func ExampleCholesky() { // Construct a symmetric positive definite matrix. tmp := mat64.NewDense(4, 4, []float64{ 2, 6, 8, -4, 1, 8, 7, -2, 2, 2, 1, 7, 8, -2, -2, 1, }) var a mat64.SymDense a.SymOuterK(1, tmp) fmt.Printf("a = %0.4v\n", mat64.Formatted(&a, mat64.Prefix(" "))) // Compute the cholesky factorization. var chol mat64.Cholesky if ok := chol.Factorize(&a); !ok { fmt.Println("a matrix is not positive semi-definite.") } // Find the determinant. fmt.Printf("\nThe determinant of a is %0.4g\n\n", chol.Det()) // Use the factorization to solve the system of equations a * x = b. b := mat64.NewVector(4, []float64{1, 2, 3, 4}) var x mat64.Vector if err := x.SolveCholeskyVec(&chol, b); err != nil { fmt.Println("Matrix is near singular: ", err) } fmt.Println("Solve a * x = b") fmt.Printf("x = %0.4v\n", mat64.Formatted(&x, mat64.Prefix(" "))) // Extract the factorization and check that it equals the original matrix. var t mat64.TriDense t.LFromCholesky(&chol) var test mat64.Dense test.Mul(&t, t.T()) fmt.Println() fmt.Printf("L * L^T = %0.4v\n", mat64.Formatted(&a, mat64.Prefix(" "))) // Output: // a = ⎡120 114 -4 -16⎤ // ⎢114 118 11 -24⎥ // ⎢ -4 11 58 17⎥ // ⎣-16 -24 17 73⎦ // // The determinant of a is 1.543e+06 // // Solve a * x = b // x = ⎡ -0.239⎤ // ⎢ 0.2732⎥ // ⎢-0.04681⎥ // ⎣ 0.1031⎦ // // L * L^T = ⎡120 114 -4 -16⎤ // ⎢114 118 11 -24⎥ // ⎢ -4 11 58 17⎥ // ⎣-16 -24 17 73⎦ }
func ExampleCholeskySymRankOne() { a := mat64.NewSymDense(4, []float64{ 1, 1, 1, 1, 0, 2, 3, 4, 0, 0, 6, 10, 0, 0, 0, 20, }) fmt.Printf("A = %0.4v\n", mat64.Formatted(a, mat64.Prefix(" "))) // Compute the Cholesky factorization. var chol mat64.Cholesky if ok := chol.Factorize(a); !ok { fmt.Println("matrix a is not positive definite.") } x := mat64.NewVector(4, []float64{0, 0, 0, 1}) fmt.Printf("\nx = %0.4v\n", mat64.Formatted(x, mat64.Prefix(" "))) // Rank-1 update the factorization. chol.SymRankOne(&chol, 1, x) // Rank-1 update the matrix a. a.SymRankOne(a, 1, x) var au mat64.SymDense au.FromCholesky(&chol) // Print the matrix that was updated directly. fmt.Printf("\nA' = %0.4v\n", mat64.Formatted(a, mat64.Prefix(" "))) // Print the matrix recovered from the factorization. fmt.Printf("\nU'^T * U' = %0.4v\n", mat64.Formatted(&au, mat64.Prefix(" "))) // Output: // A = ⎡ 1 1 1 1⎤ // ⎢ 1 2 3 4⎥ // ⎢ 1 3 6 10⎥ // ⎣ 1 4 10 20⎦ // // x = ⎡0⎤ // ⎢0⎥ // ⎢0⎥ // ⎣1⎦ // // A' = ⎡ 1 1 1 1⎤ // ⎢ 1 2 3 4⎥ // ⎢ 1 3 6 10⎥ // ⎣ 1 4 10 21⎦ // // U'^T * U' = ⎡ 1 1 1 1⎤ // ⎢ 1 2 3 4⎥ // ⎢ 1 3 6 10⎥ // ⎣ 1 4 10 21⎦ }
// NewNormalPrecision creates a new Normal distribution with the given mean and // precision matrix (inverse of the covariance matrix). NewNormalPrecision // panics if len(mu) is not equal to prec.Symmetric(). If the precision matrix // is not positive-definite, NewNormalPrecision returns nil for norm and false // for ok. func NewNormalPrecision(mu []float64, prec *mat64.SymDense, src *rand.Rand) (norm *Normal, ok bool) { if len(mu) == 0 { panic(badZeroDimension) } dim := prec.Symmetric() if dim != len(mu) { panic(badSizeMismatch) } // TODO(btracey): Computing a matrix inverse is generally numerically instable. // This only has to compute the inverse of a positive definite matrix, which // is much better, but this still loses precision. It is worth considering if // instead the precision matrix should be stored explicitly and used instead // of the Cholesky decomposition of the covariance matrix where appropriate. var chol mat64.Cholesky ok = chol.Factorize(prec) if !ok { return nil, false } var sigma mat64.SymDense sigma.InverseCholesky(&chol) return NewNormal(mu, &sigma, src) }
// ConditionNormal returns the Normal distribution that is the receiver conditioned // on the input evidence. The returned multivariate normal has dimension // n - len(observed), where n is the dimension of the original receiver. The updated // mean and covariance are // mu = mu_un + sigma_{ob,un}^T * sigma_{ob,ob}^-1 (v - mu_ob) // sigma = sigma_{un,un} - sigma_{ob,un}^T * sigma_{ob,ob}^-1 * sigma_{ob,un} // where mu_un and mu_ob are the original means of the unobserved and observed // variables respectively, sigma_{un,un} is the unobserved subset of the covariance // matrix, sigma_{ob,ob} is the observed subset of the covariance matrix, and // sigma_{un,ob} are the cross terms. The elements of x_2 have been observed with // values v. The dimension order is preserved during conditioning, so if the value // of dimension 1 is observed, the returned normal represents dimensions {0, 2, ...} // of the original Normal distribution. // // ConditionNormal returns {nil, false} if there is a failure during the update. // Mathematically this is impossible, but can occur with finite precision arithmetic. func (n *Normal) ConditionNormal(observed []int, values []float64, src *rand.Rand) (*Normal, bool) { if len(observed) == 0 { panic("normal: no observed value") } if len(observed) != len(values) { panic("normal: input slice length mismatch") } for _, v := range observed { if v < 0 || v >= n.Dim() { panic("normal: observed value out of bounds") } } ob := len(observed) unob := n.Dim() - ob obMap := make(map[int]struct{}) for _, v := range observed { if _, ok := obMap[v]; ok { panic("normal: observed dimension occurs twice") } obMap[v] = struct{}{} } if len(observed) == n.Dim() { panic("normal: all dimensions observed") } unobserved := make([]int, 0, unob) for i := 0; i < n.Dim(); i++ { if _, ok := obMap[i]; !ok { unobserved = append(unobserved, i) } } mu1 := make([]float64, unob) for i, v := range unobserved { mu1[i] = n.mu[v] } mu2 := make([]float64, ob) // really v - mu2 for i, v := range observed { mu2[i] = values[i] - n.mu[v] } n.setSigma() var sigma11, sigma22 mat64.SymDense sigma11.SubsetSym(n.sigma, unobserved) sigma22.SubsetSym(n.sigma, observed) sigma21 := mat64.NewDense(ob, unob, nil) for i, r := range observed { for j, c := range unobserved { v := n.sigma.At(r, c) sigma21.Set(i, j, v) } } var chol mat64.Cholesky ok := chol.Factorize(&sigma22) if !ok { return nil, ok } // Compute sigma_{2,1}^T * sigma_{2,2}^-1 (v - mu_2). v := mat64.NewVector(ob, mu2) var tmp, tmp2 mat64.Vector err := tmp.SolveCholeskyVec(&chol, v) if err != nil { return nil, false } tmp2.MulVec(sigma21.T(), &tmp) // Compute sigma_{2,1}^T * sigma_{2,2}^-1 * sigma_{2,1}. // TODO(btracey): Should this be a method of SymDense? var tmp3, tmp4 mat64.Dense err = tmp3.SolveCholesky(&chol, sigma21) if err != nil { return nil, false } tmp4.Mul(sigma21.T(), &tmp3) for i := range mu1 { mu1[i] += tmp2.At(i, 0) } // TODO(btracey): If tmp2 can constructed with a method, then this can be // replaced with SubSym. for i := 0; i < len(unobserved); i++ { for j := i; j < len(unobserved); j++ { v := sigma11.At(i, j) sigma11.SetSym(i, j, v-tmp4.At(i, j)) } } return NewNormal(mu1, &sigma11, src) }
// AddBatch adds a set training points to the Gp. This call updates internal // values needed for prediction, so it is more efficient to add samples // as a batch. func (g *GP) AddBatch(x mat64.Matrix, y []float64) error { // Note: The outputs are stored scaled to have a mean of zero and a variance // of 1. // Verify input parameters rx, cx := x.Dims() ry := len(y) if rx != ry { panic(badInOut) } if cx != g.inputDim { panic(badInputLength) } nSamples := len(g.outputs) // Append the new data to the list of stored data. inputs := mat64.NewDense(rx+nSamples, g.inputDim, nil) inputs.Copy(g.inputs) inputs.View(nSamples, 0, rx, g.inputDim).(*mat64.Dense).Copy(x) g.inputs = inputs // Rescale the output data to its original value, append the new data, and // then rescale to have mean 0 and variance of 1. for i, v := range g.outputs { g.outputs[i] = v*g.std + g.mean } g.outputs = append(g.outputs, y...) g.mean = stat.Mean(g.outputs, nil) g.std = stat.StdDev(g.outputs, nil) for i, v := range g.outputs { g.outputs[i] = (v - g.mean) / g.std } // Add to the kernel matrix. k := mat64.NewSymDense(rx+nSamples, nil) k.CopySym(g.k) g.k = k // Compute the kernel with the new points and the old points for i := 0; i < nSamples; i++ { for j := nSamples; j < rx+nSamples; j++ { v := g.kernel.Distance(g.inputs.RawRowView(i), g.inputs.RawRowView(j)) g.k.SetSym(i, j, v) } } // Compute the kernel with the new points and themselves for i := nSamples; i < rx+nSamples; i++ { for j := i; j < nSamples+rx; j++ { v := g.kernel.Distance(g.inputs.RawRowView(i), g.inputs.RawRowView(j)) if i == j { v += g.noise } g.k.SetSym(i, j, v) } } // Cache necessary matrix results for computing predictions. var chol mat64.Cholesky ok := chol.Factorize(g.k) if !ok { return ErrSingular } g.cholK = &chol g.sigInvY.Reset() v := mat64.NewVector(len(g.outputs), g.outputs) g.sigInvY.SolveCholeskyVec(g.cholK, v) return nil }