Supremum Strategy: Extremal Methods

• Supremum Strategy is a unified framework that transforms complex pricing, stopping, and order reduction problems into a tractable supremum formulation.
• It is applied across robust American option pricing, optimal stopping rules, and convex envelope constructions using duality or randomized models.
• The approach replaces intractable global objects with extremal representations that preserve key calibration constraints and facilitate explicit solutions.

Searching arXiv for recent and directly relevant papers on “supremum” formulations and “supremum strategy” motifs. “Supremum strategy” is not a single canonical term in the cited literature. As an Editor’s term, it denotes a family of constructions in which a pricing problem, stopping problem, geometric ordering problem, or asymptotic analysis is reduced to an extremal object: a supremum over models, stopping times, supports, paths, or admissible envelopes. In robust American option pricing, the super-hedging price is represented as a supremum over randomized martingale models (Bayraktar et al., 2016); in quasi-sure finance, a measurable essential supremum is built from quasi-sure supports (Carassus, 2021); in local path analysis, zooming in at the time of the maximum yields canonical limit processes describing the neighborhood of the supremum (Thøstesen, 2021, Ivanovs, 2016). This suggests that the unifying content of a supremum strategy is methodological rather than terminological: a hard global object is replaced by an extremal representation that is stable under duality, stopping, conditioning, or order-theoretic reduction.

1. Robust pricing, randomized models, and quasi-sure envelopes

In the discrete-time market studied by Bayraktar, Huang, and Zhou, stocks are traded dynamically, European options $g=(g_1,\dots,g_e)$ are traded statically at time $0$ with zero price, and model uncertainty is represented by a family $\mathcal P$ of probability measures. For an American payoff stream $\Phi=(\Phi_t)_{t=0}^T$, the super-hedging price is

$$\pi(\Phi):=\inf\left\{x\in\mathbb{R}:\ \exists (\tilde H,h)\in\mathcal{H}'\times\mathbb{R}^e,\ \text{s.t. } x+(\tilde H\cdot S)_T+hg\ge \Phi_t,\ \mathcal{P}\text{-q.s. for all }t\right\}.$$

The central duality theorem states that

$$\pi(\Phi)=\sup_{(c_i,Q_i)_i}\sum_i c_i \sup_{\tau\in\mathcal T}E_{Q_i}[\Phi_\tau] =\sup_{(c_i,Q_i)_i}\sum_i c_i \phi^{Q_i},$$

where the supremum is over finite families $(c_i,Q_i)_i$ with $c_i\in\mathbb R_+$, $\sum_i c_i=1$, each $Q_i\in\mathcal M$, and the mixture $0$0, meaning that the aggregate model prices the statically traded European options correctly (Bayraktar et al., 2016). The paper’s main message is that the robust super-hedging price is therefore not, in general, the highest single-model American price. The relevant dual object is a randomized model.

The paper makes that point explicit in two ways. First, Nature may select a hidden index $0$1 at time $0$2, the hedger does not observe $0$3, and the model $0$4 is then used, while the aggregate law $0$5 remains consistent with the European options. Second, the proof rewrites the problem as a min-max formula,

$0$6

and then exchanges the infimum over static positions with a supremum over finite convex combinations of models by a non-compact minimax argument. The corollary

$0$7

states that randomized models are enough; all other enlarged “Nature models” are redundant for this representation (Bayraktar et al., 2016).

A related but more abstract supremum construction appears in the quasi-sure framework of Bouchard and Nutz. Given a non-dominated set-valued prior map $0$8 and a family of universally measurable functions $0$9, the quasi-sure essential supremum is defined as the unique usa function $\mathcal P$0 satisfying quasi-sure upper-bound and minimality properties, with the explicit representation

$\mathcal P$1

This is a support-based supremum, not a pointwise one. In the corresponding one-period super-hedging problem with trading costs $\mathcal P$2, the super-hedging cost is reduced to

$\mathcal P$3

so the quasi-sure essential supremum becomes the main envelope device converting model ambiguity into an optimization in $\mathcal P$4 (Carassus, 2021).

Within this financial lineage, a supremum strategy is therefore a duality mechanism. It replaces direct pathwise domination by a maximal valuation over models or supports, while preserving the calibration constraints coming from statically traded claims or quasi-sure measurability.

2. Optimal stopping and prediction relative to the ultimate supremum

In optimal prediction, supremum strategies appear as stopping rules driven by the gap between the current state and a running or ultimate maximum. For a spectrally positive stable Lévy process $\mathcal P$5 of index $\mathcal P$6, with running supremum $\mathcal P$7, the problem

$\mathcal P$8

is reduced to a stopping problem for the reflected process $\mathcal P$9. After the deterministic time change $\Phi=(\Phi_t)_{t=0}^T$0 and the transformed process $\Phi=(\Phi_t)_{t=0}^T$1, the value function solves a fractional free-boundary problem of Riemann–Liouville/Caputo type. When the stopping region is nonempty, the optimal rule is

$\Phi=(\Phi_t)_{t=0}^T$2

with $\Phi=(\Phi_t)_{t=0}^T$3 characterized by a transcendental equation. The paper also proves a breakdown phenomenon: there exists $\Phi=(\Phi_t)_{t=0}^T$4 and a strictly increasing curve $\Phi=(\Phi_t)_{t=0}^T$5 such that the boundary rule is optimal only for $\Phi=(\Phi_t)_{t=0}^T$6 and $\Phi=(\Phi_t)_{t=0}^T$7; outside that regime it is not optimal to stop at $\Phi=(\Phi_t)_{t=0}^T$8 when the drawdown is sufficiently large (Bernyk et al., 2010).

A different extremal criterion appears in the geometric Brownian motion model of du Toit and Peskir. With

$\Phi=(\Phi_t)_{t=0}^T$9

the “supremum problem” is

$$\pi(\Phi):=\inf\left\{x\in\mathbb{R}:\ \exists (\tilde H,h)\in\mathcal{H}'\times\mathbb{R}^e,\ \text{s.t. } x+(\tilde H\cdot S)_T+hg\ge \Phi_t,\ \mathcal{P}\text{-q.s. for all }t\right\}.$$0

After rewriting the problem as

$$\pi(\Phi):=\inf\left\{x\in\mathbb{R}:\ \exists (\tilde H,h)\in\mathcal{H}'\times\mathbb{R}^e,\ \text{s.t. } x+(\tilde H\cdot S)_T+hg\ge \Phi_t,\ \mathcal{P}\text{-q.s. for all }t\right\}.$$1

the optimal strategy is bang-bang: $$\pi(\Phi):=\inf\left\{x\in\mathbb{R}:\ \exists (\tilde H,h)\in\mathcal{H}'\times\mathbb{R}^e,\ \text{s.t. } x+(\tilde H\cdot S)_T+hg\ge \Phi_t,\ \mathcal{P}\text{-q.s. for all }t\right\}.$$2 while both $$\pi(\Phi):=\inf\left\{x\in\mathbb{R}:\ \exists (\tilde H,h)\in\mathcal{H}'\times\mathbb{R}^e,\ \text{s.t. } x+(\tilde H\cdot S)_T+hg\ge \Phi_t,\ \mathcal{P}\text{-q.s. for all }t\right\}.$$3 and $$\pi(\Phi):=\inf\left\{x\in\mathbb{R}:\ \exists (\tilde H,h)\in\mathcal{H}'\times\mathbb{R}^e,\ \text{s.t. } x+(\tilde H\cdot S)_T+hg\ge \Phi_t,\ \mathcal{P}\text{-q.s. for all }t\right\}.$$4 are optimal at $$\pi(\Phi):=\inf\left\{x\in\mathbb{R}:\ \exists (\tilde H,h)\in\mathcal{H}'\times\mathbb{R}^e,\ \text{s.t. } x+(\tilde H\cdot S)_T+hg\ge \Phi_t,\ \mathcal{P}\text{-q.s. for all }t\right\}.$$5 (Toit et al., 2009). The contrast with the paper’s infimum problem is sharp: the latter has a nontrivial free boundary for $$\pi(\Phi):=\inf\left\{x\in\mathbb{R}:\ \exists (\tilde H,h)\in\mathcal{H}'\times\mathbb{R}^e,\ \text{s.t. } x+(\tilde H\cdot S)_T+hg\ge \Phi_t,\ \mathcal{P}\text{-q.s. for all }t\right\}.$$6, whereas the supremum formulation does not.

For spectrally negative Lévy processes drifting to $$\pi(\Phi):=\inf\left\{x\in\mathbb{R}:\ \exists (\tilde H,h)\in\mathcal{H}'\times\mathbb{R}^e,\ \text{s.t. } x+(\tilde H\cdot S)_T+hg\ge \Phi_t,\ \mathcal{P}\text{-q.s. for all }t\right\}.$$7, Ott studies the prediction problem of stopping at a given distance $$\pi(\Phi):=\inf\left\{x\in\mathbb{R}:\ \exists (\tilde H,h)\in\mathcal{H}'\times\mathbb{R}^e,\ \text{s.t. } x+(\tilde H\cdot S)_T+hg\ge \Phi_t,\ \mathcal{P}\text{-q.s. for all }t\right\}.$$8 from the ultimate supremum $$\pi(\Phi):=\inf\left\{x\in\mathbb{R}:\ \exists (\tilde H,h)\in\mathcal{H}'\times\mathbb{R}^e,\ \text{s.t. } x+(\tilde H\cdot S)_T+hg\ge \Phi_t,\ \mathcal{P}\text{-q.s. for all }t\right\}.$$9 under squared error $$\pi(\Phi)=\sup_{(c_i,Q_i)_i}\sum_i c_i \sup_{\tau\in\mathcal T}E_{Q_i}[\Phi_\tau] =\sup_{(c_i,Q_i)_i}\sum_i c_i \phi^{Q_i},$$0. The key state variable is the reflected gap

$$\pi(\Phi)=\sup_{(c_i,Q_i)_i}\sum_i c_i \sup_{\tau\in\mathcal T}E_{Q_i}[\Phi_\tau] =\sup_{(c_i,Q_i)_i}\sum_i c_i \phi^{Q_i},$$1

and the original objective is reduced to

$$\pi(\Phi)=\sup_{(c_i,Q_i)_i}\sum_i c_i \sup_{\tau\in\mathcal T}E_{Q_i}[\Phi_\tau] =\sup_{(c_i,Q_i)_i}\sum_i c_i \phi^{Q_i},$$2

There is a threshold dichotomy. If

$$\pi(\Phi)=\sup_{(c_i,Q_i)_i}\sum_i c_i \sup_{\tau\in\mathcal T}E_{Q_i}[\Phi_\tau] =\sup_{(c_i,Q_i)_i}\sum_i c_i \phi^{Q_i},$$3

then it is optimal to stop immediately. If

$$\pi(\Phi)=\sup_{(c_i,Q_i)_i}\sum_i c_i \sup_{\tau\in\mathcal T}E_{Q_i}[\Phi_\tau] =\sup_{(c_i,Q_i)_i}\sum_i c_i \phi^{Q_i},$$4

then there exists a unique $$\pi(\Phi)=\sup_{(c_i,Q_i)_i}\sum_i c_i \sup_{\tau\in\mathcal T}E_{Q_i}[\Phi_\tau] =\sup_{(c_i,Q_i)_i}\sum_i c_i \phi^{Q_i},$$5 such that the optimal stopping time is

$$\pi(\Phi)=\sup_{(c_i,Q_i)_i}\sum_i c_i \sup_{\tau\in\mathcal T}E_{Q_i}[\Phi_\tau] =\sup_{(c_i,Q_i)_i}\sum_i c_i \phi^{Q_i},$$6

with value function expressed explicitly in terms of the zero-scale function $$\pi(\Phi)=\sup_{(c_i,Q_i)_i}\sum_i c_i \sup_{\tau\in\mathcal T}E_{Q_i}[\Phi_\tau] =\sup_{(c_i,Q_i)_i}\sum_i c_i \phi^{Q_i},$$7 and the integrals $$\pi(\Phi)=\sup_{(c_i,Q_i)_i}\sum_i c_i \sup_{\tau\in\mathcal T}E_{Q_i}[\Phi_\tau] =\sup_{(c_i,Q_i)_i}\sum_i c_i \phi^{Q_i},$$8 (Pinto et al., 2019).

Across these stopping problems, the supremum is not merely a terminal statistic. It becomes a state variable, typically through a reflected gap process, and the strategy is encoded by a boundary in that reduced state space. The cited papers also show that such boundaries need not always behave monotonically with the drawdown: heavy-tailed upward jumps can make continued waiting optimal even at large current gaps (Bernyk et al., 2010).

3. Local path geometry and distributional structure at the supremum

A separate line of work studies the path itself in a shrinking neighborhood of the time at which the supremum is attained. For the diffusion

$$\pi(\Phi)=\sup_{(c_i,Q_i)_i}\sum_i c_i \sup_{\tau\in\mathcal T}E_{Q_i}[\Phi_\tau] =\sup_{(c_i,Q_i)_i}\sum_i c_i \phi^{Q_i},$$9

with $(c_i,Q_i)_i$0 and

$(c_i,Q_i)_i$1

Buchmann, Maller, and Müller show that

$(c_i,Q_i)_i$2

where $(c_i,Q_i)_i$3 and $(c_i,Q_i)_i$4 are independent Bessel-3 processes. Equivalently, the two-sided local limit around the supremum is a process whose left and right branches are independent downward Bessel-3 pieces scaled by $(c_i,Q_i)_i$5. The proof uses the representation of the local martingale part of $(c_i,Q_i)_i$6 as a time-changed Brownian motion, and the same zoom-in analysis yields a discretization result for the grid estimator $(c_i,Q_i)_i$7: the error is of order $(c_i,Q_i)_i$8 and is controlled by the post-supremum Bessel-3 dynamics (Thøstesen, 2021).

Müller, Rivero, and others establish the analogous zoom-in principle for a Lévy process $(c_i,Q_i)_i$9 on a finite interval. If

$c_i\in\mathbb R_+$0

for a nontrivial self-similar Lévy attractor, then at the supremum $c_i\in\mathbb R_+$1 and its time $c_i\in\mathbb R_+$2,

$c_i\in\mathbb R_+$3

where the right-hand side is built from two independent conditioned processes: one conditioned to stay negative and one conditioned to stay positive. The same fluctuation-theoretic decomposition yields a limit theorem for the discretization error of the simulated supremum and its time,

$c_i\in\mathbb R_+$4

with $c_i\in\mathbb R_+$5 (Ivanovs, 2016).

The distributional law of the supremum itself can also be reconstructed by splitting the path at the last time the maximum is attained. For a Lévy process $c_i\in\mathbb R_+$6, with

$c_i\in\mathbb R_+$7

Chaumont gives a decomposition of

$c_i\in\mathbb R_+$8

in terms of excursion entrance laws $c_i\in\mathbb R_+$9 and $\sum_i c_i=1$0, together with correction atoms governed by the ladder-time drifts $\sum_i c_i=1$1. In the type 1 case, where $\sum_i c_i=1$2 is regular for both half-lines, $\sum_i c_i=1$3 is equivalent to the positive average occupation measure $\sum_i c_i=1$4, and absolute continuity of $\sum_i c_i=1$5 for some $\sum_i c_i=1$6 is equivalent to absolute continuity of the resolvent measure $\sum_i c_i=1$7 (Chaumont, 2010).

These papers recast the supremum from a scalar maximum into a local geometric event. The relevant “strategy” is then microscopic: center at the maximizer, identify the conditioned pre- and post-supremum pieces, and use them to derive limit theorems for path shape, discretization error, or the joint law of $\sum_i c_i=1$8.

4. Pointwise supremum, envelope formulas, and supremum metrics

In several non-probabilistic settings, a supremum strategy is an envelope construction. For a family of convex functions $\sum_i c_i=1$9, with

$Q_i\in\mathcal M$0

Correa, Hantoute, and López generalize Valadier’s formula for the subdifferential of the pointwise supremum. In separated locally convex spaces they prove

$Q_i\in\mathcal M$1

and, when $Q_i\in\mathcal M$2 is finite and continuous at some point, derive the simpler Valadier-type expression involving exact nearby subgradients through the enlargement $Q_i\in\mathcal M$3 (Correa et al., 2017). The nearby-point phenomenon is central: the subdifferential of a supremum need not be recoverable from $Q_i\in\mathcal M$4 alone.

A comparable envelope problem arises for copulas with a prescribed curvilinear section $Q_i\in\mathcal M$5 along an automorphism $Q_i\in\mathcal M$6. Sánchez and Trutschnig define, using total variation of $Q_i\in\mathcal M$7, the candidate upper bound

$Q_i\in\mathcal M$8

prove that every $Q_i\in\mathcal M$9 satisfies $0$00, and then show in Theorem 3.9 that

$0$01

They further characterize when this pointwise supremum is itself a copula by conditions (5)–(7) involving $0$02, $0$03, and the geometry of rectangles crossing the curve $0$04 (Ouyang et al., 2024).

In Bayesian nonparametrics, Walker and Hjort use the supremum metric

$0$05

as the target topology for posterior consistency. Their central device is the triangle inequality

$0$06

where $0$07 is the sinc-smoothed density. Weak posterior consistency, implied by the standard Kullback–Leibler support condition, controls the third term; regularity of $0$08 controls the second; Fourier-based smoothness classes control the first. The resulting sufficient conditions for $0$09-consistency are stated to be weaker than those currently used to secure $0$10 consistency (Ho et al., 2022).

A plausible implication of these disparate results is that “supremum strategy” often means replacing a difficult object by an extremal envelope with better calculus: a subdifferential formula, an explicit least upper bound, or a triangle decomposition that transfers weak convergence into uniform control.

5. Combinatorial, control-theoretic, and morphological formulations

In combinatorics, the term arises literally in the study of supremum sections. Given a finite set $0$11 and a $0$12-representation

$0$13

Scarf’s supremum section is

$0$14

Joret, Micek, and Trotignon prove that every such $0$15 is collapsible. Their explicit discrete Morse matching is built from the function

$0$16

which uses the last order $0$17 to select a canonical witness. The matching partitions $0$18 into

$0$19

with $0$20 yielding a complete acyclic matching; by Chari’s theorem, $0$21 is collapsible (Bauer et al., 2018).

In singular control, Ferrari and Moretto formulate the supremum case by the transformation $0$22. The controlled process

$0$23

is coupled with the running supremum

$0$24

The authors introduce new integral operators for rewards integrated against the control $0$25 and against the running supremum $0$26, consistent with the HJB structure of the two-dimensional singular control problem. The supremum case is obtained by symmetry from the infimum case, and the expected optimal policy has barrier/reflection form in the $0$27-state space (Ferrari et al., 29 Jan 2025).

In color mathematical morphology, Kreyenhagen, Burgeth, and Kleefeld use a two-stage supremum construction. Colors are represented as symmetric matrices ordered by the Loewner semi-order, and the neighborhood reference is the log-exp-supremum

$0$28

Because the Loewner order is only a semi-order, the LES reference color is followed by a distance-based pre-order and then a lexicographic cascade

$0$29

to force a total order and a unique selected neighborhood color. The stated objective is to identify the original color within the structuring element that most closely resembles a supremum while avoiding false colors (Kahra et al., 14 Mar 2025).

These examples show that a supremum strategy need not be probabilistic. It may instead be a constructive rule for selecting a canonical maximal witness under incomplete order, whether the object is a face in a simplicial complex, a reflected state in singular control, or a color in a structuring element.

6. Integrability, asymptotics, and extremal phenomena

Several papers study when a supremum is finite, how its law behaves, or how large it is in random systems. In stochastic volatility models, Guo, Larsson, and Ruf propose a direct method for proving

$0$30

Using Doob’s $0$31 inequality for strictly positive submartingales under the share measure $0$32, they derive

$0$33

for constants $0$34, where $0$35 is the variance process under the transformed dynamics. For rough Bergomi with $0$36, they conclude

$0$37

which matters because an integrable supremum implies existence of an optimal stopping time for any linearly bounded American payoff (Gerhold et al., 2024).

For the perturbed multiplicative random walk

$0$38

Buraczewski, Damek, Dyszewski, and Zienkiewicz prove, under the Cramér-type condition $0$39 and a regularly varying perturbation tail $0$40, that

$0$41

with a second-order refinement

$0$42

The proof is based on a new renewal theorem for slowly varying, non-monotone kernels (Damek et al., 2018).

In Gaussian fields, Ding, Eldan, and Zhai quantify the link between superconcentration and geometry of near-maximizers. For a centered field $0$43, with $0$44, $0$45, and $0$46, they show that with high probability one can find a set $0$47 of size at least

$0$48

such that all $0$49 satisfy $0$50 and the covariances obey $0$51 for $0$52. The theorem removes the non-negative correlation assumption from earlier work and upgrades a polynomial lower bound in $0$53 to an exponential bound in $0$54. The same paper also proves a sharpened moderate deviation bound for the supremum when $0$55 is of order $0$56 (Ding et al., 2013).

In the arithmetic theory of stable processes, Kuznetsov studies the density $0$57 of

$0$58

For almost all irrational $0$59, a double series representation is absolutely convergent, but there exists a dense uncountable set $0$60 of irrational $0$61 for which, outside a zero-Hausdorff-dimension exceptional set $0$62, the series does not converge absolutely. For rational $0$63, the Mellin transform

$0$64

is given explicitly in terms of Gamma functions and a dilogarithmic factor $0$65 (Kuznetsov, 2011). This makes the analytic behavior of the supremum density depend sensitively on Diophantine properties of the stability parameter.

For random cusp forms of weight $0$66, Chhaibi, Dang, and Steiner prove two distinct supremum scales. On a fixed compact $0$67,

$0$68

with exponential concentration around the median. Globally on the modular domain,

$0$69

and the larger scale is attributed to the cusp, with the maximum typically attained near $0$70 (Huang et al., 22 Aug 2025).

Taken together, these results show that a supremum strategy can be either qualitative or quantitative. It can establish finiteness of $0$71, derive exact tail asymptotics for a supremum functional, identify arithmetic obstructions to analytic expansions of supremum densities, or reveal the geometric multiplicity of near-maximizers in random landscapes.

7. Conceptual synthesis and recurrent misconceptions

The recurring structural pattern is that the supremum is rarely treated as a bare maximum. In the cited literature it is usually accompanied by an auxiliary reduction: randomized models in robust hedging, a reflected gap process in stopping, a support operator in quasi-sure finance, a free boundary in optimal prediction, a conditioned zoom-in limit near the maximizer, or an explicit envelope in convex or copula theory. This suggests that the decisive step in a supremum strategy is not maximization alone but the choice of the state space on which the maximization becomes tractable.

Several misconceptions are explicitly contradicted by the papers. In semi-static robust hedging, the correct dual value is not always the highest single-model price; the dual object is a randomization over models compatible with static-option prices (Bayraktar et al., 2016). In stable optimal prediction, “stop when the drawdown is large” is not universally optimal; the rule breaks down outside the regime determined by $0$72 and $0$73 (Bernyk et al., 2010). In color morphology, a supremum under a semi-order is not automatically unique; the Loewner order requires an LES reference plus lexicographic refinement to produce a total order (Kahra et al., 14 Mar 2025). In copula theory, the pointwise supremum over admissible copulas need not itself be a copula unless additional variation conditions are met (Ouyang et al., 2024).

A final unifying point is that many of these constructions are dual in nature. The supremum may appear over stopping times, over models, over quasi-sure supports, over subgradients at nearby points, or over admissible envelopes. Yet the function of that supremum is similar: it isolates the worst-case or extremal geometry while preserving the structural constraints of the original problem. Under this interpretation, “supremum strategy” designates a broad extremal methodology rather than a single theorem or domain-specific algorithm.