// Copyright ©2013 The Gonum Authors. All rights reserved. // Use of this code is governed by a BSD-style // license that can be found in the LICENSE file. package floats import ( "errors" "math" "sort" "gonum.org/v1/gonum/floats/scalar" "gonum.org/v1/gonum/internal/asm/f64" ) const ( zeroLength = "floats: zero length slice" shortSpan = "floats: slice length less than 2" badLength = "floats: slice lengths do not match" badDstLength = "floats: destination slice length does not match input" ) // Add adds, element-wise, the elements of s and dst, and stores the result in dst. // It panics if the argument lengths do not match. func Add(dst, s []float64) { if len(dst) != len(s) { panic(badDstLength) } f64.AxpyUnitaryTo(dst, 1, s, dst) } // AddTo adds, element-wise, the elements of s and t and // stores the result in dst. // It panics if the argument lengths do not match. func AddTo(dst, s, t []float64) []float64 { if len(s) != len(t) { panic(badLength) } if len(dst) != len(s) { panic(badDstLength) } f64.AxpyUnitaryTo(dst, 1, s, t) return dst } // AddConst adds the scalar c to all of the values in dst. func AddConst(c float64, dst []float64) { f64.AddConst(c, dst) } // AddScaled performs dst = dst + alpha * s. // It panics if the slice argument lengths do not match. func AddScaled(dst []float64, alpha float64, s []float64) { if len(dst) != len(s) { panic(badLength) } f64.AxpyUnitaryTo(dst, alpha, s, dst) } // AddScaledTo performs dst = y + alpha * s, where alpha is a scalar, // and dst, y and s are all slices. // It panics if the slice argument lengths do not match. // // At the return of the function, dst[i] = y[i] + alpha * s[i] func AddScaledTo(dst, y []float64, alpha float64, s []float64) []float64 { if len(s) != len(y) { panic(badLength) } if len(dst) != len(y) { panic(badDstLength) } f64.AxpyUnitaryTo(dst, alpha, s, y) return dst } // argsort is a helper that implements sort.Interface, as used by // Argsort. type argsort struct { s []float64 inds []int } func (a argsort) Len() int { return len(a.s) } func (a argsort) Less(i, j int) bool { return a.s[i] < a.s[j] } func (a argsort) Swap(i, j int) { a.s[i], a.s[j] = a.s[j], a.s[i] a.inds[i], a.inds[j] = a.inds[j], a.inds[i] } // Argsort sorts the elements of dst while tracking their original order. // At the conclusion of Argsort, dst will contain the original elements of dst // but sorted in increasing order, and inds will contain the original position // of the elements in the slice such that dst[i] = origDst[inds[i]]. // It panics if the argument lengths do not match. func Argsort(dst []float64, inds []int) { if len(dst) != len(inds) { panic(badDstLength) } for i := range dst { inds[i] = i } a := argsort{s: dst, inds: inds} sort.Sort(a) } // Count applies the function f to every element of s and returns the number // of times the function returned true. func Count(f func(float64) bool, s []float64) int { var n int for _, val := range s { if f(val) { n++ } } return n } // CumProd finds the cumulative product of the first i elements in // s and puts them in place into the ith element of the // destination dst. // It panics if the argument lengths do not match. // // At the return of the function, dst[i] = s[i] * s[i-1] * s[i-2] * ... func CumProd(dst, s []float64) []float64 { if len(dst) != len(s) { panic(badDstLength) } if len(dst) == 0 { return dst } return f64.CumProd(dst, s) } // CumSum finds the cumulative sum of the first i elements in // s and puts them in place into the ith element of the // destination dst. // It panics if the argument lengths do not match. // // At the return of the function, dst[i] = s[i] + s[i-1] + s[i-2] + ... func CumSum(dst, s []float64) []float64 { if len(dst) != len(s) { panic(badDstLength) } if len(dst) == 0 { return dst } return f64.CumSum(dst, s) } // Distance computes the L-norm of s - t. See Norm for special cases. // It panics if the slice argument lengths do not match. func Distance(s, t []float64, L float64) float64 { if len(s) != len(t) { panic(badLength) } if len(s) == 0 { return 0 } if L == 2 { return f64.L2DistanceUnitary(s, t) } var norm float64 if L == 1 { for i, v := range s { norm += math.Abs(t[i] - v) } return norm } if math.IsInf(L, 1) { for i, v := range s { absDiff := math.Abs(t[i] - v) if absDiff > norm { norm = absDiff } } return norm } for i, v := range s { norm += math.Pow(math.Abs(t[i]-v), L) } return math.Pow(norm, 1/L) } // Div performs element-wise division dst / s // and stores the value in dst. // It panics if the argument lengths do not match. func Div(dst, s []float64) { if len(dst) != len(s) { panic(badLength) } f64.Div(dst, s) } // DivTo performs element-wise division s / t // and stores the value in dst. // It panics if the argument lengths do not match. func DivTo(dst, s, t []float64) []float64 { if len(s) != len(t) { panic(badLength) } if len(dst) != len(s) { panic(badDstLength) } return f64.DivTo(dst, s, t) } // Dot computes the dot product of s1 and s2, i.e. // sum_{i = 1}^N s1[i]*s2[i]. // It panics if the argument lengths do not match. func Dot(s1, s2 []float64) float64 { if len(s1) != len(s2) { panic(badLength) } return f64.DotUnitary(s1, s2) } // Equal returns true when the slices have equal lengths and // all elements are numerically identical. func Equal(s1, s2 []float64) bool { if len(s1) != len(s2) { return false } for i, val := range s1 { if s2[i] != val { return false } } return true } // EqualApprox returns true when the slices have equal lengths and // all element pairs have an absolute tolerance less than tol or a // relative tolerance less than tol. func EqualApprox(s1, s2 []float64, tol float64) bool { if len(s1) != len(s2) { return false } for i, a := range s1 { if !scalar.EqualWithinAbsOrRel(a, s2[i], tol, tol) { return false } } return true } // EqualFunc returns true when the slices have the same lengths // and the function returns true for all element pairs. func EqualFunc(s1, s2 []float64, f func(float64, float64) bool) bool { if len(s1) != len(s2) { return false } for i, val := range s1 { if !f(val, s2[i]) { return false } } return true } // EqualLengths returns true when all of the slices have equal length, // and false otherwise. It also returns true when there are no input slices. func EqualLengths(slices ...[]float64) bool { // This length check is needed: http://play.golang.org/p/sdty6YiLhM if len(slices) == 0 { return true } l := len(slices[0]) for i := 1; i < len(slices); i++ { if len(slices[i]) != l { return false } } return true } // Find applies f to every element of s and returns the indices of the first // k elements for which the f returns true, or all such elements // if k < 0. // Find will reslice inds to have 0 length, and will append // found indices to inds. // If k > 0 and there are fewer than k elements in s satisfying f, // all of the found elements will be returned along with an error. // At the return of the function, the input inds will be in an undetermined state. func Find(inds []int, f func(float64) bool, s []float64, k int) ([]int, error) { // inds is also returned to allow for calling with nil. // Reslice inds to have zero length. inds = inds[:0] // If zero elements requested, can just return. if k == 0 { return inds, nil } // If k < 0, return all of the found indices. if k < 0 { for i, val := range s { if f(val) { inds = append(inds, i) } } return inds, nil } // Otherwise, find the first k elements. nFound := 0 for i, val := range s { if f(val) { inds = append(inds, i) nFound++ if nFound == k { return inds, nil } } } // Finished iterating over the loop, which means k elements were not found. return inds, errors.New("floats: insufficient elements found") } // HasNaN returns true when the slice s has any values that are NaN and false // otherwise. func HasNaN(s []float64) bool { for _, v := range s { if math.IsNaN(v) { return true } } return false } // LogSpan returns a set of n equally spaced points in log space between, // l and u where N is equal to len(dst). The first element of the // resulting dst will be l and the final element of dst will be u. // It panics if the length of dst is less than 2. // Note that this call will return NaNs if either l or u are negative, and // will return all zeros if l or u is zero. // Also returns the mutated slice dst, so that it can be used in range, like: // // for i, x := range LogSpan(dst, l, u) { ... } func LogSpan(dst []float64, l, u float64) []float64 { Span(dst, math.Log(l), math.Log(u)) for i := range dst { dst[i] = math.Exp(dst[i]) } return dst } // LogSumExp returns the log of the sum of the exponentials of the values in s. // Panics if s is an empty slice. func LogSumExp(s []float64) float64 { // Want to do this in a numerically stable way which avoids // overflow and underflow // First, find the maximum value in the slice. maxval := Max(s) if math.IsInf(maxval, 0) { // If it's infinity either way, the logsumexp will be infinity as well // returning now avoids NaNs return maxval } var lse float64 // Compute the sumexp part for _, val := range s { lse += math.Exp(val - maxval) } // Take the log and add back on the constant taken out return math.Log(lse) + maxval } // Max returns the maximum value in the input slice. If the slice is empty, Max will panic. func Max(s []float64) float64 { return s[MaxIdx(s)] } // MaxIdx returns the index of the maximum value in the input slice. If several // entries have the maximum value, the first such index is returned. // It panics if s is zero length. func MaxIdx(s []float64) int { if len(s) == 0 { panic(zeroLength) } max := math.NaN() var ind int for i, v := range s { if math.IsNaN(v) { continue } if v > max || math.IsNaN(max) { max = v ind = i } } return ind } // Min returns the minimum value in the input slice. // It panics if s is zero length. func Min(s []float64) float64 { return s[MinIdx(s)] } // MinIdx returns the index of the minimum value in the input slice. If several // entries have the minimum value, the first such index is returned. // It panics if s is zero length. func MinIdx(s []float64) int { if len(s) == 0 { panic(zeroLength) } min := math.NaN() var ind int for i, v := range s { if math.IsNaN(v) { continue } if v < min || math.IsNaN(min) { min = v ind = i } } return ind } // Mul performs element-wise multiplication between dst // and s and stores the value in dst. // It panics if the argument lengths do not match. func Mul(dst, s []float64) { if len(dst) != len(s) { panic(badLength) } for i, val := range s { dst[i] *= val } } // MulTo performs element-wise multiplication between s // and t and stores the value in dst. // It panics if the argument lengths do not match. func MulTo(dst, s, t []float64) []float64 { if len(s) != len(t) { panic(badLength) } if len(dst) != len(s) { panic(badDstLength) } for i, val := range t { dst[i] = val * s[i] } return dst } // NearestIdx returns the index of the element in s // whose value is nearest to v. If several such // elements exist, the lowest index is returned. // It panics if s is zero length. func NearestIdx(s []float64, v float64) int { if len(s) == 0 { panic(zeroLength) } switch { case math.IsNaN(v): return 0 case math.IsInf(v, 1): return MaxIdx(s) case math.IsInf(v, -1): return MinIdx(s) } var ind int dist := math.NaN() for i, val := range s { newDist := math.Abs(v - val) // A NaN distance will not be closer. if math.IsNaN(newDist) { continue } if newDist < dist || math.IsNaN(dist) { dist = newDist ind = i } } return ind } // NearestIdxForSpan return the index of a hypothetical vector created // by Span with length n and bounds l and u whose value is closest // to v. That is, NearestIdxForSpan(n, l, u, v) is equivalent to // Nearest(Span(make([]float64, n),l,u),v) without an allocation. // It panics if n is less than two. func NearestIdxForSpan(n int, l, u float64, v float64) int { if n < 2 { panic(shortSpan) } if math.IsNaN(v) { return 0 } // Special cases for Inf and NaN. switch { case math.IsNaN(l) && !math.IsNaN(u): return n - 1 case math.IsNaN(u): return 0 case math.IsInf(l, 0) && math.IsInf(u, 0): if l == u { return 0 } if n%2 == 1 { if !math.IsInf(v, 0) { return n / 2 } if math.Copysign(1, v) == math.Copysign(1, l) { return 0 } return n/2 + 1 } if math.Copysign(1, v) == math.Copysign(1, l) { return 0 } return n / 2 case math.IsInf(l, 0): if v == l { return 0 } return n - 1 case math.IsInf(u, 0): if v == u { return n - 1 } return 0 case math.IsInf(v, -1): if l <= u { return 0 } return n - 1 case math.IsInf(v, 1): if u <= l { return 0 } return n - 1 } // Special cases for v outside (l, u) and (u, l). switch { case l < u: if v <= l { return 0 } if v >= u { return n - 1 } case l > u: if v >= l { return 0 } if v <= u { return n - 1 } default: return 0 } // Can't guarantee anything about exactly halfway between // because of floating point weirdness. return int((float64(n)-1)/(u-l)*(v-l) + 0.5) } // Norm returns the L norm of the slice S, defined as // (sum_{i=1}^N s[i]^L)^{1/L} // Special cases: // L = math.Inf(1) gives the maximum absolute value. // Does not correctly compute the zero norm (use Count). func Norm(s []float64, L float64) float64 { // Should this complain if L is not positive? // Should this be done in log space for better numerical stability? // would be more cost // maybe only if L is high? if len(s) == 0 { return 0 } if L == 2 { return f64.L2NormUnitary(s) } var norm float64 if L == 1 { for _, val := range s { norm += math.Abs(val) } return norm } if math.IsInf(L, 1) { for _, val := range s { norm = math.Max(norm, math.Abs(val)) } return norm } for _, val := range s { norm += math.Pow(math.Abs(val), L) } return math.Pow(norm, 1/L) } // Prod returns the product of the elements of the slice. // Returns 1 if len(s) = 0. func Prod(s []float64) float64 { prod := 1.0 for _, val := range s { prod *= val } return prod } // Reverse reverses the order of elements in the slice. func Reverse(s []float64) { for i, j := 0, len(s)-1; i < j; i, j = i+1, j-1 { s[i], s[j] = s[j], s[i] } } // Same returns true when the input slices have the same length and all // elements have the same value with NaN treated as the same. func Same(s, t []float64) bool { if len(s) != len(t) { return false } for i, v := range s { w := t[i] if v != w && !(math.IsNaN(v) && math.IsNaN(w)) { return false } } return true } // Scale multiplies every element in dst by the scalar c. func Scale(c float64, dst []float64) { if len(dst) > 0 { f64.ScalUnitary(c, dst) } } // ScaleTo multiplies the elements in s by c and stores the result in dst. // It panics if the slice argument lengths do not match. func ScaleTo(dst []float64, c float64, s []float64) []float64 { if len(dst) != len(s) { panic(badDstLength) } if len(dst) > 0 { f64.ScalUnitaryTo(dst, c, s) } return dst } // Span returns a set of N equally spaced points between l and u, where N // is equal to the length of the destination. The first element of the destination // is l, the final element of the destination is u. // It panics if the length of dst is less than 2. // // Span also returns the mutated slice dst, so that it can be used in range expressions, // like: // // for i, x := range Span(dst, l, u) { ... } func Span(dst []float64, l, u float64) []float64 { n := len(dst) if n < 2 { panic(shortSpan) } // Special cases for Inf and NaN. switch { case math.IsNaN(l): for i := range dst[:len(dst)-1] { dst[i] = math.NaN() } dst[len(dst)-1] = u return dst case math.IsNaN(u): for i := range dst[1:] { dst[i+1] = math.NaN() } dst[0] = l return dst case math.IsInf(l, 0) && math.IsInf(u, 0): for i := range dst[:len(dst)/2] { dst[i] = l dst[len(dst)-i-1] = u } if len(dst)%2 == 1 { if l != u { dst[len(dst)/2] = 0 } else { dst[len(dst)/2] = l } } return dst case math.IsInf(l, 0): for i := range dst[:len(dst)-1] { dst[i] = l } dst[len(dst)-1] = u return dst case math.IsInf(u, 0): for i := range dst[1:] { dst[i+1] = u } dst[0] = l return dst } step := (u - l) / float64(n-1) for i := range dst { dst[i] = l + step*float64(i) } return dst } // Sub subtracts, element-wise, the elements of s from dst. // It panics if the argument lengths do not match. func Sub(dst, s []float64) { if len(dst) != len(s) { panic(badLength) } f64.AxpyUnitaryTo(dst, -1, s, dst) } // SubTo subtracts, element-wise, the elements of t from s and // stores the result in dst. // It panics if the argument lengths do not match. func SubTo(dst, s, t []float64) []float64 { if len(s) != len(t) { panic(badLength) } if len(dst) != len(s) { panic(badDstLength) } f64.AxpyUnitaryTo(dst, -1, t, s) return dst } // Sum returns the sum of the elements of the slice. func Sum(s []float64) float64 { return f64.Sum(s) } // Within returns the first index i where s[i] <= v < s[i+1]. Within panics if: // - len(s) < 2 // - s is not sorted func Within(s []float64, v float64) int { if len(s) < 2 { panic(shortSpan) } if !sort.Float64sAreSorted(s) { panic("floats: input slice not sorted") } if v < s[0] || v >= s[len(s)-1] || math.IsNaN(v) { return -1 } for i, f := range s[1:] { if v < f { return i } } return -1 } // SumCompensated returns the sum of the elements of the slice calculated with greater // accuracy than Sum at the expense of additional computation. func SumCompensated(s []float64) float64 { // SumCompensated uses an improved version of Kahan's compensated // summation algorithm proposed by Neumaier. // See https://en.wikipedia.org/wiki/Kahan_summation_algorithm for details. var sum, c float64 for _, x := range s { // This type conversion is here to prevent a sufficiently smart compiler // from optimising away these operations. t := float64(sum + x) if math.Abs(sum) >= math.Abs(x) { c += (sum - t) + x } else { c += (x - t) + sum } sum = t } return sum + c }