Elimination
Gaussian elimination or Gauss-Jordan elimination is a foundational process for solving systems of linear equations.
Elimination can be generalized as the process of transforming an unsolvable Ax = b into a solvable Ux = c.
Contents
Introduction
Consider the below system of equations.
x + 2y + z = 2 3x + 8y + z = 12 4y + z = 2
The first step to solving this system with traditional algebra is eliminating x.
Trivially rewrite the first equation:
x + 2y + z = 2 x = 2 - 2y - z
Substitute this into the second equation:
3x + 8y + z = 12 3(2 - 2y - z) + 8y + z = 12 6 - 6y - 3z + 8y + z = 12 2y - 2z = 6
Now x has been eliminated. But can either y or z be eliminated the same way?
Linear Algebra
This system can be expressed in the form Ax = b:
A x = b ┌ ┐ ┌ ┐ ┌ ┐ │ 1 2 1│ │ x│ │ 2│ │ 3 8 1│ │ y│ = │12│ │ 0 4 1│ │ z│ │ 2│ └ ┘ └ ┘ └ ┘
Eliminating A
The solution lies in the elimination of A.
The first step is identifying pivots. By reducing the entire column to zeros and the pivot, it is eliminated into a pivot column.
Pivots are noted with boxes or circles in the matrix, as in:
┌ ┐ │ [1] 2 1│ │ 3 8 1│ │ 0 4 1│ └ ┘
Algebraically, a value is being zeroed out by expressing that variable in terms of the other equations.
Literally, we are subtracting a multiple of the pivot row from the targeted row to create a zero below the pivot. That multiple is discovered by solving part of the equation.
┌ ┐ ┌ ┐ ┌ ┐ │ 3│ │ 1│ │ 0│ │ 8│ - m│ 2│ = │ ?│ │ 1│ │ 1│ │ ?│ └ ┘ └ ┘ └ ┘ 3 - 1m = 0 m = 3 ┌ ┐ ┌ ┐ ┌ ┐ │ 3│ │ 1│ │ 0│ │ 8│ - 3│ 2│ = │ 2│ │ 1│ │ 1│ │ -2│ └ ┘ └ ┘ └ ┘
This process is then repeated for every row below. In this specific problem the multiple is zero as there is already a zero in place.
┌ ┐ ┌ ┐ ┌ ┐ │ 0│ │ 1│ │ 0│ │ 4│ - m│ 2│ = │ ?│ │ 1│ │ 1│ │ ?│ └ ┘ └ ┘ └ ┘ 0 - 1m = 0 m = 0 ┌ ┐ ┌ ┐ ┌ ┐ │ 0│ │ 1│ │ 0│ │ 4│ - 0│ 2│ = │ 4│ │ 1│ │ 1│ │ 1│ └ ┘ └ ┘ └ ┘
The matrix is reconstructed:
┌ ┐ │ [1] 2 1│ │ 0 2 -2│ │ 0 4 1│ └ ┘
The process is then repeated with a new pivot.
If a pivot is zero, rows should be swapped using permutation matrices. If all rows are zero, then this is a free column (as opposed to a pivot column) and should be skipped.
In this specific problem we do have another pivot.
┌ ┐ │ [1] 2 1│ │ 0 [2] -2│ │ 0 4 1│ └ ┘
Another iteration of elimination yields this matrix, U.
┌ ┐ │[1] 2 1 │ │ 0 [2] -2 │ │ 0 0 [5]│ └ ┘
U is the row echelon form of A. It may also be a upper triangular matrix.
Balancing the Right Hand Side
To re-balance the original equation, replicate the subtractions on b. Recall that the (non-zero) multipliers were 3 and 2:
┌ ┐ │ 2│ │12│ │ 2│ └ ┘ 12 - 2m 12 - 2(3) 6 ┌ ┐ │ 2│ │ 6│ │ 2│ └ ┘ 2 - 6m 2 - 6(2) -10 ┌ ┐ │ 2│ │ 6│ │-10│ └ ┘
Result of Elimination
The system has been reduced to Ux = c:
┌ ┐ ┌ ┐ ┌ ┐ │[1] 2 1 │ │ x│ │ 2│ │ 0 [2] -2 │ │ y│ = │ 6│ │ 0 0 [5]│ │ z│ │-10│ └ ┘ └ ┘ └ ┘
x + 2y + z = 2 2y + 2z = 6 5x = -10
Algebraic substitution can now be used to solve; x=2, y=1, z=-2.
Simplification with Augmented Matrices
Instead of eliminating A then re-balancing b, eliminate the augmented matrix in a single step.
Augmented matrix = [ A b] ┌ ┐ │ 1 2 1 2│ Augmented matrix = │ 3 8 1 12│ │ 0 4 1 2│ └ ┘
Elimination proceeds the exact same way.
┌ ┐ │[1] 2 1 2│ │ 3 8 1 12│ │ 0 4 1 2│ └ ┘ ┌ ┐ │[1] 2 1 2│ │ 0 [2] -2 6│ │ 0 4 1 2│ └ ┘ ┌ ┐ │[1] 2 1 2│ │ 0 [2] -2 6│ │ 0 0 [5] -10│ └ ┘
Which can be re-expressed as:
┌ ┐ ┌ ┐ ┌ ┐ │[1] 2 1 │ │ x│ │ 2│ │ 0 [2] -2 │ │ y│ = │ 6│ │ 0 0 [5]│ │ z│ │-10│ └ ┘ └ ┘ └ ┘
Note that this matches the reduced system that was achieved above.
Reduced Row Echelon Form
U can be further reduced with backwards elimination. Repeat the process in the opposite direction until all numbers above and below the pivots are zero, yielding the matrix R.
R is the reduced row echelon form of A. It may also be a lower triangular matrix.
Failure of Elimination
U must have n non-zero pivots for n unknowns.
As noted above, if a pivot is zero, rows should be swapped using permutation matrices. This is a temporary failure.
If rows cannot be swapped to complete elimination, the system has permanent failure. The matrix must then be non-invertible.