The important sentence/fact used here is
Any $a \in \mathcal{R}_q$ is a $\mathbb{Z}$-linear function from $\mathcal{R}$ to $\mathcal{R}_q$, so it can be represented by a square matrix $\mathbf{A}_a \in \mathbb{Z}_q^{n \times n}$.
First, note that $\mathcal{R} \leftrightarrow \mathbb{Z}^n$. For example, we can easily see $\mathcal{R} = \mathbb{Z}[x] / (x^n - 1)$ is isomorphic to $\mathbb{Z}^n$: There is a (linear algebraic) basis $\{1, x, \cdots, x^{n - 1}\}$ of $\mathcal{R}$, so we can identify their coefficients as a vector. Take $n = 4$ and $\mathcal{R} = \mathbb{Z}[x] / (x^4 - 1)$. Then, the element $a = 3 - 5x + 27x^2 + x^3$ is identified by $(3, -5, 27, 1) \in \mathcal{R}$. This is called the coefficient embedding. Note that this works for other rings, say $\mathbb{Z}[x] / (f(x))$ for any polynomial $f$.
Now, consider a fixed element $a = a_0 + a_1x + \cdots + a_{n - 1}x^{n - 1} \in \mathcal{R}$ and a varying element $b = b_0 + b_1x + \cdots + b_{n - 1}x^{n - 1} \in \mathcal{R}$. To multiply them, we can write
$$
\color{blue}{ab = \sum_i b_i (ax^i)}
$$
Here comes the important part: the term $ax^i$ is fixed! In the case of $\mathcal{R} = \mathbb{Z}[x] / (x^n - 1)$, it even has a particularly easy to see form. Indeed,
\begin{align*}
i = 0 \implies ax^i &= a_0 + a_1x + \cdots + a_{n - 1}x^{n - 1}\\
i = 1 \implies ax^i &= a_0x + a_1x^2 + \cdots + a_{n - 1}x^n \\
&= a_{n - 1} + a_0x + a_1x^2 + \cdots + a_{n - 2}x^{n - 1}\\
i = 2 \implies ax^i &= a_0x^2 + a_1x^3 + \cdots + a_{n - 1}x^{n + 1} \\ &= a_{n - 2} + a_{n - 1}x + a_0x^2 + \cdots + a_{n - 3}x^{n - 1}
\end{align*}
If the structure is not clear, perhaps writing them in vector form will be clear: performing the coefficient embedding we described, we have
\begin{align*}
i = 0 \implies ax^0 &= (a_0, a_1, \cdots, a_{n - 1}) \\
i = 1 \implies ax^1 &= (a_{n - 1}, a_0, a_1, \cdots, a_{n - 2}) \\
&\vdots \\
i = n - 2 \implies ax^{n - 2} &= (a_2, a_3, \cdots, a_{n - 1}, a_0, a_1) \\
i = n - 1 \implies ax^{n - 1} &= (a_1, a_2, \cdots, a_{n - 1}, a_0) \\
\end{align*}
Finally, to link this to $\color{blue}{ab = \sum_i b_i(ax^i)}$, we observe that by packing $(ax^0, ax^1, \cdots, ax^{n - 2}, ax^{n - 1})$ into (the columns of) matrix $\mathbf{A} = (ax^0 | ax^1 | \cdots | ax^{n - 2} | ax^{n - 1})$, then
$$
ab = \sum_i b_i \underbrace{\mathbf{A}_i}_{\text{column}} = \mathbf{A}_i \begin{pmatrix} b_0 \\ b_1 \\ \cdots \\ b_{n - 1} \end{pmatrix}
$$
The matrix $\mathbf{A}$ described here is exactly what the authors mean by "any $a \in \mathcal{R}_q$ is represented by a square matrix $\mathbf{A} \in \mathbb{Z}^{n \times n}$". As you can see, this is also why the matrix is circulant.
Hope this helps and isn't too long, I am not sure what you know and don't :)
Note 1: This can be done for any ring $\mathcal{R}$ - we say it is a $\mathbb{Z}$-module. However, the matrix $\mathbf{A}_a$ will not necessarily be circulant.
Note 2: There are embeddings apart from the coefficient embedding I described above. For example, the Minkowski embedding is used in cryptography a lot. In the case of cyclotomic rings it's also convenient, since the two embeddings are equivalent even when considering angles (orthonormal basis). But I will leave that for later. Pretty sure it's described in the survey.
Note 3: The map $\mathcal{R} \to \mathbb{Z}^n$ is usually denoted $\mathbf{rot}$ in the academia.
