---
title: Chernoff Approximations in Semigroup Theory
url: https://www.emergentmind.com/topics/chernoff-approximations
type: topic
---

# Chernoff Approximations in Semigroup Theory

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 \(L\) [2301.05284]. 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 [2301.05284].

## 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 \(C_0\)-semigroup \((e^{tL})_{t\ge 0}\) [2301.05284]. The approximation strategy replaces the usually inaccessible semigroup \(e^{tL}\) by a simpler operator family \(C(t)\) satisfying
\[
(C(t/n))^n f \to e^{tL}f,\qquad n\to\infty
\]
[2301.05284].

The standard notion of Chernoff tangency is formulated by conditions on an operator-valued map \(C:[0,\infty)\to\mathscr L(\mathcal F)\): continuity of \(t\mapsto C(t)f\), identity at \(t=0\), existence of the derivative
\[
C'(0)f=\lim_{t\to 0}\frac{C(t)f-f}{t}
\]
on a dense domain, and identification of the closure of \((C'(0),\mathcal D)\) with the target generator \((L,\operatorname{Dom}(L))\) [2301.05284]. Under the semigroup-generation assumption and the growth bound
\[
\|C(t)\|\le e^{\omega t},
\]
Chernoff’s theorem yields
\[
\lim_{n\to\infty}\sup_{t\in[0,T]}\left\|e^{tL}f - (C(t/n))^nf \right\|=0
\]
for every \(f\in\mathcal F\) and every \(T>0\) [2301.05284].

A recurrent theme in later work is that first-order tangency,
\[
C(t)=I+tL+o(t),
\]
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 \(C(t)\) and the regularity class of the initial datum [1910.09440, 2104.01249].

## 2. Higher-order tangency and convergence rates

A central refinement is the notion of Chernoff tangency of order \(k\), namely
\[
C(t)f=\left(I+tL+\frac{1}{2}t^2L^2+\dots+\frac{1}{k!}t^kL^k\right)f+o(t^{k})\textrm{ as }t\to 0.
\tag{CT3-k}
\]
This notion is explicitly introduced for convergence-rate questions in the study of the heat equation [2301.05284]. The heuristic stated there is that, for sufficiently regular initial data and fixed \(t\),
\[
\|u(t,\cdot)-u_n(t,\cdot)\|=O(1/n^k),\qquad n\to\infty,
\]
where \(u_n(t,x)=(C(t/n)^nu_0)(x)\) [2301.05284].

A general rate theorem is developed for linear semigroups in terms of local Taylor-type consistency. If
\[
\left\|
S(t)f-\sum_{k=0}^{m}\frac{t^kL^k f}{k!}
\right\|
\le
t^{m+1}\sum_{j=0}^{m+p} K_j(t)\,\|L^j f\|,
\tag{12}
\]
then one obtains
\[
\|S(t/n)^n f-e^{tL}f\| \le \frac{M_1M_2\, t^{m+1} e^{wt}}{n^m} \sum_{j=0}^{m+p} C_j(t/n)\,\|L^j f\|.
\tag{13}
\]
This gives an abstract route from one-step local defect to global product error [2104.01249]. The same paper emphasizes that, absent additional structure, convergence can be arbitrarily fast, arbitrarily slow, or merely strong without operator-norm convergence [2104.01249]. Earlier foundational work had already framed this phenomenon through approximation subspaces and proved that Chernoff convergence can be arbitrarily slow in full generality [1910.09440].

This rate sensitivity is not only abstract. In the one-dimensional heat equation on \(UC_b(\mathbb R)\), the first-order family
\[
(G(t)f)(x)=\frac{1}{2}f(x)+\frac{1}{4}f(x+2\sqrt{t})+\frac{1}{4}f(x-2\sqrt{t})
\]
and the second-order family
\[
(S(t)f)(x)=\frac{2}{3}f(x)+\frac{1}{6}f(x+\sqrt{6t})+\frac{1}{6}f(x-\sqrt{6t})
\]
exhibit the predicted distinction for smooth data [2301.05284]. For \(u_0(x)=\sin x\), the fitted slopes are approximately \(1\) for \(G\) and \(2\) for \(S\), 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 [2301.05284]. A related model-study on the heat and transport equations likewise reports \(O(n^{-1})\), \(O(n^{-2})\), and \(O(n^{-3})\) behavior for three explicit heat Chernoff families on the smooth datum \(u_0(x)=\sin x\), but only \(O(n^{-1})\) numerically for \(u_0(x)=e^{-|x|}\) across all three schemes [2012.09615].

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 \(u_0\) as a decisive mediator between formal tangency order and effective approximation order [2301.05284, 2012.09615].

## 3. Explicit constructions for parabolic and evolution equations

One of the principal motivations for Chernoff approximations is that explicit families \(C(t)\) can often be constructed even when the semigroup \(e^{tL}\) cannot [2301.05284]. This is especially prominent for second-order parabolic equations with variable coefficients.

A recent construction for
\[
u_t(t,x)=a(x)u_{xx}(t,x)+b(x)u_x(t,x)+c(x)u(t,x)=(Hu(t,\cdot))(x)
\tag{1}
\]
on \(\mathbb R\), with
\[
(H\varphi)(x)=a(x)\varphi''(x)+b(x)\varphi'(x)+c(x)\varphi(x),
\tag{9}
\]
defines first the bounded-integral Chernoff family
\[
(S(t)f)(x)=\frac12\int_{-1}^1 (1+c(x)t)\, f\bigl(x+\sqrt{6a(x)t}\,y+b(x)t\bigr)\,dy.
\tag{8}
\]
This family is Chernoff-tangent to \(H\) and yields the first-order estimate
\[
\|S(t/n)^n f-e^{tH}f\| \le e^{wt}\frac{t}{n}\sum_{j=0}^4 C_j\|f^{(j)}\|,
\qquad t>0,\ n\ge t
\]
under \(a,b,c\in HC_b^2(\mathbb R)\) and \(\inf a>0\) [2606.27232].

The main result of that paper is a corrected family
\[
(S(t)f)(x) = \frac12\int_{-1}^1 (1+c(x)t)\, f\bigl(x+\sqrt{6a(x)t}\,y+b(x)t\bigr)\,dy + \int_{-1}^1 \sum_{j=0}^6 \beta_j(t,x)y^j\, f(x+y t^{1/4})\,dy,
\tag{10}
\]
with explicit coefficients \(\beta_j(t,x)\), chosen so that
\[
S(t)f=f+tHf+\frac{t^2}{2}H^2f+o(t^2), \qquad f\in C_b^\infty(\mathbb R).
\]
This leads to the quadratic product estimate
\[
\|S(t/n)^n f-e^{tH}f\|
\le e^{wt}\frac{t^3}{n^2}\sum_{j=0}^8 C_j\|f^{(j)}\|
\]
for sufficiently smooth coefficients and data [2606.27232]. The use of proper Riemann integrals over \([-1,1]\) is emphasized there as practically convenient, and the resulting formulas are called quasi-Feynman formulas because \(S(t/n)^n\) expands into multiple bounded integrals over \([-1,1]^n\) rather than a classical path integral [2606.27232].

Shift-based constructions also remain important. For the operator
\[
(Hf)(x)=a(x)f''(x)+b(x)f'(x)+c(x)f(x),
\tag{3}
\]
a Chernoff function on \(UC_b(\mathbb R)\) is
\[
(S(t)f)(x) = \frac14 f(x+2\sqrt{a(x)t}) +\frac14 f(x-2\sqrt{a(x)t}) +\frac12 f(x+2b(x)t) +t\,c(x)f(x).
\tag{6}
\]
This family is used both for semigroup approximation and for resolvent approximation, with an explicit first-order semigroup error
\[
\|S(t/n)^n g-e^{tA}g\| \le \frac{t^2 e^{\|c\|t}}{n}
 \left( C_0\|g\|+C_1\|g'\|+C_2\|g''\|+C_3\|g'''\|+C_4\|g^{(IV)}\| \right)
\tag{9}
\]
[2301.06765].

The same architecture appears in the model transport equation, where
\[
(G(t)f)(x)=f(x+t+a t^{k+1})
\]
gives
\[
\left(\left(G\left(\frac{t}{n}\right)\right)^n u_0\right)(x) = u_0\left(x+t+\frac{a t^{k+1}}{n^k}\right),
\]
so that the convergence order is exactly tunable by the choice of \(k\) for suitable data such as \(u_0(x)=\sin x\) [2012.09615].

## 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 \(L\) generates a \(C_0\)-semigroup with
\[
R(\lambda,L)f=(\lambda I-L)^{-1}f=\int_0^\infty e^{-\lambda t}e^{tL}f\,dt,
\]
then Laplace transforms of Chernoff products approximate the resolvent:
\[
R(\lambda,L)f = \lim_{n\to\infty}\int_0^\infty e^{-\lambda t}S(t/n)^n f\,dt,
\]
for \(\operatorname{Re}\lambda>w\) under standard semigroup-growth hypotheses [2301.06765]. This yields explicit resolvent formulas and, consequently, representations of solutions of nonhomogeneous linear ODEs with variable coefficients [2301.06765].

Another extension concerns subordination. Given a semigroup \((T_t)\) known only through a Chernoff-equivalent family \((F(t))\), the subordinate semigroup
\[
T_t^f\varphi:=\int_0^\infty T_s\varphi\,\eta_t(ds)
\]
can itself be approximated by explicit Chernoff families [1512.05258]. When the transition probabilities of the subordinator are known, one constructs
\[
\mathcal F_0(t)\varphi = \int_{0+}^{\infty}\left[F\!\left(\frac{s}{m(t)}\right)\right]^{m(t)}\varphi\,\eta_t^0(ds),
\]
then
\[
\mathcal F(t) = e^{-\sigma t}\,F(\lambda t)\,\mathcal F_0(t),
\]
and obtains
\[
T_t^f=\lim_{n\to\infty}[\mathcal F(t/n)]^n
\]
[1512.05258]. When only a bounded Lévy measure is available, the paper instead uses a bounded-generator linearization
\[
F_\mu(t)\varphi := \varphi+t\int_{0+}^{\infty} \left( [F(s/m(t))]^{m(t)}\varphi-\varphi \right)\mu(ds),
\]
again producing Chernoff equivalence for the subordinate semigroup [1512.05258]. These constructions lead to Feynman formulae for subordinate Feller processes and diffusions on Euclidean spaces, star graphs, and Riemannian manifolds [1512.05258].

Killed Feller processes require another modification. If \(G\subset\mathbb R^d\) is a bounded domain and \((T_t^G)\) is the semigroup of the process killed on exit, then a Chernoff family for \(T_t^G\) is obtained from an ambient family \(F(t)\) by extension and boundary cutoffs:
\[
F_G(0):=\operatorname{Id}, \qquad
F_G(t)\varphi(x):=\varphi_{s(t)}(x)\,[F(t)E(\varphi)](x),\qquad x\in G.
\tag{11}
\]
Under compatibility assumptions on the extension operator and the core, one gets
\[
T_t^G\varphi=\lim_{n\to\infty}[F_G(t/n)]^n\varphi,
\]
which then converts into explicit \(n\)-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 [1708.02503].

On Riemannian manifolds of bounded geometry, Chernoff approximations can be built intrinsically from flow maps of vector fields. For
\[
(L_0f)(x)=\frac12\sum_{k=1}^r (A_kA_k f)(x) + A_0 f(x),
\tag{23}
\]
the explicit manifold-valued shift family
\[
(S(t)f)(x)=\frac14\sum_{j=1}^r \Big( f(Y_{x,A_j}(\sqrt{2rt})) + f(Y_{x,-A_j}(\sqrt{2rt})) \Big) +\frac12\, f(Y_{x,A_0}(2t)) +t\,c(x)f(x)
\tag{32}
\]
approximates the Feller semigroup generated by the closure of \(L_{0c}=L_0+c\) on \(C_0(M)\):
\[
\lim_{n\to\infty}\sup_{t\in[0,T]} \|S(t/n)^n f-V(t)f\|_\infty=0.
\tag{34}
\]
The same construction yields weak convergence of random walks on the manifold to the associated diffusion process [2002.06606].

The classical equal-step product formula can also be generalized to non-uniform partitions. If \(V\) is a contraction Chernoff family with pairwise commuting operators and the partition coefficients satisfy
\[
\sum_{i=1}^n \left|\frac1n-a_{n,i}\right| \to 0,
\]
then
\[
T(t)x = \lim_{n\to\infty}\prod_{i=1}^n V(a_{n,i}t)x
\]
uniformly on compact time intervals [2407.04357]. The paper stresses that this condition is stronger than the null-array condition \(\max_i a_{n,i}\to 0\), and that commutativity and contractivity are the price paid for avoiding Smolyanov’s generator-orbit consistency condition [2407.04357].

## 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 \(S(t)\) is nonlinear, acting on weighted spaces \(C_\kappa(\mathbb R^d)\), and the approximation takes the form
\[
S(t)f=\lim_{n\to\infty} I\!\left(\frac tn\right)^n f
\]
or, more generally, \(S(t)f=\lim_{n\to\infty}I(\pi_n^t)f\) for piecewise-constant partitions [2310.09830]. The one-step operators \(I(t)\) are assumed convex, monotone, exponentially stable in the weighted norm, and compatible with translation estimates and smooth test functions [2310.09830].

Unlike the linear theory, the generator is handled through \(\Gamma\)-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 \(\rho_1,\rho_2,\rho_3\). Under polynomial consistency assumptions these yield power rates
\[
\|(S(t)f-I(\pi_n^t)f)^-\|_\kappa\le c_{r,t}^-\, h_n^{\gamma^-},
\tag{2.13}
\]
and
\[
\|(S(t)f-I(\pi_n^t)f)^+\|_\kappa\le c_{r,t}^+\, h_n^{\gamma^+},
\tag{2.15}
\]
with
\[
\gamma^-= \min\left\{ \frac1{1+p}, \frac{\alpha_1^-}{1+\beta_1^-}, \ldots, \frac{\alpha_{N^-}^-}{1+\beta_{N^-}^-} \right\},
\tag{2.14}
\]
and the analogous formula for \(\gamma^+\) [2310.09830].

This nonlinear approach is applied to Nisio semigroups, where
\[
(I(t)f)(x):=\sup_{\lambda\in\Lambda}(S_\lambda(t)f)(x),
\tag{4.2}
\]
leading to rates such as \(1/2\), \(1/6\), or \(1/4\) depending on the order of consistency and additional regularity assumptions [2310.09830]. 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 [2310.09830]. 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/\(\Gamma\)-generator route [2310.09830].

## 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. \(X_i\in[0,1]\) with mean \(p\) and sample average \(\overline X\), the note on the Chernoff bound proves that the Bernoulli one-sided KL inversion
\[
p \leq \mathrm{kl}^{-1}\left(\overline{X} \, \middle| \, \frac{1}{n} \log \frac{1}{\delta} \right)
\]
remains valid without change for arbitrary \([0,1]\)-valued data, not just Bernoulli variables [2205.07880]. This is a concentration-inequality statement, not a semigroup approximation result.

A second use is Chernoff information, defined by
\[
D_C[p,q]=\max_{\alpha\in (0,1)} D_{B,\alpha}[p:q]
\]
with
\[
D_{B,\alpha}[p:q]:=-\log\int p^\alpha q^{1-\alpha}\,d\mu.
\]
Recent work interprets this through likelihood ratio exponential families, proves uniqueness of the optimizer \(\alpha^*\), and derives exact or numerical formulas for Gaussian families [2207.03745]. Earlier work on same-family exponential distributions rewrites fixed-\(\alpha\) Chernoff divergence as a skew Jensen divergence of the log-normalizer and proposes geodesic bisection to approximate the optimizing exponent [1102.2684]. 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,
\[
-\frac{3\delta^2}{6+2\delta}
\quad\text{and}\quad
-\frac{9\delta^2}{18-6\delta-\delta^2},
\]
are proposed because they remain rigorous one-sided bounds while being exactly invertible via quadratics [2109.14356]. 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 \(C_0\)-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 [2205.07880, 2207.03745, 2109.14356].

## 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 \(O(1/n)\) to \(O(1/n^2)\) or higher, at least on sufficiently regular vectors [2104.01249, 2606.27232]. 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 [2301.05284, 2012.09615]. 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 [2002.06606, 1512.05258, 2310.09830].

Several caveats recur. General convergence can be arbitrarily slow, so no universal rate follows from Chernoff’s theorem alone [1910.09440, 2104.01249]. Strong convergence need not imply operator-norm convergence, although quasi-sectorial contraction theory provides operator-norm Chernoff estimates in special settings [2205.04794]. Numerical studies often rely on short ranges such as \(n=1,\dots,11\), and some empirically fitted asymptotic slopes are explicitly described as suggestive rather than definitive [2301.05284]. In variable-coefficient problems, exact benchmark solutions are typically unavailable, making rigorous or even empirical error assessment more difficult [2301.05284, 2606.27232].

Open directions are stated quite explicitly in the cited works. These include sharper theoretical convergence-rate results for rough initial data [2301.05284], further development of fast Chernoff schemes for variable-coefficient parabolic equations [2606.27232], broader nonlinear convergence-rate theory beyond the presently treated convex monotone framework [2310.09830], and continued extension of explicit product formulas to subordinate, killed, manifold-valued, and time-fractional evolutions [1512.05258, 1708.02503, 2002.06606].

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 [2301.05284].

Source: https://www.emergentmind.com/topics/chernoff-approximations