The paper develops a new version of the Kalman filter, a classic mathematical tool used to estimate hidden states of a system from noisy measurements. The twist here is that the underlying signal evolves according to a time-fractional differential equation, meaning its rate of change depends not just on the present moment but on its entire past history. This kind of "memory" in the dynamics is more realistic for many real-world processes than the standard assumption that the future depends only on the current state. The authors derive the key equations needed to track and update best guesses about the hidden state over time, including a formula for how estimation error evolves.
The mathematical core of the work involves two integral equations. The first gives the conditional mean, meaning the best estimate of the hidden signal given all observations so far. The second is a Riccati-Volterra type equation describing how the mean-square error, essentially a measure of uncertainty in the estimate, changes over time. These equations are more complex than their classical counterparts because the fractional time derivative introduces convolution-type terms that couple the past and present in a non-trivial way.
As a concrete application, the authors apply their framework to model how children with developmental dyscalculia, a learning disability affecting numerical understanding, acquire mathematical skills over time. The idea is that learning trajectories in such children may not follow simple Markovian dynamics but instead reflect accumulated experience and persistent difficulties, making the memory-dependent fractional model a natural fit. The filter then provides a principled statistical method for tracking and predicting a child's learning state from observed performance data.