0
$\begingroup$

I come from a programming background. I am familiar with scalar calculus but not so much with vector/matrix calculus.

I am trying to understand stochastic gradient descent for multiple linear regression and needed to understand how after this step:

$\ 𝑓(W):=‖XW−Y‖^2_F=tr((XW−Y)^⊤(XW−Y)) = tr(W^⊤X^⊤XW−Y^⊤XW−W^⊤X^⊤Y+Y^⊤Y)$

follows:

$\nabla_{\mathrm W} f (\mathrm W) = 2 \, \mathrm X^{\top} \mathrm X \mathrm W - 2 \, \mathrm X^{\top} \mathrm Y = 2 \, \mathrm X^{\top} \left( \mathrm X \mathrm W - \mathrm Y \right)$

where tr represents the trace, X is the design matrix, W is the Coefficient Matrix, f(W) is the cost function and Y is the target label.

Also I know there is another approach to finding solutions for such problem using Frobenius inner product but I have absolutely no idea about Frobenius inner products.

So I just want clear steps on how the derivation follows with the rules/laws used if there are any.

$\endgroup$
  • $\begingroup$ You clearly know some of the lingo, but exactly how comfortable are you with inner products? Have you encountered abstract inner products before? What about calculus on these spaces? Do you know the definition of $\nabla_W$? $\endgroup$ – Theo Bendit Feb 16 at 0:51
  • $\begingroup$ Sorry but "No" for all of the questions you have asked except the last one. I think I do understand what the gradient with respect to the coefficient vector W means. $\endgroup$ – A-ar Feb 16 at 0:55
  • $\begingroup$ Which definition do you use? $\endgroup$ – Theo Bendit Feb 16 at 0:56
  • $\begingroup$ vector consisting of all the partial derivatives of cost function wrt each coefficient variable. $\endgroup$ – A-ar Feb 16 at 0:59
2
$\begingroup$

The Frobenius product is just a convenient notation for the trace $$A:BC = {\rm tr}(A^TBC)$$ The product distributes over addition $$A:(B+C) = A:B + A:C$$ and the cyclic property of the trace allows a product to be rearranged in various ways, e.g. $$\eqalign{ A:BC &= BC:A \cr &= A^T:(BC)^T \cr &= B^TA:C \cr &= AC^TA:B \cr }$$ Define the matrix $A=(XW-Y)$ whose full differential is $$\eqalign{dA &= dX\,W + X\,dW - dY \cr}$$ However, in the current problem, $(X,Y)$ don't change, so their differentials $(dX,dY)$ are zero, leaving $$\eqalign{dA &= X\,dW \cr}$$

One last point concerns the differential of various products.
The differential of the Frobenius product is $$\eqalign{d\,(A:B)&=dA:B + A:dB \cr&= B:dA + A:dB}$$ Similarly, the differential of a normal matrix product is $$\eqalign{d\,(AB)&=dA\,B + A\,dB}$$ However, this product cannot be rearranged because it is not commutative.

The differential of a Kronecker product is $$\eqalign{d\,(A\otimes B)&=dA\otimes B + A\otimes dB}$$ The differential of a Hadamard product is $$\eqalign{d\,(A\odot B)&=dA\odot B + A\odot dB\cr&=B\odot dA + A\odot dB }$$ The Hadamard product is commutative, and can be rearranged like the Frobenius product.



Now write the cost function in terms of the new variable.
Then calculate its differential and gradient. $$\eqalign{ f &= A:A \cr df &= 2A:dA = 2A:X\,dW = 2X^TA:dW \cr \frac{\partial f}{\partial W} &= 2X^TA = 2X^T(XW-Y) \cr }$$

$\endgroup$
  • $\begingroup$ I did not understand the part where you took the derivative. How if f = A:A then df = 2A:dA also how is dA =XdW $\endgroup$ – A-ar Feb 16 at 1:13
  • $\begingroup$ I updated the answer with some new information. $\endgroup$ – lynn Feb 16 at 3:21

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.