I am reading Boyd & Vandenberghe’s

Convex Optimization. The authors use curved greater than or equal to (`\succcurlyeq`

)f(x∗)≽α

and curved less than or equal to (

`\preccurlyeq`

)f(x∗)≼α

Can someone explain what they mean?

**Answer**

Both Chris Culter’s and Code Guru’s answers are good, and I’ve voted them both up. I hope that I’m not being inappropriate by combining and expanding upon them here.

It should be noted that the book does not use \succeq, \preceq, \succ, and \prec with *scalar* inequalities; for these, good old-fashioned inequality symbols suffice. It is only when the quantities on the left- and right-hand sides are vectors, matrices, or other multi-dimensional objects that this notation is called for.

The book refers to these relations as *generalized inequalities*, but as Code-Guru rightly points out, they have been in use for some time to represent partial orderings. And indeed, that’s exactly what they are, and the book does refer to them that way as well. But given that the text deals with convex optimization, it was apparently considered helpful to refer to them as inequalities.

Let S be a vector space, and let K\subset S be a closed, convex, and pointed cone with a non-empty interior. (By *cone*, we mean that \alpha K\equiv K for all \alpha>0; and by *pointed*, we mean that K\cap-K=\{0\}.) Such a cone K induces a *partial ordering* on the set S, and an associated set of generalized inequalities:

x \succeq_K y \quad\Longleftrightarrow\quad y \preceq_K x \quad\Longleftrightarrow\quad x – y \in K

x \succ_K y \quad\Longleftrightarrow\quad y \prec_K x \quad\Longleftrightarrow\quad x – y \in \mathop{\textrm{Int}} K

This is a *partial* ordering because, for many pairs x,y\in S, x \not\succeq_K y and y \not\succeq_K x. So that’s the primary reason why he and others prefer to use the curly inequalities to denote these orderings, reserving \geq, \leq, etc. for total orderings. But it has many of the properties of a standard inequality, such as:

x\succeq_K y \quad\Longrightarrow\quad \alpha x \succeq_K \alpha y \quad\forall \alpha>0

x\succeq_K y \quad\Longrightarrow\quad \alpha x \preceq_K \alpha y \quad\forall \alpha<0

x\succeq_K y, ~ x\preceq_K y \quad\Longrightarrow\quad x=y

x\succ_K y \quad\Longrightarrow\quad x\not\prec_K y

When the cone K is understood from context, it is often dropped, leaving only the inequality symbol \succeq. There are two cases where this is almost always done. First, when S=\mathbb{R}^n and the cone K is non-negative orthant \mathbb{R}^n_+ the generalized inequality is simply an elementwise inequality:

x \succeq_{\mathbb{R}^n_+} y \quad\Longleftrightarrow\quad x_i\geq y_i,~i=1,2,\dots,n

Second, when S is the set of symmetric n\times n matrices and K is the cone of positive semidefinite matrices \mathcal{S}^n_+=\{X\in S\,|\,\lambda_{\text{min}}(X)\geq 0\}, the inequality is a *linear matrix inequality (LMI)*:

X \succeq_{\mathcal{S}^n_+} Y \quad\Longleftrightarrow\quad \lambda_{\text{min}}(X-Y)\geq 0

In both of these cases, the cone subscript is almost always dropped.

Many texts in convex optimization don't bother with this distinction, and use \geq and \leq even for LMIs and other partial orderings. I prefer to use it whenever I can, because I think it helps people realize that this is not a standard inequality with an underlying total order. That said, I don't feel that strongly about it for \mathbb{R}^n_+; I think most people rightly assume that x\geq y is considered elementwise when x,y are vectors.

**Attribution***Source : Link , Question Author : Dinesh K. , Answer Author : Chris Culter*