An implied-volatility smile is a nice picture. That is the trap.

Traders quote a volatility for each strike because Black-Scholes gives a common language: take the observed option price, invert the Black-Scholes formula, and report the volatility that would have produced that price.1 If all the assumptions of the model were literally true, every strike at a fixed maturity would give the same number. Markets do not do that. Low strikes, high strikes, crash protection, supply and demand, jumps, stochastic volatility, and risk premia all leave shape in the cross-section.

That shape is the smile. It looks like the thing to interpret. It is not.

But the smile is not the primitive object. The primitive object is the option price as a function of strike:

\[C(K,T).\]

The smile is a reparameterization of that price curve. This matters because probabilities, state prices, and arbitrage constraints live in price space, not volatility space. A smooth-looking implied-volatility curve can still imply a bad call-price curve. A slightly noisy quote sheet can make the second derivative explode. A fitted surface can look elegant and still create a free butterfly.

The right mental move is less pretty and more useful:

do not ask what the smile looks like;
ask what density its call prices imply

Curvature Is a State Price

At a fixed maturity (T), the no-arbitrage price of a European call is the discounted risk-neutral expectation of its payoff:

\[C(K,T) = e^{-rT}\mathbb{E}^Q[(S_T-K)^+].\]

Writing the terminal density as (q_T(s)),

\[C(K,T) = e^{-rT}\int_K^\infty (s-K)q_T(s)\,ds.\]

Differentiate once with respect to strike:

\[\frac{\partial C}{\partial K} = -e^{-rT}Q(S_T>K).\]

Differentiate again:

\[\frac{\partial^2 C}{\partial K^2} = e^{-rT}q_T(K).\]

So:

\[q_T(K) = e^{rT}\frac{\partial^2 C}{\partial K^2}.\]

This is the Breeden-Litzenberger identity.2 It says that the curvature of the call-price curve is the terminal risk-neutral density.

There is an almost physical intuition behind it. A narrow butterfly spread:

\[C(K-h)-2C(K)+C(K+h)\]

pays off mostly when the terminal price lands near (K). If that butterfly has a negative price, then the market is paying you to buy a nonnegative payoff. That is static arbitrage. In the continuous limit, a strip of tiny butterflies becomes the second derivative of the call curve.

That gives three basic no-arbitrage tests for a call curve at one maturity:

price is bounded by intrinsic value and spot
price decreases as strike rises
price is convex in strike

The first two are easy to see. The third is the one the smile quietly tests. Convexity is not an aesthetic preference. It is nonnegative state-price density.

The Smile Lab

The lab below builds a synthetic option sheet. It starts with a Black-Scholes price at each strike, using a configurable implied-volatility smile. Then it can add deterministic quote roughness and optionally smooth prices before computing finite-difference curvature.

The purple bars are the Breeden-Litzenberger density estimate. Red bars are negative density. The butterfly panel shows the same information in tradeable units: each bar is the value of a one-step call butterfly. Negative bars are convexity violations.

I audited the lab before changing the prose. The default surface has total density mass 1.002, zero negative mass, a density mean of 101.87 against a 101.74 forward, and no monotonicity violations. Raising quote roughness to 40 keeps prices bounded but produces 25.6% negative density mass and a (-0.025) minimum butterfly. A 729-case grid over volatility, skew, curvature, maturity, roughness, and smoothing stayed finite; that audit also caught a synthetic quote clamp that could let far out-of-the-money calls dip below zero under roughness. The lab now clamps calls to the no-negative-price bound before checking curvature.

Implied volatility Call price curve Risk-neutral density Positive butterfly Violation

The density is estimated by central finite differences on call prices: \(q(K)\approx e^{rT}(C_{K-h}-2C_K+C_{K+h})/h^2\). The finite strike grid runs from 30 to 190 with spot fixed at 100, so extreme tail mass outside the grid is intentionally small but not modeled in detail.

Start with the default smooth surface. Probability mass should be close to one, the density mean should sit near the forward, and the minimum butterfly should be nonnegative. The smile is sloped and curved, but the price curve is still a valid collection of discounted payoffs.

Now raise Quote roughness. The implied volatility panel may still look like a market-ish smile. The density panel becomes fragile immediately. This is the curse of second derivatives: small quote wiggles become large curvature wiggles. If the red bars appear, the surface is no longer just noisy. It is asking for negative probability in some strike region.

Raise Smoothing passes. The red bars may shrink, but smoothing is not a license to average away economics. A production surface fit has to enforce arbitrage constraints, respect bid-ask spreads, and behave sensibly across maturity. A pretty spline is not enough.

Risk-Neutral Is Not Real-World

The density in the lab is risk-neutral.

That word carries a lot of weight. A risk-neutral density is the probability measure under which discounted traded prices are martingales. It is the measure that prices payoffs, not necessarily the measure that describes what investors believe will happen.

The NY Fed’s practical guide to extracting option-based risk-neutral densities states this warning plainly: option-implied risk-neutral probabilities can move because real-world probabilities moved, because risk preferences moved, or because both moved.3 A crash state can have a high risk-neutral price because investors think the crash is likely, because they hate paying off in that state, or because crash insurance is temporarily scarce.

This is why the density should not be read as:

the market's honest forecast

It is closer to:

the market's state-price-weighted forecast

That distinction is not philosophical decoration. If you use option-implied density to forecast returns, size tail hedges, or explain variance premia, you are mixing probability and marginal utility. Sometimes that is exactly what you want. Sometimes it is a category error.

Volatility Is a Coordinate System

Black-Scholes gives a one-to-one map from a European call price to an implied volatility, holding spot, strike, rate, and maturity fixed. That makes volatility convenient as a quote language. It does not make volatility the underlying state.

For example, a linearly interpolated implied-volatility curve can generate a call-price curve whose curvature oscillates. The volatility interpolation looks reasonable because volatility is smooth. But the no-arbitrage condition is not “volatility should be smooth.” It is “call price should be convex in strike.”

This is one reason volatility surface construction is a craft. The fitting procedure is doing several jobs at once:

  • denoise sparse and bid-ask-contaminated quotes;
  • interpolate liquid strikes without inventing arbitrage;
  • extrapolate tails without creating impossible digital prices;
  • align maturities so calendar spreads do not go negative;
  • produce stable Greeks for hedging and risk.

The single-maturity lab only checks vertical spread and butterfly conditions. A full surface has another dimension: maturity. Calendar no-arbitrage requires option prices to behave consistently as time to expiration changes. Dupire’s local-volatility formula goes further, using derivatives of the option surface across strike and maturity to infer an instantaneous volatility function.4 That makes the surface even more sensitive to bad derivatives. If the input surface is not arbitrage-clean, the local volatility can become negative, explosive, or meaningless.

The Derivative Amplifies the Lie

The dangerous part of the Breeden-Litzenberger identity is not the algebra. It is the numerical operation.

First derivatives amplify noise. Second derivatives amplify it harder. In the lab, quote roughness is only a small deterministic perturbation to the call sheet. The call-price panel may barely look disturbed, but the density can sprout alternating positive and negative spikes. This is exactly what happens when one tries to extract a distribution directly from stale, crossed, or illiquid option quotes.

The object one wants is smooth enough to differentiate. The object one observes is a market sheet with frictions:

discrete strikes
bid-ask spreads
stale quotes
early-exercise contamination
dividend assumptions
funding assumptions
microstructure noise

Every one of those can leak into the second derivative.

That is why an option-density workflow should look less like “differentiate the spreadsheet” and more like a constrained inference problem:

  1. Convert quotes to forward-consistent prices.
  2. Remove or downweight crossed, stale, and tiny-vega quotes.
  3. Fit a curve or surface under explicit no-arbitrage constraints.
  4. Differentiate the fitted price curve, not raw marks.
  5. Check the implied density mass, mean, tails, and butterfly prices.
  6. Compare results across smoothing assumptions.

The last step is not optional. If the density changes qualitatively when the smoother changes, the data are not telling a single story.

What the Smile Is Saying

A volatility smile is often introduced as evidence that Black-Scholes is wrong. That is true but incomplete. The more useful statement is:

the market prices different terminal states differently than a single lognormal
Black-Scholes distribution would

Low-strike equity index puts often carry high implied volatility because bad states are expensive. Out-of-the-money calls can become expensive in meme-stock, commodity squeeze, or event-driven regimes. Short maturities can kink around earnings, central-bank meetings, legal decisions, or liquidation events. The smile is a compact way of quoting all that state dependence.

But the smile only earns its interpretation after the price curve passes the static-arbitrage tests. Before that, it is a drawing with a nice accent.

The deeper lesson is almost embarrassingly geometric:

prices have shape;
shape has consequences

Slope is a digital price. Curvature is a state price. Negative curvature is a free butterfly. And the attractive volatility smile on the screen is only a safe object if the invisible price surface underneath is convex where it has to be convex.

  1. Fischer Black and Myron Scholes, “The Pricing of Options and Corporate Liabilities”, Journal of Political Economy 81(3), 1973. A public PDF copy is available from Simon Fraser University. ↩

  2. Douglas T. Breeden and Robert H. Litzenberger, “Prices of State-Contingent Claims Implicit in Option Prices”, Journal of Business 51(4), 1978. See also the SSRN record. ↩

  3. Allan M. Malz, “A Simple and Reliable Way to Compute Option-Based Risk-Neutral Distributions”, Federal Reserve Bank of New York Staff Report No. 677, 2014. ↩

  4. Bruno Dupire, “Pricing and Hedging with Smiles”, Paribas Capital Markets working paper, 1993; the related Risk article is commonly cited as “Pricing with a Smile,” Risk 7, 1994. ↩