Solve for A_n as n→∞: Difference Equation

In summary, the conversation discusses a possible approach to finding the solution of a nonlinear equation with variable coefficients using the Z-transform. The approach is similar to using Laplace or Fourier transforms for ODEs with variable coefficients and can sometimes yield an easier path to the solution. There is a discussion about the applicability of the Z-transform to problems with factorially-divergent solutions and some calculations are presented as examples.
  • #1
107
1
This is a bit of a shot in the dark, but has anybody ever encountered a theory which can tell me what the solution of this equation:

[tex]A_n = nA_{n-2} + nA_{n-3}[/tex]

behaves like, as [tex]n\to\infty[/tex]? For convenience, you can set [tex]A_n = 1[/tex] for [tex]n = 0, \ldots, 3[/tex]. Without the term on the right, it goes something like [tex]A_n \sim \Gamma(n/2)[/tex].
 
Last edited:
Physics news on Phys.org
  • #2
I'm curious where this equation popped up. It is actually pretty interesting.

I think you might be able to use the Z transform to tackle this, which assumes that the Z transform of the solution exists! I rewrote your equation so that it holds for [tex]n\geq0[/tex]:
[tex]A_{n+3}=(n+3)\left(A_{n+1}+A_n \right). [/tex]
Then using the standard unilateral Z transform:
[tex] Z(A_n) = \tilde{A}(z) = \sum_{n=0}^{\infty}A_n z^{-n}[/tex]
I get
[tex]
\left(z^2 - 2 z - 3 \right) \tilde{A}(z) + \left( z + z^2 \right) \frac{d\tilde{A}(z)}{dz} = A_0 z^3 + A_1 z^2 + \left(3 A_0 + A_2 \right) z.
[/tex]
For the homogeneous solution I get
[tex]
\tilde{A}_h (z) = \gamma \left( \frac{z^3}{(z+1)^2} \right) e^{z-z^2/2},
[/tex]
where [tex]\gamma[/tex] is a constant to be determined.

For the particular solution I try variation of parameters,
[tex]
\tilde{A}_p (z) = f(z) \tilde{A}_h (z),
[/tex]
and get,
[tex]
f = \int \, dz \, \frac{\left(A_0 z^2 + A_1 z + (3 A_0 + A_2) \right) (z+1)}{z^3} e^{-z+z^2/2}
[/tex]
where I will let you determine the constant of integration. I was unable to get Mathematica to do the integral if [tex]A_0=A_1=A_2=1[/tex]. However, if
[tex]A_0=0[/tex] and [tex]A_1=A_2=1[/tex] then mathematica gives me
[tex]
f = \frac{1}{4} \left[3 \sqrt{\frac{2 \pi}{e}} erfi \left(\frac{z-1}{\sqrt{2}} \right) - \frac{2(3z+1)e^{-z+z^2/2}}{z^2}\right]+C
[/tex]
where C is a constant to be determined and Mathematica defines
[tex]
erfi(z)=\frac{erf(iz)}{i}
[/tex]
where [tex]erf[/tex] is the error function and [tex] i=\sqrt{-1} [/tex].

This isn't the solution to your problem, as you still have to do the inverse Z transform of [tex]\tilde{A}=\tilde{A}_h+\tilde{A}_p[/tex] and determine a couple of constants I left floating around. But I think this approach is likely to lead to the solution. Please let us all know if you get further than this.

jason
 
Last edited:
  • #3
It's not clear to me how the Z-transform can be used here. This equation is obviously nonlinear, and generally the Z-transform approach is for linear systems. If there is a way to treat the equation as linear in the limit as n goes to infinity, then this should be the first step and a clear argument would need to be presented as to why this is valid.
 
Last edited:
  • #4
elect_eng said:
It's not clear to me how the Z-transform can be used here. This equation is obviously nonlinear, and generally the Z-transform approach is for linear systems. If there is a way to treat the equation as linear in the limit as n goes to infinity, then this should be the first step and a clear argument would need to be presented as to why this is valid.

While the equation has variable coefficients, it is certainly linear.

The approach I outlined is very similar to using Laplace or Fourier transforms to solve ODEs with variable coefficients - it will not always help, but sometimes it can yield an easier path to the solution.
 
  • #5
jasonRF said:
While the equation has variable coefficients, it is certainly linear.

OK, I understand the terminology and interpretation of the equation now. This equation is classified as linear with variable coefficients. To me this is a type of equation that does not directly yield to frequency transforms. However, perhaps my knowledge is limited here. I'll look at your approach in more detail.

jasonRF said:
The approach I outlined is very similar to using Laplace or Fourier transforms to solve ODEs with variable coefficients - it will not always help, but sometimes it can yield an easier path to the solution.

I'm very familiar with using Z-transforms to solve linear difference equations with constant coefficients. It's just that I have not solved an equation with variable coefficients with this technique. But, again my training may be limited, and I would certainly like to know how to apply frequency transforms to differential and difference equations with variable coefficients.
 
Last edited:
  • #6
Jason,

Thanks for your help. The Z-Transform is new to me, so I had to read up a little bit on it. For the moment, ignoring all the interesting work you did on the problem, my only question is what is the applicability of the Z-Transform to problems with factorially-divergent solutions?

In this case, we expect [tex]A_n \sim \Gamma(n/2 + \ldots)[/tex]. This has the effect that the Z-Transform is divergent everywhere in the complex plane. This is particularly troubling for the inversion process.

Suppose we have something simpler, say
[tex]A_{n+1} = n A_n[/tex]

Then the inversion of the Z-Transform gives something like
[tex]A_n = \frac{1}{2\pi i} \int Ce^{-z} z^{n-1} \ dz[/tex]

and we can see the [tex]\Gamma(n)[/tex] function somewhere in there.

Suppose we take your homogeneous Z-Transform:

[tex]
\frac{1}{2\pi i} \int \left( \frac{z^3}{(z+1)^2} \right) e^{z-z^2/2} z^{n-1} \ dz
[/tex]

I don't quite see how the inversion will be done, or how useful it'd be. There are poles at [tex]z = \pm 1[/tex], but residue contributions certainly aren't going to get you that Gamma function behaviour as [tex]n \to ]\infty[/tex].
 
Last edited:
  • #7
jasonRF said:
While the equation has variable coefficients, it is certainly linear.

The approach I outlined is very similar to using Laplace or Fourier transforms to solve ODEs with variable coefficients - it will not always help, but sometimes it can yield an easier path to the solution.

I don't mean to detract from this thread since it is not my thread and not my problem to solve. However, I find this very interesting and have learned quite a bit here.

I wanted to say thanks for setting me straight on this approach. My confusion came from the fact that all my previous training with transforms has been restricted to linear time-invariant systems, which are the most typical cases in engineering. Now that I understand this approach, I feel I have a new and useful tool in my bag of tricks. :approve:

One comment I can make is that when I did out the calculation, I came up with the following:

[tex]

\left(z^2 - 2 z - 3 \right) \tilde{A}(z) + \left( z + z^2 \right) \frac{d\tilde{A}(z)}{dz} = A_0 z^3 + A_1 z^2 + \left(-2 A_0 + A_2 \right) z [/tex]

This is slightly different from your expression. I think it is more likely that I made a mistake, but when I look at my work, I don't see where it is. Anyway, I thought I would just mention it in case the OP comes up with this answer also.
 
Last edited:
  • #8
rsq_a said:
Jason,

Thanks for your help. The Z-Transform is new to me, so I had to read up a little bit on it. For the moment, ignoring all the interesting work you did on the problem, my only question is what is the applicability of the Z-Transform to problems with factorially-divergent solutions?
.

You are absolutely correct. I was assuming a solution with a Z transform exists, and didn't follow through that one extra step to see that I was in trouble. Sorry about that!

Sorry for hte late reply - I've been out of town working 24/7.

Cheers!

jason
 

1. What is a difference equation?

A difference equation is a mathematical equation that describes the relationship between consecutive terms in a sequence or series. It is used to model discrete systems or processes where the change in the system over time is dependent on the previous state of the system.

2. How is a difference equation different from a differential equation?

A difference equation involves discrete variables, where the independent variable is defined in discrete time steps (e.g. n=1,2,3...), while a differential equation involves continuous variables, where the independent variable is defined in continuous time (e.g. t=0,1,2...).

3. How do you solve for An as n→∞ in a difference equation?

To solve for An as n→∞, you need to find the limit of the sequence or series defined by the difference equation. This can be done by finding the general solution using techniques such as iteration or substitution, and then taking the limit of the solution as n approaches infinity.

4. What are some real-world applications of difference equations?

Difference equations are commonly used in fields such as economics, physics, biology, and engineering to model various systems and processes. Examples include population growth, chemical reactions, financial markets, and control systems.

5. Can difference equations be solved numerically?

Yes, difference equations can be solved numerically using techniques such as finite differences, Euler's method, or Runge-Kutta methods. These methods involve approximating the solution using a series of discrete points and are useful for more complex difference equations that cannot be solved analytically.

Suggested for: Solve for A_n as n→∞: Difference Equation

Back
Top