Differential Calculus and Optimization Basics › Constrained Optimization and Lagrange Multipliers

Many financial optimization problems involve constraints. An unlevered, fully invested mandate may require portfolio weights to sum to 100%, while a leveraged or margin-financed portfolio can have gross exposure above 100% of capital. A risk mandate may cap modeled volatility, and a trading mandate may impose position limits. These rules lead to constrained optimization problems.

Constraints reflect the rules of a financial decision. Regulations or mandates may prohibit specified positions, risk limits may restrict concentration, and a budget constraint limits admissible spending or leverage. Constrained optimization finds the best objective value within the resulting feasible set. Sensitivity analysis can separately ask how that optimum would change if a constraint were relaxed slightly.

Lagrange multipliers turn a smooth, regular equality-constrained problem into a stationarity system for decision variables and auxiliary multipliers. This is not unrestricted minimization over both sets of variables. Consider minimizing $f(x)$ subject to $g(x)=0$.

Lagrange's approach augments the objective with a free multiplier rather than a fixed penalty. Differentiating the Lagrangian with respect to $\lambda$ recovers $g(x)=0$, while differentiation with respect to $x$ encodes the required gradient alignment. Solving these stationarity equations produces constrained candidates, not guaranteed optima; each candidate still needs feasibility, constraint-qualification, and second-order or global optimality checks.

For this regular equality surface, the objective gradient at a local minimum is a scalar multiple of the constraint gradient: it has no component along a feasible tangent direction. Otherwise one tangent orientation would decrease the objective locally. A zero objective gradient has no unique uphill direction but still satisfies this normal-space condition.

Imagine standing on a hill while constrained to walk along a smooth path. At a constrained minimum, the objective gradient has no tangential component. If a nonzero component remained, walking in its opposite direction would take you downhill, contradicting the minimum. When nonzero, the gradient is normal to the feasible path. This is the geometry captured by the Lagrange conditions.

Mathematically, this means:

$$\nabla f(x^*) = -\lambda \nabla g(x^*)$$

Together with $g(x^*)=0$, these equations form a system for candidate points and multipliers. Each candidate still needs an optimality check.

The system consists of $n+1$ equations in the $n$ components of $x$ and the multiplier $\lambda$. A nonsingular Jacobian at a solution can make that solution locally isolated, but matching the number of equations and unknowns does not guarantee existence or a unique global solution for a nonlinear system. Each candidate still requires constraint-qualification, feasibility, and second-order or global optimality checks.

Consider allocating between two assets to minimize portfolio variance, subject to achieving a target expected return.

This example studies the mean-variance trade-off within a specified asset universe and constraint set. For each feasible target expected return, the model selects the portfolio with the lowest variance. Along the upper efficient branch beyond the global-minimum-variance point, increasing the target return raises that minimum attainable variance. Outside this model and opportunity set, a portfolio can be dominated by another portfolio with both higher expected return and lower risk.

Let $w_1$ and $w_2$ be the weights in assets 1 and 2, with expected returns $\mu_1 = 0.10$ and $\mu_2 = 0.05$, variances $\sigma_1^2 = 0.04$ and $\sigma_2^2 = 0.01$, and correlation $\rho = 0.3$. The portfolio variance is:

$$\sigma_p^2 = w_1^2 \sigma_1^2 + w_2^2 \sigma_2^2 + 2 w_1 w_2 \rho \sigma_1 \sigma_2$$

where:

• $\sigma_p^2$: portfolio variance
• $w_1, w_2$: portfolio weights for assets 1 and 2
• $\sigma_1^2, \sigma_2^2$: variances of assets 1 and 2
• $\rho$: correlation between the two assets
• $2 w_1 w_2 \rho \sigma_1 \sigma_2$: covariance contribution from the interaction between assets

The first two terms scale each asset's variance by its squared weight. The cross term is weighted covariance. With same-sign nonzero weights and positive volatilities, lowering correlation reduces this term; opposite-sign weights reverse the direction, and a zero exposure gives no change. For nonnegative weights summing to one, portfolio variance cannot exceed the weighted average of component variances. This benchmark is not necessarily the variance of the lower-risk individual asset. The displayed positive-weight example exhibits a diversification benefit, while the long-short case needs its position-sign conditions.

We want to minimize $\sigma_p^2$ subject to:

1. Target return: $w_1 \mu_1 + w_2 \mu_2 = 0.08$
2. Full investment: $w_1 + w_2 = 1$

The Lagrangian is:

$$\mathcal{L} = w_1^2 \sigma_1^2 + w_2^2 \sigma_2^2 + 2 w_1 w_2 \rho \sigma_1 \sigma_2 + \lambda_1(0.08 - w_1 \mu_1 - w_2 \mu_2) + \lambda_2(1 - w_1 - w_2)$$

where:

• $\lambda_1$: Lagrange multiplier for the return constraint, representing the marginal variance cost of achieving additional return
• $\lambda_2$: multiplier for the net-budget constraint; its sign depends on the Lagrangian convention, and its magnitude is a local objective sensitivity per unit change in the constraint's right-hand side

The return equality fixes expected return. The equation $\sum_i w_i=1$ fixes net exposure. In the accounting convention where these asset weights and residual net cash exhaust the normalized budget, it leaves no residual net cash. It does not rule out leveraged long-short portfolios: weights $(2,-1)$ sum to one and have gross exposure three. Nonnegative fully invested weights or an explicit gross-exposure constraint are needed to exclude this leverage. With normalized weights, $\lambda_2$ has variance units per unit change in the budget right-hand side; it is not directly a value per dollar unless the problem is formulated in dollar holdings and currency units.

The full-investment and 8% target-return equalities determine a 60% allocation to the higher-return, higher-risk asset and a 40% allocation to the lower-return, lower-risk asset before variance enters the calculation. With two assets, this is the only feasible portfolio at that target, so it is also the minimum-variance portfolio for that target.

Under regularity conditions, a Lagrange multiplier has a local envelope interpretation: after fixing the constraint's right-hand-side parameterization and the Lagrangian sign convention, the corresponding signed derivative gives the first-order change in the optimal objective value.

For a binding constraint, the correctly signed envelope derivative describes only a local, first-order response around the current optimum. A large magnitude means the optimized objective is locally sensitive to that specified right-hand-side change; whether this is a benefit or a cost depends on whether the problem minimizes or maximizes, how relaxation is parameterized, and the multiplier sign convention. A small magnitude likewise indicates local insensitivity, not that the constraint is globally unimportant.

For our minimization problem and the convention $\lambda_1(R-w^T\mu)$, the return multiplier equals the local derivative of minimum variance with respect to decimal target return. Its sign would reverse under the opposite Lagrangian convention.

Suppose the return-constraint multiplier is 0.02 variance units per unit of decimal expected return. Raising the target from 8% to 8.1% changes decimal return by 0.001, so the first-order variance change is approximately $0.02 \times 0.001 = 0.00002$. The corresponding standard-deviation change is nonlinear and must be computed from the square root rather than assigned the same units.