Chernoff Approximations in Semigroup Theory

• Chernoff approximations are product-form schemes that recover target evolution operators by iterated application of simpler operator families in semigroup settings.
• They achieve systematic error control by matching local operator expansion with data regularity, influencing convergence rates from O(1/n) to higher orders.
• Applications span parabolic, transport, and nonlinear evolution equations with explicit constructions yielding quasi-Feynman and resolvent approximations.

Searching arXiv for recent and foundational papers on Chernoff approximations and related Chernoff formulations. Chernoff approximations are product-form approximation schemes that recover a target evolution operator from repeated application of a simpler one-step family, typically in the setting of $C_0$-semigroups and evolution equations. In the semigroup-theoretic sense, a Chernoff approximation has the form $(C(t/n))^n\to e^{tL}$ as $n\to\infty$, where $L$ is the generator of the semigroup and $C(t)$ is chosen to be Chernoff-tangent to $L$ (Dragunova et al., 2023). In the broader literature indexed here, the term “Chernoff” also appears in statistically distinct objects such as Chernoff bounds and Chernoff information; those topics share a common name but not the same mathematical role. The semigroup-based notion is the one that underlies approximation of solutions to heat, transport, parabolic, and diffusion equations, explicit Feynman- and quasi-Feynman-type formulas, resolvent approximation, subordinate semigroups, and nonlinear convex monotone semigroups (Dragunova et al., 2023).

1. Definition and operator-theoretic setting

In the semigroup setting, Chernoff approximations arise from the abstract Cauchy problem

$$\left\{ \begin{array}{ll} u'_t(t,x)=Lu(t,x) \ \mathrm{ for }\ t>0, x\in Q\ u(0,x)=u_0(x)\ \mathrm{ for } \ x\in Q \end{array} \right.$$

viewed in a Banach space $\mathcal F$, with solution

$u(t,x)=(e^{tL}u_0)(x)$

when $L$ generates a $(C(t/n))^n\to e^{tL}$0-semigroup $(C(t/n))^n\to e^{tL}$1 (Dragunova et al., 2023). The approximation strategy replaces the usually inaccessible semigroup $(C(t/n))^n\to e^{tL}$2 by a simpler operator family $(C(t/n))^n\to e^{tL}$3 satisfying

$(C(t/n))^n\to e^{tL}$4

(Dragunova et al., 2023).

The standard notion of Chernoff tangency is formulated by conditions on an operator-valued map $(C(t/n))^n\to e^{tL}$5: continuity of $(C(t/n))^n\to e^{tL}$6, identity at $(C(t/n))^n\to e^{tL}$7, existence of the derivative

$(C(t/n))^n\to e^{tL}$8

on a dense domain, and identification of the closure of $(C(t/n))^n\to e^{tL}$9 with the target generator $n\to\infty$0 (Dragunova et al., 2023). Under the semigroup-generation assumption and the growth bound

$n\to\infty$1

Chernoff’s theorem yields

$n\to\infty$2

for every $n\to\infty$3 and every $n\to\infty$4 (Dragunova et al., 2023).

A recurrent theme in later work is that first-order tangency,

$n\to\infty$5

guarantees convergence but does not determine a useful rate in full generality. This suggests a distinction between existence of a Chernoff product formula and quantitative approximation theory. A plausible implication is that the practical value of a Chernoff scheme depends less on abstract convergence than on the local expansion structure of the chosen family $n\to\infty$6 and the regularity class of the initial datum (Vedenin et al., 2019, Galkin et al., 2021).

2. Higher-order tangency and convergence rates

A central refinement is the notion of Chernoff tangency of order $n\to\infty$7, namely

$n\to\infty$8

This notion is explicitly introduced for convergence-rate questions in the study of the heat equation (Dragunova et al., 2023). The heuristic stated there is that, for sufficiently regular initial data and fixed $n\to\infty$9,

$L$0

where $L$1 (Dragunova et al., 2023).

A general rate theorem is developed for linear semigroups in terms of local Taylor-type consistency. If

$L$2

then one obtains

$L$3

This gives an abstract route from one-step local defect to global product error (Galkin et al., 2021). The same paper emphasizes that, absent additional structure, convergence can be arbitrarily fast, arbitrarily slow, or merely strong without operator-norm convergence (Galkin et al., 2021). Earlier foundational work had already framed this phenomenon through approximation subspaces and proved that Chernoff convergence can be arbitrarily slow in full generality (Vedenin et al., 2019).

This rate sensitivity is not only abstract. In the one-dimensional heat equation on $L$4, the first-order family

$L$5

and the second-order family

$L$6

exhibit the predicted distinction for smooth data (Dragunova et al., 2023). For $L$7, the fitted slopes are approximately $L$8 for $L$9 and $C(t)$0 for $C(t)$1, while for Hölder data the observed order drops substantially, and the advantage of higher-order tangency may weaken to a smaller constant rather than a higher asymptotic slope (Dragunova et al., 2023). A related model-study on the heat and transport equations likewise reports $C(t)$2, $C(t)$3, and $C(t)$4 behavior for three explicit heat Chernoff families on the smooth datum $C(t)$5, but only $C(t)$6 numerically for $C(t)$7 across all three schemes (Prudnikov, 2020).

This suggests that higher-order local consistency is not by itself sufficient to guarantee realized higher-order convergence on rough data. The numerical evidence in these model problems points to regularity of $C(t)$8 as a decisive mediator between formal tangency order and effective approximation order (Dragunova et al., 2023, Prudnikov, 2020).

3. Explicit constructions for parabolic and evolution equations

One of the principal motivations for Chernoff approximations is that explicit families $C(t)$9 can often be constructed even when the semigroup $L$0 cannot (Dragunova et al., 2023). This is especially prominent for second-order parabolic equations with variable coefficients.

A recent construction for

$L$1

on $L$2, with

$L$3

defines first the bounded-integral Chernoff family

$L$4

This family is Chernoff-tangent to $L$5 and yields the first-order estimate

$L$6

under $L$7 and $L$8 (Remizov et al., 25 Jun 2026).

The main result of that paper is a corrected family

$L$9

with explicit coefficients $$\left\{ \begin{array}{ll} u'_t(t,x)=Lu(t,x) \ \mathrm{ for }\ t>0, x\in Q\ u(0,x)=u_0(x)\ \mathrm{ for } \ x\in Q \end{array} \right.$$0, chosen so that

$$\left\{ \begin{array}{ll} u'_t(t,x)=Lu(t,x) \ \mathrm{ for }\ t>0, x\in Q\ u(0,x)=u_0(x)\ \mathrm{ for } \ x\in Q \end{array} \right.$$1

This leads to the quadratic product estimate

$$\left\{ \begin{array}{ll} u'_t(t,x)=Lu(t,x) \ \mathrm{ for }\ t>0, x\in Q\ u(0,x)=u_0(x)\ \mathrm{ for } \ x\in Q \end{array} \right.$$2

for sufficiently smooth coefficients and data (Remizov et al., 25 Jun 2026). The use of proper Riemann integrals over $$\left\{ \begin{array}{ll} u'_t(t,x)=Lu(t,x) \ \mathrm{ for }\ t>0, x\in Q\ u(0,x)=u_0(x)\ \mathrm{ for } \ x\in Q \end{array} \right.$$3 is emphasized there as practically convenient, and the resulting formulas are called quasi-Feynman formulas because $$\left\{ \begin{array}{ll} u'_t(t,x)=Lu(t,x) \ \mathrm{ for }\ t>0, x\in Q\ u(0,x)=u_0(x)\ \mathrm{ for } \ x\in Q \end{array} \right.$$4 expands into multiple bounded integrals over $$\left\{ \begin{array}{ll} u'_t(t,x)=Lu(t,x) \ \mathrm{ for }\ t>0, x\in Q\ u(0,x)=u_0(x)\ \mathrm{ for } \ x\in Q \end{array} \right.$$5 rather than a classical path integral (Remizov et al., 25 Jun 2026).

Shift-based constructions also remain important. For the operator

$$\left\{ \begin{array}{ll} u'_t(t,x)=Lu(t,x) \ \mathrm{ for }\ t>0, x\in Q\ u(0,x)=u_0(x)\ \mathrm{ for } \ x\in Q \end{array} \right.$$6

a Chernoff function on $$\left\{ \begin{array}{ll} u'_t(t,x)=Lu(t,x) \ \mathrm{ for }\ t>0, x\in Q\ u(0,x)=u_0(x)\ \mathrm{ for } \ x\in Q \end{array} \right.$$7 is

$$\left\{ \begin{array}{ll} u'_t(t,x)=Lu(t,x) \ \mathrm{ for }\ t>0, x\in Q\ u(0,x)=u_0(x)\ \mathrm{ for } \ x\in Q \end{array} \right.$$8

This family is used both for semigroup approximation and for resolvent approximation, with an explicit first-order semigroup error

$$\left\{ \begin{array}{ll} u'_t(t,x)=Lu(t,x) \ \mathrm{ for }\ t>0, x\in Q\ u(0,x)=u_0(x)\ \mathrm{ for } \ x\in Q \end{array} \right.$$9

(Remizov, 2023).

The same architecture appears in the model transport equation, where

$\mathcal F$0

gives

$\mathcal F$1

so that the convergence order is exactly tunable by the choice of $\mathcal F$2 for suitable data such as $\mathcal F$3 (Prudnikov, 2020).

4. Extensions: resolvents, subordination, manifolds, and non-uniform partitions

Chernoff approximations have been extended well beyond direct semigroup approximation. One such extension concerns resolvents. If $\mathcal F$4 generates a $\mathcal F$5-semigroup with

$\mathcal F$6

then Laplace transforms of Chernoff products approximate the resolvent: $\mathcal F$7 for $\mathcal F$8 under standard semigroup-growth hypotheses (Remizov, 2023). This yields explicit resolvent formulas and, consequently, representations of solutions of nonhomogeneous linear ODEs with variable coefficients (Remizov, 2023).

Another extension concerns subordination. Given a semigroup $\mathcal F$9 known only through a Chernoff-equivalent family $u(t,x)=(e^{tL}u_0)(x)$0, the subordinate semigroup

$u(t,x)=(e^{tL}u_0)(x)$1

can itself be approximated by explicit Chernoff families (Butko, 2015). When the transition probabilities of the subordinator are known, one constructs

$u(t,x)=(e^{tL}u_0)(x)$2

then

$u(t,x)=(e^{tL}u_0)(x)$3

and obtains

$u(t,x)=(e^{tL}u_0)(x)$4

(Butko, 2015). When only a bounded Lévy measure is available, the paper instead uses a bounded-generator linearization

$u(t,x)=(e^{tL}u_0)(x)$5

again producing Chernoff equivalence for the subordinate semigroup (Butko, 2015). These constructions lead to Feynman formulae for subordinate Feller processes and diffusions on Euclidean spaces, star graphs, and Riemannian manifolds (Butko, 2015).

Killed Feller processes require another modification. If $u(t,x)=(e^{tL}u_0)(x)$6 is a bounded domain and $u(t,x)=(e^{tL}u_0)(x)$7 is the semigroup of the process killed on exit, then a Chernoff family for $u(t,x)=(e^{tL}u_0)(x)$8 is obtained from an ambient family $u(t,x)=(e^{tL}u_0)(x)$9 by extension and boundary cutoffs: $L$0 Under compatibility assumptions on the extension operator and the core, one gets

$L$1

which then converts into explicit $L$2-fold Feynman formulae for killed diffusions and Lévy-type processes, and further into time-fractional and distributed-order time-fractional Fokker–Planck–Kolmogorov formulas after subordination by inverse subordinators (Butko, 2017).

On Riemannian manifolds of bounded geometry, Chernoff approximations can be built intrinsically from flow maps of vector fields. For

$L$3

the explicit manifold-valued shift family

$L$4

approximates the Feller semigroup generated by the closure of $L$5 on $L$6: $L$7 The same construction yields weak convergence of random walks on the manifold to the associated diffusion process (Mazzucchi et al., 2020).

The classical equal-step product formula can also be generalized to non-uniform partitions. If $L$8 is a contraction Chernoff family with pairwise commuting operators and the partition coefficients satisfy

$L$9

then

$(C(t/n))^n\to e^{tL}$00

uniformly on compact time intervals (Bernád et al., 2024). The paper stresses that this condition is stronger than the null-array condition $(C(t/n))^n\to e^{tL}$01, and that commutativity and contractivity are the price paid for avoiding Smolyanov’s generator-orbit consistency condition (Bernád et al., 2024).

5. Nonlinear Chernoff-type approximations

A substantial recent development is the extension of Chernoff approximation theory to convex monotone semigroups. In that framework, the target semigroup $(C(t/n))^n\to e^{tL}$02 is nonlinear, acting on weighted spaces $(C(t/n))^n\to e^{tL}$03, and the approximation takes the form

$(C(t/n))^n\to e^{tL}$04

or, more generally, $(C(t/n))^n\to e^{tL}$05 for piecewise-constant partitions (Blessing et al., 2023). The one-step operators $(C(t/n))^n\to e^{tL}$06 are assumed convex, monotone, exponentially stable in the weighted norm, and compatible with translation estimates and smooth test functions (Blessing et al., 2023).

Unlike the linear theory, the generator is handled through $(C(t/n))^n\to e^{tL}$07-generators and semigroup comparison principles. The main quantitative theorems provide one-sided bounds for the positive and negative parts of the error, expressed via explicit consistency functions $(C(t/n))^n\to e^{tL}$08. Under polynomial consistency assumptions these yield power rates

$(C(t/n))^n\to e^{tL}$09

and

$(C(t/n))^n\to e^{tL}$10

with

$(C(t/n))^n\to e^{tL}$11

and the analogous formula for $(C(t/n))^n\to e^{tL}$12 (Blessing et al., 2023).

This nonlinear approach is applied to Nisio semigroups, where

$(C(t/n))^n\to e^{tL}$13

leading to rates such as $(C(t/n))^n\to e^{tL}$14, $(C(t/n))^n\to e^{tL}$15, or $(C(t/n))^n\to e^{tL}$16 depending on the order of consistency and additional regularity assumptions (Blessing et al., 2023). It is also applied to nonlinear law of large numbers and central limit theorems under convex expectations, where Chernoff-type product formulas become limit approximations for nonlinear distributions (Blessing et al., 2023). This suggests that the “Chernoff approximation” concept is no longer confined to linear semigroups of PDE origin; it now includes nonlinear semigroup limits closely related to monotone approximation schemes for HJB-type equations, but obtained via a distinct semigroup/$(C(t/n))^n\to e^{tL}$17-generator route (Blessing et al., 2023).

6. Distinct uses of the name “Chernoff”

The same name appears in statistically unrelated constructions, and this distinction is essential for terminological precision.

One use is the KL-Chernoff confidence bound for means of bounded random variables. For i.i.d. $(C(t/n))^n\to e^{tL}$18 with mean $(C(t/n))^n\to e^{tL}$19 and sample average $(C(t/n))^n\to e^{tL}$20, the note on the Chernoff bound proves that the Bernoulli one-sided KL inversion

$(C(t/n))^n\to e^{tL}$21

remains valid without change for arbitrary $(C(t/n))^n\to e^{tL}$22-valued data, not just Bernoulli variables (Foong et al., 2022). This is a concentration-inequality statement, not a semigroup approximation result.

A second use is Chernoff information, defined by

$(C(t/n))^n\to e^{tL}$23

with

$(C(t/n))^n\to e^{tL}$24

Recent work interprets this through likelihood ratio exponential families, proves uniqueness of the optimizer $(C(t/n))^n\to e^{tL}$25, and derives exact or numerical formulas for Gaussian families (Nielsen, 2022). Earlier work on same-family exponential distributions rewrites fixed-$(C(t/n))^n\to e^{tL}$26 Chernoff divergence as a skew Jensen divergence of the log-normalizer and proposes geodesic bisection to approximate the optimizing exponent (Nielsen, 2011). Again, this has no direct relation to semigroup Chernoff product formulas, beyond the historical name.

A third use is inversion and approximation of Chernoff tail bounds for sums of Poisson trials. New rational approximations to the classical exponents,

$(C(t/n))^n\to e^{tL}$27

are proposed because they remain rigorous one-sided bounds while being exactly invertible via quadratics (Shiu, 2021). This belongs to tail-probability analysis rather than semigroup approximation.

A plausible implication is that “Chernoff approximation” in strict semigroup theory should be reserved for product-form approximation of $(C(t/n))^n\to e^{tL}$28-semigroups, whereas “Chernoff bound” and “Chernoff information” denote concentration and divergence concepts respectively. The literature here keeps those topics separate even when discussing them under a shared eponym (Foong et al., 2022, Nielsen, 2022, Shiu, 2021).

7. Interpretation, limitations, and open directions

Across the semigroup literature represented here, Chernoff approximations are presented less as a single method than as a design principle: construct a simple family with the correct short-time behavior and iterate it. Their effectiveness depends on three interacting factors.

The first is local consistency. First-order tangency ensures convergence, while higher-order matching of the semigroup expansion can improve the global rate from $(C(t/n))^n\to e^{tL}$29 to $(C(t/n))^n\to e^{tL}$30 or higher, at least on sufficiently regular vectors (Galkin et al., 2021, Remizov et al., 25 Jun 2026). The second is regularity of the data. Smooth initial conditions may realize the formal tangency order, whereas Hölder or merely Lipschitz data can collapse higher-order schemes back toward first-order behavior (Dragunova et al., 2023, Prudnikov, 2020). The third is the structural choice of approximating family. Shift averages, bounded-integral operators, geodesic or flow-based steps, subordinate averages, and nonlinear monotone envelopes each target different operator classes and applications (Mazzucchi et al., 2020, Butko, 2015, Blessing et al., 2023).

Several caveats recur. General convergence can be arbitrarily slow, so no universal rate follows from Chernoff’s theorem alone (Vedenin et al., 2019, Galkin et al., 2021). Strong convergence need not imply operator-norm convergence, although quasi-sectorial contraction theory provides operator-norm Chernoff estimates in special settings (Zagrebnov, 2022). Numerical studies often rely on short ranges such as $(C(t/n))^n\to e^{tL}$31, and some empirically fitted asymptotic slopes are explicitly described as suggestive rather than definitive (Dragunova et al., 2023). In variable-coefficient problems, exact benchmark solutions are typically unavailable, making rigorous or even empirical error assessment more difficult (Dragunova et al., 2023, Remizov et al., 25 Jun 2026).

Open directions are stated quite explicitly in the cited works. These include sharper theoretical convergence-rate results for rough initial data (Dragunova et al., 2023), further development of fast Chernoff schemes for variable-coefficient parabolic equations (Remizov et al., 25 Jun 2026), broader nonlinear convergence-rate theory beyond the presently treated convex monotone framework (Blessing et al., 2023), and continued extension of explicit product formulas to subordinate, killed, manifold-valued, and time-fractional evolutions (Butko, 2015, Butko, 2017, Mazzucchi et al., 2020).

In this sense, Chernoff approximations occupy a distinctive position in analysis: they are simultaneously an abstract product formula, a constructive approximation method for evolution semigroups, a source of Feynman- and quasi-Feynman-type representations, and a bridge between semigroup theory, stochastic processes, and numerical treatment of PDEs (Dragunova et al., 2023).