Interior-Point Methods for Convex Programming
Takeaway
Interior-point methods follow a central path defined by barrier functions to solve large LP/QP/SDP problems efficiently with polynomial-time guarantees.
The problem (before → after)
- Before: Simplex can be exponential in worst case; naive constraints handling is brittle.
- After: Barrier functions keep iterates strictly feasible and well-conditioned, converging rapidly.
Mental model first
Imagine sliding down a smooth canyon that avoids scraping the walls (constraints) by inflating a cushion that keeps you away from the edges while guiding you to the floor.
Just-in-time concepts
- Log barrier: φ(x) = −∑ log(−g_i(x)).
- Central path: minimize t f(x) + φ(x); increase t to tighten approximation.
- Newton steps on the barrier objective.
First-pass solution
Initialize strictly feasible x; set t; iterate Newton steps; update t (e.g., multiply by μ>1) until duality gap is small.
Iterative refinement
- Primal–dual methods improve robustness.
- Sparse linear algebra and preconditioning dominate runtime.
- Self-concordant barriers yield complexity bounds.
Principles, not prescriptions
- Maintain strict feasibility for stability.
- Use problem structure (sparsity, cones) to scale.
Common pitfalls
- Poor line search and step sizes stall progress.
- Ignoring numerical conditioning in KKT solves.
Connections and contrasts
- See also: [/blog/convex-optimization], [/blog/branch-and-bound], [/blog/benders-decomposition].
Quick checks
- Why barriers? — Turn constrained problems into smooth unconstrained ones.
- What is the duality gap? — Certificate of suboptimality; target for stopping.
- Why primal–dual? — Better conditioning and step acceptance.
Further reading
- Nesterov & Nemirovski; Wright (book above)