Srishti280992's picture
Add incentivized exploration reproduction bundle
b381c58 verified
Raw
History Blame Contribute Delete
12.3 kB
"""Executable re-derivation of the lemma chain behind Theorem 2 (Claim 1) and
Theorem 1 (Claim 2). Every step is an assertion; the script exits non-zero on failure.
Chain verified:
Lemma 7 sum_pi Q_m(pi) Reg_hat_t(pi) < (K-1)/gamma_m [IPGS kernel identity]
Cor. 1 E[(x'beta_hat - x'beta)^2] = Theta(sigma^2 d / n) [ridge, random design]
Lemma 11 R(T) <= tau_{m0-1} + 206 K sum_t 1/gamma_m(t) + Azuma
Eq. E.30 206 K sum_t 1/gamma_m(t) = Theta(sigma sqrt(K d T))
Thm 1 N(eps) ~ K^3 d / (phi0 (tau+eps)^2); L(eps) = 1 + (1-eps)/(tau rho + eps)
"""
import json, sys, os
import numpy as np
os.chdir(os.path.abspath(os.path.join(os.path.dirname(__file__), "..")))
sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
from rcb.core import ipgs, EF_ridge, N_eps, L_eps, m0_eps
RES = {}
FAIL = []
def check(name, cond, detail=""):
(RES.setdefault("checks", {}))[name] = {"pass": bool(cond), "detail": detail}
print(f"[{'PASS' if cond else 'FAIL'}] {name} {detail}")
if not cond:
FAIL.append(name)
def loglog_slope(xs, ys):
return float(np.polyfit(np.log(xs), np.log(ys), 1)[0])
# ---------------------------------------------------------------- Lemma 7
def lemma7():
"""sum_{a != b} p_t(a) (mu_hat(b) - mu_hat(a)) < (K-1)/gamma, for the IPGS kernel.
Proof step (E.12): each term is (1/gamma) * [gamma*Delta / (K + gamma*Delta)] < 1/gamma.
"""
rng = np.random.default_rng(0)
worst = -np.inf
for _ in range(200000):
K = int(rng.integers(2, 12))
gamma = float(10 ** rng.uniform(-1, 4))
mu = rng.normal(size=K)
p, b = ipgs(mu, gamma)
lhs = float(sum(p[a] * (mu[b] - mu[a]) for a in range(K) if a != b))
worst = max(worst, lhs * gamma / (K - 1))
if p[b] < 0: # kernel must stay a valid distribution
check("lemma7-valid-kernel", False, f"p(b)={p[b]}")
return
check("Lemma 7: sum_a p(a)(mu_b - mu_a) < (K-1)/gamma", worst < 1.0,
f"sup over 2e5 random (K,gamma,mu_hat) of LHS*gamma/(K-1) = {worst:.6f} < 1")
RES["lemma7_sup_ratio"] = worst
# ---------------------------------------------------------------- Corollary 1
def corollary1():
"""Excess prediction risk of ridge with random design.
The sharp identity is E_x[(x'beta_hat - x'beta)^2] = ||beta_hat - beta||^2_Sigma;
Lemma 1 + Lemma 2 bound its expectation by ~ sigma^2 Tr[(Sigma+lam)^{-1}Sigma]/n
+ lam||beta||^2, i.e. Theta(sigma^2 d / n). We measure the exponents in d, n, sigma.
"""
rng = np.random.default_rng(1)
def risk(n, d, sigma, reps=200, lam=1e-6):
out = []
Sigma = np.eye(d)
for _ in range(reps):
beta = rng.normal(size=d) / np.sqrt(d)
X = rng.normal(size=(n, d))
y = X @ beta + sigma * rng.normal(size=n)
bh = np.linalg.solve(X.T @ X + lam * np.eye(d), X.T @ y)
out.append(float((bh - beta) @ Sigma @ (bh - beta)))
return float(np.mean(out))
ns = np.array([200, 400, 800, 1600, 3200])
s_n = loglog_slope(ns, [risk(n, 10, 0.1) for n in ns])
check("Corollary 1: excess risk ~ n^-1", abs(s_n + 1.0) < 0.06,
f"fitted exponent in n = {s_n:.3f} (theory -1)")
ds = np.array([4, 8, 16, 32, 64])
s_d = loglog_slope(ds, [risk(4000, d, 0.1) for d in ds])
check("Corollary 1: excess risk ~ d^+1", abs(s_d - 1.0) < 0.06,
f"fitted exponent in d = {s_d:.3f} (theory +1)")
ss = np.array([0.02, 0.05, 0.1, 0.2, 0.4])
s_s = loglog_slope(ss, [risk(2000, 10, s) for s in ss])
check("Corollary 1: excess risk ~ sigma^+2", abs(s_s - 2.0) < 0.06,
f"fitted exponent in sigma = {s_s:.3f} (theory +2)")
RES["corollary1_exponents"] = dict(n=s_n, d=s_d, sigma=s_s)
# ---------------------------------------------------------------- Eq. E.29 / E.30
def regret_sum(T, K, d, sigma, phi0=1.0, N=1.0, c3=1.0, gamma_const=4.0, C0=206.0):
"""The learning-regret term of Lemma 11 evaluated exactly:
C0 * K * sum_{t = tau_{m0-1}+1}^{T} 1/gamma_{m(t)},
gamma_m = gamma_const * sqrt(K / E_F(2^{m-2})), E_F = c3 sigma^2 d/(phi0 n).
"""
m0 = m0_eps(N)
tot = 0.0
m = m0
while 2 ** (m - 1) < T:
lo, hi = max(2 ** (m - 1), 2 ** (m0 - 1)), min(2 ** m, T)
if hi > lo:
EF = EF_ridge(2 ** (m - 2), d, sigma, phi0, c3)
gamma = gamma_const * np.sqrt(K / EF)
tot += C0 * K * (hi - lo) / gamma
m += 1
return tot
def eq_e30():
"""Verify the closed form of Eq. (E.30): the learning term is Theta(sigma sqrt(KdT))
and obeys the paper's stated constant 151 sigma sqrt(KdT)."""
base = dict(K=5, d=5, sigma=0.05)
Ts = np.array([2 ** k for k in range(12, 22)])
sT = loglog_slope(Ts, [regret_sum(T=T, **base) for T in Ts])
check("Eq. E.30: learning regret ~ T^1/2", abs(sT - 0.5) < 0.02,
f"fitted exponent in T = {sT:.4f} (theory 1/2)")
Ks = np.array([2, 3, 5, 10, 20, 40])
sK = loglog_slope(Ks, [regret_sum(T=2 ** 20, K=K, d=5, sigma=0.05) for K in Ks])
check("Eq. E.30: learning regret ~ K^1/2", abs(sK - 0.5) < 0.02,
f"fitted exponent in K = {sK:.4f} (theory 1/2)")
ds = np.array([2, 3, 5, 10, 20, 40])
sd = loglog_slope(ds, [regret_sum(T=2 ** 20, K=5, d=d, sigma=0.05) for d in ds])
check("Eq. E.30: learning regret ~ d^1/2", abs(sd - 0.5) < 0.02,
f"fitted exponent in d = {sd:.4f} (theory 1/2)")
ss = np.array([0.01, 0.02, 0.05, 0.1, 0.2])
ss_ = loglog_slope(ss, [regret_sum(T=2 ** 20, K=5, d=5, sigma=s) for s in ss])
check("Eq. E.30: learning regret ~ sigma^1", abs(ss_ - 1.0) < 0.02,
f"fitted exponent in sigma = {ss_:.4f} (theory 1)")
# ---- the paper's explicit constant -----------------------------------
# E.30's last-but-one step replaces sum_{m=m0}^{log2 T} 2^{m/2} by
# int_{m0}^{log2 T} 2^{x/2} dx. For an INCREASING integrand the sum EXCEEDS the
# integral, so this direction is invalid: sum -> 2^{M/2}/(1 - 2^{-1/2}) = 3.414 sqrt(T)
# while the integral gives (2/ln 2) sqrt(T) = 2.885 sqrt(T). The paper's stated
# constant 151 is therefore understated by the factor 3.414/2.885 = 1.183.
M = 24
disc = sum(2 ** (m / 2) for m in range(6, M + 1))
integ = (2 ** (M / 2) - 2 ** 3) * 2 / np.log(2)
check("E.30: sum_m 2^{m/2} EXCEEDS int 2^{x/2} dx (paper bounds it the wrong way)",
disc > integ, f"discrete sum/integral = {disc/integ:.4f} > 1 "
f"(asymptotically 3.4142/2.8854 = 1.1833)")
ratios = []
for T in [2 ** k for k in range(12, 22)]:
for K in [3, 5, 10]:
for d in [2, 5, 10]:
r = regret_sum(T=T, K=K, d=d, sigma=0.05)
ratios.append(r / (0.05 * np.sqrt(K * d * T)))
mx = max(ratios)
check("E.30: paper's constant 151 is VIOLATED; correct constant is ~176",
mx > 151.0 and mx <= 177.0,
f"max over 90 (T,K,d) grid points of term/(sigma sqrt(KdT)) = {mx:.1f}; "
f"paper claims <=151, closed form 51.5/(1-2^-1/2) = {51.5/(1-2**-0.5):.1f}")
RES["e30"] = dict(T=sT, K=sK, d=sd, sigma=ss_, max_ratio=float(mx),
correct_constant=float(51.5 / (1 - 2 ** -0.5)))
# ---- E.29 line 2 typo --------------------------------------------------
# As printed, E.29 line 2 reads 52 sum_m sqrt(K E_F(...) (tau_m - tau_{m-1})).
# Lemma 11 actually yields (tau_m - tau_{m-1}) * sqrt(K E_F). The printed form is
# O(sqrt(Kd) log T); only the latter gives O(sqrt(KdT)). Distinguish by the T exponent.
def printed_sum(T, K=5, d=5, sigma=0.05):
s = 0.0
for m in range(6, int(np.log2(T)) + 1):
EF = EF_ridge(2 ** (m - 2), d, sigma, 1.0, 1.0)
s += 52 * np.sqrt(K * EF * 2 ** (m - 1))
return s
Ts = np.array([2 ** k for k in range(12, 25)])
sp = loglog_slope(Ts, [printed_sum(T) for T in Ts])
check("E.29 as printed is O(polylog T), not O(sqrt(T)) -- a typo, not the real bound",
sp < 0.15, f"fitted exponent in T of the printed expression = {sp:.4f} "
f"(log-like); the corrected expression gives {sT:.4f}")
# ---------------------------------------------------------------- Theorem 1
def theorem1():
"""N(eps) exponents (Claim 2) evaluated on the closed form of Eq. (7)."""
base = dict(sigma=0.05, eps=0.05, tau=0.01, phi0=1.0)
Ks = np.array([2, 3, 5, 10, 20, 40])
sK = loglog_slope(Ks, [N_eps(K=K, d=5, **base) for K in Ks])
check("Theorem 1: N(eps) ~ K^3", abs(sK - 3.0) < 1e-6,
f"fitted exponent in K = {sK:.6f} (paper: cubic in arms)")
# The paper advertises N(eps) as "linear in context (d)". The formula's d-factor is
# (sigma^2 d + 1), which is linear only once sigma^2 d >> 1. At the paper's own
# noise levels that crossover sits far beyond any d it uses.
ds = np.array([2, 5, 10, 20, 50, 100, 200])
sd = loglog_slope(ds, [N_eps(K=5, d=d, **base) for d in ds])
check("Theorem 1: N(eps) is NOT linear in d at the paper's own sigma=0.05",
sd < 0.2, f"fitted exponent in d over d in [2,200] = {sd:.4f}, i.e. essentially "
f"CONSTANT, because sigma^2 d < 1 until d > 1/sigma^2 = {1/0.05**2:.0f}. "
f"Every d used in the paper (2,5,10,70) lies in this flat regime.")
ds2 = np.array([1e5, 1e6, 1e7, 1e8])
sd2 = loglog_slope(ds2, [N_eps(K=5, d=d, **base) for d in ds2])
check("Theorem 1: N(eps) -> d^1 only asymptotically (sigma^2 d >> 1)",
abs(sd2 - 1.0) < 0.02,
f"fitted exponent in d over d in [1e5,1e8] = {sd2:.4f} (theory 1)")
es = np.array([0.005, 0.01, 0.02, 0.05, 0.1])
se = loglog_slope(es + 0.01, [N_eps(K=5, d=5, sigma=0.05, eps=e, tau=0.01, phi0=1.0)
for e in es])
check("Theorem 1: N(eps) ~ (tau+eps)^-2", abs(se + 2.0) < 1e-6,
f"fitted exponent in (tau+eps) = {se:.6f} (theory -2)")
ps = np.array([0.01, 0.1, 1.0, 10.0])
sp = loglog_slope(ps, [N_eps(K=5, d=5, sigma=0.05, eps=0.05, tau=0.01, phi0=p)
for p in ps])
check("Theorem 1: N(eps) ~ phi0^-1", abs(sp + 1.0) < 1e-6,
f"fitted exponent in phi0 = {sp:.6f} (theory -1)")
# L(eps) is decreasing in eps and >= 1
es = np.linspace(0.0, 1.0, 101)
Ls = np.array([L_eps(e, 0.01, 0.95) for e in es])
check("Theorem 1: L(eps) decreasing in eps and >= 1",
bool(np.all(np.diff(Ls) < 0) and np.all(Ls >= 1.0)),
f"L(0)={Ls[0]:.1f} -> L(1)={Ls[-1]:.1f}")
RES["theorem1"] = dict(K=sK, d_small=sd, d_large=sd2, eps=se, phi0=sp)
# ------------------------------- Theorem 2 statement vs its own proof
def theorem2_statement():
"""Theorem 2 states the price-of-incentives term as 'T_cold ~ m0(eps)'. But
m0(eps) = ceil(2 + log2 N(eps)) is LOGARITHMIC in N, whereas (a) the proof E.29
uses tau_{m0-1} = 2^{m0-1} ~ 2 N(eps), and (b) the body text says this cost
'scales as O(1/eps^2)'. Only the proof's reading is self-consistent."""
out = {}
for eps in [0.01, 0.02, 0.05, 0.1]:
N = N_eps(5, 5, 0.05, eps, 0.01, 1.0)
m0 = m0_eps(N)
out[eps] = dict(N=N, m0=m0, tau_m0_minus_1=2.0 ** (m0 - 1))
es = np.array(list(out))
s_m0 = loglog_slope(es + 0.01, [out[e]["m0"] for e in es])
s_tau = loglog_slope(es + 0.01, [out[e]["tau_m0_minus_1"] for e in es])
check("Theorem 2's stated T_cold ~ m0(eps) is NOT O(1/eps^2)", abs(s_m0) < 0.6,
f"exponent of m0(eps) in (tau+eps) = {s_m0:.3f}, not -2")
check("Theorem 2's proof term tau_{m0-1} IS O(1/eps^2)", abs(s_tau + 2.0) < 0.15,
f"exponent of tau_(m0-1) in (tau+eps) = {s_tau:.3f} (theory -2)")
RES["theorem2_statement"] = dict(m0_exponent=s_m0, tau_exponent=s_tau,
table={str(k): v for k, v in out.items()})
if __name__ == "__main__":
lemma7(); corollary1(); eq_e30(); theorem1(); theorem2_statement()
RES["n_fail"] = len(FAIL); RES["failed"] = FAIL
os.makedirs("outputs", exist_ok=True)
json.dump(RES, open("outputs/theory_checks.json", "w"), indent=1)
print(f"\n{len(RES['checks'])} checks, {len(FAIL)} failed")
sys.exit(1 if FAIL else 0)