diff --git a/docs/nstrophy_doc.tex b/docs/nstrophy_doc.tex
index 09efcba..9092dea 100644
--- a/docs/nstrophy_doc.tex
+++ b/docs/nstrophy_doc.tex
@@ -101,9 +101,8 @@ We truncate the Fourier modes and assume that $\hat u_k=0$ if $|k_1|>K_1$ or $|k
   \mathcal K:=\{(k_1,k_2),\ |k_1|\leqslant K_1,\ |k_2|\leqslant K_2\}
   .
 \end{equation}
-\bigskip
 
-\point{\bf Runge-Kutta methods}.
+\subsubsection{Runge-Kutta methods}.
 To solve the equation numerically, we will use Runge-Kutta methods, which compute an approximate value $\hat u_k^{(n)}$ for $\hat u_k(t_n)$.
 {\tt nstrophy} supports the 4th order Runge-Kutta ({\tt RK4}) and 2nd order Runge-Kutta ({\tt RK2}) algorithms.
 In addition, several variable step methods are implemented:
@@ -198,9 +197,8 @@ It can be made by specifying the parameter {\tt adaptive\_norm}.
   \end{equation}
   This norm is selected by choosing {\tt adaptive\_norm=enstrophy}.
 \end{itemize}
-\bigskip
 
-\point{\bf Reality}.
+\subsubsection{Reality}.
 Since $U$ is real, $\hat U_{-k}=\hat U_k^*$, and so
 \begin{equation}
   \hat u_{-k}=\hat u_k^*
@@ -222,9 +220,8 @@ Thus,
   \label{realT}
 \end{equation}
 In order to keep the computation as quick as possible, we only compute and store the values for $k_1\geqslant 0$.
-\bigskip
 
-\point{\bf FFT}. We compute T using a fast Fourier transform, defined as
+\subsubsection{FFT}. We compute T using a fast Fourier transform, defined as
 \begin{equation}
   \mathcal F(f)(n):=\sum_{m\in\mathcal N}e^{-\frac{2i\pi}{N_1}m_1n_1-\frac{2i\pi}{N_2}m_2n_2}f(m_1,m_2)
 \end{equation}
@@ -267,9 +264,8 @@ Therefore,
     \mathcal F\left(q_x|q|\hat u_q\right)(n)
   \right)(k)
 \end{equation}
-\bigskip
 
-\point{\bf Energy}.
+\subsubsection{Energy}.
 We define the energy as
 \begin{equation}
   E(t)=\frac12\int\frac{dx}{L^2}\ U^2(t,x)=\frac12\sum_{k\in\mathbb Z^2}|\hat U_k|^2
@@ -370,18 +366,16 @@ and
   .
   \label{enstrophy}
 \end{equation}
-\bigskip
 
-\point{\bf Enstrophy}.
+\subsubsection{Enstrophy}.
 The enstrophy is defined as
 \begin{equation}
   \mathcal En(t)=\int\frac{dx}{L^2}\ |\nabla U|^2
   =\frac{4\pi^2}{L^2}\sum_{k\in\mathbb Z^2}k^2|\hat U_k|^2
   .
 \end{equation}
-\bigskip
 
-\point{\bf Numerical instability}.
+\subsubsection{Numerical instability}.
 In order to prevent the algorithm from blowing up, it is necessary to impose the reality of $u(x)$ by hand, otherwise, truncation errors build up, and lead to divergences.
 It is sufficient to ensure that the convolution term $T(\hat u,k)$ satisfies $T(\hat u,-k)=T(\hat u,k)^*$.
 After imposing this condition, the algorithm no longer blows up, but it is still unstable (for instance, increasing $K_1$ or $K_2$ leads to very different results).