线性系统

2x + y = 5          Two lines in 2D.
x - y  = 1          They intersect at x=2, y=1.

A = | 2  1 |    b = | 5 |
    | 1 -1 |        | 1 |

Row picture: solve 2x + y = 5 and x - y = 1 simultaneously.

Column picture: find x1, x2 such that:
  x1 * [2, 1] + x2 * [1, -1] = [5, 1]
  2 * [2, 1] + 1 * [1, -1] = [4+1, 2-1] = [5, 1]   check.

1. For each column k (the pivot column):
   a. Find the largest entry in column k at or below row k (partial pivoting).
   b. Swap that row with row k.
   c. For each row i below k:
      - Compute multiplier m = A[i][k] / A[k][k]
      - Subtract m times row k from row i.
2. Back substitute: solve from the last equation upward.

Original:
| 2  1  1 | 8 |       R2 = R2 - (2)R1     | 2  1   1 |  8 |
| 4  3  3 |20 |  -->  R3 = R3 - (1)R1 --> | 0  1   1 |  4 |
| 2  3  1 |12 |                            | 0  2   0 |  4 |

                       R3 = R3 - (2)R2     | 2  1   1 |  8 |
                                       --> | 0  1   1 |  4 |
                                           | 0  0  -2 | -4 |

Back substitute:
  -2 * x3 = -4    -->  x3 = 2
  x2 + 2  = 4     -->  x2 = 2
  2*x1 + 2 + 2 = 8 --> x1 = 2

Bad pivot:                       With partial pivoting:
| 0.001  1 | 1.001 |            Swap rows first:
| 1      1 | 2     |            | 1      1 | 2     |
                                 | 0.001  1 | 1.001 |
m = 1/0.001 = 1000              m = 0.001/1 = 0.001
R2 = R2 - 1000*R1               R2 = R2 - 0.001*R1
| 0.001  1     | 1.001   |      | 1      1     | 2     |
| 0     -999   | -999.0  |      | 0      0.999 | 0.999 |

x2 = 1.000 (correct)            x2 = 1.000 (correct)
x1 = (1.001 - 1)/0.001          x1 = (2 - 1)/1 = 1.000 (correct)
   = 0.001/0.001 = 1.000        Stable because the multiplier is small.

A = L @ U

| 2  1  1 |   | 1  0  0 |   | 2  1   1 |
| 4  3  3 | = | 2  1  0 | @ | 0  1   1 |
| 2  3  1 |   | 1  2  1 |   | 0  0  -2 |

A = Q @ R

Q has orthonormal columns: Q^T Q = I
R is upper triangular

To solve Ax = b:
  QRx = b
  Rx = Q^T b    (just multiply by Q^T, no inversion needed)
  Back substitute to get x.

Given columns a1, a2, ... of A:

q1 = a1 / ||a1||

q2 = a2 - (a2 . q1) * q1        (subtract projection onto q1)
q2 = q2 / ||q2||                (normalize)

q3 = a3 - (a3 . q1) * q1 - (a3 . q2) * q2
q3 = q3 / ||q3||

R[i][j] = qi . aj    for i <= j

A = L @ L^T

| 4  2 |   | 2  0 |   | 2  1 |
| 2  5 | = | 1  2 | @ | 0  2 |

L[i][i] = sqrt(A[i][i] - sum(L[i][k]^2 for k < i))
L[i][j] = (A[i][j] - sum(L[i][k]*L[j][k] for k < j)) / L[j][j]    for i > j

minimize ||Ax - b||^2

This is the sum of squared residuals:
  sum((A[i,:] @ x - b[i])^2 for i in range(m))

Original system (overdetermined, 4 equations, 2 unknowns):
| 1  1 |         | 3 |
| 1  2 | x     = | 5 |       No exact x satisfies all 4 equations.
| 1  3 |         | 6 |
| 1  4 |         | 8 |

Normal equations:
A^T A = | 4  10 |    A^T b = | 22 |
        | 10 30 |            | 63 |

Solve: x = [1.5, 1.7]

This is linear regression. x[0] is the intercept, x[1] is the slope.

A = U Sigma V^T        (SVD)

Sigma = | 5  0 |       Sigma+ = | 1/5  0  0 |
        | 0  2 |                | 0  1/2  0 |
        | 0  0 |

A+ = V Sigma+ U^T

Well-conditioned (kappa ~ 1):        Ill-conditioned (kappa ~ 10^15):
Small change in b -->                Small change in b -->
small change in x                    huge change in x

| 2  0 |   kappa = 2/1 = 2          | 1   1          |   kappa ~ 10^15
| 0  1 |   safe to solve            | 1   1+10^(-15) |   solution is garbage

Algorithm sketch:
  x0 = initial guess (often zero)
  r0 = b - A x0           (residual)
  p0 = r0                 (search direction)

  For k = 0, 1, 2, ...:
    alpha = (rk . rk) / (pk . A pk)
    x_{k+1} = xk + alpha * pk
    r_{k+1} = rk - alpha * A pk
    beta = (r_{k+1} . r_{k+1}) / (rk . rk)
    p_{k+1} = r_{k+1} + beta * pk
    if ||r_{k+1}|| < tolerance: stop

import numpy as np

def gaussian_elimination(A, b):
    n = len(b)
    Ab = np.hstack([A.astype(float), b.reshape(-1, 1).astype(float)])

    for k in range(n):
        max_row = k + np.argmax(np.abs(Ab[k:, k]))
        Ab[[k, max_row]] = Ab[[max_row, k]]

        if abs(Ab[k, k]) < 1e-12:
            raise ValueError(f"Matrix is singular or nearly singular at pivot {k}")

        for i in range(k + 1, n):
            m = Ab[i, k] / Ab[k, k]
            Ab[i, k:] -= m * Ab[k, k:]

    x = np.zeros(n)
    for i in range(n - 1, -1, -1):
        x[i] = (Ab[i, -1] - Ab[i, i+1:n] @ x[i+1:n]) / Ab[i, i]

    return x

def lu_decompose(A):
    n = A.shape[0]
    L = np.eye(n)
    U = A.astype(float).copy()
    P = np.eye(n)

    for k in range(n):
        max_row = k + np.argmax(np.abs(U[k:, k]))
        if max_row != k:
            U[[k, max_row]] = U[[max_row, k]]
            P[[k, max_row]] = P[[max_row, k]]
            if k > 0:
                L[[k, max_row], :k] = L[[max_row, k], :k]

        for i in range(k + 1, n):
            L[i, k] = U[i, k] / U[k, k]
            U[i, k:] -= L[i, k] * U[k, k:]

    return P, L, U

def lu_solve(P, L, U, b):
    n = len(b)
    Pb = P @ b.astype(float)

    y = np.zeros(n)
    for i in range(n):
        y[i] = Pb[i] - L[i, :i] @ y[:i]

    x = np.zeros(n)
    for i in range(n - 1, -1, -1):
        x[i] = (y[i] - U[i, i+1:] @ x[i+1:]) / U[i, i]

    return x

def cholesky(A):
    n = A.shape[0]
    L = np.zeros_like(A, dtype=float)

    for i in range(n):
        for j in range(i + 1):
            s = A[i, j] - L[i, :j] @ L[j, :j]
            if i == j:
                if s <= 0:
                    raise ValueError("Matrix is not positive definite")
                L[i, j] = np.sqrt(s)
            else:
                L[i, j] = s / L[j, j]

    return L

def least_squares_normal(A, b):
    AtA = A.T @ A
    Atb = A.T @ b
    return gaussian_elimination(AtA, Atb)

def ridge_regression(A, b, lam):
    n = A.shape[1]
    AtA = A.T @ A + lam * np.eye(n)
    Atb = A.T @ b
    L = cholesky(AtA)
    y = np.zeros(n)
    for i in range(n):
        y[i] = (Atb[i] - L[i, :i] @ y[:i]) / L[i, i]
    x = np.zeros(n)
    for i in range(n - 1, -1, -1):
        x[i] = (y[i] - L.T[i, i+1:] @ x[i+1:]) / L.T[i, i]
    return x

def condition_number(A):
    U, S, Vt = np.linalg.svd(A)
    return S[0] / S[-1]

np.random.seed(42)
X_raw = np.random.randn(100, 3)
w_true = np.array([2.0, -1.0, 0.5])
y = X_raw @ w_true + np.random.randn(100) * 0.1

X = np.column_stack([np.ones(100), X_raw])

w_ols = least_squares_normal(X, y)
print(f"OLS weights (ours):    {w_ols}")

w_np = np.linalg.lstsq(X, y, rcond=None)[0]
print(f"OLS weights (numpy):   {w_np}")
print(f"Max difference: {np.max(np.abs(w_ols - w_np)):.2e}")

w_ridge = ridge_regression(X, y, lam=1.0)
print(f"Ridge weights (ours):  {w_ridge}")

from sklearn.linear_model import Ridge
ridge_sk = Ridge(alpha=1.0, fit_intercept=False)
ridge_sk.fit(X, y)
print(f"Ridge weights (sklearn): {ridge_sk.coef_}")