Main Content

bicgstab

Solve system of linear equations — stabilized biconjugate gradients method

Description

example

x = bicgstab(A,b) attempts to solve the system of linear equations A*x = b for x using the Biconjugate Gradients Stabilized Method. When the attempt is successful, bicgstab displays a message to confirm convergence. If bicgstab fails to converge after the maximum number of iterations or halts for any reason, it displays a diagnostic message that includes the relative residual norm(b-A*x)/norm(b) and the iteration number at which the method stopped.

example

x = bicgstab(A,b,tol) specifies a tolerance for the method. The default tolerance is 1e-6.

example

x = bicgstab(A,b,tol,maxit) specifies the maximum number of iterations to use. bicgstab displays a diagnostic message if it fails to converge within maxit iterations.

example

x = bicgstab(A,b,tol,maxit,M) specifies a preconditioner matrix M and computes x by effectively solving the system AM1y=b for y, where y=Mx. Using a preconditioner matrix can improve the numerical properties of the problem and the efficiency of the calculation.

example

x = bicgstab(A,b,tol,maxit,M1,M2) specifies factors of the preconditioner matrix M such that M = M1*M2.

example

x = bicgstab(A,b,tol,maxit,M1,M2,x0) specifies an initial guess for the solution vector x. The default is a vector of zeros.

example

[x,flag] = bicgstab(___) returns a flag that specifies whether the algorithm successfully converged. When flag = 0, convergence was successful. You can use this output syntax with any of the previous input argument combinations. When you specify the flag output, bicgstab does not display any diagnostic messages.

example

[x,flag,relres] = bicgstab(___) also returns the relative residual norm(b-A*x)/norm(b). If flag is 0, then relres <= tol.

example

[x,flag,relres,iter] = bicgstab(___) also returns the iteration number iter at which x was computed.

example

[x,flag,relres,iter,resvec] = bicgstab(___) also returns a vector of the residual norms at each iteration, including the first residual norm(b-A*x0).

Examples

collapse all

Solve a square linear system using bicgstab with default settings, and then adjust the tolerance and number of iterations used in the solution process.

Create a random sparse matrix A with 50% density. Also create a random vector b for the right-hand side of Ax=b.

rng default
A = sprand(400,400,.5);
A = A'*A;
b = rand(400,1);

Solve Ax=b using bicgstab. The output display includes the value of the relative residual error b-Axb.

x = bicgstab(A,b);
bicgstab stopped at iteration 20 without converging to the desired tolerance 1e-06
because the maximum number of iterations was reached.
The iterate returned (number 20) has relative residual 0.12.

By default bicgstab uses 20 iterations and a tolerance of 1e-6, and the algorithm is unable to converge in those 20 iterations for this matrix. Since the residual is still large, it is a good indicator that more iterations (or a preconditioner matrix) are needed. You also can use a larger tolerance to make it easier for the algorithm to converge.

Solve the system again using a tolerance of 1e-4 and 100 iterations.

x = bicgstab(A,b,1e-4,100);
bicgstab stopped at iteration 100 without converging to the desired tolerance 0.0001
because the maximum number of iterations was reached.
The iterate returned (number 100) has relative residual 0.044.

Even with a looser tolerance and more iterations, the residual error does not improve much. When an iterative algorithm stalls in this manner, it is a good indication that a preconditioner matrix is needed.

Calculate the incomplete Cholesky factorization of A, and use the L' factor as a preconditioner input to bicgstab.

L = ichol(A);
x = bicgstab(A,b,1e-4,100,L');
bicgstab converged at iteration 30.5 to a solution with relative residual 5.3e-05.

Using a preconditioner improves the numerical properties of the problem enough that bicgstab is able to converge.

Examine the effect of using a preconditioner matrix with bicgstab to solve a linear system.

Load west0479, a real 479-by-479 nonsymmetric sparse matrix.

load west0479
A = west0479;

Define b so that the true solution to Ax=b is a vector of all ones.

b = sum(A,2);

Set the tolerance and maximum number of iterations.

tol = 1e-12;
maxit = 20;

Use bicgstab to find a solution at the requested tolerance and number of iterations. Specify five outputs to return information about the solution process:

  • x is the computed solution to A*x = b.

  • fl0 is a flag indicating whether the algorithm converged.

  • rr0 is the relative residual of the computed answer x.

  • it0 is the iteration number when x was computed.

  • rv0 is a vector of the residual history for b-Ax.

[x,fl0,rr0,it0,rv0] = bicgstab(A,b,tol,maxit); 
fl0
fl0 = 1
rr0
rr0 = 1
it0
it0 = 0

fl0 is 1 because bicgstab does not converge to the requested tolerance 1e-12 within the requested 20 iterations. In fact, the behavior of bicgstab is so poor that the initial guess x0 = zeros(size(A,2),1) is the best solution and is returned, as indicated by it0 = 0.

To aid with the slow convergence, you can specify a preconditioner matrix. Since A is nonsymmetric, use ilu to generate the preconditioner M=LU. Specify a drop tolerance to ignore nondiagonal entries with values smaller than 1e-6. Solve the preconditioned system AM-1Mx=b by specifying L and U as inputs to bicgstab.

setup = struct('type','ilutp','droptol',1e-6);
[L,U] = ilu(A,setup);
[x1,fl1,rr1,it1,rv1] = bicgstab(A,b,tol,maxit,L,U);
fl1
fl1 = 0
rr1
rr1 = 3.8661e-14
it1
it1 = 3

The use of an ilu preconditioner produces a relative residual less than the prescribed tolerance of 1e-12 at the third iteration. The output rv1(1) is norm(b), and the output rv1(end) is norm(b-A*x1).

You can follow the progress of bicgstab by plotting the relative residuals at each iteration. Plot the residual history of each solution with a line for the specified tolerance.

semilogy(0:length(rv0)-1,rv0/norm(b),'-o')
hold on
semilogy(0:length(rv1)-1,rv1/norm(b),'-o')
yline(tol,'r--');
legend('No preconditioner','ILU preconditioner','Tolerance','Location','East')
xlabel('Iteration number')
ylabel('Relative residual')

Examine the effect of supplying bicgstab with an initial guess of the solution.

Create a tridiagonal sparse matrix. Use the sum of each row as the vector for the right-hand side of Ax=b so that the expected solution for x is a vector of ones.

n = 900;
e = ones(n,1);
A = spdiags([e 2*e e],-1:1,n,n);
b = sum(A,2);

Use bicgstab to solve Ax=b twice: one time with the default initial guess, and one time with a good initial guess of the solution. Use 50 iterations and the default tolerance for both solutions. Specify the initial guess in the second solution as a vector with all elements equal to 0.99.

maxit = 50;
x1 = bicgstab(A,b,[],maxit);
bicgstab converged at iteration 20.5 to a solution with relative residual 9.3e-07.
x0 = 0.99*e;
x2 = bicgstab(A,b,[],maxit,[],[],x0);
bicgstab converged at iteration 4 to a solution with relative residual 8.7e-07.

In this case supplying an initial guess enables bicgstab to converge more quickly.

Returning Intermediate Results

You also can use the initial guess to get intermediate results by calling bicgstab in a for-loop. Each call to the solver performs a few iterations and stores the calculated solution. Then you use that solution as the initial vector for the next batch of iterations.

For example, this code performs 100 iterations four times and stores the solution vector after each pass in the for-loop:

x0 = zeros(size(A,2),1);
tol = 1e-8;
maxit = 100;
for k = 1:4
    [x,flag,relres] = bicgstab(A,b,tol,maxit,[],[],x0);
    X(:,k) = x;
    R(k) = relres;
    x0 = x;
end

X(:,k) is the solution vector computed at iteration k of the for-loop, and R(k) is the relative residual of that solution.

Solve a linear system by providing bicgstab with a function handle that computes A*x in place of the coefficient matrix A.

One of the Wilkinson test matrices generated by gallery is a 21-by-21 tridiagonal matrix. Preview the matrix.

A = gallery('wilk',21)
A = 21×21

    10     1     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0
     1     9     1     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0
     0     1     8     1     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0
     0     0     1     7     1     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0
     0     0     0     1     6     1     0     0     0     0     0     0     0     0     0     0     0     0     0     0     0
     0     0     0     0     1     5     1     0     0     0     0     0     0     0     0     0     0     0     0     0     0
     0     0     0     0     0     1     4     1     0     0     0     0     0     0     0     0     0     0     0     0     0
     0     0     0     0     0     0     1     3     1     0     0     0     0     0     0     0     0     0     0     0     0
     0     0     0     0     0     0     0     1     2     1     0     0     0     0     0     0     0     0     0     0     0
     0     0     0     0     0     0     0     0     1     1     1     0     0     0     0     0     0     0     0     0     0
      ⋮

The Wilkinson matrix has a special structure, so you can represent the operation A*x with a function handle. When A multiplies a vector, most of the elements in the resulting vector are zeros. The nonzero elements in the result correspond with the nonzero tridiagonal elements of A. Moreover, only the main diagonal has nonzeros that are not equal to 1.

The expression Ax becomes:

Ax=[1010001910001810017100161001510014100130001000110][x1x2x3x4x5x21]=[10x1+x2x1+9x2+x3x2+8x3+x4x19+9x20+x21x20+10x21].

The resulting vector can be written as the sum of three vectors:

Ax=[0+10x1+x2x1+9x2+x3x2+8x3+x4x19+9x20+x21x20+10x21+0]=[0x1x20]+[10x19x210x21]+[x2x210].

In MATLAB®, write a function that creates these vectors and adds them together, thus giving the value of A*x:

function y = afun(x)
y = [0; x(1:20)] + ...
    [(10:-1:0)'; (1:10)'].*x + ...
    [x(2:21); 0];
end

(This function is saved as a local function at the end of the example.)

Now, solve the linear system Ax=b by providing bicgstab with the function handle that calculates A*x. Use a tolerance of 1e-12 and 50 iterations.

b = ones(21,1);
tol = 1e-12;  
maxit = 50;
x1 = bicgstab(@afun,b,tol,maxit)
bicgstab converged at iteration 11.5 to a solution with relative residual 5.2e-13.
x1 = 21×1

    0.0910
    0.0899
    0.0999
    0.1109
    0.1241
    0.1443
    0.1544
    0.2383
    0.1309
    0.5000
      ⋮

Check that afun(x1) produces a vector of ones.

afun(x1)
ans = 21×1

    1.0000
    1.0000
    1.0000
    1.0000
    1.0000
    1.0000
    1.0000
    1.0000
    1.0000
    1.0000
      ⋮

Local Functions

function y = afun(x)
y = [0; x(1:20)] + ...
    [(10:-1:0)'; (1:10)'].*x + ...
    [x(2:21); 0];
end

Input Arguments

collapse all

Coefficient matrix, specified as a square matrix or function handle. This matrix is the coefficient matrix in the linear system A*x = b. Generally, A is a large sparse matrix or a function handle that returns the product of a large sparse matrix and column vector.

Specifying A as a Function Handle

You can optionally specify the coefficient matrix as a function handle instead of a matrix. The function handle returns matrix-vector products instead of forming the entire coefficient matrix, making the calculation more efficient.

To use a function handle, use the function signature function y = afun(x). Parameterizing Functions explains how to provide additional parameters to the function afun, if necessary. The function call afun(x) must return the value of A*x.

Data Types: double | function_handle
Complex Number Support: Yes

Right-hand side of linear equation, specified as a column vector. The length of b must be equal to size(A,1).

Data Types: double
Complex Number Support: Yes

Method tolerance, specified as a positive scalar. Use this input to trade-off accuracy and runtime in the calculation. bicgstab must meet the tolerance within the number of allowed iterations to be successful. A smaller value of tol means the answer must be more precise for the calculation to be successful.

Data Types: double

Maximum number of iterations, specified as a positive scalar integer. Increase the value of maxit to allow more iterations for bicgstab to meet the tolerance tol. Generally, a smaller value of tol means more iterations are required to successfully complete the calculation.

Preconditioner matrices, specified as separate arguments of matrices or function handles. You can specify a preconditioner matrix M or its matrix factors M = M1*M2 to improve the numerical aspects of the linear system and make it easier for bicgstab to converge quickly. You can use the incomplete matrix factorization functions ilu and ichol to generate preconditioner matrices. You also can use equilibrate prior to factorization to improve the condition number of the coefficient matrix. For more information on preconditioners, see Iterative Methods for Linear Systems.

bicgstab treats unspecified preconditioners as identity matrices.

Specifying M as a Function Handle

You can optionally specify any of M, M1, or M2 as function handles instead of matrices. The function handle performs matrix-vector operations instead of forming the entire preconditioner matrix, making the calculation more efficient.

To use a function handle, use the function signature function y = mfun(x). Parameterizing Functions explains how to provide additional parameters to the function mfun, if necessary. The function call mfun(x) must return the value of M\x or M2\(M1\x).

Data Types: double | function_handle
Complex Number Support: Yes

Initial guess, specified as a column vector with length equal to size(A,2). If you can provide bicgstab with a more reasonable initial guess x0 than the default vector of zeros, then it can save computation time and help the algorithm converge faster.

Data Types: double
Complex Number Support: Yes

Output Arguments

collapse all

Linear system solution, returned as a column vector. This output gives the approximate solution to the linear system A*x = b. If the calculation is successful (flag = 0), then relres is less than or equal to tol.

Whenever the calculation is not successful (flag ~= 0), the solution x returned by bicgstab is the one with minimal residual norm computed over all the iterations.

Convergence flag, returned as one of the scalar values in this table. The convergence flag indicates whether the calculation was successful and differentiates between several different forms of failure.

Flag Value

Convergence

0

Success — bicgstab converged to the desired tolerance tol within maxit iterations.

1

Failure — bicgstab iterated maxit iterations but did not converge.

2

Failure — The preconditioner matrix M or M = M1*M2 is ill conditioned.

3

Failure — bicgstab stagnated after two consecutive iterations were the same.

4

Failure — One of the scalar quantities calculated by the bicgstab algorithm became too small or too large to continue computing.

Relative residual error, returned as a scalar. The relative residual error relres = norm(b-A*x)/norm(b) is an indication of how accurate the answer is. If the calculation converges to the tolerance tol within maxit iterations, then relres <= tol.

Data Types: double

Iteration number, returned as a scalar. This output indicates the iteration number at which the computed answer for x was calculated. Each outer iteration of bicgstab includes two inner iterations, so iter can be returned as a decimal number of iterations.

Data Types: double

Residual error, returned as a vector. The residual error norm(b-A*x) reveals how close the algorithm is to converging for a given value of x. The number of elements in resvec is equal to the number of iterations. You can examine the contents of resvec to help decide whether to change the values of tol or maxit.

Data Types: double

More About

collapse all

Biconjugate Gradients Stabilized Method

The biconjugate gradients stabilized (BiCGSTAB) algorithm was developed to improve on the BiCG algorithm by using restarted GMRES steps to mitigate the irregular convergence behavior in the BiCG steps.

The GMRES(1) repair steps in BiCGSTAB only utilize degree-one minimum residual polynomials, so the method can stagnate when these low-order steps are insufficient. Use of higher order polynomials led to the development of other methods such as BiCGSTABL [1].

Tips

  • Convergence of most iterative methods depends on the condition number of the coefficient matrix, cond(A). You can use equilibrate to improve the condition number of A, and on its own this makes it easier for most iterative solvers to converge. However, using equilibrate also leads to better quality preconditioner matrices when you subsequently factor the equilibrated matrix B = R*P*A*C.

  • You can use matrix reordering functions such as dissect and symrcm to permute the rows and columns of the coefficient matrix and minimize the number of nonzeros when the coefficient matrix is factored to generate a preconditioner. This can reduce the memory and time required to subsequently solve the preconditioned linear system.

References

[1] Barrett, R., M. Berry, T.F. Chan, et al., Templates for the Solution of Linear Systems: Building Blocks for Iterative Methods, SIAM, Philadelphia, 1994.

[2] van der Vorst, H.A., "BI-CGSTAB: A fast and smoothly converging variant of BI-CG for the solution of nonsymmetric linear systems," SIAM J. Sci. Stat. Comput., March 1992, Vol. 13, No. 2, pp. 631–644.

Extended Capabilities

Introduced before R2006a