First of all thank you for your time and help. I am more on programming side and have average knowledge of calculus so less formal on mathematical notation. I have encountered a problem in non linear equality constraint optimization

The objective function which has to be minimized $$ ax_1^2 + bx_2^2 + cx_3^2 $$

subject to constraint

$$ \frac {1}{x_1} + \frac {1}{x_2} + \frac {1}{x_3} = T $$

So the Lagrangian is

$$ L(..)=ax_1^2 + bx_2^2 + cx_3^2 + \lambda (\frac {1}{x_1} + \frac {1}{x_2} + \frac {1}{x_3} -T) $$

at times the coefficient c takes the value c = o, which results in

$$ ax_1^2 + bx_2^2 $$

subject to constraint

$$ \frac {1}{x_1} + \frac {1}{x_2} + \frac {1}{x_3} = T $$

So the new Lagrangian is

$$ L_1(..)=ax_1^2 + bx_2^2 + \lambda (\frac {1}{x_1} + \frac {1}{x_2} + \frac {1}{x_3} -T) $$

Question: My understanding is that $L_1$ has no valid analytical solution nor numerical solution. I am looking for concrete theoretical reasons to convince my senior that in case of c=0 the optimization problem is no longer valid minimization problem. Preferably independent of $L_1$, considering only the objective function and constraint. I will appreciate the multiple explanations from the perspective of geometry, calculus e.t.c.


Let's suppose that the objective function attains a minimum on $x_1,x_2,x_3>0$. Say that minimum is $M$, achieved at $(x_1, x_2, x_3)$.

Then $(x_1, x_2, x_3)$ lies on the surface $\frac{1}{x_1} + \frac{1}{x_2} + \frac{1}{x_3} = T$ and also on $a x_1^2 + b x_2^2 + c x_3^2 = M$.

That latter surface is an ellipsoid when $c \not = 0$; the former surface is like so. (The first image is viewed from afar; the second is looking directly down on the point $(1,1,1)$ from the origin.)


End on

We fixed $M$ to be a specific value; now let it vary. It's plausible (indeed, obvious!) that if we make the ellipsoid too small (decreasing $M$ sufficiently) then it will stop intersecting with that surface eventually: it'll start fitting below the bend you see in the picture. So actually there is a minimum of the objective function.

But when $c = 0$, the objective function is now plotting a cylindrical prism rather than an ellipsoid, and now decreasing $M$ just makes the cylinder more squashed. The cylinder will still extend all the way up to infinity, and eventually it'll meet the ever-so-gradually-curving surface, no matter how small $M$ is. So whatever value of $M$ we picked, we could decrease $M$ and still find a solution to the system.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.