Slide 1

Slide 1 text

Transition Matrix Estimation in High Dimensional Time Series http://proceedings.mlr.press/v28/han13a.pdf Presenter: Elizabeth Ramirez - @eramirem

Slide 2

Slide 2 text

whoami Electrical Engineer Digital Signal Processing Applied Mathematician Computational Science and Engineering Applied Scientist Commodities. Spend a lot of my time worrying about how to process remote sensing data computationally efficient.

Slide 3

Slide 3 text

No content

Slide 4

Slide 4 text

what is a transition matrix Matrix whose product with the initial state vector produces the state vector at a later time. This is not the same as a stochastic matrix, which represents the transition probabilities in a Markov chain. The transition matrix is TIME INVARIANT.

Slide 5

Slide 5 text

what is a vector autoregressive (VAR) model Model that capture linear interdependencies among mutivariable timeseries. Each variable evolves based on its own lagged values, lagged values of other variables and error terms.

Slide 6

Slide 6 text

what is spectral norm The maximum singular value of a matrix, i.e. maximum scale by which the matrix can stretch a vector. A.K.A. natural norm induced by L2-norm

Slide 7

Slide 7 text

what is a stationary process Stochastic process in whose unconditional joint probability distribution does not change when shifted in time, i.e. mean and variance don't change over time. As you might imagine, many stochastic processes are non-stationary, but as we do in CSE world: non-linear becomes linear non-stationary becomes stationary Most common cause of stationarity violation: trend

Slide 8

Slide 8 text

what is marginal / lagged covariance Marginal: the opposite to conditional. Direct representation of the covariance of the timeseries, without the effects of other variables. Relies on the marginal distribution of the variables. Lagged: Covariance of time series with time-lagged time series

Slide 9

Slide 9 text

why we wanna estimate a transition matrix. You might know the equations that govern your system. If you know the dynamics of your system you should be able to calculate your matrix in a closed form. If you have non-linearities in your system you try to linearize and calculate the matrix. But if you HAVE NO IDEA of the dynamics of your system, you need a good estimator.

Slide 10

Slide 10 text

what we want to prove 1) The error of the proposed method calculating the transition matrix of a VAR model is smaller than the one obtained using LSE and ridge/lasso penalty 2) Than this error is bounded

Slide 11

Slide 11 text

1. Introduction Time series where Follows a stationary process given by ,

Slide 12

Slide 12 text

1. Introduction Optimization problem using Least Squares Estimator

Slide 13

Slide 13 text

1. Introduction Build a new estimator that satisfies:

Slide 14

Slide 14 text

2. Background Induced Matrix Norm

Slide 15

Slide 15 text

Proposition 2.1. Suppose we have Let: Then:

Slide 16

Slide 16 text

Proposition 2.1. Proof

Slide 17

Slide 17 text

3. Methods and Algorithms Marginal and lag one covariance matrices:

Slide 18

Slide 18 text

3. Methods and Algorithms Using Proposition 2.1. results subject to

Slide 19

Slide 19 text

3. Methods and Algorithms Equivalent to calculate column by column: subject to

Slide 20

Slide 20 text

3. Methods and Algorithms Vector decomposition subject to and

Slide 21

Slide 21 text

3. Methods and Algorithms subject to

Slide 22

Slide 22 text

4.1. Main Result Classes of Matrices

Slide 23

Slide 23 text

Theorem 4.1. With

Slide 24

Slide 24 text

thank you!