Deriving the Adjoint / Tangent Linear Model for Nonlinear PDE

In summary: But I want to know how to derive the adjoint matrix from the continuous adjoint equation.In summary, the book tells you to integrate by parts, then integrate with respect to space (y), then integrate with respect to space (x). However, the Lagrange multipliers are not explained and I do not know how to get to them.
  • #1
finite_diffidence
2
0
Homework Statement
Deriving the Adjoint / Tangent Linear Model for a non-linear PDE
Relevant Equations
please see below as latex is not rendering
I am trying to derive the adjoint / tangent linear model matrix for this partial differential equation, but cannot follow the book's steps as I do not know the math. This technique will be used to solve another homework question. Rather than posting the homework question, I would like to understand the technique generally so I can forever use it. Here is the equation:

$$\frac{\partial u}{\partial t} = \frac{\partial u}{\partial y} \frac{\partial u}{\partial x}$$

Now in the book the derivation continues as follows:

1. First substitute in $$ u \rightarrow u + \delta u $$.

Then get rid of all the terms which have $$ \delta u \delta u $$

We are left with:

$$ \frac{\partial \delta u}{\partial t} = \frac{\partial \delta u}{\partial y} \frac{\partial u}{\partial x} + \frac{\partial u}{\partial y} \frac{\partial \delta u}{\partial x}$$

2. Now we can integrate by parts to move the derivatives over from the perturbation to the Lagrange multiplier. We ignore the surface term picked up as it will vanish at the boundaries:

$$ \int \lambda \left(\frac{\partial \delta u}{\partial t} - \frac{\partial \delta u}{\partial y} \frac{\partial u}{\partial x} + \frac{\partial u}{\partial y} \frac{\partial \delta u}{\partial x} \right)d\Omega dt
$$

to:

$$ \int \delta u \left( -\frac{\partial \lambda}{\partial t} + \frac{\partial}{\partial y} (\lambda \frac{\partial u}{\partial x}) + \frac{\partial u}{\partial x} (\lambda \frac{\partial u}{\partial y}) \right)d\Omega dt
$$

> How did we do this trick? I was told it is integration by parts, but could someone do it explicitly step by step?3. This gives us the continuous adjoint equation:

$$
\frac{\partial \lambda}{\partial t} = \frac{\partial}{\partial y} (\lambda \frac{\partial u}{\partial x}) + \frac{\partial u}{\partial x} (\lambda \frac{\partial u}{\partial y})
$$

4. From that equation we can discretize in time and write out the longhand summation of all the lagrange multiplies. We pick up a load of terms:

$$
\lambda_0 - \lambda_1 \left(u_1 - u_0 - \Delta t \frac{\partial u_0}{\partial y} \frac{\partial u_0}{\partial x}\right) + \lambda_1 \left(u_2 - u_1 - \Delta t \frac{\partial u_1}{\partial y} \frac{\partial u_1}{\partial x}\right)
$$

Or playing the same trick :

$$ \lambda_0 - \lambda_1 + \Delta t\left(\frac{\partial}{\partial y}\lambda_1\frac{\partial u_1}{\partial x} + \frac{\partial}{\partial x}\lambda_1\frac{\partial u_1}{\partial y} \right)
$$

> How do we get these two lines? I have no idea. If someone could do it very explicitly that would be helpful in me understanding what I am missing.

If someone with infinitely better math skills than mine could illuminate the way, that would be much appreciated.
 
Physics news on Phys.org
  • #2
Hello @finite_diffidence , :welcome: !

finite_diffidence said:
Rather than posting the homework question
Humor us and do it anyway. The context is not as irrelevant as you seem to think .

##u(x,y,t) = x+y+t## is a solution to the DE.

Where do the Lagrange multipliers come from ? Out of the blue ?
finite_diffidence said:
Now in the book
finite_diffidence said:
I was told it is integration by parts
A talking book ?
 
  • #3
BvU said:
Hello @finite_diffidence , :welcome: !

Humor us and do it anyway. The context is not as irrelevant as you seem to think .

##u(x,y,t) = x+y+t## is a solution to the DE.

Where do the Lagrange multipliers come from ? Out of the blue ?
A talking book ?

Hello BvU,

The other question is basically continuing this same equation, but I must derive the adjoint matrix.

My attempt of the problem starting at step 2. Looking at each term individually as I multiply through by ##\lambda##:

Integrating with respect to ##t##:

$$
\int \lambda \frac{\partial \delta u}{\partial t} = \lambda \delta u - \int \frac{\partial \lambda}{\partial t}\delta u
$$

Now looking at the second term, integrating with respect to space (##y##):

$$
\int -\lambda \frac{\partial \delta u}{\partial y} \frac{\partial u}{\partial x} = -\lambda \frac{\partial u}{\partial x} \delta u + \int \frac{\lambda}{\partial y} \frac{\partial u}{\partial x}
$$

Now looking at the third term, integrating with respect to space (##x##):

$$
\int -\lambda \frac{\partial \delta u}{\partial x} \frac{\partial u}{\partial y} = -\lambda \frac{\partial u}{\partial y} \delta u + \int \frac{\lambda}{\partial x} \frac{\partial u}{\partial y}
$$

The terms without the integrals disappear as we are dealing with Dirichlet boundary conditions, so we are left with the continuous adjoint equation i.e 3..

However, now I do not know how to get to 4. What are the steps to go from my continuous adjoint to the multiplied out version. I am not sure where the Lagrange multipliers come from. The book stated integration by parts which I figured out after much staring.
 

1. What is the purpose of deriving the adjoint/tangent linear model for nonlinear PDE?

The purpose of deriving the adjoint/tangent linear model for nonlinear PDE is to better understand the behavior of the system and to make predictions about its future state. This model allows for the analysis of the sensitivity of the system to changes in its initial conditions or parameters, which can be useful in various applications such as optimization and control.

2. How is the adjoint/tangent linear model derived for nonlinear PDE?

The adjoint/tangent linear model is derived by taking the linearization of the nonlinear PDE around a given solution. This involves computing the Jacobian matrix of the nonlinear PDE and solving a set of linear equations to obtain the adjoint/tangent linear model.

3. What are the main differences between the adjoint and tangent linear models for nonlinear PDE?

The main difference between the adjoint and tangent linear models is their respective roles in sensitivity analysis. The adjoint model is used to compute the sensitivity of the system's output to changes in its inputs, while the tangent linear model is used to compute the sensitivity of the system's state to changes in its initial conditions or parameters.

4. Can the adjoint/tangent linear model be used for any type of nonlinear PDE?

Yes, the adjoint/tangent linear model can be derived for any type of nonlinear PDE. However, the complexity of the model and the difficulty of its derivation may vary depending on the specific form of the PDE.

5. What are some applications of the adjoint/tangent linear model for nonlinear PDE?

The adjoint/tangent linear model has various applications in fields such as fluid dynamics, weather forecasting, and quantum mechanics. It can be used for data assimilation, optimization, and parameter estimation, among others. Additionally, this model has also been applied in machine learning and artificial intelligence for training neural networks.

Similar threads

  • Advanced Physics Homework Help
Replies
3
Views
394
Replies
8
Views
1K
  • Advanced Physics Homework Help
Replies
10
Views
583
  • Advanced Physics Homework Help
Replies
9
Views
2K
  • Advanced Physics Homework Help
Replies
6
Views
311
  • Advanced Physics Homework Help
Replies
10
Views
1K
Replies
27
Views
2K
  • Advanced Physics Homework Help
Replies
1
Views
989
  • Advanced Physics Homework Help
Replies
5
Views
2K
  • Advanced Physics Homework Help
Replies
1
Views
900
Back
Top