14
$\begingroup$

I'm looking at some formulas involving matrices (in the context of machine learning, but I'm not sure it's relevant) and I came across $\odot$. What could this mean? The context is $M \odot N$, where $M$ is a matrix and $N$ might be a vector, or a matrix, or a scalar, it's a bit dense so it's hard to tell. I have reason to believe it may be the Hadamard product, is there anything else it could mean?

$\endgroup$
  • 5
    $\begingroup$ Could you give any more context? It could just mean $M$ acts on $v$ by multiplication. $\endgroup$ – Ian Coley May 30 '14 at 20:17
  • 2
    $\begingroup$ I think you might have meant "dot in a circle", rather than "circle in a dot", in your title. $\endgroup$ – Robert Lewis May 30 '14 at 20:21
  • 1
    $\begingroup$ I have seen it used to define simple "non-standard" operators when learning group theory so that there is no confusion with $+$ or $\times$. e.g. $M\odot v=Mv+(v^TMv)I$ is something that I remember as an example. The context of your book should clarify it though. $\endgroup$ – Daryl May 30 '14 at 20:28
  • $\begingroup$ here is an exampe, at 46. min youtu.be/iX5V1WpxxkY definition of LSTM $\endgroup$ – jhegedus Sep 1 '16 at 19:49
  • $\begingroup$ The Sun, maybe? en.wikipedia.org/wiki/Solar_symbol $\endgroup$ – Oscar Lanzi Oct 27 '16 at 1:09
14
$\begingroup$

In the LSTM equations, the circled dot operator is typically used to represent element-wise multiplication.

$\endgroup$
  • $\begingroup$ Is it possible to reconcile this interpretation with the OP's claim that "$M$ is a matrix and $N$ might be a vector, or a matrix, or a scalar"? $\endgroup$ – hardmath Sep 21 '16 at 1:33
4
$\begingroup$

When I researched about the symbol ⊙. I got two things:

$1$. In the book Meaning, Logic and Ludics-By Alain Lecomte, The writer says that: Let $U$ and $B$ be two positive designs we denote tensor product of $U$ and $B$ by $U$⊙$B$.

$2$. In the book Dag Prawitz on Proofs and Meaning: The writer says that $\alpha⊙\beta$ means that either $\beta $ is derivable from $\alpha$ or is $\alpha$ derivable from $\beta $.

I think that second on e makes sense, as for tensor product I have always seen the symbol ⊗ being used. I provided you the info which I had, hope it helps.

$\endgroup$
  • $\begingroup$ This is the correct answer! $\endgroup$ – Yahya Nov 7 '19 at 10:43
4
$\begingroup$

We use the circle with a dot notation in nonlinear continuum mechanics. It is defined in index notation as \begin{align} \left(\boldsymbol{M} ⊙ \boldsymbol{N}\right)_{ABCD}= \dfrac{1}{2}\left(M_{AC}N_{BD}+M_{AD}N_{BC}\right) \end{align} where $\boldsymbol{M}$ and $\boldsymbol{N}$ are second order tensors. The result of this operation is a 4th order tensor. You can think of it as related to the outer product of two matrices or $\boldsymbol{M}\otimes \boldsymbol{N}$ in that it takes two 2nd order tensors and generates a 4th order tensor. Remember the outer product is defined for two second order tensors as

\begin{align} \left(\boldsymbol{M} \otimes \boldsymbol{N}\right)_{ABCD}= M_{AB}N_{CD} \end{align}

The circle with a dot operation $⊙$ occurs when calculating the elastic modulus $\mathbf{\underline{C}}$ (a 4th order tensor) from constitutive laws. Examples include the Mooney-Rivlin or Neohookean material models. To get the elastic modulus you must take derivatives of tensors with respect to other tensors. For instance,

\begin{align} \underline{\mathbf{C}} = 2\dfrac{d\boldsymbol{S}}{d\boldsymbol{C}} \end{align} or in index notation as \begin{align} \text{C}_{ABCD} = 2\dfrac{\partial S_{AB}}{\partial C_{CD}} \end{align} where $\boldsymbol{C}$ is the left Cauchy-Green deformation tensor and $\boldsymbol{S}$ is the 2nd Piola-Kirchhoff stress tensor. Based on the constitutive models, which relate stress in a material to the strain, it turns out $\boldsymbol{S}$ is a function of the inverse of $\boldsymbol{C}$. If you work it out you will find that you will have to compute $\frac{\partial \boldsymbol{C}^{-1}}{\partial \boldsymbol{C}}$. The proof is a headache but the result comes out to \begin{align} \dfrac{\partial \boldsymbol{C}^{-1}}{\partial \boldsymbol{C}} = -\boldsymbol{C}^{-1} ⊙ \boldsymbol{C}^{-1} \end{align} In index notation \begin{align} \dfrac{\partial C^{-1}_{AB}}{\partial C_{CD}} = -\dfrac{1}{2}\left(C^{-1}_{AC}C^{-1}_{BD}+C^{-1}_{AD}C^{-1}_{BC}\right)=-\left(\boldsymbol{C}^{-1} ⊙ \boldsymbol{C}^{-1}\right)_{ABCD} \end{align} The circle with a dot operation only arises because $\boldsymbol{C}$ is a symmetric matrix, i.e., $\boldsymbol{C}=\boldsymbol{C}^{T}$ and $\boldsymbol{C}_{sym}=\dfrac{1}{2}\left(\boldsymbol{C}+\boldsymbol{C}^{T}\right) = \boldsymbol{C}$. Note that if taking the derivative of an inverse of a nonsymmetric tensor with respect to itself yields \begin{align} \dfrac{\partial A_{AB}^{-1}}{\partial A_{CD}}=-A^{-1}_{AC}A^{-1}_{DB} \end{align} and this is not the outer product. This operation has not yet been given a symbol.

Note:

  • the outer product $\otimes$ is also called the tensor product.
  • The indices are capital letters ABCD since in continuum mechanics capital letters denote Lagrangian/material (or reference configuration) coordinates. Lower case indices ijkl denote spatial/Eulerian coordinates.

Addendum: Other uses I have seen for $⊙$ include

  1. In physics, I have seen it mean a point source such as a point charge or gravity source like a planet.
  2. In physics, I have seen it mean the vector points out of the page $⊙$. And $\otimes$ means the direction of the vector is into the page. I have seen this in E&M for B-fields and E-fields and mechanics for torques.
  3. In mathematics it could mean a function composition operator, which maps functions to functions, e.g., $\,f⊙g$.

This is what I think it is used for in that Long Short-Term Memory neural networking lecture https://www.youtube.com/watch?v=iX5V1WpxxkY&feature=youtu.be at 46:00.

In mathematics, functional compositions are usually denoted by a small circle $\circ$. For example, Eulerian $f(\boldsymbol{x},t)$ and Lagrangian $F(\boldsymbol{X},t)$ descriptions are related to each other by a function composition: \begin{align} F(\boldsymbol{X},t)=f(\boldsymbol{\Phi}(\boldsymbol{X},t),t) \textrm{ or } F = f \circ \boldsymbol{\Phi} \end{align}

$\endgroup$

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.