采样方法

U ~ Uniform(0, 1)

P(a <= U <= b) = b - a    for 0 <= a <= b <= 1

Properties:
  E[U] = 0.5
  Var(U) = 1/12

F(x) = P(X <= x)

Properties:
  F is non-decreasing
  F(-inf) = 0
  F(+inf) = 1
  F maps the real line to [0, 1]

Algorithm:
  1. Generate u ~ Uniform(0, 1)
  2. Return F_inverse(u)

Why it works:
  P(X <= x) = P(F_inverse(U) <= x) = P(U <= F(x)) = F(x)

PDF: f(x) = lambda * exp(-lambda * x),   x >= 0
CDF: F(x) = 1 - exp(-lambda * x)

Solve F(x) = u for x:
  u = 1 - exp(-lambda * x)
  exp(-lambda * x) = 1 - u
  x = -ln(1 - u) / lambda

Since (1 - U) and U have the same distribution:
  x = -ln(u) / lambda

Target distribution: p(x)  (can evaluate, possibly unnormalized)
Proposal distribution: q(x)  (can sample from)
Bound: M such that p(x) <= M * q(x) for all x

Algorithm:
  1. Sample x ~ q(x)
  2. Sample u ~ Uniform(0, 1)
  3. If u < p(x) / (M * q(x)), accept x
  4. Otherwise, reject and go to step 1

Acceptance rate = Z/M, where Z is the integral of p(x)
                (Z = 1, so the rate is 1/M, when p is normalized)

Goal: estimate E_p[f(x)] = integral of f(x) * p(x) dx

Rewrite:
  E_p[f(x)] = integral of f(x) * (p(x)/q(x)) * q(x) dx
            = E_q[f(x) * w(x)]

where w(x) = p(x) / q(x)  are the importance weights.

Estimator:
  E_p[f(x)] ~ (1/N) * sum(f(x_i) * w(x_i))    where x_i ~ q(x)

Goal: estimate I = integral of g(x) dx over domain D

Method:
  1. Sample x_1, ..., x_N uniformly from D
  2. I ~ (Volume of D / N) * sum(g(x_i))

Error: O(1 / sqrt(N))   regardless of dimension

Sample (x, y) uniformly from [-1, 1] x [-1, 1]
Count how many fall inside the unit circle: x^2 + y^2 <= 1
pi ~ 4 * (count inside) / (total count)

E[f(X)] ~ (1/N) * sum(f(x_i))    where x_i ~ p(x)

The sample mean converges to the true expectation.
Variance of the estimator = Var(f(X)) / N

Target: p(x)  (known up to a normalizing constant)
Proposal: q(x'|x)  (how to propose the next state given the current state)

Metropolis-Hastings algorithm:
  1. Start at some x_0
  2. For t = 1, 2, ..., T:
     a. Propose x' ~ q(x'|x_t)
     b. Compute acceptance ratio:
        alpha = [p(x') * q(x_t|x')] / [p(x_t) * q(x'|x_t)]
     c. Accept with probability min(1, alpha):
        - If u < alpha (u ~ Uniform(0,1)): x_{t+1} = x'
        - Otherwise: x_{t+1} = x_t
  3. Discard first B samples (burn-in)
  4. Return remaining samples

Target: p(x_1, x_2, ..., x_d)

Algorithm:
  For each iteration t:
    Sample x_1^{t+1} ~ p(x_1 | x_2^t, x_3^t, ..., x_d^t)
    Sample x_2^{t+1} ~ p(x_2 | x_1^{t+1}, x_3^t, ..., x_d^t)
    ...
    Sample x_d^{t+1} ~ p(x_d | x_1^{t+1}, x_2^{t+1}, ..., x_{d-1}^{t+1})

p_i = exp(z_i / T) / sum(exp(z_j / T))

T = 1.0: standard softmax (original distribution)
T -> 0:  argmax (deterministic, always picks highest logit)
T -> inf: uniform (all tokens equally likely)
T < 1.0: sharpens the distribution (more confident, less diverse)
T > 1.0: flattens the distribution (less confident, more diverse)

Algorithm:
  1. Compute softmax probabilities for all V tokens
  2. Sort tokens by probability (descending)
  3. Keep only the top k tokens
  4. Renormalize: p_i' = p_i / sum(p_j for j in top-k)
  5. Sample from the renormalized distribution

k = 1:  greedy decoding
k = V:  no filtering (standard sampling)
k = 40: typical setting, removes long tail of unlikely tokens

Algorithm:
  1. Compute softmax probabilities for all V tokens
  2. Sort tokens by probability (descending)
  3. Find smallest k such that sum of top-k probabilities >= p
  4. Keep only those k tokens
  5. Renormalize and sample

p = 0.9:  keeps tokens covering 90% of probability mass
p = 1.0:  no filtering
p = 0.1:  very restrictive, nearly greedy

Standard sampling (not differentiable):
  z ~ N(mu, sigma^2)

  The randomness blocks gradient flow.
  d/d_mu [sample from N(mu, sigma^2)] = ???

Reparameterized sampling:
  epsilon ~ N(0, 1)          (fixed random noise, no parameters)
  z = mu + sigma * epsilon   (deterministic function of parameters)

  Now z is a deterministic, differentiable function of mu and sigma.
  d(z)/d(mu) = 1
  d(z)/d(sigma) = epsilon

  Gradients flow through mu and sigma.

To sample from a categorical distribution with log-probabilities log(p_1), ..., log(p_k):
  1. Sample g_i ~ Gumbel(0, 1) for each category
     (g = -log(-log(u)), where u ~ Uniform(0, 1))
  2. Return argmax(log(p_i) + g_i)

This produces exact categorical samples.

Replace the hard argmax with a soft softmax:
  y_i = exp((log(p_i) + g_i) / tau) / sum(exp((log(p_j) + g_j) / tau))

tau (temperature) controls the approximation:
  tau -> 0:  approaches a one-hot vector (hard categorical)
  tau -> inf: approaches uniform (1/k, 1/k, ..., 1/k)
  tau = 1.0: soft approximation

Standard Monte Carlo:
  Sample N points uniformly from [0, 1]
  Some regions may have clusters, others gaps

Stratified sampling:
  Divide [0, 1] into N equal strata: [0, 1/N), [1/N, 2/N), ..., [(N-1)/N, 1)
  Sample one point uniformly within each stratum
  x_i = (i + u_i) / N   where u_i ~ Uniform(0, 1),  i = 0, ..., N-1

Var(stratified) <= Var(standard Monte Carlo)

The improvement is largest when f(x) varies smoothly.
For piecewise-constant functions, stratified sampling is exact.

Forward process (known):
  x_t = sqrt(alpha_t) * x_{t-1} + sqrt(1 - alpha_t) * epsilon
  where epsilon ~ N(0, I)

  After T steps: x_T ~ N(0, I)  (pure noise)

Reverse process (learned):
  x_{t-1} = (1/sqrt(alpha_t)) * (x_t - (1 - alpha_t)/sqrt(1 - alpha_bar_t) * epsilon_theta(x_t, t)) + sigma_t * z
  where z ~ N(0, I)

  Each denoising step is a sampling step.

import math
import random

def sample_uniform(a, b):
    return a + (b - a) * random.random()

def sample_exponential_inverse_cdf(lam):
    u = random.random()
    return -math.log(u) / lam

def rejection_sample(target_pdf, proposal_sample, proposal_pdf, M):
    while True:
        x = proposal_sample()
        u = random.random()
        if u < target_pdf(x) / (M * proposal_pdf(x)):
            return x

def importance_sampling_estimate(f, target_pdf, proposal_pdf, proposal_sample, n):
    total = 0
    for _ in range(n):
        x = proposal_sample()
        w = target_pdf(x) / proposal_pdf(x)
        total += f(x) * w
    return total / n

def monte_carlo_pi(n):
    inside = 0
    for _ in range(n):
        x = random.uniform(-1, 1)
        y = random.uniform(-1, 1)
        if x*x + y*y <= 1:
            inside += 1
    return 4 * inside / n

def metropolis_hastings(target_log_pdf, proposal_sample, proposal_log_pdf, x0, n_samples, burn_in):
    samples = []
    x = x0
    for i in range(n_samples + burn_in):
        x_new = proposal_sample(x)
        log_alpha = (target_log_pdf(x_new) + proposal_log_pdf(x, x_new)
                     - target_log_pdf(x) - proposal_log_pdf(x_new, x))
        if math.log(random.random()) < log_alpha:
            x = x_new
        if i >= burn_in:
            samples.append(x)
    return samples

def gibbs_sampling_2d(conditional_x_given_y, conditional_y_given_x, x0, y0, n_samples, burn_in):
    x, y = x0, y0
    samples = []
    for i in range(n_samples + burn_in):
        x = conditional_x_given_y(y)
        y = conditional_y_given_x(x)
        if i >= burn_in:
            samples.append((x, y))
    return samples

def softmax(logits):
    max_l = max(logits)
    exps = [math.exp(z - max_l) for z in logits]
    total = sum(exps)
    return [e / total for e in exps]

def temperature_sample(logits, temperature):
    scaled = [z / temperature for z in logits]
    probs = softmax(scaled)
    return sample_from_probs(probs)

def top_k_sample(logits, k):
    indexed = sorted(enumerate(logits), key=lambda x: -x[1])
    top = indexed[:k]
    top_logits = [l for _, l in top]
    probs = softmax(top_logits)
    idx = sample_from_probs(probs)
    return top[idx][0]

def top_p_sample(logits, p):
    probs = softmax(logits)
    indexed = sorted(enumerate(probs), key=lambda x: -x[1])
    cumsum = 0
    selected = []
    for token_idx, prob in indexed:
        cumsum += prob
        selected.append((token_idx, prob))
        if cumsum >= p:
            break
    sel_probs = [pr for _, pr in selected]
    total = sum(sel_probs)
    sel_probs = [pr / total for pr in sel_probs]
    idx = sample_from_probs(sel_probs)
    return selected[idx][0]

def reparam_sample(mu, sigma):
    epsilon = random.gauss(0, 1)
    return mu + sigma * epsilon

def reparam_gradient(mu, sigma, epsilon):
    dz_dmu = 1.0
    dz_dsigma = epsilon
    return dz_dmu, dz_dsigma

def gumbel_sample():
    u = random.random()
    return -math.log(-math.log(u))

def gumbel_softmax(logits, temperature):
    gumbels = [math.log(p) + gumbel_sample() for p in logits]
    return softmax([g / temperature for g in gumbels])

import numpy as np

rng = np.random.default_rng(42)

exponential_samples = rng.exponential(scale=2.0, size=10000)
print(f"Exponential mean: {exponential_samples.mean():.4f} (expected 2.0)")

from scipy import stats
normal = stats.norm(loc=0, scale=1)
print(f"CDF at 1.96: {normal.cdf(1.96):.4f}")
print(f"Inverse CDF at 0.975: {normal.ppf(0.975):.4f}")

logits = np.array([2.0, 1.0, 0.5, 0.1, -1.0])
temperature = 0.7
scaled = logits / temperature
probs = np.exp(scaled - scaled.max()) / np.exp(scaled - scaled.max()).sum()
token = rng.choice(len(logits), p=probs)
print(f"Sampled token index: {token}")