Merge pull request #27 from dingraha/brent_cs_safe

Make `ksmax` and friends complex-step safe

Project.toml

@@ -1,7 +1,7 @@
 name = "FLOWMath"
 uuid = "6cb5d3fb-0fe8-4cc2-bd89-9fe0b19a99d3"
 authors = ["Andrew Ning <[email protected]>"]
-version = "0.4.1"
+version = "0.4.2"
 
 [deps]
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"

src/smooth.jl

@@ -23,7 +23,7 @@ goes to infinity the maximum function is returned. Is mathematically guaranteed
 overestimate the maximum function, i.e. `maximum(x) <= ksmax(x, hardness)`.
 """
 function ksmax(x, hardness=50)
-    k = maximum(x)
+    k = maximum(real(x))
     return 1.0 / hardness * log(sum(exp.(hardness * (x .- k)))) .+ k
 end
 
@@ -83,7 +83,7 @@ Computes the derivative of the Kreisselmeier–Steinhauser constraint aggregatio
 function with respect to `hardness`.
 """
 function ksmax_h(x, hardness)
-    k = maximum(x)
+    k = maximum(real(x))
     tmp1 = exp.(hardness * (x .- k))
     tmp2 = sum((x .- k) .* tmp1)
     tmp3 = sum(tmp1)
@@ -98,7 +98,7 @@ Computes the second derivative of the Kreisselmeier–Steinhauser constraint agg
 function with respect to `hardness`.
 """
 function ksmax_hh(x, hardness)
-    k = maximum(x)
+    k = maximum(real(x))
     tmp1 = exp.(hardness * (x .- k))
     tmp2 = sum((x .- k) .* tmp1)
     tmp2_h = sum((x .- k) .^ 2 .* tmp1)

test/runtests.jl

@@ -260,6 +260,19 @@ hardness = 100.0
 x_max_smooth = ksmin(x, hardness)
 @test isapprox(x_max_smooth, -0.006931471805599453)
 
+# Test we can diff through ksmax and ksmin:
+ksmax_wrapper(x) = sin(ksmax(x.^2 .+ 2))
+x0 = [1.0, 1.5, 2.0, 2.5, 3.0]
+g1 = ForwardDiff.gradient(ksmax_wrapper, x0)
+g2 = FiniteDiff.finite_difference_gradient(ksmax_wrapper, x0, Val(:complex))
+@test maximum(abs.(g2 .- g1)) < 1e-12
+
+ksmin_wrapper(x) = sin(ksmin(x.^2 .+ 2))
+x0 = [1.0, 1.5, 2.0, 2.5, 3.0]
+g1 = ForwardDiff.gradient(ksmin_wrapper, x0)
+g2 = FiniteDiff.finite_difference_gradient(ksmin_wrapper, x0, Val(:complex))
+@test maximum(abs.(g2 .- g1)) < 1e-12
+
 # -------------------------
 
 # ------ ksmax_adaptive ---------
@@ -348,6 +361,19 @@ smoothing_fraction = 0.2
 x_max_smooth = ksmin_adaptive(x, smoothing_fraction=smoothing_fraction)
 @test isapprox(x_max_smooth, -5.2856933329025475e-6)
 
+# Test we can diff through ksmax_adaptive and ksmin_adaptive:
+ksmax_adaptive_wrapper(x) = sin(ksmax_adaptive(x.^2 .+ 2))
+x0 = [1.0, 1.5, 2.0, 2.5, 3.0]
+g1 = ForwardDiff.gradient(ksmax_adaptive_wrapper, x0)
+g2 = FiniteDiff.finite_difference_gradient(ksmax_adaptive_wrapper, x0, Val(:complex))
+@test maximum(abs.(g2 .- g1)) < 1e-12
+
+ksmin_adaptive_wrapper(x) = sin(ksmin_adaptive(x.^2 .+ 2))
+x0 = [1.0, 1.5, 2.0, 2.5, 3.0]
+g1 = ForwardDiff.gradient(ksmin_adaptive_wrapper, x0)
+g2 = FiniteDiff.finite_difference_gradient(ksmin_adaptive_wrapper, x0, Val(:complex))
+@test maximum(abs.(g2 .- g1)) < 1e-12
+
 # -------------------------
 
 # ------ sigmoid ---------