diff --git a/include/hpcombi/epu8.hpp b/include/hpcombi/epu8.hpp
index 36587c8..b13301b 100644
--- a/include/hpcombi/epu8.hpp
+++ b/include/hpcombi/epu8.hpp
@@ -51,7 +51,7 @@ operator"" _u8(unsigned long long arg) noexcept {  // NOLINT
 epu8 stands for *Extended Packed Unsigned, grouped by 8 bits*;
 this is the low level type chosen by Intel for their API to intrinsics,
 ie a SIMD vector of 16 unsigned bytes (16×8 = 128bits).
-Functions using this type uses semantically equivalent types,
+Functions using this type use semantically equivalent types,
 eg a _m128 which is 2 vect of 64bits.
 a flag tells the compiler to silently consider those types equivalent.
  */
diff --git a/include/hpcombi/hpcombi.hpp b/include/hpcombi/hpcombi.hpp
index 352c10d..fcb094a 100644
--- a/include/hpcombi/hpcombi.hpp
+++ b/include/hpcombi/hpcombi.hpp
@@ -83,5 +83,6 @@ Data structure should preserve locality. You might want to compute some stats on
 This lib is implemented with speed in mind, not code safety.
 Eg. there are no checks when building a permutation, which could be invalid (like non injective).
 
-We now suggest to have a look, in the menus above, at Classes → [Class list](annotated.html).
+We now suggest to have a look, in the menus above, at Classes → [Class list](annotated.html),
+esp. at classes are HPCombi::Perm16 and HPCombi::BMat8.
 */
\ No newline at end of file
diff --git a/include/hpcombi/perm16.hpp b/include/hpcombi/perm16.hpp
index 0b17751..f4098f0 100644
--- a/include/hpcombi/perm16.hpp
+++ b/include/hpcombi/perm16.hpp
@@ -195,6 +195,7 @@ struct PPerm16 : public PTransf16 {
      * @verbatim {0,0xFF,2,1,3,5,6,0xFF,8,9,0xFF,10,12,0xFF,0xFF,0xFF}
      * @endverbatim
      */
+
     /** @copydoc common_inverse_pperm
      *  @par Algorithm:
      *  @f$O(n)@f$ algorithm using reference cast to arrays
@@ -253,42 +254,50 @@ struct Perm16 : public Transf16 /* public PPerm : diamond problem */ {
      * Returns
      * @verbatim {0,4,2,1,3,5,6,7,8,9,10,11,12,13,14,15} @endverbatim
      */
+
     /** @copydoc common_inverse
      *  @par Algorithm:
      *  Reference @f$O(n)@f$ algorithm using loop and indexed access
      */
     Perm16 inverse_ref() const;
+
     /** @copydoc common_inverse
      *  @par Algorithm:
      *  @f$O(n)@f$ algorithm using reference cast to arrays
      */
     Perm16 inverse_arr() const;
+
     /** @copydoc common_inverse
      *  @par Algorithm:
      *  Insert the identity in the least significant bits and sort using a
-     *  sorting network. The number of round of the optimal sorting network is
+     *  sorting network. The number of rounds of the optimal sorting network is
      *  open as far as I know, therefore the complexity is unknown.
      */
     Perm16 inverse_sort() const;
+
     /** @copydoc common_inverse
      *  @par Algorithm:
      *  @f$O(\log n)@f$ algorithm using some kind of vectorized dichotomic
      * search.
      */
     Perm16 inverse_find() const { return permutation_of(v, one()); }
+
     /** @copydoc common_inverse
      *  @par Algorithm:
      *
-     * Raise \e *this to power @f$\text{LCM}(1, 2, ..., n) - 1@f$ so complexity
-     * is in @f$O(log (\text{LCM}(1, 2, ..., n) - 1)) = O(n)@f$
+     * Use HPCombi::pow to
+     * raise \e *this to power @f$\text{LCM}(1, 2, ..., n) - 1@f$ so complexity
+     * is @f$O(log (\text{LCM}(1, 2, ..., n) - 1)) = O(n)@f$
      */
     Perm16 inverse_pow() const;
+
     /** @copydoc common_inverse
      *  @par Algorithm:
      *  Compute power from @f$n/2@f$ to @f$n@f$, when @f$\sigma^k(i)=i@f$ then
      *  @f$\sigma^{-1}(i)=\sigma^{k-1}(i)@f$. Complexity @f$O(n)@f$
      */
     Perm16 inverse_cycl() const;
+
     /** @copydoc common_inverse
      *
      *  Frontend method: currently aliased to #inverse_cycl */