Improve taylor series calculation of exp and sin by binary splitting method

exp and sin becomes orders of magnitude faster.
To make log and atan also fast, log and atan now depends on exp and sin.
log(x): solve exp(y)-x=0 by Newton's method
atan(x): solve tan(y)-x=0 by Newton's method

Author: tompng

lib/bigdecimal.rb

@@ -49,6 +49,46 @@ def self.nan_computation_result # :nodoc:
       end
       BigDecimal::NAN
     end
+
+    # Iteration for Newton's method with increasing precision
+    def self.newton_loop(prec, initial_precision: 8, safe_margin: 2) # :nodoc:
+      precs = []
+      while prec > initial_precision
+        precs << prec
+        prec = (precs.last + 1) / 2 + safe_margin
+      end
+      precs.reverse_each do |p|
+        yield p
+      end
+    end
+
+    # Calculates Math.log(x.to_f) considering large or small exponent
+    def self.float_log(x) # :nodoc:
+      Math.log(x._decimal_shift(-x.exponent).to_f) + x.exponent * Math.log(10)
+    end
+
+    # Calculating Taylor series sum using binary splitting method
+    # Calculates f(x) = (x/d0)*(1+(x/d1)*(1+(x/d2)*(1+(x/d3)*(1+...))))
+    # x.n_significant_digits or ds.size must be small to be performant.
+    def self.taylor_sum_binary_splitting(x, ds, prec) # :nodoc:
+      fs = ds.map {|d| [0, BigDecimal(d)] }
+      # fs = [[a0, a1], [b0, b1], [c0, c1], ...]
+      # f(x) = a0/a1+(x/a1)*(1+b0/b1+(x/b1)*(1+c0/c1+(x/c1)*(1+d0/d1+(x/d1)*(1+...))))
+      while fs.size > 1
+        # Merge two adjacent fractions
+        # from: (1 + a0/a1 + x/a1 * (1 + b0/b1 + x/b1 * rest))
+        # to:   (1 + (a0*b1+x*(b0+b1))/(a1*b1) + (x*x)/(a1*b1) * rest)
+        xn = xn ? xn.mult(xn, prec) : x
+        fs = fs.each_slice(2).map do |(a, b)|
+          b ||= [0, BigDecimal(1)._decimal_shift([xn.exponent, 0].max + 2)]
+          [
+            (a[0] * b[1]).add(xn * (b[0] + b[1]), prec),
+            a[1].mult(b[1], prec)
+          ]
+        end
+      end
+      BigDecimal(fs[0][0]).div(fs[0][1], prec)
+    end
   end
 
   #  call-seq:
@@ -202,9 +242,7 @@ def sqrt(prec)
     ex = exponent / 2
     x = _decimal_shift(-2 * ex)
     y = BigDecimal(Math.sqrt(x.to_f), 0)
-    precs = [prec + BigDecimal.double_fig]
-    precs << 2 + precs.last / 2 while precs.last > BigDecimal.double_fig
-    precs.reverse_each do |p|
+    Internal.newton_loop(prec + BigDecimal.double_fig) do |p|
       y = y.add(x.div(y, p), p).div(2, p)
     end
     y._decimal_shift(ex).mult(1, prec)
@@ -239,59 +277,32 @@ def self.log(x, prec)
     return BigDecimal(0) if x == 1
 
     prec2 = prec + BigDecimal.double_fig
-    BigDecimal.save_limit do
-      BigDecimal.limit(0)
-      if x > 10 || x < 0.1
-        log10 = log(BigDecimal(10), prec2)
-        exponent = x.exponent
-        x = x._decimal_shift(-exponent)
-        if x < 0.3
-          x *= 10
-          exponent -= 1
-        end
-        return (log10 * exponent).add(log(x, prec2), prec)
-      end
-
-      x_minus_one_exponent = (x - 1).exponent
 
-      # log(x) = log(sqrt(sqrt(sqrt(sqrt(x))))) * 2**sqrt_steps
-      sqrt_steps = [Integer.sqrt(prec2) + 3 * x_minus_one_exponent, 0].max
-
-      lg2 = 0.3010299956639812
-      sqrt_prec = prec2 + [-x_minus_one_exponent, 0].max + (sqrt_steps * lg2).ceil
-
-      sqrt_steps.times do
-        x = x.sqrt(sqrt_prec)
-      end
-
-      # Taylor series for log(x) around 1
-      # log(x) = -log((1 + X) / (1 - X)) where X = (x - 1) / (x + 1)
-      # log(x) = 2 * (X + X**3 / 3 + X**5 / 5 + X**7 / 7 + ...)
-      x = (x - 1).div(x + 1, sqrt_prec)
-      y = x
-      x2 = x.mult(x, prec2)
-      1.step do |i|
-        n = prec2 + x.exponent - y.exponent + x2.exponent
-        break if n <= 0 || x.zero?
-        x = x.mult(x2.round(n - x2.exponent), n)
-        y = y.add(x.div(2 * i + 1, n), prec2)
-      end
+    if x < 0.1 || x > 10
+      exponent = (3 * x).exponent - 1
+      x = x._decimal_shift(-exponent)
+      return log(10, prec2).mult(exponent, prec2).add(log(x, prec2), prec)
+    end
 
-      y.mult(2 ** (sqrt_steps + 1), prec)
+    # Solve exp(y) - x = 0 with Newton's method
+    # Repeat: y -= (exp(y) - x) / exp(y)
+    y = BigDecimal(BigDecimal::Internal.float_log(x), 0)
+    exp_additional_prec = [-(x - 1).exponent, 0].max
+    BigDecimal::Internal.newton_loop(prec2) do |p|
+      expy = exp(y, p + exp_additional_prec)
+      y = y.sub(expy.sub(x, p).div(expy, p), p)
     end
+    y.mult(1, prec)
   end
 
-  # Taylor series for exp(x) around 0
-  private_class_method def self._exp_taylor(x, prec) # :nodoc:
-    xn = BigDecimal(1)
-    y = BigDecimal(1)
-    1.step do |i|
-      n = prec + xn.exponent
-      break if n <= 0 || xn.zero?
-      xn = xn.mult(x, n).div(i, n)
-      y = y.add(xn, prec)
-    end
-    y
+  private_class_method def self._exp_binary_splitting(x, prec) # :nodoc:
+    return BigDecimal(1) if x.zero?
+    # Find k that satisfies x**k / k! < 10**(-prec)
+    log10 = Math.log(10)
+    logx = BigDecimal::Internal.float_log(x.abs)
+    step = (1..).bsearch { |k| Math.lgamma(k + 1)[0] - k * logx > prec * log10 }
+    # exp(x)-1 = x*(1+x/2*(1+x/3*(1+x/4*(1+x/5*(1+...)))))
+    1 + BigDecimal::Internal.taylor_sum_binary_splitting(x, [*1..step], prec)
   end
 
   # call-seq:
@@ -316,11 +327,21 @@ def self.exp(x, prec)
     prec2 = prec + BigDecimal.double_fig + cnt
     x = x._decimal_shift(-cnt)
 
-    # Calculation of exp(small_prec) is fast because calculation of x**n is fast
-    # Calculation of exp(small_abs) converges fast.
-    # exp(x) = exp(small_prec_part + small_abs_part) = exp(small_prec_part) * exp(small_abs_part)
-    x_small_prec = x.round(Integer.sqrt(prec2))
-    y = _exp_taylor(x_small_prec, prec2).mult(_exp_taylor(x.sub(x_small_prec, prec2), prec2), prec2)
+    # Decimal form of bit-burst algorithm
+    # Calculate exp(x.xxxxxxxxxxxxxxxx) as
+    # exp(x.xx) * exp(0.00xx) * exp(0.0000xxxx) * exp(0.00000000xxxxxxxx)
+    x = x.mult(1, prec2)
+    n = 2
+    y = BigDecimal(1)
+    BigDecimal.save_limit do
+      BigDecimal.limit(0)
+      while x != 0 do
+        partial_x = x.truncate(n)
+        x -= partial_x
+        y = y.mult(_exp_binary_splitting(partial_x, prec2), prec2)
+        n *= 2
+      end
+    end
 
     # calculate exp(x * 10**cnt) from exp(x)
     # exp(x * 10**k) = exp(x * 10**(k - 1)) ** 10

lib/bigdecimal/math.rb

@@ -73,6 +73,37 @@ def sqrt(x, prec)
     end
   end
 
+  private_class_method def _sin_binary_splitting(x, prec) # :nodoc:
+    return x if x.zero?
+    x2 = x.mult(x, prec)
+    # Find k that satisfies x2**k / (2k+1)! < 10**(-prec)
+    log10 = Math.log(10)
+    logx = BigDecimal::Internal.float_log(x.abs)
+    step = (1..).bsearch { |k| Math.lgamma(2 * k + 1)[0] - 2 * k * logx > prec * log10 }
+    # Construct denominator sequence for binary splitting
+    # sin(x) = x*(1-x2/(2*3)*(1-x2/(4*5)*(1-x2/(6*7)*(1-x2/(8*9)*(1-...)))))
+    ds = (1..step).map {|i| -(2 * i) * (2 * i + 1) }
+    x.mult(1 + BigDecimal::Internal.taylor_sum_binary_splitting(x2, ds, prec), prec)
+  end
+
+  private_class_method def _sin_around_zero(x, prec) # :nodoc:
+    # Divide x into several parts
+    # sin(x.xxxxxxxx...) = sin(x.xx + 0.00xx + 0.0000xxxx + ...)
+    # Calculate sin of each part and restore sin(0.xxxxxxxx...) using addition theorem.
+    sin = BigDecimal(0)
+    cos = BigDecimal(1)
+    n = 2
+    while x != 0 do
+      partial_x = x.truncate(n)
+      x -= partial_x
+      s = _sin_binary_splitting(partial_x, prec)
+      c = (1 - s * s).sqrt(prec)
+      sin, cos = (sin * c).add(cos * s, prec), (cos * c).sub(sin * s, prec)
+      n *= 2
+    end
+    sin
+  end
+
   # call-seq:
   #   sin(decimal, numeric) -> BigDecimal
   #
@@ -88,26 +119,9 @@ def sin(x, prec)
     BigDecimal::Internal.validate_prec(prec, :sin)
     x = BigDecimal::Internal.coerce_to_bigdecimal(x, prec, :sin)
     return BigDecimal::Internal.nan_computation_result if x.infinite? || x.nan?
-    n    = prec + BigDecimal.double_fig
-    one  = BigDecimal("1")
-    two  = BigDecimal("2")
+    n = prec + BigDecimal.double_fig
     sign, x = _sin_periodic_reduction(x, n)
-    x1   = x
-    x2   = x.mult(x,n)
-    y    = x
-    d    = y
-    i    = one
-    z    = one
-    while d.nonzero? && ((m = n - (y.exponent - d.exponent).abs) > 0)
-      m = BigDecimal.double_fig if m < BigDecimal.double_fig
-      x1  = -x2.mult(x1,n)
-      i  += two
-      z  *= (i-one) * i
-      d   = x1.div(z,m)
-      y  += d
-    end
-    y = BigDecimal("1") if y > 1
-    y.mult(sign, prec)
+    _sin_around_zero(x, n).mult(sign, prec)
   end
 
   # call-seq:
@@ -125,8 +139,9 @@ def cos(x, prec)
     BigDecimal::Internal.validate_prec(prec, :cos)
     x = BigDecimal::Internal.coerce_to_bigdecimal(x, prec, :cos)
     return BigDecimal::Internal.nan_computation_result if x.infinite? || x.nan?
-    sign, x = _sin_periodic_reduction(x, prec + BigDecimal.double_fig, add_half_pi: true)
-    sign * sin(x, prec)
+    n = prec + BigDecimal.double_fig
+    sign, x = _sin_periodic_reduction(x, n, add_half_pi: true)
+    _sin_around_zero(x, n).mult(sign, prec)
   end
 
   # call-seq:
@@ -164,28 +179,21 @@ def atan(x, prec)
     x = BigDecimal::Internal.coerce_to_bigdecimal(x, prec, :atan)
     return BigDecimal::Internal.nan_computation_result if x.nan?
     n = prec + BigDecimal.double_fig
-    pi = PI(n)
+    return PI(n).div(2 * x.infinite?, prec) if x.infinite?
+
     x = -x if neg = x < 0
-    return pi.div(neg ? -2 : 2, prec) if x.infinite?
-    return pi.div(neg ? -4 : 4, prec) if x.round(prec) == 1
-    x = BigDecimal("1").div(x, n) if inv = x > 1
-    x = (-1 + sqrt(1 + x.mult(x, n), n)).div(x, n) if dbl = x > 0.5
-    y = x
-    d = y
-    t = x
-    r = BigDecimal("3")
-    x2 = x.mult(x,n)
-    while d.nonzero? && ((m = n - (y.exponent - d.exponent).abs) > 0)
-      m = BigDecimal.double_fig if m < BigDecimal.double_fig
-      t = -t.mult(x2,n)
-      d = t.div(r,m)
-      y += d
-      r += 2
+    x = BigDecimal(1).div(x, n) if inv = x < -1 || x > 1
+
+    # Solve tan(y) - x = 0 with Newton's method
+    # Repeat: y -= (tan(y) - x) * cos(y)**2
+    y = BigDecimal(Math.atan(x.to_f), 0)
+    BigDecimal::Internal.newton_loop(n) do |p|
+      s = sin(y, p)
+      c = (1 - s * s).sqrt(p)
+      y = y.sub(c * (s.sub(c * x.mult(1, p), p)), p)
     end
-    y *= 2 if dbl
-    y = pi / 2 - y if inv
-    y = -y if neg
-    y.mult(1, prec)
+    y = PI(n) / 2 - y if inv
+    y.mult(neg ? -1 : 1, prec)
   end
 
   # call-seq:

test/bigdecimal/test_bigmath.rb

@@ -138,27 +138,19 @@ def test_exp
 
   def test_log
     assert_equal(0, BigMath.log(BigDecimal("1.0"), 10))
-    assert_in_epsilon(Math.log(10)*1000, BigMath.log(BigDecimal("1e1000"), 10))
+    assert_in_epsilon(1000 * Math.log(10), BigMath.log(BigDecimal("1e1000"), 10))
+    assert_in_epsilon(19999999999999 * Math.log(10), BigMath.log(BigDecimal("1E19999999999999"), 10))
+    assert_in_epsilon(-19999999999999 * Math.log(10), BigMath.log(BigDecimal("1E-19999999999999"), 10))
     assert_in_exact_precision(
       BigDecimal("2.3025850929940456840179914546843642076011014886287729760333279009675726096773524802359972050895982983419677840422862"),
       BigMath.log(BigDecimal("10"), 100),
       100
     )
     assert_converge_in_precision {|n| BigMath.log(BigDecimal("2"), n) }
-    assert_converge_in_precision {|n| BigMath.log(BigDecimal("1e-30") + 1, n) }
-    assert_converge_in_precision {|n| BigMath.log(BigDecimal("1e-30"), n) }
+    assert_converge_in_precision {|n| BigMath.log(1 + SQRT2 * BigDecimal("1e-30"), n) }
+    assert_converge_in_precision {|n| BigMath.log(SQRT2 * BigDecimal("1e-30"), n) }
     assert_converge_in_precision {|n| BigMath.log(BigDecimal("1e30"), n) }
     assert_converge_in_precision {|n| BigMath.log(SQRT2, n) }
     assert_raise(Math::DomainError) {BigMath.log(BigDecimal("-0.1"), 10)}
-    assert_separately(%w[-rbigdecimal], <<-SRC)
-    begin
-      x = BigMath.log(BigDecimal("1E19999999999999"), 10)
-    rescue FloatDomainError
-    else
-      unless x.infinite?
-        assert_in_epsilon(Math.log(10)*19999999999999, x)
-      end
-    end
-    SRC
   end
 end