Mixing
Explanation of generalization of Newton's Method for multiple dimensions
Clash Royale CLAN TAG#URR8PPP
.everyoneloves__top-leaderboard:empty,.everyoneloves__mid-leaderboard:empty margin-bottom:0;
up vote
2
down vote
favorite
I've been following the CS 229 lecture videos for machine learning, and in lecture 4 (~14:00), Ng explains Newton's Method for optimization to maximize an objective function ($f$), but doesn't clearly explain the derivation of the higher dimension generalization:
$$
theta := H^-1 nabla_theta f(theta)
$$
I've read that the Hessian matrix ($H$) is a multiple dimension generalization of the second order derivative, but other than that I'm not sure I understand the formulation of the Hessian or why we multiply by its inverse; perhaps I need to brush up on linear algebra.
What's an intuitive explanation of the Hessian and its inverse in this formula?
optimization linear-algebra matrix-inverse hessian
asked 5 hours ago
Michael Yang
112
New contributor
Michael Yang is a new contributor to this site. Take care in asking for clarification, commenting, and answering.
Check out our Code of Conduct.
add a comment |Â
up vote
2
down vote
favorite
I've been following the CS 229 lecture videos for machine learning, and in lecture 4 (~14:00), Ng explains Newton's Method for optimization to maximize an objective function ($f$), but doesn't clearly explain the derivation of the higher dimension generalization:
$$
theta := H^-1 nabla_theta f(theta)
$$
I've read that the Hessian matrix ($H$) is a multiple dimension generalization of the second order derivative, but other than that I'm not sure I understand the formulation of the Hessian or why we multiply by its inverse; perhaps I need to brush up on linear algebra.
What's an intuitive explanation of the Hessian and its inverse in this formula?
optimization linear-algebra matrix-inverse hessian
asked 5 hours ago
Michael Yang
112
New contributor
Michael Yang is a new contributor to this site. Take care in asking for clarification, commenting, and answering.
Check out our Code of Conduct.
add a comment |Â
up vote
2
down vote
favorite
up vote
2
down vote
favorite
I've been following the CS 229 lecture videos for machine learning, and in lecture 4 (~14:00), Ng explains Newton's Method for optimization to maximize an objective function ($f$), but doesn't clearly explain the derivation of the higher dimension generalization:
$$
theta := H^-1 nabla_theta f(theta)
$$
I've read that the Hessian matrix ($H$) is a multiple dimension generalization of the second order derivative, but other than that I'm not sure I understand the formulation of the Hessian or why we multiply by its inverse; perhaps I need to brush up on linear algebra.
What's an intuitive explanation of the Hessian and its inverse in this formula?
optimization linear-algebra matrix-inverse hessian
asked 5 hours ago
Michael Yang
112
New contributor
Michael Yang is a new contributor to this site. Take care in asking for clarification, commenting, and answering.
Check out our Code of Conduct.
I've been following the CS 229 lecture videos for machine learning, and in lecture 4 (~14:00), Ng explains Newton's Method for optimization to maximize an objective function ($f$), but doesn't clearly explain the derivation of the higher dimension generalization:
$$
theta := H^-1 nabla_theta f(theta)
$$
I've read that the Hessian matrix ($H$) is a multiple dimension generalization of the second order derivative, but other than that I'm not sure I understand the formulation of the Hessian or why we multiply by its inverse; perhaps I need to brush up on linear algebra.
What's an intuitive explanation of the Hessian and its inverse in this formula?
optimization linear-algebra matrix-inverse hessian
optimization linear-algebra matrix-inverse hessian
asked 5 hours ago
Michael Yang
112
New contributor
Michael Yang is a new contributor to this site. Take care in asking for clarification, commenting, and answering.
Check out our Code of Conduct.
asked 5 hours ago
Michael Yang
112
New contributor
Michael Yang is a new contributor to this site. Take care in asking for clarification, commenting, and answering.
Check out our Code of Conduct.
asked 5 hours ago
Michael Yang
112
New contributor
Michael Yang is a new contributor to this site. Take care in asking for clarification, commenting, and answering.
Check out our Code of Conduct.
asked 5 hours ago
Michael Yang
112
asked 5 hours ago
Michael Yang
112
112
New contributor
Michael Yang is a new contributor to this site. Take care in asking for clarification, commenting, and answering.
Check out our Code of Conduct.
New contributor
Michael Yang is a new contributor to this site. Take care in asking for clarification, commenting, and answering.
Check out our Code of Conduct.
Michael Yang is a new contributor to this site. Take care in asking for clarification, commenting, and answering.
Check out our Code of Conduct.
add a comment |Â
add a comment |Â
1 Answer
1
active
oldest
votes
up vote
3
down vote
Overview
Suppose we want to minimize an objective function $f$ that maps a parameter vector $x in mathbbR^d$ to a scalar value. The idea behind Newton's method is to locally approximate $f$ with a quadratic function, then solve for the point that minimizes this approximation. We then jump to this new point, where we build a new approximation, and repeat the process until convergence.
Local quadratic approximation
Say our current location in parameter space is the vector $x_0$. We can approximate $f$ in the vicinity of $x_0$ using the second order Taylor series expansion:
$$f(x) approx f_T(x) =
f(x_0) + (x-x_0)^T nabla f(x_0) + frac12(x-x_0)^T H(x_0) (x-x_0)$$
This is a quadratic function whose value at $x_0$, as well as its first and second partial derivatives, match those of $f$. $nabla f(x_0)$ is the gradient of $f$ evaluated at $x_0$. This is a vector containing the partial derivatives of $f$, and is analogous to the derivative for functions of single variables. Similarly, $H(x_0)$ is the Hessian of $f$ evaluated at $x_0$. It's a matrix containing second partial derivatives of $f$, and is analogous to the second derivative for functions of single variables.
$$nabla f(x_0) = left [ matrix
fracpartial fpartial x_1 (x_0) \
vdots \
fracpartial fpartial x_d (x_0) \
right ]
quad
H(x_0) = left [ matrix
fracpartial^2 fpartial x_1^2 (x_0) &
cdots &
fracpartial^2 fpartial x_1 partial x_d (x_0) \
vdots & ddots & vdots \
fracpartial^2 fpartial x_d partial x_1 (x_0) &
cdots &
fracpartial^2 fpartial x_d^2 (x_0) \
right ]$$
Minimizing the local approximation
Given the local approximation $f_T$ at the current point $x_0$, we want to find a new point $x$ that minimizes $f_T$. This can be done in closed form by taking the gradient of $f_T$ with respect to $x$, setting it to zero, then solving for $x$.
The gradient of $f_T$ can be found by differentiating the expression above (keeping in mind that $nabla f(x_0)$ is just a fixed vector and $H(x_0)$ is just a fixed matrix):
$$nabla f_T(x) = nabla f(x_0) + H(x_0) (x - x_0)$$
Setting it to zero, we have:
$$H(x_0) (x - x_0) = -nabla f(x_0)$$
We need to isolate $x$, so multiply both sides by the inverse of the Hessian, then move $x_0$ to the righthand side:
$$x = x_0 - Big( H(x_0) Big)^-1 nabla f(x_0)$$
answered 51 mins ago
user20160
14.5k12351
add a comment |Â
1 Answer
1
active
oldest
votes
1 Answer
1
active
oldest
votes
active
oldest
votes
active
oldest
votes
up vote
3
down vote
Overview
Suppose we want to minimize an objective function $f$ that maps a parameter vector $x in mathbbR^d$ to a scalar value. The idea behind Newton's method is to locally approximate $f$ with a quadratic function, then solve for the point that minimizes this approximation. We then jump to this new point, where we build a new approximation, and repeat the process until convergence.
Local quadratic approximation
Say our current location in parameter space is the vector $x_0$. We can approximate $f$ in the vicinity of $x_0$ using the second order Taylor series expansion:
$$f(x) approx f_T(x) =
f(x_0) + (x-x_0)^T nabla f(x_0) + frac12(x-x_0)^T H(x_0) (x-x_0)$$
This is a quadratic function whose value at $x_0$, as well as its first and second partial derivatives, match those of $f$. $nabla f(x_0)$ is the gradient of $f$ evaluated at $x_0$. This is a vector containing the partial derivatives of $f$, and is analogous to the derivative for functions of single variables. Similarly, $H(x_0)$ is the Hessian of $f$ evaluated at $x_0$. It's a matrix containing second partial derivatives of $f$, and is analogous to the second derivative for functions of single variables.
$$nabla f(x_0) = left [ matrix
fracpartial fpartial x_1 (x_0) \
vdots \
fracpartial fpartial x_d (x_0) \
right ]
quad
H(x_0) = left [ matrix
fracpartial^2 fpartial x_1^2 (x_0) &
cdots &
fracpartial^2 fpartial x_1 partial x_d (x_0) \
vdots & ddots & vdots \
fracpartial^2 fpartial x_d partial x_1 (x_0) &
cdots &
fracpartial^2 fpartial x_d^2 (x_0) \
right ]$$
Minimizing the local approximation
Given the local approximation $f_T$ at the current point $x_0$, we want to find a new point $x$ that minimizes $f_T$. This can be done in closed form by taking the gradient of $f_T$ with respect to $x$, setting it to zero, then solving for $x$.
The gradient of $f_T$ can be found by differentiating the expression above (keeping in mind that $nabla f(x_0)$ is just a fixed vector and $H(x_0)$ is just a fixed matrix):
$$nabla f_T(x) = nabla f(x_0) + H(x_0) (x - x_0)$$
Setting it to zero, we have:
$$H(x_0) (x - x_0) = -nabla f(x_0)$$
We need to isolate $x$, so multiply both sides by the inverse of the Hessian, then move $x_0$ to the righthand side:
$$x = x_0 - Big( H(x_0) Big)^-1 nabla f(x_0)$$
answered 51 mins ago
user20160
14.5k12351
add a comment |Â
up vote
3
down vote
Overview
Suppose we want to minimize an objective function $f$ that maps a parameter vector $x in mathbbR^d$ to a scalar value. The idea behind Newton's method is to locally approximate $f$ with a quadratic function, then solve for the point that minimizes this approximation. We then jump to this new point, where we build a new approximation, and repeat the process until convergence.
Local quadratic approximation
Say our current location in parameter space is the vector $x_0$. We can approximate $f$ in the vicinity of $x_0$ using the second order Taylor series expansion:
$$f(x) approx f_T(x) =
f(x_0) + (x-x_0)^T nabla f(x_0) + frac12(x-x_0)^T H(x_0) (x-x_0)$$
This is a quadratic function whose value at $x_0$, as well as its first and second partial derivatives, match those of $f$. $nabla f(x_0)$ is the gradient of $f$ evaluated at $x_0$. This is a vector containing the partial derivatives of $f$, and is analogous to the derivative for functions of single variables. Similarly, $H(x_0)$ is the Hessian of $f$ evaluated at $x_0$. It's a matrix containing second partial derivatives of $f$, and is analogous to the second derivative for functions of single variables.
$$nabla f(x_0) = left [ matrix
fracpartial fpartial x_1 (x_0) \
vdots \
fracpartial fpartial x_d (x_0) \
right ]
quad
H(x_0) = left [ matrix
fracpartial^2 fpartial x_1^2 (x_0) &
cdots &
fracpartial^2 fpartial x_1 partial x_d (x_0) \
vdots & ddots & vdots \
fracpartial^2 fpartial x_d partial x_1 (x_0) &
cdots &
fracpartial^2 fpartial x_d^2 (x_0) \
right ]$$
Minimizing the local approximation
Given the local approximation $f_T$ at the current point $x_0$, we want to find a new point $x$ that minimizes $f_T$. This can be done in closed form by taking the gradient of $f_T$ with respect to $x$, setting it to zero, then solving for $x$.
The gradient of $f_T$ can be found by differentiating the expression above (keeping in mind that $nabla f(x_0)$ is just a fixed vector and $H(x_0)$ is just a fixed matrix):
$$nabla f_T(x) = nabla f(x_0) + H(x_0) (x - x_0)$$
Setting it to zero, we have:
$$H(x_0) (x - x_0) = -nabla f(x_0)$$
We need to isolate $x$, so multiply both sides by the inverse of the Hessian, then move $x_0$ to the righthand side:
$$x = x_0 - Big( H(x_0) Big)^-1 nabla f(x_0)$$
answered 51 mins ago
user20160
14.5k12351
add a comment |Â
up vote
3
down vote
up vote
3
down vote
Overview
Suppose we want to minimize an objective function $f$ that maps a parameter vector $x in mathbbR^d$ to a scalar value. The idea behind Newton's method is to locally approximate $f$ with a quadratic function, then solve for the point that minimizes this approximation. We then jump to this new point, where we build a new approximation, and repeat the process until convergence.
Local quadratic approximation
Say our current location in parameter space is the vector $x_0$. We can approximate $f$ in the vicinity of $x_0$ using the second order Taylor series expansion:
$$f(x) approx f_T(x) =
f(x_0) + (x-x_0)^T nabla f(x_0) + frac12(x-x_0)^T H(x_0) (x-x_0)$$
This is a quadratic function whose value at $x_0$, as well as its first and second partial derivatives, match those of $f$. $nabla f(x_0)$ is the gradient of $f$ evaluated at $x_0$. This is a vector containing the partial derivatives of $f$, and is analogous to the derivative for functions of single variables. Similarly, $H(x_0)$ is the Hessian of $f$ evaluated at $x_0$. It's a matrix containing second partial derivatives of $f$, and is analogous to the second derivative for functions of single variables.
$$nabla f(x_0) = left [ matrix
fracpartial fpartial x_1 (x_0) \
vdots \
fracpartial fpartial x_d (x_0) \
right ]
quad
H(x_0) = left [ matrix
fracpartial^2 fpartial x_1^2 (x_0) &
cdots &
fracpartial^2 fpartial x_1 partial x_d (x_0) \
vdots & ddots & vdots \
fracpartial^2 fpartial x_d partial x_1 (x_0) &
cdots &
fracpartial^2 fpartial x_d^2 (x_0) \
right ]$$
Minimizing the local approximation
Given the local approximation $f_T$ at the current point $x_0$, we want to find a new point $x$ that minimizes $f_T$. This can be done in closed form by taking the gradient of $f_T$ with respect to $x$, setting it to zero, then solving for $x$.
The gradient of $f_T$ can be found by differentiating the expression above (keeping in mind that $nabla f(x_0)$ is just a fixed vector and $H(x_0)$ is just a fixed matrix):
$$nabla f_T(x) = nabla f(x_0) + H(x_0) (x - x_0)$$
Setting it to zero, we have:
$$H(x_0) (x - x_0) = -nabla f(x_0)$$
We need to isolate $x$, so multiply both sides by the inverse of the Hessian, then move $x_0$ to the righthand side:
$$x = x_0 - Big( H(x_0) Big)^-1 nabla f(x_0)$$
answered 51 mins ago
user20160
14.5k12351
Overview
Suppose we want to minimize an objective function $f$ that maps a parameter vector $x in mathbbR^d$ to a scalar value. The idea behind Newton's method is to locally approximate $f$ with a quadratic function, then solve for the point that minimizes this approximation. We then jump to this new point, where we build a new approximation, and repeat the process until convergence.
Local quadratic approximation
Say our current location in parameter space is the vector $x_0$. We can approximate $f$ in the vicinity of $x_0$ using the second order Taylor series expansion:
$$f(x) approx f_T(x) =
f(x_0) + (x-x_0)^T nabla f(x_0) + frac12(x-x_0)^T H(x_0) (x-x_0)$$
This is a quadratic function whose value at $x_0$, as well as its first and second partial derivatives, match those of $f$. $nabla f(x_0)$ is the gradient of $f$ evaluated at $x_0$. This is a vector containing the partial derivatives of $f$, and is analogous to the derivative for functions of single variables. Similarly, $H(x_0)$ is the Hessian of $f$ evaluated at $x_0$. It's a matrix containing second partial derivatives of $f$, and is analogous to the second derivative for functions of single variables.
$$nabla f(x_0) = left [ matrix
fracpartial fpartial x_1 (x_0) \
vdots \
fracpartial fpartial x_d (x_0) \
right ]
quad
H(x_0) = left [ matrix
fracpartial^2 fpartial x_1^2 (x_0) &
cdots &
fracpartial^2 fpartial x_1 partial x_d (x_0) \
vdots & ddots & vdots \
fracpartial^2 fpartial x_d partial x_1 (x_0) &
cdots &
fracpartial^2 fpartial x_d^2 (x_0) \
right ]$$
Minimizing the local approximation
Given the local approximation $f_T$ at the current point $x_0$, we want to find a new point $x$ that minimizes $f_T$. This can be done in closed form by taking the gradient of $f_T$ with respect to $x$, setting it to zero, then solving for $x$.
The gradient of $f_T$ can be found by differentiating the expression above (keeping in mind that $nabla f(x_0)$ is just a fixed vector and $H(x_0)$ is just a fixed matrix):
$$nabla f_T(x) = nabla f(x_0) + H(x_0) (x - x_0)$$
Setting it to zero, we have:
$$H(x_0) (x - x_0) = -nabla f(x_0)$$
We need to isolate $x$, so multiply both sides by the inverse of the Hessian, then move $x_0$ to the righthand side:
$$x = x_0 - Big( H(x_0) Big)^-1 nabla f(x_0)$$
answered 51 mins ago
user20160
14.5k12351
answered 51 mins ago
user20160
14.5k12351
answered 51 mins ago
user20160
14.5k12351
answered 51 mins ago
user20160
14.5k12351
14.5k12351
add a comment |Â
add a comment |Â
Michael Yang is a new contributor. Be nice, and check out our Code of Conduct.
Michael Yang is a new contributor. Be nice, and check out our Code of Conduct.
Michael Yang is a new contributor. Be nice, and check out our Code of Conduct.
Michael Yang is a new contributor. Be nice, and check out our Code of Conduct.
Sign up or log in
StackExchange.ready(function ()
StackExchange.helpers.onClickDraftSave('#login-link');
);
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
StackExchange.ready(
function ()
StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fstats.stackexchange.com%2fquestions%2f369848%2fexplanation-of-generalization-of-newtons-method-for-multiple-dimensions%23new-answer', 'question_page');
);
Post as a guest
Sign up or log in
StackExchange.ready(function ()
StackExchange.helpers.onClickDraftSave('#login-link');
);
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
Sign up or log in
StackExchange.ready(function ()
StackExchange.helpers.onClickDraftSave('#login-link');
);
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
Sign up or log in
StackExchange.ready(function ()
StackExchange.helpers.onClickDraftSave('#login-link');
);
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
