An edge is a claim about curvature, not just direction.
A backtest reports a level. The second-order structure around that level is what says whether it is stable, whether the parameters were ever identified, and how the risk behaves when two things move at once.
Most work stops at first order. Stopping there is where the failures come from.
Why second order
The gradient tells you which way. The Hessian tells you the shape you are standing on.
The gradient of a function points downhill. It is first-order information, and it is what almost every method uses. It says nothing about how quickly the slope is changing, in which directions it changes fastest, or whether the point you have arrived at is a genuine minimum or just somewhere the ground happens to be flat.
The Hessian, the matrix of second derivatives, answers exactly those questions. It is the local shape. First-order information is what everyone uses; second-order information is what tells you how wrong you can be.
The same object turns up in four places this firm works, and they are the same mathematics wearing different clothes.
The gamma matrix
The diagonal is what every book measures. The off-diagonal is where multi-asset books break.
The diagonal is the risk most books already measure. The rest of the matrix is where the surprises live.
For a book written on several underlyings, the Hessian of value with respect to spot is the gamma matrix. Its diagonal is gamma per underlying, which every risk system reports. Its off-diagonal entries are cross-gamma: how the delta in one name moves when a different name moves.
A book can be flat on every diagonal and still carry large exposure to two things moving together. Hedging the diagonal and ignoring the rest is a common and expensive way to be wrong.
Curvature chooses the step
Newton's method reads the curvature. Gradient descent guesses, and pays for it in an elongated valley.
Where curvature is badly conditioned (a long, narrow valley, which is the normal shape of a calibration problem), first-order methods crawl along the flat direction and overshoot the steep one. The step size that is stable across the narrow axis is far too small for the long one.
Newton's method uses the Hessian to choose the step instead of guessing it, which is why it does not zigzag. On a quadratic it arrives in one step. Real objectives are not quadratic, but they are locally close enough that the same information is what makes the difference between a fit that converges and one that wanders.
Curvature is the uncertainty
A flat optimum means the parameter was never identified, however good the in-sample number looked.
The Hessian of the log-likelihood at its maximum is the observed Fisher information, and its inverse is the asymptotic covariance of the parameter estimates. Stated plainly: the curvature of the objective is the uncertainty on what was fitted.
This is the overfitting argument made exactly rather than rhetorically. Two calibrations can report the same point estimate and the same in-sample fit while one of them has genuinely identified a parameter and the other has not. Nothing in the reported level distinguishes them. The curvature does.
Minimum, or saddle
Eigenvalues of the same sign are a genuine optimum. Mixed signs are a ridge you happen to be standing on.
Classifying a critical point is a statement about the eigenvalues of the Hessian there. All positive is a minimum; all negative a maximum; mixed signs a saddle: a point that is uphill in one direction and downhill in another, and therefore not a place anything stays.
A parameter set that sits on a sharp ridge in backtest space is a saddle. It looked like an optimum because the search only ever moved along directions where it was one. It will not be there next year, and no amount of in-sample performance says otherwise.
What this commits to
An edge is a claim about curvature, not just about direction.
A firm without a track record is asking to be judged on something. This is the something: a stated method, precise enough to be argued with.
In practice it means the second-order structure gets computed and reported rather than assumed: the full risk matrix rather than its diagonal, the curvature of a fit alongside its point estimate, the conditioning of an optimisation rather than only its final objective value. None of that is exotic. It is standard mathematics that is routinely skipped because first order is cheaper and usually looks fine.
It is skipped right up until two things move at once.