Second-Order Cone Programming

A linear program can cap a portfolio's total capital or any single position's size, but it cannot naturally express "keep the portfolio's volatility under 10%" — volatility involves a square root of a sum of squares, which is not a straight-line relationship no matter how you slice it. Second-order cone programming (SOCP) is exactly the extension of linear programming built to handle constraints and objectives shaped like a length or distance — a Euclidean norm — while staying just as tractable to solve reliably and fast. It's the natural home for risk-constrained portfolio optimization, robust optimization under uncertainty, and any problem where "keep some vector's size under control" is a real constraint.

The analogy: capping how far you can wander, not just where

A linear constraint is like a set of straight fences: don't cross this line, don't cross that one. A second-order cone constraint is different — it's like being tied to a stake by a rope of fixed length: you can move in any direction, but your total distance from the stake (measured the ordinary way, as a straight-line distance, not city-block distance) can't exceed some limit. That "distance from a stake" shape — a norm — is precisely the geometry of an ice-cream cone in higher dimensions, which is where the name comes from: the constraint traces out a cone-shaped region.

The mechanics, one symbol at a time

A second-order cone constraint takes the form

$$\|A\mathbf{x} + \mathbf{b}\|_2 \le \mathbf{c}^\top \mathbf{x} + d,$$

where $\|\cdot\|_2$ denotes the ordinary Euclidean norm (length) of a vector. In words: the length of some linear combination of the decision variables must stay below some other linear combination of them — a "budget" for how big a certain vector quantity is allowed to get, expressed through a square root of a sum of squares rather than a plain straight-line bound.

A key example: a portfolio variance constraint $\mathbf{w}^\top \Sigma \mathbf{w} \le \sigma_{\max}^2$ (weights $\mathbf{w}$, covariance matrix $\Sigma$, volatility cap $\sigma_{\max}$) can be rewritten, using $\Sigma = L L^\top$ (a Cholesky factorization), as

$$\|L^\top \mathbf{w}\|_2 \le \sigma_{\max},$$

exactly a second-order cone constraint. In plain English: instead of capping variance directly (a quadratic, awkward form), you cap the length of a transformed weight vector — same constraint, geometrically cleaner, and solvable by the same efficient algorithms used for linear programs (interior-point methods), because SOCP retains the crucial convexity property that makes those solvers reliable.

Worked example 1: a volatility-capped two-asset portfolio

Two assets have volatilities $\sigma_1=20\%,\ \sigma_2=10\%$, uncorrelated ($\rho=0$), so $\Sigma = \begin{pmatrix}0.04 & 0\\0&0.01\end{pmatrix}$ and $L = \begin{pmatrix}0.2&0\\0&0.1\end{pmatrix}$. Cap portfolio volatility at $\sigma_{\max}=12\%$ with weights $w_1+w_2=1$. The SOCP constraint is $\|(0.2w_1, 0.1w_2)\|_2 \le 0.12$. Try $w_1=0.4,\ w_2=0.6$: $\|(0.08, 0.06)\|_2 = \sqrt{0.0064+0.0036} = \sqrt{0.01} = 0.10 \le 0.12$ — feasible, with headroom. Try $w_1=0.6,\ w_2=0.4$: $\|(0.12, 0.04)\|_2 = \sqrt{0.0144+0.0016}=\sqrt{0.016}\approx0.1265 > 0.12$ — infeasible, breaching the volatility cap.

Worked example 2: comparing to a naive linear cap

A tempting shortcut is capping each weight's contribution linearly, e.g. $0.2w_1 + 0.1w_2 \le 0.12$. At $w_1=0.6,w_2=0.4$: $0.2(0.6)+0.1(0.4)=0.12+0.04=0.16 > 0.12$ — this linear proxy also flags infeasibility here, but the two constraints don't agree in general. At $w_1=0.5,w_2=0.5$: the linear proxy gives $0.1+0.05=0.15>0.12$ (flagged infeasible), while the true SOCP volatility is $\|(0.1,0.05)\|_2=\sqrt{0.0125}\approx0.1118 \le 0.12$ — actually feasible. The linear shortcut is systematically too conservative because it ignores diversification (the square-root-of-sum-of-squares shrinks total risk below the sum of individual risks); only the genuine SOCP constraint captures that correctly.

A norm constraint like $\|x\|_2 \le r$ traces a circle (or, in the general linear-combination form above, a tilted ellipse) rather than a straight boundary — picture the boundary curving smoothly rather than kinking at corners, which is exactly what separates a cone constraint from a linear one.

What this means in practice

SOCP is the standard formulation for volatility-constrained and risk-parity-flavored portfolio optimization, robust optimization (where uncertainty in expected returns is itself modeled as a norm ball), and any problem needing a "keep this vector's size bounded" constraint alongside otherwise linear objectives and constraints. Because it remains convex, SOCP inherits the reliable global-optimality guarantees and fast interior-point solvers of linear programming, unlike general nonconvex quadratic problems.