Preconditioners in SciPy - Step-by-Step Execution

Concept Flow - Preconditioners

Start: Linear System Ax=b

↓

Choose Preconditioner M

↓

Transform System: M^-1 A x = M^-1 b

↓

Iterative Solver (e.g., CG) uses M to speed up

↓

Check Convergence

↓

Solution x

Preconditioners transform a linear system to help iterative solvers find solutions faster by improving convergence.

Execution Sample

SciPy

import numpy as np
from scipy.sparse.linalg import cg, LinearOperator

A = np.array([[4,1],[1,3]])
b = np.array([1,2])

M_inv = np.linalg.inv(np.diag(np.diag(A)))

x, info = cg(A, b, M=LinearOperator(A.shape, matvec=lambda x: M_inv @ x))

Solves Ax=b using Conjugate Gradient with a diagonal preconditioner to speed up convergence.

Execution Table

Step	Residual Norm	Preconditioned Residual	Direction Vector	Solution x	Info
0	1.8028	[0.4507, 0.6009]	[0.4507, 0.6009]	[0.0, 0.0]	Starting CG
1	0.2236	[0.1118, 0.1491]	[0.1118, 0.1491]	[0.0909, 0.1818]	Iteration 1
2	0.0	[0.0, 0.0]	[0.0, 0.0]	[0.0909, 0.5455]	0

💡 Residual norm reached zero, CG converged in 2 iterations with preconditioning

Variable Tracker

Variable	Start	After 1	After 2	Final
Residual Norm	1.8028	0.2236	0.0	0.0
Solution x	[0.0, 0.0]	[0.0909, 0.1818]	[0.0909, 0.5455]	[0.0909, 0.5455]
Direction Vector	[0.4507, 0.6009]	[0.1118, 0.1491]	[0.0, 0.0]	[0.0, 0.0]

Key Moments - 2 Insights

Why do we apply M^-1 to the residual instead of directly to A or b?

Applying M^-1 to the residual improves the search direction in the iterative solver, making convergence faster, as shown in the 'Preconditioned Residual' column in the execution_table.

What does the 'Info' value represent in the CG solver output?

'Info' indicates the solver status: 0 means convergence achieved, as seen in the last row of the execution_table where 'Info' is '0'.

Visual Quiz - 3 Questions

Look at the execution_table at Step 1, what is the approximate value of the Residual Norm?

A 0.0

B 1.8028

C 0.2236

D 0.5

💡 Hint

Check the 'Residual Norm' column at Step 1 in the execution_table.

At which step does the CG solver converge according to the execution_table?

A Step 0

B Step 2

C Step 1

D Never converges

💡 Hint

Look at the 'Info' column and residual norm reaching zero in the execution_table.

If we remove the preconditioner M, how would the number of iterations likely change?

A Increase, more iterations needed

B Stay the same

C Decrease to 1 iteration

D Solver would fail immediately

💡 Hint

Preconditioners improve convergence speed; without M, iterative solvers usually take more steps.

Concept Snapshot

Preconditioners help iterative solvers solve Ax=b faster.
They transform the system using M^-1 to improve convergence.
Common preconditioners include diagonal or incomplete factorizations.
Use with solvers like Conjugate Gradient (cg) in scipy.
Preconditioning reduces iterations and computation time.

Full Transcript

Preconditioners are tools used to make solving linear systems faster. When we have a system Ax=b, we choose a preconditioner M to transform it into M^-1 A x = M^-1 b. This helps iterative solvers like Conjugate Gradient find the solution quicker by improving the system's properties. In the example, we used a diagonal preconditioner and saw the residual norm drop quickly, converging in just two iterations. The preconditioned residual guides the solver's search direction. Without preconditioning, the solver would take more steps. The 'Info' output tells us if the solver converged. Preconditioners are important for efficient data science computations involving large linear systems.

Practice

1.

What is the main purpose of a preconditioner in scipy when solving linear systems?

easy

A. To speed up the convergence of iterative solvers

B. To increase the size of the matrix

C. To change the solution of the system

D. To make the matrix non-square

Solution

1. Step 1: Understand the role of preconditioners

   Preconditioners are used to improve the efficiency of iterative methods by transforming the system into an easier one to solve.

2. Step 2: Identify the effect on convergence

   They help iterative solvers like Conjugate Gradient converge faster by approximating the inverse of the matrix.

3. Final Answer:

   To speed up the convergence of iterative solvers -> Option A

4. Quick Check:

   Preconditioner purpose = speed up convergence [OK]

Hint: Preconditioners help iterative solvers run faster [OK]

Common Mistakes:

• Thinking preconditioners change the solution
• Believing preconditioners increase matrix size
• Confusing preconditioners with matrix transformations that alter shape

2.

Which of the following is the correct way to create a simple Jacobi preconditioner using scipy.sparse.linalg.LinearOperator?

import numpy as np
from scipy.sparse.linalg import LinearOperator

A = np.array([[4, 1], [1, 3]])
M = LinearOperator(shape=A.shape, matvec=lambda x: ...)

easy

A. matvec=lambda x: x / np.diag(A)

B. matvec=lambda x: np.dot(A, x)

C. matvec=lambda x: x * np.diag(A)

D. matvec=lambda x: np.linalg.solve(A, x)

Solution

1. Step 1: Recall Jacobi preconditioner definition

   Jacobi preconditioner uses the inverse of the diagonal elements of matrix A.

2. Step 2: Implement matvec for Jacobi

   Applying the preconditioner means dividing each element of x by the corresponding diagonal element of A.

3. Final Answer:

   matvec=lambda x: x / np.diag(A) -> Option A

4. Quick Check:

   Jacobi preconditioner = divide by diagonal [OK]

Hint: Jacobi preconditioner divides vector by matrix diagonal [OK]

Common Mistakes:

• Using matrix multiplication instead of division
• Trying to solve full system instead of diagonal scaling
• Multiplying by diagonal instead of dividing

3.

Given the following code, what will be the output of print(M.matvec(b))?

import numpy as np
from scipy.sparse.linalg import LinearOperator

A = np.array([[2, 0], [0, 5]])
b = np.array([4, 10])
M = LinearOperator(shape=A.shape, matvec=lambda x: x / np.diag(A))
print(M.matvec(b))

medium

A. [0.5, 2.0]

B. [8.0, 50.0]

C. [2.0, 2.0]

D. [4.0, 10.0]

Solution

1. Step 1: Calculate diagonal of A

   Diagonal elements are [2, 5].

2. Step 2: Apply matvec function

   Divide each element of b by corresponding diagonal: [4/2, 10/5] = [2.0, 2.0].

3. Final Answer:

   [2.0, 2.0] -> Option C

4. Quick Check:

   Vector divided by diagonal = [2.0, 2.0] [OK]

Hint: Divide vector elements by diagonal elements to get output [OK]

Common Mistakes:

• Multiplying instead of dividing
• Confusing vector and matrix multiplication
• Using wrong diagonal values

4.

Identify the error in the following code that attempts to create a Jacobi preconditioner:

import numpy as np
from scipy.sparse.linalg import LinearOperator

A = np.array([[3, 1], [1, 4]])
M = LinearOperator(shape=A.shape, matvec=lambda x: np.diag(A) * x)
print(M.matvec(np.array([1, 2])))

medium

A. Using np.diag(A) incorrectly as a matrix

B. Multiplying by diagonal instead of dividing

C. Shape of LinearOperator is wrong

D. Input vector has wrong size

Solution

1. Step 1: Understand Jacobi preconditioner operation

   Jacobi preconditioner divides vector elements by diagonal elements of A.

2. Step 2: Check given matvec function

   Code multiplies vector by diagonal instead of dividing, which is incorrect.

3. Final Answer:

   Multiplying by diagonal instead of dividing -> Option B

4. Quick Check:

   Jacobi requires division, not multiplication [OK]

Hint: Jacobi preconditioner divides vector by diagonal, not multiply [OK]

Common Mistakes:

• Confusing multiplication with division
• Ignoring element-wise operations
• Not verifying mathematical definition

5.

You want to speed up solving a large sparse system Ax = b using Conjugate Gradient in scipy. Which approach best uses a preconditioner?

from scipy.sparse.linalg import cg, LinearOperator
import numpy as np

# A is large sparse matrix
# b is known vector

# Option 1: Use identity preconditioner
M1 = LinearOperator(A.shape, matvec=lambda x: x)

# Option 2: Use Jacobi preconditioner
diag = A.diagonal()
M2 = LinearOperator(A.shape, matvec=lambda x: x / diag)

# Option 3: Use incomplete Cholesky (not shown)

x, info = cg(A, b, M=M2)

Why is Option 2 preferred over Option 1?

hard

A. Because identity preconditioner is not a LinearOperator

B. Because identity preconditioner changes the solution

C. Because Jacobi preconditioner makes matrix larger

D. Because Jacobi preconditioner approximates inverse and speeds convergence

Solution

1. Step 1: Understand identity preconditioner effect

   Identity preconditioner does nothing; it returns the vector unchanged, so no speedup.

2. Step 2: Understand Jacobi preconditioner effect

   Jacobi approximates the inverse of the diagonal, improving convergence speed of iterative solver.

3. Final Answer:

   Because Jacobi preconditioner approximates inverse and speeds convergence -> Option D

4. Quick Check:

   Jacobi preconditioner = faster convergence [OK]

Hint: Jacobi preconditioner speeds up solver by approximating inverse [OK]

Common Mistakes:

• Thinking identity preconditioner changes solution
• Believing Jacobi increases matrix size
• Confusing LinearOperator requirements