Curvature in the Jump: Geometrizing Financial Lévy Models

TL;DR for operators

Jaehyung Choi’s paper does not offer a new trading strategy, volatility forecast, or backtest that makes the Sharpe ratio stand up and sing.¹ Its contribution is more structural: it builds an information-geometric framework for Lévy processes, the family of stochastic processes often used when financial returns refuse to behave like polite Gaussian increments.

The mechanism is simple in outline and technical in execution. A Lévy process is described by a triplet: diffusion, jump measure, and drift. Choi derives an $\alpha$-divergence between two such processes directly from their triplets, under mutual absolute-continuity conditions. From that divergence, the paper constructs the Fisher information matrix and the $\alpha$-connection. In plainer operational language: it turns jump-model families into geometric surfaces where distance, curvature, and local sensitivity can be computed.

For quant teams, the business value is model diagnosis, not market clairvoyance. The framework can support model comparison, calibration sensitivity analysis, bias-reduced estimation, and Bayesian predictive priors for Lévy-based models used in option pricing, risk measurement, portfolio modelling, or non-Gaussian return modelling.

The important boundary is that this is not a plug-and-play empirical result. The examples are mathematical demonstrations, not live-market validation. Some clean estimation benefits depend on e-flat geometry. Tempered stable, CTS/CGMY, and variance gamma cases behave nicely. The Merton jump-diffusion case is messier: its geometry is not e-flat, so the most convenient penalised likelihood route is unavailable. The map is useful, but it is still a map. One must still check whether the terrain has agreed to be mapped.

The business problem is not jumps. It is comparing jump models without superstition

Financial modelling already knows that asset returns jump. That is not the revelation. Black-Scholes-style diffusion is elegant, but markets have a vulgar habit of producing abrupt moves, asymmetric tails, and option smiles that make simple Gaussian assumptions look like a Victorian etiquette manual at a street fight.

Hence Lévy models. They let returns include discontinuities, heavy tails, asymmetric jumps, and infinite divisibility. The practical menu includes tempered stable processes, CGMY or classical tempered stable models, variance gamma processes, and Merton’s jump-diffusion model. These models are useful because they encode different stories about how jumps arrive, how large they are, and how quickly extreme tails are damped.

The harder question is not “can we add jumps?” It is:

How do we compare two jump models in a way that respects their stochastic structure?

A quant desk can calibrate several models to option prices. A risk team can fit jump models to historical returns. A portfolio group can run stress scenarios under alternative tail assumptions. But after that, the usual questions arrive, wearing different hats:

• Are these two calibrated models meaningfully different, or only numerically different?
• Which parameters are locally fragile?
• Which direction in parameter space changes the distribution most?
• Can estimation be regularised using the natural geometry of the model?
• Can Bayesian priors be built from the model structure rather than chosen because a spreadsheet needed a cell filled?

Choi’s paper attacks these questions by moving from probability distributions to differential geometry. That sounds abstract because it is. But the abstraction is doing actual work: it gives model space a metric.

The mechanism: triplet to divergence to geometry

A Lévy process is described by its Lévy triplet $(\sigma, \nu, \gamma)$:

• $\sigma$ captures the diffusion component;
• $\nu$ is the Lévy measure, describing jump intensity and jump-size structure;
• $\gamma$ is the drift term.

The triplet is not merely notation. It is the process’s operating manual. If two models have different Lévy measures, they have different jump behaviour. If their drift compensation differs, they may behave differently under a martingale or risk-neutral condition. If their diffusion coefficient differs, mutual absolute continuity may fail under the paper’s setup.

Choi’s framework begins from the Radon–Nikodym derivative between two Lévy processes, relying on known conditions under which two such processes are mutually absolutely continuous. That condition matters. The framework is not saying every Lévy process can be compared geometrically to every other Lévy process without constraint. The diffusion coefficients must match, the Lévy measures must be mutually absolutely continuous, and additional drift conditions apply when $\sigma = 0$.

Once that comparability condition is in place, the paper derives an $\alpha$-divergence. For $\alpha \neq \pm 1$, the result can be written compactly through an intermediate quantity $\Delta_T^{(\alpha)}(P|Q)$:

At $\alpha=-1$, the divergence recovers the Kullback–Leibler divergence. At $\alpha=1$, it gives the dual direction. At $\alpha=0$, it corresponds to the self-dual Hellinger case. So $\alpha$ is not a decorative Greek letter. It controls the geometry’s viewpoint: which model is being treated as the reference, how asymmetry is handled, and how local structure is extracted.

The next move is the important one. Once the divergence is available, information geometry tells us how to derive the local metric and connection. The metric is the Fisher information matrix. The connection describes how the geometry bends under a chosen $\alpha$-geometry.

For pure-jump cases with $\sigma=0$, the Fisher information matrix takes the clean form:

This is the operational heart of the paper. The model’s sensitivity is measured through derivatives of the log-density of the Lévy measure. In business terms: the jump structure itself determines which parameter directions are fragile, stable, redundant, or expensive to estimate.

For risk-neutral exponential Lévy models with $\sigma \neq 0$, an additional term appears from the martingale-compensated jump component. That is not a nuisance. It is precisely what makes option-pricing models different from generic return models: risk-neutral consistency changes the geometry.

What the paper actually proves, and what it merely demonstrates

The paper contains no empirical experiment section, no market dataset, no ablation table, and no backtest. Its evidence is mathematical. That is not a weakness, but it does change how one should read it.

This distinction matters because mathematical finance papers often get misread twice. First by people who see equations and assume the result is irrelevant. Then by people who see “financial models” and assume the result is a trading edge. Both camps are being dramatic.

The paper is useful because it gives a coherent structure for model comparison and estimation. It does not tell anyone what to buy on Monday.

Why divergence is the right first object

A model space cannot have useful geometry until it has a notion of difference. In ordinary Euclidean space, distance is obvious. In probability space, it is not. Two distributions can differ in tail behaviour, jump intensity, drift compensation, skewness, or local variance while looking similar under a coarse summary statistic.

The $\alpha$-divergence solves this by measuring dissimilarity between probability laws. For Lévy processes, the challenge is to express that dissimilarity using the components that define the process: the triplet. Choi’s main theoretical result is to do exactly that.

For a quant team, this is where the practical intuition begins. If two calibrated models produce similar vanilla option prices but have different Lévy measures, the divergence can reveal that the models are not close in jump space. Conversely, if two parameter vectors look numerically different but sit near each other under the information metric, the modelling difference may be operationally small.

That is the difference between comparing parameter values and comparing model behaviour. Finance has wasted plenty of time confusing the two. One more spreadsheet column is not a geometry.

Fisher information turns calibration into terrain

After deriving divergence, the paper constructs the Fisher information matrix. This is where the framework becomes useful for estimation.

The Fisher information matrix tells us how sharply the model distribution changes when parameters move. High curvature means small parameter changes can produce large distributional differences. Low curvature means the model is locally insensitive. If calibration is an optimisation problem, Fisher geometry describes the terrain under the optimiser’s feet.

In jump models, this matters because parameters often have uneven identifiability. Tail decay parameters, jump intensities, and asymmetry parameters may not be equally visible in market data. Option prices may strongly identify one combination and weakly identify another. Historical returns may reveal jump frequency but not tail damping. High-frequency returns may show discontinuities while still leaving risk-neutral jump pricing ambiguous.

The metric does not solve these problems by magic. It reveals where they live.

A useful operational reading is:

The paper’s strongest business implication is not that a particular Lévy model is better. It is that model selection can be made less folkloric. That is already progress. Folklore is expensive when it is levered.

The pure-jump cases behave cleanly

The examples begin with tempered stable processes. These are pure-jump Lévy processes, meaning $\sigma=0$, so the Fisher geometry depends directly on the Lévy measure. Choi shows that the general formula recovers earlier results on the information geometry of tempered stable processes. That is an important consistency check: the new framework does not bulldoze prior results; it absorbs them.

For generalized tempered stable processes, the Lévy measure has separate positive and negative jump components, with tail tempering controlled by $\lambda_+$ and $\lambda_-$. Under the paper’s admissibility conditions, the comparison varies these tempering parameters while other parameters are constrained. That restriction is not a footnote to politely ignore. It is part of the price of mutual absolute continuity.

The resulting Fisher matrix is diagonal in the positive and negative tempering coordinates:

The diagonal form is operationally pleasant. It says that, in this parameterisation, positive-tail and negative-tail tempering sensitivities decouple. A modeller can inspect how each side of the return distribution contributes to local information without pretending upward and downward jumps are the same creature wearing different shoes.

The CTS/CGMY case is a specialised version where both tails share a common scale and tail index structure. Its Fisher matrix has the same block-diagonal character:

The $\alpha$-connection components vanish for $\alpha=1$, making the geometry e-flat. That detail is not decorative. E-flatness is what allows the paper to connect the geometry to bias-reduced penalised likelihood methods.

The penalised likelihood has the form:

where $\mathcal{J}$ is the Jeffreys prior derived from the Fisher metric. For CTS processes, the paper obtains:

This is the cleanest part of the paper for applied teams: a model family used in finance receives a geometry-derived prior and a path toward bias-reduced estimation. No one had to invent a regulariser because it looked smooth in a notebook.

Variance gamma is where the boundary becomes visible

Variance gamma processes are also widely used in financial modelling, but they are not simply another tempered stable example. The paper treats VG as a limiting case, introducing a small positive parameter $a$ to regularise the Lévy measure and then taking $a \to 0$.

This is an implementation-relevant moment. The original VG Lévy measure creates a divergence issue when integrated over the real line. Choi’s solution is to modify the measure so that standard divergence calculations become valid, then recover the original structure in the limit.

That procedure produces a particularly clean Fisher matrix:

The corresponding Jeffreys prior is:

This result is elegant, and elegance is where applied teams should become slightly suspicious. The derivation depends on a regularisation-and-limit argument involving the Gamma function near zero. That does not invalidate the result. It does mean implementation should respect the limiting procedure rather than casually substituting $a=0$ into formulas where the original family does not permit it.

In operational terms: VG gets a usable geometry, but not for free. The derivation includes a bridge. Do not pretend the bridge is the ground.

Merton’s jump-diffusion enters the room and ruins the symmetry

The Merton model is different from the previous pure-jump examples because it includes both diffusion and compound Poisson jumps. Its triplet has $\sigma \neq 0$, and the Lévy measure uses normally distributed jump sizes with intensity $\lambda$, mean jump size $m$, and jump volatility $\delta$.

The paper derives the $\alpha$-divergence, Fisher information matrix, and $\alpha$-connection for risk-neutral Merton models with the same diffusion volatility $\sigma$. The formulas are more complicated, as one would expect when diffusion and jump compensation both enter the structure.

The key practical result is not the formula’s length. It is the geometric classification: the Merton model’s geometry is not e-flat, and neither are its submanifolds. Therefore, the convenient penalised likelihood method available in the e-flat pure-jump cases is not available here.

This does not make Merton useless. It makes Merton less geometrically cooperative.

That distinction matters. Many business readers want rankings: model A good, model B bad. The paper does not offer that. It says that certain model families admit cleaner information-geometric estimation machinery than others. A model may still be valuable because it captures a feature of the market, prices a product well, or aligns with desk intuition. But if its geometry is curved in the wrong way, some estimation shortcuts disappear. The model has not failed. It has merely declined to be convenient.

The practical value is cheaper diagnosis, not automatic alpha

The most realistic business use of this paper is as infrastructure for quantitative model governance.

A bank, hedge fund, market maker, or risk platform using Lévy models can use the framework to improve three workflows.

First, model comparison. Instead of comparing parameter vectors directly, teams can compare the process laws induced by those parameters. This matters when different parameterisations produce similar market prices but different tail risks.

Second, calibration diagnostics. The Fisher matrix identifies locally sensitive and insensitive directions. If a parameter direction has low information under the model and data setup, a calibration routine may produce a precise-looking but fragile number. Finance loves precise-looking fragile numbers. They make excellent audit findings later.

Third, estimation design. In e-flat cases, Jeffreys priors and penalised likelihood corrections offer a principled path to bias reduction. Bayesian predictive priors can also be built from superharmonic functions on the geometric manifold. This is not merely academic tidiness. In noisy calibration settings, structure-aware priors can reduce overfitting to transient market surfaces.

Here is the clean separation:

The paper’s business relevance is therefore not “use Lévy geometry and make money”. That would be adorable. The relevance is: if your organisation already uses jump models, this gives you a principled way to understand how those models differ and where their calibrations are unstable.

The hidden strength is unification

A major virtue of the paper is that it treats several financial Lévy models inside one framework.

Without this geometry, model families are often handled as separate species. CGMY gets one calibration routine, variance gamma gets another, Merton gets a third. Documentation then becomes a museum of special cases. Every model has its own notation, priors, numerical tricks, and risk caveats. Lovely, in the same way a drawer full of unlabelled adapters is lovely.

The paper’s framework provides a common route:

1. specify the Lévy triplet;
2. check absolute-continuity and martingale conditions where required;
3. derive the $\alpha$-divergence;
4. obtain the Fisher information matrix;
5. inspect the $\alpha$-connection;
6. identify whether e-flat estimation tools apply;
7. derive Jeffreys or predictive priors where the geometry supports them.

That sequence is the mechanism-first value. It turns the question from “which formula did this model happen to come with?” into “what geometry does this model induce?”

For model governance, that matters. A common geometry can make model inventory less arbitrary. It can support consistent documentation, consistent sensitivity analysis, and more disciplined comparisons between desks or product classes.

The assumptions are not paperwork

The most important limitations are structural, not rhetorical.

First, the paper assumes the Lévy processes satisfy mutual absolute-continuity conditions. In particular, the diffusion coefficient must match between the two processes under comparison. The Lévy measures must be mutually absolutely continuous, and the Radon–Nikodym density must satisfy integrability conditions. When $\sigma=0$, the drift terms are constrained by the difference in small-jump measures.

Second, the examples often restrict which parameters vary. In the GTS and CTS settings, admissibility conditions mean not every parameter can be freely compared across two processes. If a practitioner treats the formulas as full-space distances across all possible parameter changes, the mathematics will file a quiet complaint.

Third, the variance gamma case depends on regularisation. The paper introduces a positive parameter $a$, performs calculations in a well-defined family, and takes $a\to0$. That is a valid mathematical manoeuvre, but it is still a manoeuvre. Numerical implementations should check stability near the limit.

Fourth, the clean bias-reduction route depends on e-flatness. CTS/CGMY and VG geometries behave well in this respect. Merton does not. This is not a minor technicality. It decides whether a whole class of convenient estimation methods is available.

Finally, the paper does not provide empirical validation. There are no market data experiments showing improved option calibration, lower pricing error, better VaR performance, or stronger predictive accuracy. Those would be separate studies. This paper builds the instrument panel. It does not fly the aircraft.

How a quant team should read this paper

The best audience for this work is not a trader looking for a signal. It is a quant researcher, model validation team, or risk infrastructure group already working with non-Gaussian process models.

A reasonable internal follow-up would look like this:

The paper also suggests a useful research programme: combine this geometry with empirical calibration studies. For example, one could compare standard maximum likelihood, penalised likelihood using Jeffreys priors, and Bayesian predictive priors on historical option surfaces. The paper does not do that. But it makes the comparison mathematically coherent.

That is enough for one paper. We need not demand that every theorem also bring coffee.

Conclusion: curvature is a model-risk signal

Choi’s paper reframes Lévy financial models as geometric objects. The key contribution is not poetic; it is mechanical. Start from the Lévy triplet. Derive $\alpha$-divergence. From divergence, build Fisher information and $\alpha$-connections. Then use the geometry to reason about distance, sensitivity, bias correction, and predictive priors.

The strongest practical takeaway is that jump-model choice can be treated less like a catalogue decision and more like a geometric calibration problem. Different models induce different terrain. Some are flat enough to support clean estimation corrections. Some, like Merton’s jump-diffusion, are structurally more complicated. Some require careful limiting arguments before they become usable.

For financial institutions, the business case is not glamour. It is disciplined model infrastructure: better diagnostics, clearer assumptions, more principled priors, and fewer mysterious calibration artefacts pretending to be insight.

That is a modest promise, which is usually the kind worth taking seriously.

Cognaptus: Automate the Present, Incubate the Future.