Barrier method Newton complexity
DisprovedConvexOptimization.barrier_method_sqrt_m_complexity⚠️ DEPRECATED — this statement is false and has been disproved.
The replacement is
ConvexOptimization.barrier_method_sqrt_m_complexity_on, which states the same theorem with a feasibility-confined line search.Why it is false. Each centering run is required to be an
IsDampedNewtonSequenceOnrun on the strictly feasible set, so every iterate must satisfy ; butIsBacktrackingSteptests the Armijo inequality at regardless of feasibility. Off the strictly feasible set the barrier does not blow up:logBarrieris real-valued and Mathlib'sReal.logreturns at and on the negatives, so stays finite there. Here the defect needs no adversarial choice of data — the definition oflogBarriersupplies it.Counterexample. , , , , , . The barrier objective is on , self-concordant for every with , and the central path is . Since , the first centering run must start at and minimise ; there , so and is infeasible. The Armijo test passes at (where , using ) and at every with (because on ), so is forced and the iterate leaves the feasible set.
How the replacement fixes it.
IsBacktrackingStepOnaccepts the largest whose trial point is both strictly feasible and Armijo-acceptable — exactly B&V's convention off (§9.1, p. 457). A closed-sublevel-set hypothesis for is also stated explicitly, as the Dikin-ellipsoid argument behind (9.55) needs.
The barrier method with reaches duality gap in centering steps — the goal of this mission.
Consider minimizing a convex subject to with each convex, and let . Assume the self-concordance hypothesis of §11.5.1 — the centering objective is self-concordant on the strictly feasible set for every , with positive definite Hessian — and let denote the central point for parameter . Fix backtracking parameters , , an initial , a target gap and a centering accuracy , and run the barrier method with the schedule
Then each centering step, started at the previous central point, is completed by a damped Newton run to accuracy in a number of iterations bounded uniformly in ,
and after
outer steps the duality gap is at most : .
Multiplying the two bounds gives a total of Newton steps — the celebrated square-root complexity of interior-point methods. The mechanism is a balance: shrinking toward makes each centering step cost Newton iterations, because the per-centering objective gap is then , while the outer count grows only as . Notably the bound involves no condition number and no dimension — only the number of constraints — which is precisely what self-concordance buys.
Formalization Note The theorem is existential in the runs, matching the book's idealized analysis in which each outer step begins exactly at the previous central point; the central path is supplied as a hypothesis (xc t is a strictly feasible minimizer for each ) rather than constructed, and the gradient and Hessian fields are given as explicit -indexed families with HasGradientAt / HasFDerivAt hypotheses. The problem is formalized without equality constraints, as in Chapter 11's development. Source: B&V §11.5.3, pp. 590–591, eqs. (11.28)–(11.29).
import Mathlib import Definitions.Def_ConvexOptimization_selfConcordance import Definitions.Def_ConvexOptimization_IsBacktrackingStep import Definitions.Def_ConvexOptimization_logBarrier open scoped RealInnerProductSpace ENNReal open MeasureTheory
theorem ConvexOptimization.barrier_method_sqrt_m_complexity {n mI : ℕ} (hmI : 0 < mI)
(α β t0 ε εnt : ℝ)
(hα0 : 0 < α) (hα : α < 1 / 2) (hβ0 : 0 < β) (hβ1 : β < 1)
(ht0 : 0 < t0) (hε : 0 < ε) (hεnt0 : 0 < εnt) (hεnt : εnt < 1 / 4)
(f₀ : EuclideanSpace ℝ (Fin n) → ℝ) (hf₀ : ConvexOn ℝ Set.univ f₀)
(fc : Fin mI → EuclideanSpace ℝ (Fin n) → ℝ)
(hfc : ∀ i, ConvexOn ℝ Set.univ (fc i))
-- self-concordance assumption of §11.5.1: t·f₀ + φ is SC for every t ≥ 0
(hSC : ∀ t : ℝ, 0 ≤ t →
IsSelfConcordantOn {x | ∀ i, fc i x < 0}
(fun x => t * f₀ x + logBarrier fc x))
-- gradient and Hessian fields of x ↦ t·f₀(x) + φ(x), and nondegeneracy
(g : ℝ → EuclideanSpace ℝ (Fin n) → EuclideanSpace ℝ (Fin n))
(hg : ∀ t : ℝ, 0 < t → ∀ x, (∀ i, fc i x < 0) →
HasGradientAt (fun y => t * f₀ y + logBarrier fc y) (g t x) x)
(H : ℝ → EuclideanSpace ℝ (Fin n) →
EuclideanSpace ℝ (Fin n) →L[ℝ] EuclideanSpace ℝ (Fin n))
(hH : ∀ t : ℝ, 0 < t → ∀ x, (∀ i, fc i x < 0) →
HasFDerivAt (g t) (H t x) x)
(hHpd : ∀ t : ℝ, 0 < t → ∀ x, (∀ i, fc i x < 0) →
∀ v, v ≠ 0 → 0 < ⟪H t x v, v⟫)
-- the central path, as exact minimizers over the strictly feasible set
(xc : ℝ → EuclideanSpace ℝ (Fin n))
(hxc_str : ∀ t : ℝ, 0 < t → ∀ i, fc i (xc t) < 0)
(hxc_min : ∀ t : ℝ, 0 < t →
IsMinOn (fun x => t * f₀ x + logBarrier fc x) {x | ∀ i, fc i x < 0} (xc t)) :
-- outer schedule μ = 1 + 1/√m, tᵢ = μⁱ t0; Nout outer steps suffice
∃ (w : ℕ → ℕ → EuclideanSpace ℝ (Fin n)) (K : ℕ → ℕ),
(∀ i, w i 0 = xc ((1 + 1 / Real.sqrt mI) ^ i * t0)) ∧
(∀ i, IsDampedNewtonSequenceOn {x | ∀ i', fc i' x < 0}
(fun x => (1 + 1 / Real.sqrt mI) ^ (i + 1) * t0 * f₀ x + logBarrier fc x)
(g ((1 + 1 / Real.sqrt mI) ^ (i + 1) * t0))
(H ((1 + 1 / Real.sqrt mI) ^ (i + 1) * t0)) α β (w i)) ∧
(∀ i, ((1 + 1 / Real.sqrt mI) ^ (i + 1) * t0 * f₀ (w i (K i)) +
logBarrier fc (w i (K i))) -
((1 + 1 / Real.sqrt mI) ^ (i + 1) * t0 *
f₀ (xc ((1 + 1 / Real.sqrt mI) ^ (i + 1) * t0)) +
logBarrier fc (xc ((1 + 1 / Real.sqrt mI) ^ (i + 1) * t0))) ≤ εnt) ∧
(∀ i, (K i : ℝ) ≤
(20 - 8 * α) / (α * β * (1 - 2 * α) ^ 2) / 2 +
Real.logb 2 (Real.logb 2 (1 / εnt)) + 2) ∧
(mI : ℝ) / ((1 + 1 / Real.sqrt mI) ^
⌈Real.sqrt mI * Real.logb 2 (mI / (t0 * ε))⌉₊ * t0) ≤ ε := by
sorry
Read-back
What the Lean code literally says, in plain math · claude-fable-5
Fix naturals with , and reals with , , , , and . Hypotheses on the problem data: is convex on all of , and each of the constraint functions is convex on all of — note that, unlike the two central-path theorems, NO differentiability hypotheses on or the appear here. Writing for the log-barrier (total real logarithm: , for ; genuine on the strictly feasible set) and : (H1) for EVERY real (including ), the function is self-concordant on , meaning is convex with the convexity inequality for this function on , the function is three times continuously differentiable on (within-set sense), and for every and every its line restriction satisfies at (total possibly-junk derivatives; real power with for ); (H2) a two-parameter field such that for every and every , the function is differentiable at with gradient ; (H3) a field of continuous linear operators such that for every and , the map is Fréchet-differentiable at with derivative ; (H4) for every , , and : ; (H5) a map (defined for all reals, constrained only at ) such that for every : (strictly feasible) and for every — a central point is hypothesized to exist for every positive parameter; if this hypothesis is unsatisfiable and the theorem is vacuous. Conclusion (existential): writing (real square root; so and ), there exist a doubly-indexed family (outer stage , inner iterate ) and a map such that ALL FIVE of the following hold. (1) For every : — each stage's inner run starts at the EXACT central point for the pre-update parameter (exponent ); in particular , and stage starts at , not at the terminal iterate of stage . (2) For every : the sequence is a damped Newton sequence on for the stage objective (exponent ), with gradient field , operator field , and parameters ; unfolded: for every , , and there exist and with , for some natural , , ( or the candidate fails this inequality), and . (3) For every :
i.e. at the designated index the stage- objective value of the inner iterate is within of its value at the exact center for parameter . (4) For every , as reals:
a bound uniform in whose leading term is HALF of — no factor of an initial gap appears — and whose additive constant is (compare in the single-Newton-phase iteration bound); is total (here makes , so both logs are genuine and positive). (5)
where is the natural-number ceiling (smallest natural the real value, and if that value is , e.g. when , in which case the claim reduces to ); under the hypotheses so its is genuine (possibly negative if the ratio is ). Note that conjunct (5) mentions only — none of , , , — it is a standalone arithmetic inequality asserting that after multiplicative parameter updates the quantity is at most . The theorem does not assert any inequality of the form ; no suboptimality bound on the iterates, and no link between the stage count of conjunct (5) and the per-stage statements (1)–(4) (which hold for every ), is part of the statement. The whole conclusion is existential: SOME such family of runs and index map exists.
Confirmed by the mission captain (proposal self-audit).