- Constrained optimization introduction
- Lagrange multipliers, using tangency to solve constrained optimization
- Finishing the intro lagrange multiplier example
- Lagrange multiplier example, part 1
- Lagrange multiplier example, part 2
- The Lagrangian
- Meaning of the Lagrange multiplier
- Proof for the meaning of Lagrange multipliers
Lagrange multiplier example, part 1
A Lagrange multipliers example of maximizing revenues subject to a budgetary constraint. Created by Grant Sanderson.
Want to join the conversation?
- Grant Sanderson is da best.(23 votes)
- If the equation 20h + 2,000s = 20,000 had immediately been simplified (to h + 100s = 1000), would that have any downstream effect of our optimization? I ask because the gradient of g(h,s) would be different, and therefore the gradient of R would equal something different.(3 votes)
- The gradient of g would actually be different. It would've changed from [20, 2000] to [1, 100]. And it would've gotten you the same equations but lambda would've been different.
The unsimplified equations were
200/3 * (s/h)^1/3 = 20 * lambda
100/3 * (h/s)^2/3 = 20000 * lambda
The simplified equations would be the same thing except it would be 1 and 100 instead of 20 and 20000.
But it would be the same equations because essentially, simplifying the equation would have made the vector shorter by 1/20th. But lambda would have compensated for that because the Langrage Multiplier makes the vectors the same length, so the lambda would have been 20 times as big. You would have gotten the same maximized hours of labor and tons of steel.(3 votes)
- Not sure about this but I was thinking..
Solving the problem was based on the assumption that the contours going to the origin won't give a maximum. But what if at those points the contours of the graph is higher?
Wouldn't that mean that it's not necessarily the max point and would depend on whether the graph curves up or down?(2 votes)
- You would need to have an understanding of the behavior of your function. In the given example, the function R(h,s) will increase as h and s increase. R(h,s) will be equal to zero at the origin. Thus our maximum point will be where the contour is tangent to our constraint function.(1 vote)
- How would someone generate a multivariable revenue function like that?(2 votes)
- How can REVENUE be a function of labor and steel COSTS? It seems to me that revenue should be the number of widgets sold times the selling price. PROFIT would be revenue minus the costs of production (labor and steel).(2 votes)
- for example if i have a given problem that looks like this z=4x^2+3xy+6y^2 constraint to x + y =56(1 vote)
- How did he get the revenue function?(1 vote)
- How to get the extremum of a multivariable function?(1 vote)
- Have a look at the respective video https://www.khanacademy.org/math/multivariable-calculus/applications-of-multivariable-derivatives/optimizing-multivariable-functions-videos/v/multivariable-maxima-and-minima.(1 vote)
- I am solving a similar problem about budget constraints.I have a bunch of x,y,z values and I want to model an equation based on those datapoints.
In the video he uses direct equation but I have like 20 points and I want to model a best fit equation. Can someone help me(1 vote)
- at5:56, when finding the gradient of g, why do we write the partial derivative of g with respect to s when the equation has no s in it?(0 votes)
- The function g(h,s) is 20*h + 2,000*s.
(I can understand some confusion happening here because the 's' does look like a 'g').
So, writing it out, it's 20 times however many [hours], and 2,000 times however much [steel].(4 votes)
- [Instructor] So let's say you're running some kind of company and you guys produce widgets. You produce some little trinket that people enjoy buying and the main costs that you have are labor, the workers that you have creating these and steel. And let's just say that your labor costs are $20 per hour, $20 each hour and then your steel costs are $2,000. Keep the numbers kind of related to each other. $2,000 for every ton of steel. And then you've had your analysts work a little bit on trying to model the revenues you can make with your widgets as a function of hours of labor and tons of steel. Now let's say the revenue model that they've come up with. The revenue as a function of hours of labor and then S for steel, let's say, is equal to about 100 times the hours of labor to the power of 2/3 multiplied by the tons of steel to the power of 1/3. If you put in a given amount of labor and a given amount of steel, this is about how much money you're gonna expect to earn. And of course you wanna earn as much as you can plus you actually have a budget for how much you're able to spend on all these things. And your budget is $20,000. You're willing to spend $20,000 and you wanna make as much money as you can, according to this model based on that. Now this is exactly the kind of problem that the Lagrange multiplier technique is made for. We're trying to maximize some kind of function and we have a constraint. Now right now the constraint isn't written as a formula, but we can pretty easily write as a formula. Because what makes up our budget? Well, it's gonna be the number of hours of labor multiplied by 20, so that's gonna be $20 per hour multiplied by the number of hours you put in plus $2000 per tons of steel times the tons of steel that you put in. So the constraint is basically that you have to have these guys equal $20,000. I mean you could say less than, you could say you're not willing to go any more than that, but intuitively and in reality, it's gonna be the case that in order to maximize your revenues, you should squeeze every dollar that you have available and actually hit this constraint. So this right here is the constraint of our problem. And let's go ahead and give this guy a name, the function that we're dealing with a name. And I'm gonna call it g of h, s, which is gonna be that guy. And now if you'll remember in the last few videos, the way we visualize something like this, is to think about the set of all possible input, so in this case, you might be thinking about the h s plane, the number of hours of labor on one axis, the number of tons of steel on another. And this constraint, well, in this case, it's a linear function, so this constraint is gonna give us some kind of line that tells us which pairs of s and h are gonna achieve that constraint. And then the revenue function that we're dealing with will have certain contours. Maybe revenues of $10,000 have a certain contour that looks like this and revenues of $100,000 have a certain contour that looks like this. But what we want is to find which value is barely touching the constraint curve. Just tangent to it at a given point cause that's gonna be the contour line, where if you up the value by just a little bit, it would no longer intersect with that curve, there would no longer be values of h and s that satisfy this constraint. And the way to think about finding that tangency is to consider the vector perpendicular to the tangent line to the curve at that point, which fortunately is represented by... (mumbles) Make some room for myself here. Represented by the gradient, the gradient of our r function, the function whose contour's this is, the revenue. And what it means for this to be tangent to the constraint line is that there's gonna be another vector, the gradient of g, of our constraint function, that points in the same direction, that's proportional to that. And typically the way you write this is to say that the gradient of this function is proportional to the gradient of g and this proportionality constant is called our Lagrange multiplier. It's called the Lagrange multiplier. So, let's go ahead and start working it out. Let's first compute the gradient of r. So the gradient of r is gonna be the partial derivative of r with respect to its first variable, which is h, so partial derivative with respect to h. And the second component is its partial derivative with respect to that second variable, s, with respect to s. And in this case, that first partial derivative, if we treat h as a variable and s as a constant, then that 2/3 gets brought down, so that'll be 100 times 2/3 times h to the power of, well, we've gotta subtract one from 2/3 when we bring it down, so that'll be negative 1/3 multiplied by s to the 1/3. And then the second component here, the partial derivative with respect to s, is gonna be 100 times, well now, by treating s as the variable, we take down that 1/3, so that's 1/3, h to the 2/3 just looks like a constant as far as s is concerned and then we take s to the 1/3 minus 1, which is negative 2/3. Great, so that's the gradient of R. And now we need the gradient of g. And that one's a lot easier, actually, cause g is just a linear function. So when we take the gradient of g, which is its partial derivative, with respect to h, partial h, and its partial derivative with respect to s, partial s. Well, the partial with respect to h is just 20. The function looks like 20 times h plus something that's a constant so that ends up being 20. And then the partial with respect to s, likewise is just 2,000, cause it's just some constant multiplied by s plus a bunch of other stuff that looks like constants. So that's great, and this means when we set the gradient of r equal to the gradient of g, the pair of equations that we get, here let me just write it all out again, is we have this top one, which I'll call 200/3 times, and let's go ahead and do a little simplifying while I'm rewriting things here. So h to the 1/3 is really one over, h to the negative 1/3, sorry, is one over h to the 1/3, that's s to the 1/3. So all of this, that first component, is being set equal to the first component of the gradient of g, which is 20 times lambda, times this Lagrange multiplier, cause we're not setting the gradients equal to each other, we're just setting them proportional to each other. So that's the first equation and then the second one, I'll go ahead and do some simplifying while I rewrite that one also. That's gonna be 100/3 and then h to the 2/3 so times h to the 2/3 divided by s to the 2/3 cause s to the negative 2/3 is the same as 1 over s to the 2/3. All of that is equal to 2,000 times lambda. And the important thing it's that same lambda, because the entire vector has to be proportional. And I think right here's probably a pretty good point to stop and in the next video, I'll go ahead and work through the details and we'll land on a solution.