Bubeck Monograph Lab arXiv:1405.4980 v2.4 Core

Convex Optimization: Algorithms & Complexity Workbench

A visual, interactive oracle complexity laboratory based on Sébastien Bubeck's foundational text.

Loss Landscape & Trajectories Interactive Phase Space

Click / drag to relocate x₀
Click landscape to place initial point x₀
x: 0.00, y: 0.00 | f(x): 0.000
Step t = 0 / 50

Convergence Bounds & Decay log₁₀(f(x_t) - f*)

Theorem 3.7: Nesterov Accelerated Gradient Method
f(x_t) - f(x^*) \le \frac{2 \beta ||x_0 - x^*||^2}{(t + 1)^2} = \mathcal{O}\left(\frac{1}{t^2}\right)
Matches Nemirovski & Yudin (1983) minimax lower bound for the first-order black-box oracle model.
Method f(x_t) - f* ||∇f(x_t)|| ||x_t - x*|| Empirical Rate

Bubeck Oracle Complexity Matrix Foundations and Trends in ML Monograph

Minimax optimal iterations to achieve ε-suboptimality: f(x_T) - f* ≤ ε
Function Class & Assumptions Oracle Model Minimax Lower Bound Optimal Algorithm Bound Attaining Method
L-Lipschitz Convex ||f(x) - f(y)|| ≤ L||x - y|| First-Order Subgradient Oracle (f(x), g ∈ ∂f(x)) Ω( (L R / ε)² ) O( (L R / ε)² ) Subgradient Method (η_t = R / (L√t))
β-Smooth Convex ||∇f(x) - ∇f(y)|| ≤ β||x - y|| First-Order Gradient Oracle (f(x), ∇f(x)) Ω( √(β R² / ε) ) O( √(β R² / ε) ) Nesterov Accelerated Gradient (NAG)
α-Strongly Convex & β-Smooth Condition number κ = β / α First-Order Gradient Oracle (f(x), ∇f(x)) Ω( √κ · log(1/ε) ) O( √κ · log(1/ε) ) Nesterov AGD with Restart / Heavy-Ball
Constrained on Polytope X β-smooth over bounded domain diam(X)=D Linear Optimization Oracle (arg min_{s∈X} ⟨∇f(x), s⟩) Ω( (β D² / ε) ) O( (β D² / ε) ) Frank-Wolfe (Conditional Gradient)
Simplex Δ_n Domain (Non-Euclidean) Dimension-free log(n) geometry First-Order Oracle with Entropic Divergence Ω( (L_∞² log n) / ε² ) O( (L_∞² log n) / ε² ) Mirror Descent (Bregman KL Regularizer)
General Convex (Dimension-dependent) X ⊂ Rⁿ bounded convex body Zeroth / Subgradient Separation Oracle Ω( n · log(1/ε) ) O( n² · log(1/ε) ) Ellipsoid Method / Center of Gravity