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 𝟙whereS(x)is the diagonal matrix holding the slack vector on its diagonal, and𝟙is the all-ones vector34:33,34:59. - Hessian
34:59:∇²f_λ(x) = Aᵀ S(x)⁻² A - Crucially, the Hessian does not depend on
cor onλat all — only on the slacks34:59,43:37.
Centrality
-
Definition. The centrality of
xforf_{λ_k}is37:12:δ_k(x) = ‖∇f_{λ_k}(x)‖_{(∇²f_{λ_k}(x))⁻¹} -
If
xis the exact minimizer off_λ, thenδ_k(x) = 0, because the gradient vanishes — so centrality measures how close you are to the minimizer, in a funny norm37:12,38:07. -
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
| Term | Condition | Timestamp |
|---|---|---|
Finely central for f_{λ_k} | δ_k(x) ≤ 1/100 | 38:35 |
Coarsely central for f_{λ_k} | δ_k(x) ≤ 1/3 | 38:35 |
The IPM invariant
- Start with an approximate optimum for
λ₀, bump to a slightly largerλ₁, warm-start Newton's method there, run a few steps, then bump toλ₂, and so on36:25,36:51. - The loop invariant: hold a finely central point for
λ_k; ifλ_{k+1}is not too much bigger, that point is still coarsely central forλ_{k+1}39:24. - Going from
1/3back down to1/100costs only a constant number of Newton steps, since each step (or each constant number of steps) roughly halves the error while the Hessians behave39:47,40:05,40:18. - The famous
1 + 1/√mgrowth rate forλis precisely the largest increment that preserves coarse centrality40:40,40:55. - On quantifiers:
εneed not be uniform acrossλ's. Newton's method is re-instantiated for eachf_{λ_k}separately, so it suffices thatεbe a fixed small constant within each instantiation40:55,41:21.
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