Main content

# Laplacian intuition

A visual understanding for how the Laplace operator is an extension of the second derivative to multivariable functions. Created by Grant Sanderson.

## Want to join the conversation?

- If gradient means what direction you should walk to get higher, shouldn't it therefore mean the "slope of steepest
*ascent*", not descent?(20 votes)- Yes I think Grant got it wrong there. Because what he said would contradict the behaviour of vectors near those hills and valleys. (vectors pointing towards hills and outward away from valleys)(9 votes)

- I searched countless hours into my mind and into the interent, to define divergence of a Scalar Function. Does it even exist? Finally, found it and it is not customary to call it a divergence of Scalar Function, and rather a Laplacian, which is rather a divergence of its gradient Vector field. Thank you.(2 votes)
- the divergence of a scalar function does not make sense from what I have gathered about what divergence represents. Divergence is an operator applied to a vector field that produces a scalar field. When evaluated at a point P= (P0, P1, P2, ...,Pn), the divergence tells you the density of outward flux from an infinitesimal volume centered at P.(10 votes)

- So the Laplacian indicates how much of a local minimum or máximum a point is and the video says that it is the analogous to the second derivative test in single variable calculus. But.. we don't use the Laplacian to find local min/max, we use the Hessian determinant in the second partial derivative test to find local Min, Max.. So.. What is it good for??(3 votes)
- There are many applications, such as in physics (physics is all about differential equations). There are also a bunch of stuff it can be used for in math, chemistry, and computer science.(3 votes)

- What is a Laplacian output of a saddle point ?(4 votes)
- I'm posting this as "food for thought" rather than a hard & fast answer, but I think that (for a scalar function f of two scalar variables
*x,y*) it's arbitrary unless the up direction has exactly the same curvature as the down direction, in which case it would be zero.(2 votes)

- Why do they use Nabla so much in so many different equations in multi-variable calc? It's confusing and hard to remember which variation of Nabla means what. :/(3 votes)
- the Nabla means ==> [∂/∂b1, ∂/∂b2, ∂/∂b3, ..., ∂/∂bn] where bi is the ith vector of the basis making up the vector space F is defined in

normally you just see: [∂/∂x, ∂/∂y, ∂/∂z] because the basis is i, j, k, which are just unit vectors that are all orthogonal to one another. This notation is not to be taken to seriously however, because multiplication is not the same as applying an operator to a function(4 votes)

- At3:31the arrow above being red suggests that it's bigger than the yellow under it so the divergence at this point isn't 0 right?(2 votes)
- Doesn't make a lot of sense for me this concept in functions where the divergence is constant for example:

*f(x,y) = x^2 + y^2, div(grad(f)) = 4.*

Does it mean that in every point the function diverges from the same point(0,0) and then thats the minima?(2 votes) - that moment you think you reached a video about the Laplace transform so you go ahead and finish the differential equation course and then learn the Laplacian operator has nothing to do with it.(1 vote)
- Is this related to find some sort of max/min of a scarlar field? Are there natural phenomena related to this concept?(1 vote)

## Video transcript

- [Voiceover] So here I'm
gonna talk about the Laplacian. Laplacian. And the Laplacian is a certain operator in the same way that the
divergence, or the gradient, or the curl, or even just
the derivative are operators. The things that take in
some kind of function and give you another function. So in this case, let's say we
have a multivariable function like F, that just takes in
a two-dimensional input, F of X Y. So you might imagine its graph
as being something like this, where the input space
is this X-Y plane here, so each of the points,
X Y, is a point here, and then the output is just given by the height of that graph. So the Laplacian of F is denoted with a right-side-up triangle, and it's gonna give you a
new scalar valued function of X and Y. So it's gonna give you a new function that takes in a two-dimensional input and just outputs a number. And it's kind of like a second derivative, because the way that it's defined is that you take the divergence of the gradient of your function F. So that's kind of how it's defined, the divergence of the gradient of F. And a more common notation
that you might see here is to take that upside-down
triangle, nabla, dot product, with nabla of F. So remember, if F is a
scalar valued function, then the gradient of F
gives you a vector field, a certain vector field. But the divergence of a vector field gives you another scalar valued function. So this is the sense in which
it's a second derivative. But let's see if we can kind
of understand intuitively what this should mean. 'Cause the gradient, if you remember, gives you the slope of steepest descent. So it's a vector field
in the input space of X. And each one of the vectors
points in the direction that you should walk, such that this graph is kind of a hill on top of you, it tells you the direction you should go to increase your direction
the most rapidly. And if that seems unfamiliar,
it doesn't make sense, maybe go take a look at that
video on gradients and graphs and how they relate to each other. So with the specific graph
that I have pictured here, when you have kind of the top of a hill, all of the points around it, the direction that you should walk, is towards the top of that hill. Whereas when you have
kind of like the bottom, a little goalie here, all of the directions you should walk to increase the value of
the function most rapidly, are directly away from that value, which you might call a local minimum. So let's temporarily get rid of the graph, just so we can look at the gradient field here pretty clearly. And now let's think
about what the divergence is supposed to represent. So now the divergence, and
again if this feels unfamiliar, maybe go back and take a look
at the divergence videos, but that divergence has you
imagining that this vector field corresponds to some kind of fluid flow. So you imagine little
like water molecules, and at any given moment,
they're moving along the vector that they're attached to. So for example, if you
had a water molecule that started off kind of here, you would start by going along that vector and then kind of follow the ones near it, and it looks like it kind
of ends up in this spot. And a lot of the water molecules seem to kind of converge over there. Whereas over here, the water
molecules tend to go away when they're following those vectors away from this point. And when they go away like that, when you have a whole bunch of vectors kind of pointed away, that's an indication that
divergence is positive, because they're diverging away. So over here, divergence is positive. Whereas the opposite case,
where all of the water molecules are kind of coming in towards a point, that's where divergence is negative. And in another area, let's
say it was kinda like this center point, where you
have some water molecules that looks like they're coming in, but other ones are going out, and at least from this
picture, it doesn't seem like the ones going out are
doing so at a faster rate or slower than they are here. This would be roughly zero divergence. So now let's think
about what it might mean when you take the divergence
of the gradient field of F. So let me kind of clear
up the markings I made on top of it here. Points of high divergence, points where it diverges a lot here, why are those vectors pointing away? And if we pull up the graph again, the reason they're pointing away is 'cause the direction of steepest descent is kind of uphill everywhere;
you are in a valley. Whereas in the opposite circumstance, where divergence is highly negative, 'cause points are converging towards it, why are they pointing towards it? Well this is a gradient field, so they're pointing towards that spot, because that's where anywhere around it, you should walk towards
that spot to go uphill. So in other words, the
divergence of the gradient is very high at points that are kind of like minima, at points
where everyone around them tends to be higher. But the divergence of the gradient is low at points that look more
like maximum points, where when you evaluate the function at all of the points
around that input point, they give something smaller. So this Laplacian operator
is kind of like a measure of how much of a minimum
point is this X Y. You will be very positive when F evaluated at that point tends to give a smaller value than F evaluated at
neighbors of that point. But it'll be very negative if when you evaluate F at that point it tends to be bigger than its neighbors. And this should feel kind of analogous to the second derivative
in ordinary calculus, when you have some kind of graph of just a single variable function, the second derivative of X will be low, it'll be negative at points where it kind of
looks like a local maximum. But over here, the second derivative of x would be positive at points that kind of look like a local minimum. SO in that way, the Laplacian is sort of an analog of the second derivative for scalar valued multivariable functions. And in the next video, I'll
go through an example of that.