- [Voiceover] Hello everyone, what I wanna talk about here

is how to interpret the directional derivative

in terms of graphs.

I have here the graph of a function,

a multivariable function,

it's f of xy is equal to x squared times y.

In the last couple videos I talked

about what the directional derivative is,

how you can formally define it,

how you can compute it using the gradient.

Generally the setup that you might have is,

you have some kind of vector, and this is a vector

in the input space so in this case

it's gonna be in the xy plane.

In this case I'll just take the vector 1 1.

Okay? And the directional derivative, which we

denote by kind of taking the gradient symbol,

except you stick the name of that vector down in the lower

part there, the directional derivative of your function,

it'll still take the same input.

This is kind of a measure of how the function changes

when the input moves in that direction.

So I'll show you what I mean, I mean

you could imagine slicing this graph by some kind of plane

but that plane doesn't necessarily have to be parallel

to the x or y axes.

That's what we did for the partial derivative, we took

a plane that represented the constant x value or

the constant y value, but this is gonna be a plane that

kind of tells you what movement in the direction of

your vector looks like, and like I have a number of other

times I'm gonna go ahead and slice the graph along that

plane, and just to make it clear, I'm gonna color in where

the graph intersects that slice.

This vector here, this little v, you'll be thinking of it

as living on the xy plane and it's determining the direction

of this plane that we're slicing things with.

On the xy plane you've got this vector,

it's 1 1, it kind of points to that diagonal direction,

and then you take the whole plane and you slice your graph.

And if we want to interpret the directional derivative here,

I'm gonna go ahead and fill it an actual value,

so let's say we wanted to do it at -1 1,

- 1, -1

'cause I guess I chose a plane that passes through the

origin, so I've got to make sure that the point

I'm evaluating actually goes along this plane,

but you could imagine one that points in the same direction,

but you kind of slide it back and forth,

if we're doing this, we can interpret this as a slope,

but you have to be very careful, if you're gonna interpret

this as a slope, it has to be the case that you're dealing

with a unit vector, that the magnitude of your vector

is equal to 1.

I mean, it doesn't have to be, you can kind of account

for it later but it makes it more easy to think about.

If we're just thinking of a unit vector.

When I go over here instead of saying that it's 1 1,

I'm gonna say it's whatever vector points in the same

direction but has a unit length, and in this case that

happens to be square root of 2 over 2,

for each of the components.

You can kind of think about why that would be true

by diagonal but this is a vector with unit length,

and its magnitude is 1, and it points in that direction.

If we're evaluating this negative point like 1 1,

we can draw that on the graph, see where it actually is,

and in this case it'll be, oops, moving things about when

I had a point.

It'll be this point and if you kinda look from above,

you see that's -1, -1,

and if we want the slope at that

point, you're kinda thinking of the tangent line here.

Tangent line to that curve, and we're wondering what its

slope is, so, the reason that the directional derivative

is gonna give us this slope, is because, another notation

that might be kinda helpful for what this directional

derivative is, some people will write partial f,

and partial v.

You can think about that as taking a slight

nudge in the direction of v, so this would be a little

nudge, a little partial nudge in the direction of v.

And then you're saying "what changed in the value

of the function that's then resulting?"

"The height of the graph, does it devalue the function?".

As this initial change approaches zero and the resulting

change approaches zero as well, that ratio, the ratio

of the partial f to partial v, is going to give the slope

of this tangent line.

Conceptually, that's kind of a nicer notation, but the

reason we use this other notation is nabla sub v 1,

is it's very indicative of how you compute things

once you need it computed.

You take the gradient of f, just the vector value function

gradient of f, and take the dot product with the vector.

Let's actually do that, just to see what this would

look like, and I'll go ahead and write it over here,

use a different color.

The gradient of f, first of all, is a vector full of partial

derivatives, it'll be the partial derivative of f with

respect to x and the partial derivative of f with respect

to y.

When we actually evaluate this, we take a look, partial

derivative of f with respect to x, x looks like the variable

y I just a constant, so its partial derivative is 2 times x

times y.

2 times x times y.

but when we take the partial with

respect to y, y now looks like a variable, and x looks

like a constant, derivative of a constant times a variable,

is just that constant x squared.

And if we were to evaluate this at the point -1, -1,

then you can plug that in, 2 times -1 times -1 would be 2,

and then negative 1 squared, would be 1.

So that would be our gradient at that point, which means

if we want to evaluate gradient of f times v, we could go

over here, and say that's 2 1,

'cause we evaluate the gradient at the point we care about.

And then the dot product, with v itself in this case,

root 2 over 2, and root 2 over 2.

The answer that we get, we multiply the fist two components

together, 2 times root 2 over 2, then square it to 2,

and then here we multiply the second components together,

and that's gonna be 1 times root 2 over 2, root 2 over 2,

and that would be our answer, that would be our slope.

But this only works if your vector is a unit vector,

and I showed this in the last video where we talked about

the formal definition of the directional derivative.

If you scale v by 2, and I can do it here if instead of v

you're talking about 2 v, so I'll go ahead and make myself

some room here.

If you're taking the directional derivative along 2 v

of f, the way that we're computing that, we're still taking

the gradient of f, dot product with 2 times your vector,

and dot product, you can pull out that too.

This is just gonna double the value of the entire thing.

V, this started with v, it's gonna be twice the value,

the derivative will become twice the value, but you don't

necessarily want that because you'd see this plane you

sliced with, if instead of doing it in the direction of v,

the unit vector, you did in the direction of 2 times v,

it's the same plane, it's the same slice you're taking,

and you'd want that same slope, so that's gonna mess

everything up.

This is super important if you're thinking about things in

the context of slope, one thing that you could say is

your formula for the slope of a graph in the direction of v,

is you take your directional derivative, that dot product

between f and v, and you just always make sure to divide it

by the magnitude of v, divide it by that magnitude.

That will always take care of what you want, that's

basically a way of making sure that really, you're taking

the directional derivative in the direction of a certain

unit vector.

Some people even go so far as to define the directional

derivative to be this, to be something where you normalize

out the length of that vector.

I don't really like that, but I think that's because they're

thinking of the slope context, they're thinking of rates of

change as being the slope of a graph.

One thing I'd like to emphasize as always, graphical

intuition is good, and visual intuition is always great,

you should always be trying to find a way to think about

things visually, but with multivariable functions,

the graph isn't the only way.

You can kind of more generally think about just a nudge

in the v direction, and in the context where v doesn't

have a length 1, the nudge doesn't represent an actual size

but it's a certain scaling constant times that vector,

you can look at the video on the formal definition for the

directional derivative, if you want more details on that.

But I do think this is actually a good way to get a feel for

what the directional derivative is all about.