome spectral bias under independent inputs.
- Time series inputs are lagged observations with strong temporal autocorrelation.
- Temporal autocorrelation reintroduces spectral bias in KANs.
- Bias becomes more pronounced as autocorrelation increases.
- Standard KANs face substantial difficulties with strongly autocorrelated inputs.
- Discrete Cosine Transform (DCT) reduces correlations among network inputs.
- DCT preprocessing substantially reduces low-frequency preference in TSF.
- Study published on arXiv with ID 2604.23518.
Entities
Institutions
- arXiv