diff --git a/src/gradient.cpp b/src/gradient.cpp
index deee6347..4a58dcd7 100644
--- a/src/gradient.cpp
+++ b/src/gradient.cpp
@@ -22,7 +22,8 @@
 // UMAP
 
 umap_gradient::umap_gradient(const double a, const double b, const double gamma) :
-  a(a), b(b), a_b_m2(-2.0 * a * b), gamma_b_2(2.0 * gamma * b) {  }
+  clip_max(4.0), a(a), b(b), a_b_m2(-2.0 * a * b), gamma_b_2(2.0 * gamma * b)
+  { }
 
 // The Python implementation does two pow calls: one for d^b, and one for
 // d^(b-1) I have replaced the second call with (d^b) / d, which saves on the
@@ -73,7 +74,8 @@ double fastPrecisePow(double a, double b) {
 
 // Approximate-Power UMAP: UMAP with an approximation to the slow pow calculation
 apumap_gradient::apumap_gradient(const double a, const double b, const double gamma) :
-  a(a), b(b), a_b_m2(-2.0 * a * b), gamma_b_2(2.0 * gamma * b) {  }
+  clip_max(4.0), a(a), b(b), a_b_m2(-2.0 * a * b), gamma_b_2(2.0 * gamma * b)
+  {  }
 
 const double apumap_gradient::grad_attr(const double dist_squared) const {
   const double pd2b = fastPrecisePow(dist_squared, b);
@@ -87,7 +89,7 @@ const double apumap_gradient::grad_rep(const double dist_squared) const {
 
 // t-UMAP
 
-tumap_gradient::tumap_gradient() {}
+tumap_gradient::tumap_gradient() : clip_max(4.0) {}
 
 const double tumap_gradient::grad_attr(const double dist_squared) const {
   return -2.0 / (dist_squared + 1.0);
@@ -100,7 +102,8 @@ const double tumap_gradient::grad_rep(const double dist_squared) const {
 
 // LargeVis
 
-largevis_gradient::largevis_gradient(const double gamma) : gamma_2(gamma * 2.0) {}
+largevis_gradient::largevis_gradient(const double gamma) : clip_max(5.0),
+gamma_2(gamma * 2.0) {}
 
 const double largevis_gradient::grad_attr(const double dist_squared) const {
   return -2.0 / (dist_squared + 1.0);
diff --git a/src/gradient.h b/src/gradient.h
index 3aa9cc3c..27c2dc67 100644
--- a/src/gradient.h
+++ b/src/gradient.h
@@ -27,7 +27,8 @@ class umap_gradient {
   umap_gradient(const double a, const double b, const double gamma);
   const double grad_attr(const double dist_squared) const;
   const double grad_rep(const double dist_squared) const;
-  static const constexpr double clip_max = 4.0;
+  // static const constexpr double clip_max = 4.0;
+  const double clip_max;
 private:
   const double a;
   const double b;
@@ -41,7 +42,8 @@ class apumap_gradient {
   apumap_gradient(const double a, const double b, const double gamma);
   const double grad_attr(const double dist_squared) const;
   const double grad_rep(const double dist_squared) const;
-  static const constexpr double clip_max = 4.0;
+  // static const constexpr double clip_max = 4.0;
+  const double clip_max;
 private:
   const double a;
   const double b;
@@ -60,7 +62,8 @@ class tumap_gradient {
   tumap_gradient();
   const double grad_attr(const double dist_squared) const;
   const double grad_rep(const double dist_squared) const;
-  static const constexpr double clip_max = 4.0;
+  // static const constexpr double clip_max = 4.0;
+  const double clip_max;
 };
 
 class largevis_gradient {
@@ -68,7 +71,8 @@ class largevis_gradient {
   largevis_gradient(const double gamma);
   const double grad_attr(const double dist_squared) const;
   const double grad_rep(const double dist_squared) const;
-  static const constexpr double clip_max = 5.0;
+  // static const constexpr double clip_max = 5.0;
+  const double clip_max;
 
 private:
   const double gamma_2;