The Efficient Frontier

The efficient frontier is the geometric heart of MPT (Harry Markowitz): out of the infinitely many portfolios you can build from a set of assets, only a one-dimensional curve is worth considering. Every portfolio on it delivers the minimum possible variance for its expected return, everything below and to the right is dominated. Understanding the frontier means being able to derive it, not just draw it, because the derivation exposes the two-fund theorem and the tangency portfolio that the whole edifice of asset pricing is built on.

The optimization problem

Take $N$ risky assets with expected excess-return vector $\mu \in \mathbb{R}^N$ and covariance matrix $\Sigma \in \mathbb{R}^{N\times N}$ (assumed positive definite, so invertible). A fully-invested portfolio is a weight vector $w$ with $w^\top \mathbf{1} = 1$. Its expected return is $w^\top \mu$ and its variance is $w^\top \Sigma w$. The frontier is traced by solving, for each target return $\mu_p$,

$\min_w \; \tfrac12\, w^\top \Sigma w \quad \text{s.t.} \quad w^\top \mu = \mu_p, \;\; w^\top \mathbf{1} = 1.$

Deriving the frontier

Form the Lagrangian with multipliers $\lambda, \gamma$:

$\mathcal{L} = \tfrac12 w^\top \Sigma w - \lambda(w^\top \mu - \mu_p) - \gamma(w^\top \mathbf{1} - 1).$

The first-order condition $\partial\mathcal{L}/\partial w = \Sigma w - \lambda\mu - \gamma\mathbf{1} = 0$ gives the optimal weights as a linear function of the two multipliers:

$w = \lambda\,\Sigma^{-1}\mu + \gamma\,\Sigma^{-1}\mathbf{1}.$

This single equation is the whole story: every frontier portfolio is a linear combination of the two fixed vectors $\Sigma^{-1}\mu$ and $\Sigma^{-1}\mathbf{1}$. To pin down $\lambda, \gamma$, impose the two constraints. Define the three scalars

$A = \mathbf{1}^\top \Sigma^{-1}\mathbf{1}, \qquad B = \mathbf{1}^\top \Sigma^{-1}\mu, \qquad C = \mu^\top \Sigma^{-1}\mu,$

and $D = AC - B^2 > 0$ (positive by Cauchy–Schwarz whenever $\mu$ is not proportional to $\mathbf{1}$). Substituting $w$ into the constraints yields a $2\times 2$ linear system whose solution is $\lambda = (A\mu_p - B)/D$ and $\gamma = (C - B\mu_p)/D$. Plugging back, the minimum variance attainable at return $\mu_p$ is

$\boxed{\;\sigma_p^2 = \frac{A\,\mu_p^2 - 2B\,\mu_p + C}{D}.\;}$

This is a parabola in $(\mu_p, \sigma_p^2)$ space and a hyperbola in $(\sigma_p, \mu_p)$ space. The upper branch, where higher return comes with higher variance, is the efficient frontier; the lower branch is dominated.

The global minimum-variance portfolio

The apex of the hyperbola, the portfolio of least variance regardless of return, comes from setting $d\sigma_p^2/d\mu_p = 0$, giving $\mu_p = B/A$ and

$w_{\text{gmv}} = \frac{\Sigma^{-1}\mathbf{1}}{\mathbf{1}^\top \Sigma^{-1}\mathbf{1}}, \qquad \sigma^2_{\text{gmv}} = \frac{1}{A}.$

Notice it depends only on $\Sigma$, not on $\mu$, which is why the GMV portfolio is the practitioner's refuge when expected returns are too noisy to trust (see Pitfalls of Mean-Variance Optimization).

Two-fund theorem

Because $w = \lambda\Sigma^{-1}\mu + \gamma\Sigma^{-1}\mathbf{1}$ is affine in $\mu_p$, any two distinct frontier portfolios span the entire frontier: a convex combination of them is again a frontier portfolio. Hence all investors, whatever their risk appetite, can be served by just two mutual funds. This is the two-fund separation theorem, the origin of index-plus-bond investing.

Adding a risk-free asset: the tangency portfolio and CML

Introduce a risk-free rate $r_f$ (lend or borrow freely). Now an investor splits between the risk-free asset and one portfolio of risky assets. Maximizing the Sharpe ratio $\text{SR}(w) = (w^\top\mu - r_f)/\sqrt{w^\top\Sigma w}$ over the budget constraint gives the tangency portfolio

$w_{\text{tan}} = \frac{\Sigma^{-1}(\mu - r_f\mathbf{1})}{\mathbf{1}^\top \Sigma^{-1}(\mu - r_f\mathbf{1})}.$

(Note the family resemblance to multi-asset The Kelly Criterion sizing, $\Sigma^{-1}\mu$, same object, differently normalized.) Combining the risk-free asset with $w_{\text{tan}}$ traces a straight line in $(\sigma, \mu)$ space, the capital market line:

$\mathbb{E}[R_p] = r_f + \frac{\mathbb{E}[R_{\text{tan}}] - r_f}{\sigma_{\text{tan}}}\,\sigma_p.$

Its slope is the maximum attainable Sharpe ratio, and it is tangent to the risky-asset hyperbola at $w_{\text{tan}}$. Every efficient portfolio is now a mix of one risky fund and cash, the sharper form of two-fund separation, and the launchpad for the The Capital Asset Pricing Model (CAPM).

Worked example

Two assets: $\mu = (0.08,\ 0.12)^\top$, volatilities $0.15$ and $0.25$, correlation $\rho = 0.30$, so

$\Sigma = \begin{pmatrix} 0.0225 & 0.011250 \\ 0.011250 & 0.0625 \end{pmatrix}.$

The GMV weight in asset 1 is $w_1 = (\sigma_2^2 - \rho\sigma_1\sigma_2)/(\sigma_1^2 + \sigma_2^2 - 2\rho\sigma_1\sigma_2)$. Numerically $w_1 = (0.0625 - 0.01125)/(0.0225 + 0.0625 - 0.0225) = 0.05125/0.0625 = 0.82$, so $w_2 = 0.18$. That portfolio has variance $w^\top\Sigma w \approx 0.0189$, i.e. volatility $\approx 13.7\%$, below the 15% vol of asset 1 and far below asset 2's 25%. Diversification has produced a portfolio less risky than either constituent: the frontier bulges to the left of the individual assets precisely because $\rho < 1$. With $r_f = 0.03$, the tangency portfolio tilts toward the higher-Sharpe asset and defines the CML the investor levers up or down.

Failure modes

• Garbage inputs. The frontier is a deterministic function of $(\mu, \Sigma)$; estimation error in $\mu$ makes the efficient frontier itself a random object, and the "optimal" portfolio can be wildly wrong. This is Pitfalls of Mean-Variance Optimization in full.
• Ill-conditioned $\Sigma$. When assets number near the sample length, $\Sigma^{-1}$ amplifies noise; see Covariance Matrix Estimation and Ledoit-Wolf Covariance Shrinkage.
• No constraints. Unconstrained frontiers routinely demand enormous long-short positions; adding no-short or turnover constraints bends the frontier inward but stabilizes it.
• Non-Gaussian returns. Variance is only the right risk measure under elliptical returns or quadratic utility; fat tails argue for Expected Shortfall (CVaR) instead.

In interviews

Be ready to set up the Lagrangian and show that the optimal weights are $w = \lambda\Sigma^{-1}\mu + \gamma\Sigma^{-1}\mathbf{1}$, the step that instantly gives two-fund separation. State the GMV portfolio $\Sigma^{-1}\mathbf{1}/(\mathbf{1}^\top\Sigma^{-1}\mathbf{1})$ and note it ignores $\mu$. Explain why the frontier is a hyperbola and where the tangency portfolio sits. A classic follow-up: "why does adding a risk-free asset turn the curve into a line?", because you are now mixing a point (cash, zero variance) with a single fund, and mixing is linear in $(\sigma,\mu)$. See Sharpe Ratio and The Capital Asset Pricing Model (CAPM) for where this leads.