func (bed BinaryVarintEncoderDecoder) Decode(r io.Reader, n *big.Float) error {
var isInteger int8
var f float64
var exponent int32
n.SetUint64(0)
if err := binary.Read(r, binary.BigEndian, &isInteger); err != nil {
return err
}
if isInteger <= 0 {
var x int64
var err error
if x, err = binary.ReadVarint(miniByteReader{r}); err != nil {
return err
}
n.SetInt64(x)
n.SetPrec(ENCODER_DECODER_PREC)
return nil
} else {
if err := binary.Read(r, binary.BigEndian, &f); err != nil {
return err
}
if err := binary.Read(r, binary.BigEndian, &exponent); err != nil {
return err
}
bed.tmp.SetFloat64(f)
bed.tmp.SetPrec(ENCODER_DECODER_PREC)
n.SetMantExp(bed.tmp, int(exponent))
return nil
}
}
func mandelbrotFloat(a, b *big.Float) color.Color {
var x, y, nx, ny, x2, y2, f2, f4, r2, tmp big.Float
f2.SetInt64(2)
f4.SetInt64(4)
x.SetInt64(0)
y.SetInt64(0)
defer func() { recover() }()
for n := uint8(0); n < iterations; n++ {
// Not update x2 and y2
// because they are already updated in the previous loop
nx.Sub(&x2, &y2)
nx.Add(&nx, a)
tmp.Mul(&x, &y)
ny.Mul(&f2, &tmp)
ny.Add(&ny, b)
x.Set(&nx)
y.Set(&ny)
x2.Mul(&x, &x)
y2.Mul(&y, &y)
r2.Add(&x2, &y2)
if r2.Cmp(&f4) > 0 {
return color.Gray{255 - contrast*n}
}
}
return color.Black
}
func ExampleFloat_Add() {
// Operating on numbers of different precision.
var x, y, z big.Float
x.SetInt64(1000) // x is automatically set to 64bit precision
y.SetFloat64(2.718281828) // y is automatically set to 53bit precision
z.SetPrec(32)
z.Add(&x, &y)
fmt.Printf("x = %s (%s, prec = %d, acc = %s)\n", &x, x.Format('p', 0), x.Prec(), x.Acc())
fmt.Printf("y = %s (%s, prec = %d, acc = %s)\n", &y, y.Format('p', 0), y.Prec(), y.Acc())
fmt.Printf("z = %s (%s, prec = %d, acc = %s)\n", &z, z.Format('p', 0), z.Prec(), z.Acc())
// Output:
// x = 1000 (0x.fap10, prec = 64, acc = Exact)
// y = 2.718281828 (0x.adf85458248cd8p2, prec = 53, acc = Exact)
// z = 1002.718282 (0x.faadf854p10, prec = 32, acc = Below)
}
// Given a vector of real numbers x = [x_0, x_1, ..., x_n], this
// uses the PSLQ algorithm to find a list of integers
// [c_0, c_1, ..., c_n] such that
//
// |c_1 * x_1 + c_2 * x_2 + ... + c_n * x_n| < tolerance
//
// and such that max |c_k| < maxcoeff. If no such vector exists, Pslq
// returns one of the errors in this package depending on whether it
// has run out of iterations, precision or explored up to the
// maxcoeff. The tolerance defaults to 3/4 of the precision.
//
// This is a fairly direct translation of the pseudocode given by
// David Bailey, "The PSLQ Integer Relation Algorithm":
// http://www.cecm.sfu.ca/organics/papers/bailey/paper/html/node3.html
//
// If a result is returned, the first non-zero element will be positive
func (e *Pslq) Run(x []big.Float) ([]big.Int, error) {
n := len(x)
if n <= 1 {
return nil, ErrorBadArguments
}
// At too low precision, the algorithm becomes meaningless
if e.prec < 64 {
return nil, ErrorPrecisionTooLow
}
if e.verbose && int(e.prec)/max(2, int(n)) < 5 {
log.Printf("Warning: precision for PSLQ may be too low")
}
if e.verbose {
log.Printf("PSLQ using prec %d and tol %g", e.prec, e.tol)
}
if e.tol.Sign() == 0 {
return nil, ErrorToleranceRoundsToZero
}
// Temporary variables
tmp0 := new(big.Float).SetPrec(e.prec)
tmp1 := new(big.Float).SetPrec(e.prec)
bigTmp := new(big.Int)
// Convert to use 1-based indexing to allow us to be
// consistent with Bailey's indexing.
xNew := make([]big.Float, len(x)+1)
minx := new(big.Float).SetPrec(e.prec)
minxFirst := true
for i, xk := range x {
p := &xNew[i+1]
p.Set(&xk)
tmp0.Abs(p)
if minxFirst || tmp0.Cmp(minx) < 0 {
minxFirst = false
minx.Set(tmp0)
}
}
x = xNew
if debug {
printVector("x", x)
}
// Sanity check on magnitudes
if minx.Sign() == 0 {
return nil, ErrorZeroArguments
}
tmp1.SetInt64(128)
tmp0.Quo(&e.tol, tmp1)
if minx.Cmp(tmp0) < 0 { // minx < tol/128
return nil, ErrorArgumentTooSmall
}
tmp0.SetInt64(4)
tmp1.SetInt64(3)
tmp0.Quo(tmp0, tmp1)
var γ big.Float
e.Sqrt(tmp0, &γ) // sqrt(4<<prec)/3)
if debug {
fmt.Printf("γ = %f\n", &γ)
}
A := newBigIntMatrix(n+1, n+1)
B := newBigIntMatrix(n+1, n+1)
H := newMatrix(n+1, n+1)
// Initialization Step 1
//
// Set the n×n matrices A and B to the identity.
for i := 1; i <= n; i++ {
for j := 1; j <= n; j++ {
if i == j {
A[i][j].SetInt64(1)
B[i][j].SetInt64(1)
} else {
A[i][j].SetInt64(0)
B[i][j].SetInt64(0)
}
H[i][j].SetInt64(0)
}
}
if debug {
printBigIntMatrix("A", A)
printBigIntMatrix("B", B)
printMatrix("H", H)
}
// Initialization Step 2
//
// For k := 1 to n
// compute s_k := sqrt( sum_j=k^n x_j^2 )
// endfor.
// Set t = 1/s1.
// For k := 1 to n:
// y_k := t * x_k
// s_k := t * s_k
// endfor.
s := make([]big.Float, n+1)
for i := 1; i <= n; i++ {
s[i].SetInt64(0)
}
for k := 1; k <= n; k++ {
var t big.Float
t.SetInt64(0)
for j := k; j <= n; j++ {
tmp0.Mul(&x[j], &x[j])
t.Add(&t, tmp0)
}
e.Sqrt(&t, &s[k])
}
if debug {
fmt.Println("Init Step 2")
printVector("s", s)
}
var t big.Float
t.Set(&s[1])
y := make([]big.Float, len(x))
copy(y, x)
for k := 1; k <= n; k++ {
// y[k] = (x[k] << prec) / t
y[k].Quo(&x[k], &t)
// s[k] = (s[k] << prec) / t
s[k].Quo(&s[k], &t)
}
if debug {
printVector("y", y)
printVector("s", s)
}
// Init Step 3
//
// Compute the n×(n−1) matrix H as follows:
// For i := 1 to n:
// for j := i + 1 to n − 1:
// set Hij := 0
// endfor
// if i ≤ n − 1 then set Hii := s_(i+1)/s_i
// for j := 1 to i−1:
// set Hij := −y_i * y_j / (s_j * s_(j+1))
// endfor
// endfor
for i := 1; i <= n; i++ {
for j := i + 1; j < n; j++ {
H[i][j].SetInt64(0)
}
if i <= n-1 {
if s[i].Sign() == 0 {
// Precision probably exhausted
return nil, ErrorPrecisionExhausted
}
// H[i][i] = (s[i+1] << prec) / s[i]
H[i][i].Quo(&s[i+1], &s[i])
}
for j := 1; j < i; j++ {
var sjj1 big.Float
sjj1.Mul(&s[j], &s[j+1])
if debug {
fmt.Printf("sjj1 = %f\n", &sjj1)
}
if sjj1.Sign() == 0 {
// Precision probably exhausted
return nil, ErrorPrecisionExhausted
}
// H[i][j] = ((-y[i] * y[j]) << prec) / sjj1
tmp0.Mul(&y[i], &y[j])
tmp0.Neg(tmp0)
H[i][j].Quo(tmp0, &sjj1)
}
}
if debug {
fmt.Println("Init Step 3")
printMatrix("H", H)
}
// Init Step 4
//
// Perform full reduction on H, simultaneously updating y, A and B:
//
// For i := 2 to n:
// for j := i−1 to 1 step−1:
// t := nint(Hij/Hjj)
// y_j := y_j + t * y_i
// for k := 1 to j:
// Hik := Hik − t * Hjk
// endfor
// for k := 1 to n:
// Aik := Aik − t * Ajk
// Bkj := Bkj + t * Bki
// endfor
// endfor
// endfor
for i := 2; i <= n; i++ {
for j := i - 1; j > 0; j-- {
//t = floor(H[i][j]/H[j,j] + 0.5)
var t big.Int
var tFloat big.Float
if H[j][j].Sign() == 0 {
// Precision probably exhausted
return nil, ErrorPrecisionExhausted
}
tmp0.Quo(&H[i][j], &H[j][j])
e.NearestInt(tmp0, &t)
tFloat.SetInt(&t).SetPrec(e.prec)
if debug {
fmt.Printf("H[i][j]=%f\n", &H[i][j])
fmt.Printf("H[j][j]=%f\n", &H[j][j])
fmt.Printf("tmp=%f\n", tmp0)
fmt.Printf("t=%d\n", &t)
}
// y[j] = y[j] + (t * y[i] >> prec)
tmp0.Mul(&y[i], &tFloat)
y[j].Add(&y[j], tmp0)
for k := 1; k <= j; k++ {
// H[i][k] = H[i][k] - (t * H[j][k] >> prec)
tmp0.Mul(&H[j][k], &tFloat)
H[i][k].Sub(&H[i][k], tmp0)
}
for k := 1; k <= n; k++ {
bigTmp.Mul(&t, &A[j][k])
A[i][k].Sub(&A[i][k], bigTmp)
bigTmp.Mul(&t, &B[k][i])
B[k][j].Add(&B[k][j], bigTmp)
}
}
}
if debug {
fmt.Println("Init Step 4")
printBigIntMatrix("A", A)
printBigIntMatrix("B", B)
printMatrix("H", H)
}
// Main algorithm
var REP int
var norm big.Int
vec := make([]big.Int, n)
for REP = 0; REP < e.maxsteps; REP++ {
// Step 1
//
// Select m such that γ^i * |Hii| is maximal when i = m.
m := -1
var szmax big.Float
szmax.SetInt64(-1)
var γPower big.Float
γPower.Set(&γ)
for i := 1; i < n; i++ {
var absH big.Float
absH.Abs(&H[i][i])
var sz big.Float
sz.Mul(&γPower, &absH)
// sz := (g**i * abs(h)) >> (prec * (i - 1))
if sz.Cmp(&szmax) > 0 {
m = i
szmax.Set(&sz)
}
γPower.Mul(&γPower, &γ)
}
if debug {
fmt.Println("Step 1")
fmt.Printf("szmax=%f\n", &szmax)
fmt.Printf("m=%d\n", m)
}
// Step 2
//
// Exchange entries m and m+1 of y, corresponding rows
// of A and H, and corresponding columns of B.
y[m], y[m+1] = y[m+1], y[m]
for i := 1; i < n+1; i++ {
H[m][i], H[m+1][i] = H[m+1][i], H[m][i]
}
for i := 1; i < n+1; i++ {
A[m][i], A[m+1][i] = A[m+1][i], A[m][i]
}
for i := 1; i < n+1; i++ {
B[i][m], B[i][m+1] = B[i][m+1], B[i][m]
}
if debug {
fmt.Println("Step 2")
printVector("y", y)
printBigIntMatrix("A", A)
printBigIntMatrix("B", B)
printMatrix("H", H)
}
// Step 3
//
// If m ≤ n−2 then update H as follows:
//
// t0 := sqrt( Hmm^2 + H(m,m+1)^2 )
// t1 := Hmm/t0
// t2 := H(m,m+1)/t0.
// for i := m to n:
// t3 := Him
// t4 := Hi,m+1
// Him := t1t3 +t2t4
// Hi,m+1 := −t2t3 +t1t4
// endfor.
if m <= n-2 {
tmp0.Mul(&H[m][m], &H[m][m])
tmp1.Mul(&H[m][m+1], &H[m][m+1])
tmp0.Add(tmp0, tmp1)
var t0 big.Float
e.Sqrt(tmp0, &t0)
// Precision probably exhausted
if t0.Sign() == 0 {
return nil, ErrorPrecisionExhausted
}
var t1, t2 big.Float
t1.Quo(&H[m][m], &t0)
t2.Quo(&H[m][m+1], &t0)
for i := m; i <= n; i++ {
var t3, t4 big.Float
t3.Set(&H[i][m])
t4.Set(&H[i][m+1])
// H[i][m] = (t1*t3 + t2*t4) >> prec
tmp0.Mul(&t1, &t3)
tmp1.Mul(&t2, &t4)
H[i][m].Add(tmp0, tmp1)
// H[i][m+1] = (-t2*t3 + t1*t4) >> prec
tmp0.Mul(&t2, &t3)
tmp1.Mul(&t1, &t4)
H[i][m+1].Sub(tmp1, tmp0)
}
}
if debug {
fmt.Println("Step 3")
printMatrix("H", H)
}
// Step 4
// Perform block reduction on H, simultaneously updating y, A and B:
//
// For i := m+1 to n:
// for j := min(i−1, m+1) to 1 step −1:
// t := nint(Hij/Hjj)
// yj := yj + t * yi
// for k := 1 to j:
// Hik := Hik − tHjk
// endfor
// for k := 1 to n:
// Aik := Aik −tAjk
// Bkj := Bkj +tBki
// endfor
// endfor
// endfor.
for i := m + 1; i <= n; i++ {
var t big.Int
var tFloat big.Float
for j := min(i-1, m+1); j > 0; j-- {
if H[j][j].Sign() == 0 {
// Precision probably exhausted
return nil, ErrorPrecisionExhausted
}
tmp0.Quo(&H[i][j], &H[j][j])
e.NearestInt(tmp0, &t)
tFloat.SetInt(&t).SetPrec(e.prec)
// y[j] = y[j] + ((t * y[i]) >> prec)
tmp0.Mul(&y[i], &tFloat)
y[j].Add(&y[j], tmp0)
for k := 1; k <= j; k++ {
// H[i][k] = H[i][k] - (t * H[j][k] >> prec)
tmp0.Mul(&H[j][k], &tFloat)
H[i][k].Sub(&H[i][k], tmp0)
}
for k := 1; k <= n; k++ {
bigTmp.Mul(&t, &A[j][k])
A[i][k].Sub(&A[i][k], bigTmp)
bigTmp.Mul(&t, &B[k][i])
B[k][j].Add(&B[k][j], bigTmp)
}
}
}
if debug {
fmt.Println("Step 4")
printBigIntMatrix("A", A)
printBigIntMatrix("B", B)
printMatrix("H", H)
}
// Step 6
//
// Termination test: If the largest entry of A exceeds
// the level of numeric precision used, then precision
// is exhausted. If the smallest entry of the y vector
// is less than the detection threshold, a relation
// has been detected and is given in the corresponding
// column of B.
//
// Until a relation is found, the error typically decreases
// slowly (e.g. a factor 1-10) with each step TODO: we could
// compare err from two successive iterations. If there is a
// large drop (several orders of magnitude), that indicates a
// "high quality" relation was detected. Reporting this to
// the user somehow might be useful.
maxAPrecision := 0
for i := 1; i <= n; i++ {
for j := 1; j <= n; j++ {
precision := A[i][j].BitLen()
if precision > maxAPrecision {
maxAPrecision = precision
}
}
}
if debug {
log.Printf("Max A precision = %d, precision = %d, tolerance %d, ratio = %.3f\n", maxAPrecision, e.prec, e.target, float64(maxAPrecision)/float64(e.target))
}
if float64(maxAPrecision)/float64(e.target) > 0.85 {
if e.verbose {
log.Printf("CANCELLING after step %d/%d.", REP, e.maxsteps)
}
return nil, ErrorPrecisionExhausted
}
var best_err big.Float
best_err.Set(&e.maxcoeff_fp)
for i := 1; i <= n; i++ {
var err big.Float
err.Abs(&y[i])
// Maybe we are done?
if err.Cmp(&e.tol) < 0 {
// We are done if the coefficients are acceptable
var maxc big.Int
for j := 1; j <= n; j++ {
if debug {
fmt.Printf("vec[%d]=%d\n", j-1, &B[j][i])
}
t := B[j][i]
if debug {
fmt.Printf("vec[%d]=%d\n", j-1, t)
}
vec[j-1] = t
if t.Sign() < 0 {
t.Neg(&t)
}
if t.Cmp(&maxc) > 0 {
maxc.Set(&t)
}
}
if debug {
fmt.Printf("maxc = %d, maxcoeff = %d\n", maxc, e.maxcoeff)
}
if maxc.Cmp(&e.maxcoeff) < 0 {
if e.verbose {
log.Printf("FOUND relation at iter %d/%d, error: %g", REP, e.maxsteps, &err)
}
// Find sign of first non zero item
sign := 0
for i := range vec {
sign = vec[i].Sign()
if sign != 0 {
break
}
}
// Normalise vec making first non-zero argument positive
if sign < 0 {
for i := range vec {
vec[i].Neg(&vec[i])
}
}
return vec, nil
}
}
if err.Cmp(&best_err) < 0 {
best_err = err
}
}
// Step 5
//
// Norm bound: Compute M := 1/maxj |Hj|, where Hj
// denotes the j-th row of H.
//
// Then there can exist no relation vector whose
// Euclidean norm is less than M.
//
// Calculate a lower bound for the norm. We could do this
// more exactly (using the Euclidean norm) but there is probably
// no practical benefit.
var recnorm big.Float
recnorm.SetInt64(0)
for i := 1; i <= n; i++ {
for j := 1; j <= n; j++ {
tmp0.Abs(&H[i][j])
if tmp0.Cmp(&recnorm) > 0 {
recnorm.Set(tmp0)
}
}
}
norm.Set(&e.maxcoeff)
if recnorm.Sign() != 0 {
// norm = ((1 << (2 * prec)) / recnorm) >> prec
tmp0.Quo(&e.one, &recnorm)
tmp0.Int(&norm)
}
if e.verbose {
log.Printf("%2d/%2d: Error: %g Norm: %d", REP, e.maxsteps, &best_err, &norm)
}
if norm.Cmp(&e.maxcoeff) >= 0 {
if e.verbose {
log.Printf("CANCELLING after step %d/%d.", REP, e.maxsteps)
log.Printf("Could not find an integer relation. Norm bound: %d", &norm)
}
return nil, ErrorNoRelationFound
}
}
if e.verbose {
log.Printf("CANCELLING after step %d/%d.", REP, e.maxsteps)
log.Printf("Could not find an integer relation. Norm bound: %d", &norm)
}
return nil, ErrorIterationsExceeded
}
func renderFloat(img *image.RGBA) {
var yminF, ymaxMinF, heightF big.Float
yminF.SetInt64(ymin)
ymaxMinF.SetInt64(ymax - ymin)
heightF.SetInt64(height)
var xminF, xmaxMinF, widthF big.Float
xminF.SetInt64(xmin)
xmaxMinF.SetInt64(xmax - xmin)
widthF.SetInt64(width)
var y, x big.Float
for py := int64(0); py < height; py++ {
// y := float64(py)/height*(ymax-ymin) + ymin
y.SetInt64(py)
y.Quo(&y, &heightF)
y.Mul(&y, &ymaxMinF)
y.Add(&y, &yminF)
for px := int64(0); px < width; px++ {
// x := float64(px)/width*(xmax-xmin) + xmin
x.SetInt64(px)
x.Quo(&x, &widthF)
x.Mul(&x, &xmaxMinF)
x.Add(&x, &xminF)
c := mandelbrotFloat(&x, &y)
if c == nil {
c = color.Black
}
img.Set(int(px), int(py), c)
}
}
}