---
title: "Laplacian in 2D"
subtitle: "Module 4: Space and Pattern Formation"
format: html
---
::: {.callout-note}
## Reading
- **Compulsory:** @pearson1993complex.
- **Recommended:** @murray2003mathematical, Section 2.5 (scale and geometry effects).
:::
This laboratory lifts the 1D workflow to a rectangle. The main change is the Laplacian. The Euler loop, the boundary logic, and the validation steps stay the same. If your 1D code was modular, you mostly change the array shape.
## Rectangular Domains
You work on $[0, L_x] \times [0, L_y]$. After discretization each field becomes a matrix.
- The row index stores the $x$ position, the column index the $y$ position.
- The full state has shape `(2, num_x, num_y)`, one matrix for $u$ and one for $v$.
```python
num_x = int(length_x / dx)
num_y = int(length_y / dx)
uv = np.ones((2, num_x, num_y))
```
You still update the whole matrix at once. Shift it up, down, left, and right, and combine the shifted copies.
## The 5-Point Stencil
$$
\Delta u = \frac{\partial^2 u}{\partial x^2} + \frac{\partial^2 u}{\partial y^2}
\approx \frac{u_{i+1,j} + u_{i-1,j} + u_{i,j+1} + u_{i,j-1} - 4u_{i,j}}{dx^2}.
$$ {#eq-lap2-stencil}
Applying the 1D centered difference in each direction gives this stencil, with error $\mathcal{O}(dx^2)$ [@leveque2007finite, Chapter 3].
::: {.callout-note}
## Exercise
Implement the stencil with `np.roll()`. Each call shifts the whole matrix by one grid cell along one direction.
```python
import numpy as np
def laplacian_2d(field, dx):
# TODO: combine the four nearest neighbours and the centre point with np.roll
lap =
return lap / dx**2
```
:::
::: {.callout-tip collapse="true"}
## Solution
```python
def laplacian_2d(field, dx):
lap = -4 * field
lap += np.roll(field, shift=1, axis=0)
lap += np.roll(field, shift=-1, axis=0)
lap += np.roll(field, shift=1, axis=1)
lap += np.roll(field, shift=-1, axis=1)
return lap / dx**2
```
:::
## Boundary Conditions on a Rectangle
The same three options as in [Boundary Conditions](boundary-conditions.qmd), now on four edges.
- **Periodic:** leave the array unchanged; `np.roll` already wraps each side onto the opposite one.
- **Neumann:** copy the nearest interior row or column onto each edge.
- **Dirichlet:** overwrite each edge with prescribed values.
```python
def apply_boundary_conditions(uv, bc="neumann", dirichlet_values=(1.4, 1.96)):
if bc == "periodic":
return uv
if bc == "neumann":
uv[:, 0, :] = uv[:, 1, :]
uv[:, -1, :] = uv[:, -2, :]
uv[:, :, 0] = uv[:, :, 1]
uv[:, :, -1] = uv[:, :, -2]
return uv
if bc == "dirichlet":
for species, value in enumerate(dirichlet_values):
uv[species, 0, :] = value
uv[species, -1, :] = value
uv[species, :, 0] = value
uv[species, :, -1] = value
return uv
raise ValueError("bc must be 'periodic', 'neumann', or 'dirichlet'")
```
## Validate the Stencil
Test on $u = \cos(2\pi x/L_x)\cos(4\pi y/L_y)$ with periodic wrapping, where $\Delta u = -\left[(2\pi/L_x)^2 + (4\pi/L_y)^2\right] u$.
```{python}
import numpy as np
def laplacian_2d(field, dx):
lap = -4 * field
lap += np.roll(field, shift=1, axis=0)
lap += np.roll(field, shift=-1, axis=0)
lap += np.roll(field, shift=1, axis=1)
lap += np.roll(field, shift=-1, axis=1)
return lap / dx**2
length_x, length_y = 20.0, 50.0
for dx in (1.0, 0.5, 0.25):
x = np.arange(int(length_x / dx)) * dx
y = np.arange(int(length_y / dx)) * dx
xx, yy = np.meshgrid(x, y, indexing="ij")
u = np.cos(2 * np.pi * xx / length_x) * np.cos(4 * np.pi * yy / length_y)
exact = -((2 * np.pi / length_x) ** 2 + (4 * np.pi / length_y) ** 2) * u
print(f"dx = {dx:.2f} max error = {np.max(np.abs(laplacian_2d(u, dx) - exact)):.2e}")
```
The error again drops by four for each halving of $dx$.
::: {.callout-important}
## Stability in 2D
The 5-point stencil halves the stable time step: $dt \le dx^2/(4D)$ instead of $dx^2/(2D)$. With $d = 30$ and $dx = 1$ that is $dt \le 0.0083$, so the 1D choice $dt = 0.01$ fails in 2D ([Time-Step Stability](time-step-stability.qmd)).
:::
::: {.callout-tip}
## Optional extension
A 9-point stencil that also uses the diagonal neighbours gives a more isotropic Laplacian. `amlab.pdes_2d.gray_scott.laplacian_9pt` implements one. Compare the two stencils on the test function above.
:::