Exponential Lifting: Theory and Applications

• Exponential lifting is the process of embedding a base model into a richer ambient structure, often incurring exponential increases in resource or complexity.
• Different fields interpret exponential lifting variously, from exponential memory blow-up in stochastic games to geometric path lifting via the exponential map in differential geometry.
• The technique extends to complexity theory, mixed-integer optimization, and category theory, highlighting challenges in transferring lower bounds and preserving structural properties.

Exponential lifting denotes a family of lifting constructions whose common feature is a passage from a base model to a richer ambient structure, with the qualifier “exponential” taking different meanings in different fields. In current arXiv usage, it can refer to an unavoidable exponential increase in strategy memory when lifting optimal strategies from MDPs to 2-player stochastic games, the exponential growth of cover degree needed to simplify short curves on punctured hyperbolic surfaces, the replacement of rough Volterra kernels by finite sums of exponentials in Markovian volatility models, path lifting through the exponential map in affine and semi-Riemannian geometry, the transfer of exponential large deviation estimates between product measures, or the lifting of the linear-logic exponential modality to Eilenberg–Moore categories (Dantam et al., 17 Jun 2026, Gaster, 2015, Jaber, 2018, Silva et al., 2021, Hurtado, 2023, Lemay, 2018). A common thread is that the lifted object is not merely embedded into a larger space; it acquires new structural constraints, and the word “exponential” may indicate resource blow-up, exponential kernels, exponential decay, or the geometric map $\exp$.

1. Terminological scope and usage landscape

Across the cited literature, the phrase does not designate a single standardized construction. Instead, it organizes several technical motifs that are only partially analogous. The following table summarizes the main usages represented in current arXiv work.

Domain	Meaning of “exponential lifting”	Representative paper
Stochastic games	lifting memoryless MDP strategies to 2-player stochastic games with an exponential increase in memory modes	(Dantam et al., 17 Jun 2026)
Hyperbolic surfaces	exponential lower bound on the cover degree needed for a simple lift of a curve of length at most $L$ on a punctured surface	(Gaster, 2015)
Rough volatility	replacing a rough kernel by a finite sum of exponentials, one Markovian factor per exponential term	(Jaber, 2018)
Affine and Lorentzian geometry	lifting paths through the exponential map, controlled by continuation properties	(Silva et al., 2021, Silva et al., 30 Aug 2025)
Anderson localization	lifting exponential large deviation estimates from stationary to non-stationary product measures	(Hurtado, 2023)
Mixed-integer formulations and proof complexity	lifting lower bounds from one model to a richer one, often with exponential consequences	(Cevallos et al., 2017, Beame et al., 2022)
Linear logic and category theory	lifting the exponential modality $!$ and MELL structure to an Eilenberg–Moore category	(Lemay, 2018)
Neural network activations	the cited paper does not define the term, but discusses a hypothetical extension from spline lifting to exponential-type bases	(Ochs et al., 2018)

This diversity matters because statements about “exponential lifting” are often domain-specific. In one strand the phrase concerns asymptotic lower bounds; in another it concerns a concrete representation by exponential kernels; in another it refers to the exponential map $\exp_p$ rather than exponential growth.

2. Strategy lifting in stochastic games

In turn-based 2-player zero-sum stochastic games, “exponential lifting” refers to lifting optimal strategies from MDPs to the 2-player setting. Results of Gimbert and Kelmendi, as reported in the later analysis of mean-payoff-parity and related objectives, show that for shift-invariant inverse-submixing objectives, memoryless optimality in MDPs can be lifted to finite-memory optimality in 2-player stochastic games, but with an exponential increase in the number of memory modes. The central contribution of the 2026 paper is that this exponential blow-up is tight: there exists a shift-invariant inverse-submixing objective for which optimal strategies in all MDPs are memoryless, yet in a family of 2-player stochastic games every optimal strategy for Max requires at least $2^{\Omega(n)}$ memory states, even when randomized strategies are allowed (Dantam et al., 17 Jun 2026).

The same paper also isolates a natural objective for which the worst-case exponential phenomenon collapses to a smaller bound. For mean-payoff-parity, Maximizer has optimal memoryless randomized strategies in MDPs, while optimal deterministic strategies require exponential memory. In 2-player stochastic games, however, optimal randomized strategies require, at least and at most, linear memory equal to the number of even colors. The paper further proves an impossibility result for a different lifting construction: there exists a shift-invariant objective where both players have optimal memoryless randomized strategies in all MDPs, but optimal randomized Max strategies still require infinite memory in deterministic 2-player games (Dantam et al., 17 Jun 2026). This sharply separates “finite-memory liftability” from “memoryless behavior in the 1-player case.”

3. Complexity-theoretic and polyhedral lifting

In communication complexity, proof complexity, and mixed-integer optimization, lifting usually means transferring lower bounds or hardness from a simpler base model to a richer one. In the mixed-integer setting, the 2017 framework on extended formulations starts from exponential or sub-exponential lower bounds for linear extended formulations and lifts them to the mixed-integer setting. Its main consequence is that any mixed-integer extended formulation of sub-exponential size for the matching polytope, cut polytope, traveling salesman polytope, or dominant of the odd-cut polytope needs $\Omega(n/\log n)$ many integer variables, while the same polyhedra admit polynomial-size mixed-integer formulations with only $O(n)$ or $O(n \log n)$ many integer variables. The technical engine is a decomposition technique that approximates any mixed-integer description of a convex set by the intersection of that set with the union of a small number of affine subspaces (Cevallos et al., 2017).

A related but distinct usage appears in query-to-communication lifting theorems. The 2022 study of Index and Inner-Product gadgets describes “exponential lifting” as the program of converting decision-tree lower bounds into exponential lower bounds in richer models such as communication protocols or proof systems. It proves that the conjecture of Lovett, Meka, Mertz, Pitassi and Zhang is false when the size of the Index gadget is $\log N-\omega(1)$, and that Inner-Product also loses the relevant disperser property at size $\log N-\omega(1)$. At the same time, it establishes a positive lifting theorem for a restricted class of communication protocols using Index gadgets of size at least $L$0, proves a lifting theorem from decision tree size to parity decision tree size, and derives a proof-complexity lifting theorem from tree-resolution size to tree-like $L$1 refutation size, yielding many new exponential lower bounds on such proofs (Beame et al., 2022). In this strand, exponential lifting is fundamentally about model transfer of lower bounds, not about exponential maps or exponential kernels.

4. Covering spaces and the exponential map in geometry

In low-dimensional and global differential geometry, the phrase has two prominent meanings. On punctured hyperbolic surfaces, Gaster’s study of simple lifts quantifies the worst-case degree needed to lift a closed curve to a simple curve in a finite cover. For a fixed hyperbolic surface $L$2, the function $L$3 is the minimum degree needed so that every curve of length at most $L$4 has a simple lift. In the closed case, $L$5 is asymptotically linear in $L$6. In the punctured case, however, Gaster proves that for any $L$7, there is $L$8 such that for all $L$9,

$!$0

and explains the mechanism through the explicit family $!$1, for which $!$2 while the hyperbolic length grows like $!$3 in a cusp (Gaster, 2015). Here “exponential lifting” means that the worst-case cover degree required to simplify curves of length at most $!$4 grows at least exponentially in $!$5.

A second geometric meaning concerns path lifting under the exponential map. The 2021 analysis of affine and Lorentzian manifolds imports Browder–Rheinboldt path-lifting and continuation theory to the exponential map. For local homeomorphisms, the path-lifting property is equivalent to the continuation property, and for smooth local diffeomorphisms with continuation for piecewise smooth paths, the map is a smooth covering map. Applied to $!$6, this framework yields generalized Hadamard–Cartan type results, existence and multiplicity theorems for geodesics, and causal geodesic results in Lorentzian geometry under pseudoconvexity and disprisonment hypotheses (Silva et al., 2021).

The 2025 work on exponential-type maps pushes this viewpoint past the singularities of $!$7. It proves that every smooth path, up to a nondecreasing reparametrization, can be partially lifted to a curve that is inextensible in the domain of definition of the exponential map. Under the path-continuation property for the exponential map, the paper proves a global path-lifting theorem and derives alternative proofs of the Hopf–Rinow theorem, Serre’s multiplicity theorem for geodesics, and the Avez–Seifert theorem for globally hyperbolic spacetimes (Silva et al., 30 Aug 2025). In this geometric strand, “exponential” refers literally to the exponential map, not to exponential asymptotics.

5. Markovian and probabilistic lifting in stochastic analysis

In mathematical finance, the 2018 lifted Heston construction uses “exponential lifting” to convert a non-Markovian rough volatility model into a finite-dimensional Markovian one. Starting from the rough Heston kernel

$!$8

the model replaces it by a finite sum of exponentials

$!$9

and introduces one variance factor per exponential term. The resulting lifted Heston model is Markovian for fixed $\exp_p$0, recovers classical Heston at $\exp_p$1, and converges to rough Heston as $\exp_p$2 under an explicit geometric grid in the Laplace variable. The paper emphasizes that this “exponential lifting” is precisely the replacement of a non-Markovian power-law kernel by a finite sum of exponentials, with one Markovian factor per exponential term (Jaber, 2018).

The same paper also supplies a quantitative numerical interpretation. For a benchmark rough Heston surface with $\exp_p$3, the lifted model with $\exp_p$4 has MSE $\exp_p$5 and evaluation time about $\exp_p$6 seconds, compared with about $\exp_p$7 seconds for the rough Heston surface computation. It further adopts $\exp_p$8 and $\exp_p$9 as a practical parametrization and reports that with this fixed exponential grid the lifted model consistently reproduces rough Heston volatility surfaces for maturities from $2^{\Omega(n)}$0 week to $2^{\Omega(n)}$1 years (Jaber, 2018).

A different probabilistic meaning appears in one-dimensional Anderson localization. The 2023 paper introduces a method to lift exponential large deviation estimates from a stationary product measure $2^{\Omega(n)}$2 to a non-stationary product measure $2^{\Omega(n)}$3. Under the condition

$2^{\Omega(n)}$4

if a family of adapted events satisfies $2^{\Omega(n)}$5 uniformly in $2^{\Omega(n)}$6, then $2^{\Omega(n)}$7 for some $2^{\Omega(n)}$8. Applied to transfer matrices, determinants, and Green’s functions, this yields Anderson localization for certain non-stationary one-dimensional lattice models: there is a full-measure set on which the operator has pure point spectrum and all eigenfunctions are exponentially decaying (Hurtado, 2023). Here the lifted object is not a strategy or a path but an exponential large deviation estimate.

6. Exponential modalities and lifted feature maps

In categorical logic, “exponential lifting” has a precise structural meaning tied to linear logic. A linear category is a symmetric monoidal closed category equipped with a monoidal coalgebra modality $2^{\Omega(n)}$9, which models the multiplicative and exponential fragment of intuitionistic linear logic. The 2018 paper studies when this exponential or “of course” modality can be lifted through a monad $\Omega(n/\log n)$0 to the Eilenberg–Moore category $\Omega(n/\log n)$1. It defines a MELL lifting monad as a symmetric Hopf monad endowed with a symmetric monoidal mixed distributive law $\Omega(n/\log n)$2, and proves that under these conditions the whole linear category structure lifts to $\Omega(n/\log n)$3. In particular, $\Omega(n/\log n)$4 again becomes a linear category with its own exponential $\Omega(n/\log n)$5, and the forgetful functor $\Omega(n/\log n)$6 strictly preserves the symmetric monoidal structure, the closed structure, and the monoidal coalgebra modality (Lemay, 2018).

That paper also identifies concrete sources of such liftings. Groups in the category of coalgebras of the monoidal coalgebra modality induce MELL lifting monads, and enrichment over abelian groups supplies a source of such groups. It additionally defines mixed distributive laws of symmetric comonoidal monads over symmetric monoidal comonads and shows how differential category structure can be lifted along the same pattern (Lemay, 2018). In this context, “exponential” refers to the linear-logic modality $\Omega(n/\log n)$7, not to an asymptotic growth rate.

A more speculative usage appears in the neural-network paper on lifting layers. That work introduces lifting as a nonlinear transfer function based on barycentric coordinates, showing that lifting followed by a linear layer yields a linear spline and that certain training problems become convex in the lifted parameters. The paper explicitly states that it does not define “exponential lifting,” but it does discuss how one could conceptualize such a variant by replacing the linear barycentric basis by exponential-type basis functions. It notes that convexity in the coefficients would remain as long as the feature map is fixed, while the simplex geometry used for the loss-lifting convexification would be lost (Ochs et al., 2018). This suggests a possible extension of lifting methodology, but not an established technical term in that paper.

7. Cross-domain distinctions and recurrent themes

A common misconception is that “exponential lifting” always denotes exponential complexity. The cited literature shows otherwise. In stochastic games and mixed-integer formulations, the exponential feature is indeed a lower bound or blow-up in memory, size, or integer variables (Dantam et al., 17 Jun 2026, Cevallos et al., 2017). In rough volatility, it is the use of exponential kernels $\Omega(n/\log n)$8 and exponentially spaced grids in a Markovian approximation (Jaber, 2018). In affine and semi-Riemannian geometry, it is the exponential map $\Omega(n/\log n)$9 and the possibility of lifting curves through it (Silva et al., 2021, Silva et al., 30 Aug 2025). In Anderson localization, it is exponential decay of probabilities and eigenfunctions (Hurtado, 2023). In category theory, it is the exponential modality of linear logic (Lemay, 2018).

Another recurrent theme is that lifting rarely preserves favorable structure for free. In games, memoryless optimality in MDPs can lift only at exponential memory cost, and some randomized constructions fail so badly that infinite memory becomes necessary (Dantam et al., 17 Jun 2026). In communication lifting, standard disperser-based methods fail for Index and Inner-Product gadgets below $O(n)$0 (Beame et al., 2022). In geometry, global path lifting through the exponential map requires a continuation property precisely because singularities and incomplete domains obstruct naive covering-space arguments (Silva et al., 30 Aug 2025). This suggests that the most robust interpretation of exponential lifting is not a single definition but a structural pattern: a lifting step exposes latent geometric, probabilistic, algebraic, or adversarial complexity that is invisible in the base model.