3
$\begingroup$

Consider a double sequence $a_{m,n}$ defined by the difference equation

$$ a_{m,n} = \alpha(m,n) a_{m,n-1} + \beta(m,n) a_{m-1,n} $$

with known boundary conditions $a_{m,0}$ and $a_{0,n}$, and where $\alpha,\beta : {\Bbb Z}^2 \to {\Bbb R}$. How to obtain a closed-form solution for $a_{m,n}$ in terms of $\alpha,\beta$?


Example: $\alpha=\mathrm{const.}$ and $\beta(m,n)=m$.

$\endgroup$
7
  • 1
    $\begingroup$ Yes, there is a general strategy: interpret the recurrence as a weighted lattice-path sum. $\endgroup$ Commented 20 hours ago
  • $\begingroup$ Can you elaborate for everybody? $\endgroup$ Commented 19 hours ago
  • $\begingroup$ I am on phone. May be later. $\endgroup$ Commented 19 hours ago
  • 2
    $\begingroup$ You have not told us $\alpha(m,n)$ and $\beta(m,n)$ so there is no reason to suppose in general that there is a closed form beyond performing the recursion across the lattice. $\endgroup$ Commented 11 hours ago
  • 1
    $\begingroup$ What is the motivation for this question? Does this come from image processing? Numerical PDEs? $\endgroup$ Commented 8 hours ago

3 Answers 3

0
$\begingroup$

Hint.

Assuming $\alpha(m,n)=\alpha, \beta(m,n) = m$ we have

$$ a_{m,n} = \alpha a_{m,n-1}+m a_{m-1,n} $$

or by the characteristic functions method with $u(x,y) = \sum_{m,n}a_{m,n}x^my^n$

$$ u(x,y) - \alpha x u(x,y) - \partial_x u(x,y) = \phi_0(x,y) $$

with $\phi_0(x,y)$ to account for the initial conditions. After solving for $u(x,y)$ we can recover the $a_{m,n}$ after a series expansion.

$\endgroup$
0
$\begingroup$

I hope now it completes it for you now. I am not very good in recurrence relations but I tried. If it does not work, let me know I will delete my answer.

Interpret the recurrence as a weighted lattice–path problem.

Consider the integer lattice $\mathbb{Z}_{\ge 0}^2. $

To compute $a_{m,n}$, the recurrence allows two possibilities:

  • $ (m,n-1) \rightarrow (m,n) $ with weight $\alpha(m,n), $

  • $ (m-1,n) \rightarrow (m,n)$ with weight $\beta(m,n). $

Thus the value $ a_{m,n}$ is obtained by summing weighted contributions coming from these two predecessor points.

This is the key idea.

Here is the next hint, building directly on the lattice--path picture you already have.

Hint 2. Unroll the recurrence completely. Any value $a_{m,n}$ can be written as a sum over all monotone lattice paths from the boundary to $(m,n)$.

More precisely, every path $\gamma$ from either $(k,0)\to(m,n)$ or $(0,\ell)\to(m,n)$, using only steps $(1,0)$ and $(0,1)$, contributes a weight $ \prod_{(i,j)\in\gamma} \begin{cases} \alpha(i,j) & \text{if the step is }(0,1),\\ \beta(i,j) & \text{if the step is }(1,0), \end{cases} $ multiplied by the boundary value at the starting point.

So $a_{m,n}$ is a path sum: $ a_{m,n} =\sum_{\gamma:(k,0)\to(m,n)} a_{k,0}\,w(\gamma) +\sum_{\gamma:(0,\ell)\to(m,n)} a_{0,\ell}\,w(\gamma). $

The recurrence disappears; what remains is combinatorics plus products of $\alpha,\beta$.

The next move after this hint is to reorganize these sums, either by fixing the number of horizontal steps first or by grouping paths with the same step pattern.

Fix $k\ge 0$ and consider monotone lattice paths $\gamma:(k,0)\to(m,n).$ Such a path consists of exactly $ m-k$ horizontal steps and $n$ vertical steps.

Hence these paths are in bijection with subsets $S\subseteq{1,2,\dots,m-k+n} \text{ satisfying } |S|=n, \text{ where } S \text{ records the positions of the vertical steps.} $

$ \text{For a given } S, \text{ the corresponding path } \gamma_S$ is uniquely determined, and its weight is

$w(\gamma_S)$

$\prod_{r=1}^{m-k+n} \begin{cases} \alpha(i_r,j_r) & \text{if } r\in S,\ \beta(i_r,j_r) & \text{if } r\notin S, \end{cases} $

$ \text{where } (i_r,j_r) \text{ is the lattice point entered at the } r\text{-th step.} $

Therefore the path sum can be written explicitly as

$\sum_{\gamma:(k,0)\to(m,n)} w(\gamma)$

$\sum_{\substack{S\subseteq{1,\dots,m-k+n}\ |S|=n}} \prod_{r=1}^{m-k+n} \begin{cases} \alpha(i_r,j_r) & r\in S,\ \beta(i_r,j_r) & r\notin S. \end{cases} $

An analogous expression holds for paths starting at $(0,\ell)$.

Then use induction on $m+n$

$\endgroup$
4
  • 1
    $\begingroup$ Alright but I don't see how this reformulation helps in finding a closed form expression: For example, in the one dimensional case, having something like $a_m=\alpha(m)a_{m-1} + \beta(m)$ allows to write down a closed form expression for $a_m$ almost immediately. Then one proves its validity via induction. I don't understand why there seem to be obstructions in dimensions $>1$. After all, the cardinality of $\mathbb N$ is equal to the one of $\mathbb N\times\mathbb N$ which is equal to the one of $\mathbb N\times \mathbb N\times\mathbb N$, etc. $\endgroup$ Commented 8 hours ago
  • $\begingroup$ @Mathematicsenthusiast That is why I asked you want me to further elaborate? $\endgroup$ Commented 8 hours ago
  • $\begingroup$ Please go ahead $\endgroup$ Commented 8 hours ago
  • 1
    $\begingroup$ So far, we are just shifting the problem from one position to the next and 'interpreting' along the way. I appreciate the interpretations but obviously I am interested in precise expressions for $\sum_{\gamma:(k,0)\to(m,n)}$ and $\sum_{\gamma:(0,l)\to(m,n)}$. And to prove these via induction. $\endgroup$ Commented 7 hours ago
-2
$\begingroup$

One way to write down this summation of weights along paths should be \begin{align*} a_{m,n} &= \sum_{1 \leq m_0 \leq \dots \leq m_n = m} \left(a_{m_0,0} \prod_{1 \leq s \leq n} \left(a(m_{s-1}, s) \prod_{m_{s-1} < k \leq m_s} b(k,s)\right)\right) \\ &+ \sum_{1 \leq n_0 \leq \dots \leq n_m = n} \left(a_{0,n_0} \prod_{1 \leq s \leq m} \left(b(s, n_{s-1}) \prod_{n_{s-1} < k \leq n_s} a(s, k)\right)\right), \end{align*} not sure if that helps you.

$\endgroup$

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.