The Fenton distribution: 60 years, and the answer was grade 2

In 1960, Lawrence Fenton published an approximation for the distribution of the sum of lognormal random variables. His moment-matching approach was elegant, and it worked well enough for engineering applications. Over the next 60 years, dozens of researchers improved on it: Schwartz-Yeh (1982), Beaulieu (2004), Mehta et al. (2007), and many others.

None of them found the exact answer. Not because the problem was too hard in any absolute sense, but because the right framework didn't exist.

Why lognormals don't add

If \(X \sim \text{LogNormal}(\mu_1, \sigma_1^2)\) and \(Y \sim \text{LogNormal}(\mu_2, \sigma_2^2)\), their product \(XY\) is also lognormal. That's easy.

But \(X + Y\) is not lognormal, and its distribution has no known closed form. The moment-generating function doesn't factor. The characteristic function is intractable. Every standard tool in probability theory hits a wall.

This matters in practice: portfolio VaR, credit risk aggregation, insurance loss totals, and wireless channel modeling all involve sums of correlated lognormals. The entire quantitative finance industry uses Monte Carlo simulation for this, running millions of samples for a number that should have a formula.

The Latent insight

Represent each lognormal as an element in a grade-2 tensor algebra:

\[X_i = \exp(\mu_i + \sigma_i Z_i) \mapsto \ell_i \in \mathcal{H}_0 \oplus \mathcal{H}_1 \oplus \mathcal{H}_2\]

The sum \(S = \sum X_i\) lives in the same algebra. The crucial observation: all grades above 2 vanish identically for lognormal sums. This isn't an approximation — it's an algebraic identity.

This means the distribution of \(S\) is completely determined by its grade-0, grade-1, and grade-2 components. Grade 0 is a constant. Grade 1 encodes the marginal structure. Grade 2 encodes all pairwise correlations.

The result: a closed-form expression for the CDF and PDF of \(S\), valid for any dimension, any correlation structure.

What Fenton-Wilkinson was actually doing

The Fenton-Wilkinson approximation matches the first two moments of \(S\) to a lognormal. In our framework, this is the grade-1 truncation — keeping \(\mathcal{H}_0 \oplus \mathcal{H}_1\) and discarding \(\mathcal{H}_2\).

That's why it works well when correlations are weak (\(\mathcal{H}_2\) is small) and fails badly when correlations are strong (\(\mathcal{H}_2\) dominates). The approximation error is precisely the grade-2 residual, and we can compute it exactly.

Broader lesson

The Fenton problem was "unsolvable" for 60 years because everyone worked within probability theory's standard toolkit: moment methods, saddlepoint approximations, series expansions. The algebraic framework gives you a different lens, and through that lens the problem is trivially grade 2.

I suspect many "unsolvable" problems in applied mathematics are similarly waiting for the right decomposition. The difficulty isn't the mathematics — it's seeing the structure.

The result is Lean-verified and has a Zenodo DOI. The code is reproducible.