Skip to main content

3.10 Linearization and Multivariate Taylor Series

The gradient $\triangledown f$ of a function $f$ is often used for a locally linear approximation of $f$ around $x_0$.
$f(x) \approx f(x_0)+(\bigtriangledown_xf)(x_0)(x-x_0)$

Here $(\bigtriangledown_xf)(x_0)$ is the gradient of $f$ with respect to $x$, evaluated at $x_0$.
The following Figure illustrates the linear approximation of a function $f$ at an input $x_0$. The original function is approximated by a straight line.This approximation is locally accurate, but the farther we move away from $x_0$, the worse the approximation gets.This is a special case of multivariate Taylor series expansion of $f$ at $x_0$, where we consider only the first two terms.




Multivariate Taylor Series: Definition
We consider a function
$f:\mathbb{R}^D \to \mathbb{R}$
$x \to f(x), \quad x \in \mathbb{R}^D$
that is smooth at $x_0$. When we define the difference vector $\delta= x- x_0$, the multivariate Taylor series of $f$ at $x_0$ is defined as

$f(x)=\sum_{k=0}^{\infty } \frac{D_x^kf(x_0)}{k!} \delta^k$
where $D^k_xf(x_0)$ is the $k-th$ derivative of $f$ with respect to $x$, evaluated at $x_0$.

Taylor Polynomial: Definition
The Taylor polynomial of degree $n$ of $f$ at $x_0$ contains the first $n+1$ components of the Taylor series and is defined as
$T_n(x)=\sum_{k=0}^{n } \frac{D_x^kf(x_0)}{k!} \delta^k$

Note that both $D_x^kf$ and $\delta^k$ are $k-th$ order tensors.i.e., $k$ dimensional arrays. The $k$th order tensor $\delta^k \in \mathbb{R}^{D \times D \times \ldots \times D}$ is obtained as a $k$ fold outer product denoted by $\bigotimes$, of the vector $\delta \in \mathbb{R}^D$. For example

$\delta^2=\delta \bigotimes \delta=\delta \delta^T,\quad \delta^2[i,j]=\delta[i][\delta[j]$

In general we obtain the terms
$D_x^kf(x_0)\delta^k=\sum_{i_1=1}^D\cdots\sum_{i_k=1}^DD_x^kf(x_0)[i_1,\ldots,i_k]\delta[i_1]\cdots\delta[i_k]$

Lets write down the first few terms of $D_x^kf(x_0)\delta^k$ of the Taylor Series expansion for $k=0,\ldots,3$ and $\delta=x-x_0$.


Example: ( university question)
Consider the function $f(x,y)=x^2+2xy+y^3$. Find the Taylor series expansion of  $f$ at $(x_0,y_0)=(1,2)$
Note that we are looking for a Taylor series expansion, which itself is a linear combination of Polynomials.To express the polynomial of degree 3 , the Taylor series expansion must contain terms of 4th or higher order.

$f(1,2)=13$
$\frac{\partial f}{\partial x}=2x+2y=\frac{\partial f}{\partial x}(1,2)=6$
$\frac{\partial f}{\partial y}=2x+3y^2=\frac{\partial f}{\partial y}(1,2)=14$
Therefore we obtain
$D_{x,y}^1f(1,2)=\bigtriangledown_{x,y}f(1,2)=\left[ \frac{\partial f}{\partial x} \quad \frac{\partial f}{\partial y}\right]=[6 \quad 14] \in \mathbb{R}^{1 \times 2}$

$\frac{D_{x,y}^1f(1,2)}{1!}\delta=\left[6 \quad 14 \right]\begin{bmatrix}
x-1\\
y-2\end{bmatrix}=6(x-1)+14(y-2)$

Note that this contains only linear terms.i.e., first-order polynomials.
The second order partial derivatives are given by
 $\frac{\partial^2 f}{\partial x^2}=2 \Rightarrow \frac{\partial^2 f}{\partial x^2}(1,2)=2$
 $\frac{\partial^2 f}{\partial y^2}=6y \Rightarrow \frac{\partial^2 f}{\partial y^2}(1,2)=12$
 $\frac{\partial^2 f}{\partial y \partial x}=2 \Rightarrow \frac{\partial^2 f}{\partial y \partial x}(1,2)=2$
$\frac{\partial^2 f}{\partial x \partial y}=2 \Rightarrow \frac{\partial^2 f}{\partial x \partial y}(1,2)=2$

$H=\begin{bmatrix}
2 & 2\\
2& 6y
\end{bmatrix} \in \mathbb{R}^{2 \times 2}$

So the Hessian at $(1,2)$ is
$H(1,2)=\begin{bmatrix}
2 & 2\\
2& 12
\end{bmatrix} \in \mathbb{R}^{2 \times 2}$

Therefore the next term of the Taylor series expansion is given by
$\frac{D_{x,y}^2f(1,2)}{2!}\delta^2=\frac{1}{2}\delta^TH(1,2)\delta$
$\quad=\frac{1}{2}\begin{bmatrix}
x-1 & y-2
\end{bmatrix}\begin{bmatrix}
2 & 2\\
2& 12
\end{bmatrix} \begin{bmatrix}
x-1\\
y-2
\end{bmatrix}$
$=(x-1)^2+2(x-1)(y-2)+6(y-2)^2$
This is a second order polynomial

The third order derivatives are obtained as
$D^3_{x,y}f=\left[ \frac{\partial H}{\partial x}\quad \frac{\partial H}{\partial y}\right] \in \mathbb{R}^{2 \times 2 \times 2}$

$D^3_{x,y}f[:,:,1]=\frac{\partial H}{\partial x}=\begin{bmatrix}
\frac{\partial ^3f}{\partial x^3} & \frac{\partial ^3f}{\partial x^2 \partial y}\\
\frac{\partial ^3f}{\partial x \partial y \partial x}& \frac{\partial ^3f}{\partial x\partial y^2}
\end{bmatrix}$


$D^3_{x,y}f[:,:,2]=\frac{\partial H}{\partial y}=\begin{bmatrix}
\frac{\partial ^3f}{\partial y \partial x^2} & \frac{\partial ^3f}{\partial y \partial x \partial y}\\
\frac{\partial ^3f}{\partial y^2 \partial x}& \frac{\partial ^3f}{\partial y^3}
\end{bmatrix}$

Since most of the terms are constant in second order Hessian, the only non zero third order partial derivative  is
$\frac{\partial ^3f}{\partial y^3} =6 \Rightarrow \frac{\partial ^3f}{\partial y^3}(1,2)=6$
All other derivatives will vanish
$D^3_{x,y}f[:,:,1]=\begin{bmatrix}0 & 0\\
0& 0
\end{bmatrix}$
$D^3_{x,y}f[:,:,2]=\begin{bmatrix}0 & 0\\
0& 6
\end{bmatrix}$
$\frac{D^3_{x,y}f(1,2)}{3!}\delta^3=(y-2)^3$

Overall, the exact Taylor series expansion of $f$ at $(x_0,y_0)=(1,2)$ is
$f(x,y)=f(1,2)+D^1_{x,y}f(1,2)\delta+\frac{D^2_{x,y}f(1,2)}{2!}\delta^2+\frac{D^3_{x,y}f(1,2)}{3!}\delta^3$

$=13+6(x-1)+14(y-2)+(x-1)^2+2(x-1)(y-2)+6(y-2)^2+(y-2)^3 $
$=13+6x-6+14y-28+x^2+1-2x+2xy-4x-2y+4+6y^2+24-24y+y^3-8+12y-6y^2 $
$=x^2+2xy+y^3+(6x-6x)+(14y-2y-24y+12y)+(13-6-28+1+4+24-8)$
$=x^2+2xy+y^3$

Find the second order Taylor series expansion for $f(x,y) = e^{-(x^2+y^2)} cos(xy)$ about $(0 , 0)$.

$f(0,0)=1$

First order partial derivatives
$\frac{\partial f}{\partial x}=-2xe^{-(x^2+y^2)}cos(xy)-ye^{-(x^2+y^2)}sin(xy)$
$\frac{\partial f}{\partial x}=-e^{-(x^2+y^2)}(2x.cos(xy)+y.sin(xy))$
$\frac{\partial f}{\partial x}(0,0)=0$

$\frac{\partial f}{\partial y}=-2ye^{-(x^2+y^2)}cos(xy)-xe^{-(x^2+y^2)}sin(xy)$
$\frac{\partial f}{\partial y}=-e^{-(x^2+y^2)}(2y.cos(xy)+x.sin(xy))$
$\frac{\partial f}{\partial y}(0,0)=0$
$D_{x,y}^1f(0,0)=\bigtriangledown_{x,y}f(0,0)=\left[ \frac{\partial f}{\partial x} \quad \frac{\partial f}{\partial y}\right]=[0 \quad 0] \in \mathbb{R}^{1 \times 2}$
$\frac{D_{x,y}^1f(0,0)}{1!}\delta=\left[0 \quad 0 \right]\begin{bmatrix}
x\\
y\end{bmatrix}=0$

The second order partial derivatives are given by
$\frac{\partial^2 f}{\partial x^2}=e^{-(x^2+y^2)}.2x(2x.cos(xy)+y.sin(xy))-e^{(x^2+y^2)}(2.cos(xy)-2xy.sin(xy)+y^2.cos(xy))$
$\frac{\partial^2 f}{\partial x^2}=e^{-(x^2+y^2)}((4x^2-y^2-2)cos(xy)+4xy.sin(xy))$
$\frac{\partial^2 f}{\partial x^2}(0,0)=-2$
similarly
$\frac{\partial^2 f}{\partial y^2}=e^{-(x^2+y^2)}((4xy.sin(xy)-(x^2-4y^2+2)cos(xy))$
$\frac{\partial^2 f}{\partial y^2}(0,0)=-2$

$\frac{\partial^2 f}{\partial x \partial y}=\frac{\partial^2 f}{\partial y \partial x}=e^{-(x^2+y^2)}((2x^2+2y^2-1)sin(xy)+3xy.cos(xy))$
$\frac{\partial^2 f}{\partial x \partial y}=\frac{\partial^2 f}{\partial y \partial x}(0,0)=0$
$\frac{D_{x,y}^2f(0,0)}{2!}\delta^2=\frac{1}{2}\left[x \quad y \right]\begin{bmatrix}
-2 & 0\\
0 & -2
\end{bmatrix}\begin{bmatrix}x\\
y\end{bmatrix}=-x^2-y^2$

the second degree Taylor polynomial at point (0,0) is
$f(0,0)+\frac{D_{x,y}^1f(0,0)}{1!}\delta+\frac{D_{x,y}^2f(0,0)}{2!}\delta^2$
$=1+0-x^2-y^2$
$=1-x^2-y^2$

Find the linear approximation to the function $f(x,y) = 2 - sin(-x -3y)$ at the point (0 , π), and then use your answer to estimate $f(0.001 , π)$.
$f(0,\pi)=2-sin(0-3\pi)=2+sin(3\pi)=2$
First order partial derivatives
$\frac{\partial f}{\partial x}=cos(-x-3y)$
$\frac{\partial f}{\partial x}(0,\pi)=cos(-3\pi)=-1$

$\frac{\partial f}{\partial y}=3cos(-x-3y)$
$\frac{\partial f}{\partial y}(0,\pi)=3.cos(-3\pi)=-3$
$\frac{D_{x,y}^1f(0,0)}{1!}\delta=\left[-1\quad -3 \right]\begin{bmatrix}
x\\
y-\pi\end{bmatrix}=-x-3y+3\pi$

Linear approximation at $(0,\pi)$ is
$f(0,\pi)+\frac{D_{x,y}^1f(0,0)}{1!}\delta$
$=2-x-3y+3\pi$
value at $f(0.001 , π)$ is
$2-0.001-3\pi+3\pi$
$=1.999$

Find the second order Taylor series expansion for $f(x, y) = (x + y)^2$ about $(0 , 0)$

$f(0,0)=0$

First order partial derivatives
$\frac{\partial f}{\partial x}=2(x+y)$
$\frac{\partial f}{\partial x}(0,0)=0$
$\frac{\partial f}{\partial y}=2(x+y)$
$\frac{\partial f}{\partial y}(0,0)=0$
$\frac{D_{x,y}^1f(0,0)}{1!}\delta=\left[0 \quad 0 \right]
\begin{bmatrix}
x\\
y
\end{bmatrix}=0$


Second order partial derivatives
$\frac{\partial^2 f}{\partial x^2}=2$
$\frac{\partial^2 f}{\partial x^2}(0,0)=2$
$\frac{\partial^2 f}{\partial y^2}=2$
$\frac{\partial^2 f}{\partial y^2}(0,0)=2$

$\frac{\partial^2 f}{\partial x \partial y}=2$
$\frac{\partial^2 f}{\partial x \partial y}(0,0)=2$
$\frac{\partial^2 f}{\partial y \partial x}=2$
$\frac{\partial^2 f}{\partial y\partial x}(0,0)=2$
$\frac{D_{x,y}^2f(0,0)}{2!}\delta^2=\frac{1}{2}\left[x \quad y \right]\begin{bmatrix}
2 & 2\\
2 & 2
\end{bmatrix}\begin{bmatrix}x\\
y\end{bmatrix}=x^2+2xy+y^2$

the second order Taylor series expansion about(0,0) is
$f(0,0)+\frac{D_{x,y}^1f(0,0)}{1!}\delta+\frac{D_{x,y}^2f(0,0)}{2!}\delta^2$
$=0+0+x^2+2xy+y^2$
$=x^2+2xy+y^2$

Compute the Taylor series expansion of $f(x,y)=x^2+2xy+y^3$ at $(x_0,y_0)=(1,1)$

$f(1,1)=4$
First order partial derivatives
$\frac{\partial f}{\partial x}=2x+2y$
$\frac{\partial f}{\partial x}(1,1)=4$
$\frac{\partial f}{\partial y}=2x+3y^2$
$\frac{\partial f}{\partial y}(1,1)=5$
$\frac{D_{x,y}^1f(1,1)}{1!}\delta=\left[4 \quad 5 \right]
\begin{bmatrix}
x-1\\
y-1
\end{bmatrix}=4x+5y-9$

Second order partial derivatives
$\frac{\partial^2 f}{\partial x^2}=2$
$\frac{\partial^2 f}{\partial x^2}(1,1)=2$
$\frac{\partial^2 f}{\partial y^2}=6y$
$\frac{\partial^2 f}{\partial y^2}(1,1)=6$

$\frac{\partial^2 f}{\partial x \partial y}=2$
$\frac{\partial^2 f}{\partial x \partial y}(1,1)=2$
$\frac{\partial^2 f}{\partial y \partial x}=2$
$\frac{\partial^2 f}{\partial y\partial x}(1,1)=2$

$\frac{D_{x,y}^2f(1,1)}{2!}\delta^2=\frac{1}{2}\left[x-1 \quad y-1 \right]\begin{bmatrix}
2 & 2\\
2 & 6
\end{bmatrix}
\begin{bmatrix}
x-1\\
y-1\end{bmatrix}=x^2+3y^2-4x-8y+2xy+6$

the second order Taylor series expansion about(1,1) is
$f(1,1)+\frac{D_{x,y}^1f(1,1)}{1!}\delta+\frac{D_{x,y}^2f(1,1)}{2!}\delta^2$
$=4+(4x+5y-9)+x^2+3y^2-4x-8y+2xy+6$
$=x^2+3y^2-3y+2xy+1$

Compute the Taylor series expansion of $f(x,y)=x^2y+3y-2$ at $(x_0,y_0)=(1,-2)$

$f(x,y)=x^2y+3y-2$
$f(1,-2)=-10$
First order partial derivatives
$\frac{\partial f}{\partial x}=2xy$
$\frac{\partial f}{\partial x}(1,-2)=-4$
$\frac{\partial f}{\partial y}=x^2+3$
$\frac{\partial f}{\partial y}(1,-2)=4$
$\frac{D_{x,y}^1f(1,1)}{1!}\delta=\left[-4 \quad 4 \right]
\begin{bmatrix}
x-1\\
y+2
\end{bmatrix}=-4x+4y+12$

Second order partial derivatives
$\frac{\partial^2 f}{\partial x^2}=2y$
$\frac{\partial^2 f}{\partial x^2}(1,-2)=-4$
$\frac{\partial^2 f}{\partial y^2}=0$
$\frac{\partial^2 f}{\partial y^2}(1,-2)=0$

$\frac{\partial^2 f}{\partial x \partial y}=2x$
$\frac{\partial^2 f}{\partial x \partial y}(1,-2)=2$
$\frac{\partial^2 f}{\partial y \partial x}=2x$
$\frac{\partial^2 f}{\partial y\partial x}(1,-2)=2$

$\frac{D_{x,y}^2f(1,-2)}{2!}\delta^2=\frac{1}{2}\left[x-1 \quad y+2 \right]\begin{bmatrix}
-4 & 2\\
2 & 0
\end{bmatrix}
\begin{bmatrix}
x-1\\
y+2\end{bmatrix}=-2x^2+12x-3y+4xy-8$

the second order Taylor series expansion about(1,-2) is
$f(1,-2)+\frac{D_{x,y}^1f(1,-2)}{1!}\delta+\frac{D_{x,y}^2f(1,-2)}{2!}\delta^2$
$=-10+(-4x+4y+12)-2x^2+12x-3y+4xy-8$
$=-2x^2+8x+y-6$

Comments

Popular posts from this blog

Mathematics for Machine Learning- CST 284 - KTU Minor Notes - Dr Binu V P

  Introduction About Me Syllabus Course Outcomes and Model Question Paper University Question Papers and Evaluation Scheme -Mathematics for Machine learning CST 284 KTU Overview of Machine Learning What is Machine Learning (video) Learn the Seven Steps in Machine Learning (video) Linear Algebra in Machine Learning Module I- Linear Algebra 1.Geometry of Linear Equations (video-Gilbert Strang) 2.Elimination with Matrices (video-Gilbert Strang) 3.Solving System of equations using Gauss Elimination Method 4.Row Echelon form and Reduced Row Echelon Form -Python Code 5.Solving system of equations Python code 6. Practice problems Gauss Elimination ( contact) 7.Finding Inverse using Gauss Jordan Elimination  (video) 8.Finding Inverse using Gauss Jordan Elimination-Python code Vectors in Machine Learning- Basics 9.Vector spaces and sub spaces 10.Linear Independence 11.Linear Independence, Basis and Dimension (video) 12.Generating set basis and span 13.Rank of a Matrix 14.Linear Mapping and Matr

4.3 Sum Rule, Product Rule, and Bayes’ Theorem

 We think of probability theory as an extension to logical reasoning Probabilistic modeling  provides a principled foundation for designing machine learning methods. Once we have defined probability distributions corresponding to the uncertainties of the data and our problem, it turns out that there are only two fundamental rules, the sum rule and the product rule. Let $p(x,y)$ is the joint distribution of the two random variables $x, y$. The distributions $p(x)$ and $p(y)$ are the corresponding marginal distributions, and $p(y |x)$ is the conditional distribution of $y$ given $x$. Sum Rule The addition rule states the probability of two events is the sum of the probability that either will happen minus the probability that both will happen. The addition rule is: $P(A∪B)=P(A)+P(B)−P(A∩B)$ Suppose $A$ and $B$ are disjoint, their intersection is empty. Then the probability of their intersection is zero. In symbols:  $P(A∩B)=0$  The addition law then simplifies to: $P(A∪B)=P(A)+P(B)$  wh

5.1 Optimization using Gradient Descent

Since machine learning algorithms are implemented on a computer, the mathematical formulations are expressed as numerical optimization methods.Training a machine learning model often boils down to finding a good set of parameters. The notion of “good” is determined by the objective function or the probabilistic model. Given an objective function, finding the best value is done using optimization algorithms. There are two main branches of continuous optimization constrained and unconstrained. By convention, most objective functions in machine learning are intended to be minimized, that is, the best value is the minimum value. Intuitively finding the best value is like finding the valleys of the objective function, and the gradients point us uphill. The idea is to move downhill (opposite to the gradient) and hope to find the deepest point. For unconstrained optimization, this is the only concept we need,but there are several design choices. For constrained optimization, we need to intr