diff --git a/internal/scalar/scalar.go b/internal/scalar/scalar.go
index 7b75593..fa44c2b 100644
--- a/internal/scalar/scalar.go
+++ b/internal/scalar/scalar.go
@@ -968,7 +968,7 @@ func (s *Scalar) SignedRadix16() [64]int8 {
 	// Compute unsigned radix-16 digits:
 	for i := 0; i < 32; i++ {
 		digits[2*i] = int8(s[i] & 15)
-		digits[2*i+1] = int8((s[i]>>4) & 15)
+		digits[2*i+1] = int8((s[i] >> 4) & 15)
 	}
 
 	// Recenter coefficients:
@@ -980,3 +980,114 @@ func (s *Scalar) SignedRadix16() [64]int8 {
 
 	return digits
 }
+
+// Given k > 0, set s = s**(2*i).
+func (s *Scalar) pow2k(k int) {
+	for i := 0; i < k; i++ {
+		s.Mul(s, s)
+	}
+}
+
+// Inv sets s to the inverse of a nonzero scalar v and returns s.
+func (s *Scalar) Inv(t *Scalar) *Scalar {
+	// Uses a hardcoded sliding window of width 4.
+	var table [8]Scalar
+	var tt Scalar
+	tt.Mul(t, t)
+	table[0] = *t
+	for i := 0; i < 7; i++ {
+		table[i+1].Mul(&table[i], &tt)
+	}
+	// Now table = [t**1, t**3, t**7, t**11, t**13, t**15]
+	// so t**k = t[k/2] for odd k
+
+	// To compute the sliding window digits, use the following Sage script:
+
+	// sage: import itertools
+	// sage: def sliding_window(w,k):
+	// ....:     digits = []
+	// ....:     while k > 0:
+	// ....:         if k % 2 == 1:
+	// ....:             kmod = k % (2**w)
+	// ....:             digits.append(kmod)
+	// ....:             k = k - kmod
+	// ....:         else:
+	// ....:             digits.append(0)
+	// ....:         k = k // 2
+	// ....:     return digits
+
+	// Now we can compute s roughly as follows:
+
+	// sage: s = 1
+	// sage: for coeff in reversed(sliding_window(4,l-2)):
+	// ....:     s = s*s
+	// ....:     if coeff > 0 :
+	// ....:         s = s*t**coeff
+
+	// This works on one bit at a time, with many runs of zeros.
+	// The digits can be collapsed into [(count, coeff)] as follows:
+
+	// sage: [(len(list(group)),d) for d,group in itertools.groupby(sliding_window(4,l-2))]
+
+	// Entries of the form (k, 0) turn into pow2k(k)
+	// Entries of the form (1, coeff) turn into a squaring and then a table lookup.
+	// We can fold the squaring into the previous pow2k(k) as pow2k(k+1).
+
+	*s = table[1/2]
+	s.pow2k(127 + 1)
+	s.Mul(s, &table[1/2])
+	s.pow2k(4 + 1)
+	s.Mul(s, &table[9/2])
+	s.pow2k(3 + 1)
+	s.Mul(s, &table[11/2])
+	s.pow2k(3 + 1)
+	s.Mul(s, &table[13/2])
+	s.pow2k(3 + 1)
+	s.Mul(s, &table[15/2])
+	s.pow2k(4 + 1)
+	s.Mul(s, &table[7/2])
+	s.pow2k(4 + 1)
+	s.Mul(s, &table[15/2])
+	s.pow2k(3 + 1)
+	s.Mul(s, &table[5/2])
+	s.pow2k(3 + 1)
+	s.Mul(s, &table[1/2])
+	s.pow2k(4 + 1)
+	s.Mul(s, &table[15/2])
+	s.pow2k(4 + 1)
+	s.Mul(s, &table[15/2])
+	s.pow2k(4 + 1)
+	s.Mul(s, &table[7/2])
+	s.pow2k(3 + 1)
+	s.Mul(s, &table[3/2])
+	s.pow2k(4 + 1)
+	s.Mul(s, &table[11/2])
+	s.pow2k(5 + 1)
+	s.Mul(s, &table[11/2])
+	s.pow2k(9 + 1)
+	s.Mul(s, &table[9/2])
+	s.pow2k(3 + 1)
+	s.Mul(s, &table[3/2])
+	s.pow2k(4 + 1)
+	s.Mul(s, &table[3/2])
+	s.pow2k(4 + 1)
+	s.Mul(s, &table[3/2])
+	s.pow2k(4 + 1)
+	s.Mul(s, &table[9/2])
+	s.pow2k(3 + 1)
+	s.Mul(s, &table[7/2])
+	s.pow2k(3 + 1)
+	s.Mul(s, &table[3/2])
+	s.pow2k(3 + 1)
+	s.Mul(s, &table[13/2])
+	s.pow2k(3 + 1)
+	s.Mul(s, &table[7/2])
+	s.pow2k(4 + 1)
+	s.Mul(s, &table[9/2])
+	s.pow2k(3 + 1)
+	s.Mul(s, &table[15/2])
+	s.pow2k(4 + 1)
+	s.Mul(s, &table[11/2])
+
+	return s
+}
diff --git a/internal/scalar/scalar_test.go b/internal/scalar/scalar_test.go
index ea167c9..5262ffe 100644
--- a/internal/scalar/scalar_test.go
+++ b/internal/scalar/scalar_test.go
@@ -82,3 +82,17 @@ func TestNonAdjacentForm(t *testing.T) {
 		}
 	}
 }
+
+func TestInvert(t *testing.T) {
+	invertWorks := func(x Scalar) bool {
+		var xInv, check Scalar
+		xInv.Inv(&x)
+		check.Mul(&x, &xInv)
+
+		return check.Equal(&scOne) == 1
+	}
+
+	if err := quick.Check(invertWorks, quickCheckConfig); err != nil {
+		t.Error(err)
+	}
+}