Gradient and Hessian of the barrier objective

With f_λ(x) = λcᵀx − Σᵢ₌₁ᵐ log(s(x)ᵢ) 33:39:

  • Gradient 34:33: ∇f_λ(x) = λc − Aᵀ S(x)⁻¹ 𝟙 where S(x) is the diagonal matrix holding the slack vector on its diagonal, and 𝟙 is the all-ones vector 34:33, 34:59.
  • Hessian 34:59: ∇²f_λ(x) = Aᵀ S(x)⁻² A
  • Crucially, the Hessian does not depend on c or on λ at all — only on the slacks 34:59, 43:37.

Centrality

  • Definition. The centrality of x for f_{λ_k} is 37:12:

    δ_k(x) = ‖∇f_{λ_k}(x)‖_{(∇²f_{λ_k}(x))⁻¹}

  • If x is the exact minimizer of f_λ, then δ_k(x) = 0, because the gradient vanishes — so centrality measures how close you are to the minimizer, in a funny norm 37:12, 38:07.

  • This is exactly the quantity Newton's progress theorem shrinks, which is why it is chosen as the IPM progress measure 35:56, 36:51.

Two thresholds

TermConditionTimestamp
Finely central for f_{λ_k}δ_k(x) ≤ 1/10038:35
Coarsely central for f_{λ_k}δ_k(x) ≤ 1/338:35

The IPM invariant

  • Start with an approximate optimum for λ₀, bump to a slightly larger λ₁, warm-start Newton's method there, run a few steps, then bump to λ₂, and so on 36:25, 36:51.
  • The loop invariant: hold a finely central point for λ_k; if λ_{k+1} is not too much bigger, that point is still coarsely central for λ_{k+1} 39:24.
  • Going from 1/3 back down to 1/100 costs only a constant number of Newton steps, since each step (or each constant number of steps) roughly halves the error while the Hessians behave 39:47, 40:05, 40:18.
  • The famous 1 + 1/√m growth rate for λ is precisely the largest increment that preserves coarse centrality 40:40, 40:55.
  • On quantifiers: ε need not be uniform across λ's. Newton's method is re-instantiated for each f_{λ_k} separately, so it suffices that ε be a fixed small constant within each instantiation 40:55, 41:21.
graph LR
    A["finely central for λ_k<br/>δ ≤ 1/100"] -->|"λ_{k+1} = (1+α)λ_k"| B["coarsely central for λ_{k+1}<br/>δ ≤ 1/3"]
    B -->|"O(1) Newton steps"| C["finely central for λ_{k+1}<br/>δ ≤ 1/100"]
    C -.->|repeat| A