An evolution equation asks how a state changes with time. In finite dimensions this usually leads to an ordinary differential equation on $\mathbb R^n$; in analysis the state may instead be a function, a distribution, or a vector in an infinite-dimensional space. The abstract Cauchy problem is the language that lets us treat many differential equations, partial differential equations, and operator evolution equations as one problem on a [Banach space](/page/Banach%20Space). The main shift is conceptual: the time derivative remains a derivative in one real variable, while all spatial structure is placed inside an operator $A$. For the [heat equation](/page/Heat%20Equation), $A$ may be the Laplacian with boundary conditions; for a transport equation, $A$ may be a first-order differential operator; for a system of linear ODEs, $A$ is a matrix. This abstraction separates the evolution mechanism from the specific representation of the state. A strongly continuous semigroup, also called a $C_0$-semigroup, is a family $(T(t))_{t\ge 0}$ of bounded linear operators on $X$ such that $T(0)=I$, $T(t+s)=T(t)T(s)$ for $s,t\ge 0$, and $\|T(t)x-x\|_X\to 0$ as $t\downarrow 0$ for each $x\in X$. It represents the finite-time evolution associated with an infinitesimal operator. The concept is central to strongly continuous semigroups, generators, the spectrum of an operator, and well-posedness. It also gives a precise way to distinguish classical solutions, which satisfy the equation pointwise in the Banach space, from mild solutions, which are obtained by integrating the evolution law. ## Definition The initial-value problem becomes abstract only after the ambient space and the operator are specified. The state space records what kind of objects are evolving, while the operator records the spatial or structural rule that drives the evolution. This first definition isolates the simplest autonomous linear problem, where the future is determined by one fixed operator and one initial state. ### Linear and Inhomogeneous Forms The linear homogeneous problem is the reference case because it isolates the role of the operator without external forcing. Once this case is fixed, the inhomogeneous version adds a source term without changing the underlying state space. [definition: Linear Abstract Cauchy Problem] Let $X$ be a Banach space over $\mathbb K$, where $\mathbb K$ is either $\mathbb R$ or $\mathbb C$. Let $A: D(A) \subset X \to X$ be a linear operator with domain $D(A)$, let $u_0 \in X$, and let $u: [0,\infty) \to X$ be the unknown solution curve. The homogeneous linear abstract Cauchy problem associated to $(A,u_0)$ is \begin{align*} \frac{du}{dt}(t) = Au(t), \qquad t > 0. \end{align*} The initial condition is \begin{align*} u(0) = u_0. \end{align*} [/definition] Here $u$ is not a scalar function in general. It is a curve in the Banach space $X$, so the derivative $du/dt$ is a norm derivative in $X$. When $A$ is unbounded, the condition $Au(t)$ only makes sense when $u(t) \in D(A)$, which is why solution concepts matter. Before the domain subtleties appear, the bounded-operator case shows the shape of the abstraction. It is the Banach-space analogue of the matrix equation $\dot{x}=Ax$, and it is the model against which the unbounded theory is compared. ### Bounded Operator Model Here $\mathcal L(X)$ denotes the Banach algebra of bounded linear operators from $X$ to itself, equipped with the operator norm $\|\cdot\|_{\mathcal L(X)}$. [example: Bounded Operator Model] Let $X$ be a Banach space, let $A \in \mathcal L(X)$, and fix $u_0 \in X$. Since $\|A^n u_0\|_X \le \|A\|_{\mathcal L(X)}^n\|u_0\|_X$, for every $t \ge 0$ we have \begin{align*} \left\|\frac{t^n A^n u_0}{n!}\right\|_X \le \frac{(t\|A\|_{\mathcal L(X)})^n}{n!}\|u_0\|_X. \end{align*} The scalar series $\sum_{n=0}^\infty (t\|A\|_{\mathcal L(X)})^n/n!$ converges, so the series \begin{align*} e^{tA}u_0=\sum_{n=0}^\infty \frac{t^n A^n u_0}{n!} \end{align*} converges absolutely in $X$. On each interval $0 \le t \le T$, the same estimate with $T$ in place of $t$ gives [uniform convergence](/page/Uniform%20Convergence). Define $u(t)=e^{tA}u_0$. Differentiating the partial sums gives \begin{align*} \frac{d}{dt}\sum_{n=0}^N \frac{t^n A^n u_0}{n!}=\sum_{n=1}^N \frac{n t^{n-1}A^n u_0}{n!}. \end{align*} For $n \ge 1$, \begin{align*} \frac{n t^{n-1}A^n u_0}{n!}=\frac{t^{n-1}A^n u_0}{(n-1)!}. \end{align*} Hence the derivative series is \begin{align*} \sum_{n=1}^\infty \frac{t^{n-1}A^n u_0}{(n-1)!}. \end{align*} On $0 \le t \le T$ its terms satisfy \begin{align*} \left\|\frac{t^{n-1}A^n u_0}{(n-1)!}\right\|_X \le \|A\|_{\mathcal L(X)}\frac{(T\|A\|_{\mathcal L(X)})^{n-1}}{(n-1)!}\|u_0\|_X, \end{align*} so this derivative series also converges uniformly on compact time intervals. Therefore termwise differentiation is valid, and \begin{align*} u'(t)=\sum_{n=1}^\infty \frac{t^{n-1}A^n u_0}{(n-1)!}. \end{align*} Reindexing with $m=n-1$ gives \begin{align*} u'(t)=\sum_{m=0}^\infty \frac{t^m A^{m+1}u_0}{m!}. \end{align*} Because $A$ is bounded, it may be passed through the norm-convergent series: \begin{align*} Au(t)=A\sum_{m=0}^\infty \frac{t^m A^m u_0}{m!}=\sum_{m=0}^\infty \frac{t^m A^{m+1}u_0}{m!}. \end{align*} Thus $u'(t)=Au(t)$ for every $t \ge 0$. At $t=0$, \begin{align*} u(0)=\sum_{n=0}^\infty \frac{0^n A^n u_0}{n!}=u_0, \end{align*} where the $n=0$ term is $u_0$ and every term with $n \ge 1$ is $0$. Since $D(A)=X$ for a bounded operator $A \in \mathcal L(X)$, the curve $u(t)=e^{tA}u_0$ is a classical solution of the homogeneous abstract Cauchy problem. This is the Banach-space version of the matrix formula for $\dot{x}=Ax$. [/example] Many physical and analytic models are not closed systems: heat may be injected, a force may act on a mechanical system, or a PDE may acquire a source term after boundary data are transformed. The abstract framework needs a version that records this external input without changing the role of $A$ as the internal evolution operator. [definition: Inhomogeneous Abstract Cauchy Problem] Let $X$ be a Banach space over $\mathbb K$, let $A: D(A) \subset X \to X$ be a linear operator, let $u_0 \in X$, let $f: [0,\infty) \to X$ be a function, and let $u: [0,\infty) \to X$ be the unknown solution curve. The inhomogeneous linear abstract Cauchy problem associated to $(A,f,u_0)$ is \begin{align*} \frac{du}{dt}(t) = Au(t) + f(t), \qquad t > 0. \end{align*} The initial condition is \begin{align*} u(0) = u_0. \end{align*} [/definition] The forcing term may represent heat sources, external forces, boundary data after lifting, or lower-order terms. Its regularity determines which solution concept is appropriate and how much time regularity the solution can have. The most direct solution concept asks the equation to hold as an equality in $X$ for every positive time. This is the infinite-dimensional analogue of a differentiable solution of a linear ODE, and it is the correct notion when the evolving state stays in the domain of the operator. ### Classical and Mild Solutions The first distinction is whether the differential equation is imposed pointwise in the Banach-space norm or recovered from an integrated evolution formula. This separates smooth data that remain in the operator domain from rough data that still evolve through a semigroup. [definition: Classical Solution of an Abstract Cauchy Problem] Let $X$ be a Banach space, let $A: D(A) \subset X \to X$ be a linear operator, let $u_0 \in X$, and let $f: [0,\infty) \to X$. A classical solution of the inhomogeneous abstract Cauchy problem on an interval $[0,T]$ is a function $u: [0,T] \to X$ such that $u \in C([0,T];X) \cap C^1((0,T];X)$, $u(t) \in D(A)$ for every $t \in (0,T]$, $u(0)=u_0$, and \begin{align*} \frac{du}{dt}(t) = Au(t)+f(t) \end{align*} for every $t \in (0,T]$. [/definition] Classical solutions are strong enough to encode the equation directly, but in PDE they can be too restrictive. Initial data in $L^2(U)$ for the heat equation need not lie in the domain of the Dirichlet Laplacian, even though the heat flow still has a meaningful evolution. To include rough initial states, the equation should be read through its accumulated time evolution. If the operator $A$ generates operators $T(t)$ that advance the free system, then a source inserted at time $s$ is carried from $s$ to $t$ by $T(t-s)$. This viewpoint leads to the solution concept used most often in [semigroup theory](/page/Semigroup%20Theory). [definition: Mild Solution of an Abstract Cauchy Problem] Let $X$ be a Banach space, let $A: D(A) \subset X \to X$ be a linear operator that generates a strongly continuous semigroup $(T(t))_{t \ge 0}$ with $T(t): X \to X$ for every $t \ge 0$, let $u_0 \in X$, and let $f \in C([0,T];X)$. A mild solution of the inhomogeneous abstract Cauchy problem on $[0,T]$ is a function $u \in C([0,T];X)$ satisfying \begin{align*} u(t) = T(t)u_0 + \int_0^t T(t-s)f(s)\, ds \end{align*} for every $t \in [0,T]$, where the integral is the Bochner integral in $X$. [/definition] The formula in the definition is the infinite-dimensional version of variation of constants. It is often meaningful even when $u(t)$ does not belong to $D(A)$ at every time and even when $du/dt$ does not exist as a norm derivative at $t=0$. A concept of well-posedness records whether the equation behaves as a genuine evolution problem rather than only a formal expression. It packages existence, uniqueness, and continuous dependence on initial data, because all three are needed before an initial-value problem can support approximation, perturbation, or modelling arguments. ### Well-Posedness Well-posedness records when the abstract equation is a genuine evolution law rather than only a formal differential expression. The semigroup formulation packages existence, uniqueness, and continuous dependence in one operator family. [definition: Semigroup Well-Posed Abstract Cauchy Problem] Let $X$ be a Banach space and let $A: D(A) \subset X \to X$ be a linear operator. The homogeneous abstract Cauchy problem for $A$ is well-posed in the semigroup sense on $X$ if there exists a strongly continuous semigroup $(T(t))_{t \ge 0}$ with $T(t): X \to X$ for every $t \ge 0$ such that, for every $u_0 \in X$, the function $u: [0,\infty) \to X$ defined by \begin{align*} u(t)=T(t)u_0 \end{align*} is the unique mild solution with initial value $u_0$. [/definition] This definition fixes the solution class rather than leaving it implicit. A problem may be well-posed for mild solutions on $X$ while classical solutions exist only for a smaller set of initial data such as $D(A)$. ## Solution Concepts and Operators The operator in an abstract Cauchy problem is often unbounded. This is not a defect; differential operators on infinite-dimensional spaces are naturally unbounded when they remember boundary conditions and differentiability requirements. A semigroup describes finite-time evolution, while a differential equation describes infinitesimal change. To connect the two, we need a definition that extracts the infinitesimal operator from the short-time behaviour of $T(t)$. That operator is the generator. [definition: Generator of a Strongly Continuous Semigroup] Let $X$ be a Banach space and let $(T(t))_{t \ge 0}$ be a strongly continuous semigroup with $T(t): X \to X$ for every $t \ge 0$. Its generator is the operator $A: D(A) \subset X \to X$ defined by \begin{align*} D(A) = \left\{x \in X : \lim_{h \downarrow 0} \frac{T(h)x-x}{h} \text{ exists in } X\right\}. \end{align*} For $x \in D(A)$, \begin{align*} Ax = \lim_{h \downarrow 0} \frac{T(h)x-x}{h}. \end{align*} [/definition] This definition turns a family of time-evolution maps into an infinitesimal law. The abstract Cauchy problem then asks whether the infinitesimal law reconstructs the whole evolution. Unbounded operators can misbehave if their graphs are not stable under limits. Since solutions are commonly built as limits of approximations, the theory needs a condition guaranteeing that a limiting state and limiting image still belong to the same operator graph. [definition: Closed Operator] Let $X$ and $Y$ be Banach spaces. A linear operator $A: D(A) \subset X \to Y$ is closed if whenever $x_n \in D(A)$, $x_n \to x$ in $X$, and $Ax_n \to y$ in $Y$, then $x \in D(A)$ and $Ax=y$. [/definition] Closedness is the operator-theoretic replacement for having a well-behaved graph. Generators of strongly continuous semigroups are closed, and this prevents pathological limiting behaviour in the differential equation. The remaining issue is whether the infinitesimal equation $u'(t)=Au(t)$ is really recovered from the time-evolution family. A semigroup can advance states by $T(t)x$, but this only solves the homogeneous abstract Cauchy problem when differentiating the orbit reproduces the generator action on the allowed initial data. [quotetheorem:3145] This theorem explains why semigroups are the natural engine behind the abstract Cauchy problem. The operator $A$ gives the infinitesimal rule, while $T(t)$ gives the actual time-$t$ evolution. For inhomogeneous equations, the solution must remember when each part of the forcing was added. A contribution inserted at time $s$ should evolve freely for the remaining time $t-s$, so the semigroup must appear inside an integral over past times. [quotetheorem:7057] The formula separates two effects: free evolution from the initial state and accumulated evolution from the forcing. In PDE language, it is the abstract Duhamel formula. ## Equivalent Characterisations A major reason for introducing the abstract Cauchy problem is that well-posedness can be recognised through operator-theoretic conditions. Instead of solving every PDE directly, one studies whether its spatial operator generates a semigroup. [quotetheorem:7981] This characterisation turns an evolution question into a generator question. In applications, it is often easier to prove resolvent estimates, dissipativity, or sectoriality for $A$ than to construct solutions directly. The most famous generator criterion is the [Hille-Yosida theorem](/theorems/3139). It relates generation to bounds on resolvents, linking the abstract Cauchy problem to spectral theory. For a closed operator $A$, the resolvent set $\rho(A)$ consists of those $\lambda$ for which $\lambda I-A$ is bijective with bounded inverse. The resolvent operator is written $R(\lambda,A)=(\lambda I-A)^{-1}$. For an unbounded operator, constructing $T(t)$ directly can be difficult because the formal exponential series need not converge on the whole space. Resolvent estimates give a substitute: they control the inverses of $\lambda I-A$ uniformly enough to recover the finite-time evolution. [quotetheorem:3139] The theorem shows that solving an evolution equation is tied to invertibility of $\lambda I-A$. Thus the abstract Cauchy problem forms a bridge between time evolution and the resolvent methods of functional analysis. ## PDE Models and Domain Effects The bounded case is useful as a model, but it hides the main analytical difficulty. Differential operators usually have domains smaller than the whole function space. In the heat-equation example, $L^2(U)$ is the [Hilbert space](/page/Hilbert%20Space) of square-integrable functions on $U$, $H^2(U)$ is the Sobolev space with weak derivatives up to order two in $L^2(U)$, and $H^1_0(U)$ is the closure of compactly supported smooth functions in the $H^1$ norm. The trace of a Sobolev function is its boundary value in the Sobolev sense, so $H^1_0(U)$ encodes zero boundary values. The Dirichlet Laplacian is the Laplace operator $\Delta$ with this zero-boundary domain. [example: Heat Equation as an Abstract Cauchy Problem] Let $U \subset \mathbb R^n$ be a bounded [open set](/page/Open%20Set) with $C^2$ boundary and set $X=L^2(U)$. Define the Dirichlet Laplacian $A:D(A)\subset X\to X$ by \begin{align*} D(A)=H^2(U)\cap H^1_0(U). \end{align*} For $v\in D(A)$, \begin{align*} Av=\Delta v. \end{align*} With this choice of operator, the homogeneous abstract Cauchy problem $u'(t)=Au(t)$ becomes \begin{align*} u'(t)=\Delta u(t). \end{align*} The initial condition is \begin{align*} u(0)=u_0. \end{align*} The zero boundary condition is encoded in the domain: if $u(t)\in D(A)$, then $u(t)\in H^1_0(U)$, so its trace on $\partial U$ is $0$. Thus the abstract equation is exactly the heat equation with zero Dirichlet boundary condition, read as an equation in $L^2(U)$. The Dirichlet Laplacian generates a strongly continuous semigroup $(T(t))_{t\ge 0}$ on $L^2(U)$ by the standard generation theorem for the heat semigroup. Hence, for every $u_0\in L^2(U)$, the mild solution is \begin{align*} u(t)=T(t)u_0. \end{align*} If $u_0\in D(A)$, then *Semigroup Solution of the Homogeneous Abstract Cauchy Problem* gives $u(t)\in D(A)$ for every $t\ge 0$ and \begin{align*} u'(t)=Au(t)=\Delta u(t). \end{align*} So rough initial data in $L^2(U)$ still give a mild heat evolution, while initial data in $H^2(U)\cap H^1_0(U)$ give a classical $L^2(U)$-valued solution. [/example] This example is one of the main reasons the abstraction matters. Initial data in $L^2(U)$ can be physically meaningful for heat flow even when the Laplacian of the initial state is not an $L^2$ function. The next example shows a boundary of the theory: a densely defined operator is not automatically a generator. Domain information and resolvent estimates are part of the problem. [example: A Formal Operator Need Not Generate an Evolution] Let $X=C([0,1])$ with the sup norm, and define $A:D(A)\subset X\to X$ by \begin{align*} D(A)=\{u\in C^1([0,1]):u(0)=u(1)=0\}. \end{align*} For $u\in D(A)$, set \begin{align*} Au=\frac{du}{dx}. \end{align*} The formal equation $du/dt=Au$ is the transport equation \begin{align*} \frac{\partial u}{\partial t}(t,x)=\frac{\partial u}{\partial x}(t,x). \end{align*} Along a characteristic with $x(t)=x_0-t$, the chain rule gives \begin{align*} \frac{d}{dt}u(t,x_0-t)=\frac{\partial u}{\partial t}(t,x_0-t)-\frac{\partial u}{\partial x}(t,x_0-t)=0. \end{align*} Hence the characteristic value is constant, so whenever $x+t\le 1$ the natural translation formula is \begin{align*} u(t,x)=u(0,x+t)=u_0(x+t). \end{align*} Now take the admissible initial state $u_0(x)=x(1-x)$. Then $u_0\in C^1([0,1])$, and \begin{align*} u_0(0)=0(1-0)=0. \end{align*} Also, \begin{align*} u_0(1)=1(1-1)=0. \end{align*} Thus $u_0\in D(A)$. If the translation flow preserved the imposed domain condition $u(t)\in D(A)$, then it would have to satisfy $u(t,0)=0$ for every $t>0$. But the characteristic formula gives, for $0<t<1$, \begin{align*} u(t,0)=u_0(t)=t(1-t). \end{align*} Since $0<t<1$ implies $t>0$ and $1-t>0$, we have \begin{align*} t(1-t)>0. \end{align*} Therefore $u(t,0)\ne 0$, contradicting the requirement $u(t)\in D(A)$. So the endpoint conditions are not harmless decorations on the differential expression $d/dx$: with this domain, the transport law immediately carries interior values into a boundary where the domain demands zero. This operator-domain pair therefore does not generate the natural translation evolution on $C([0,1])$ with both endpoint conditions imposed. [/example] The failure is not caused by differentiating functions. It is caused by choosing an operator-domain pair that is incompatible with a forward time evolution. ## Semigroup Stability and Regularity A well-posed abstract Cauchy problem has a family of solution operators. These operators inherit the algebra of time: evolving for time $s$ and then for time $t$ is the same as evolving for time $s+t$. [quotetheorem:7982] The semigroup law gives the algebra of evolution, but well-posedness also requires stability under changes in the initial state. The next estimate supplies this quantitative control over finite time intervals. [quotetheorem:7983] This is the continuous-dependence part of well-posedness. It is also the estimate that makes approximation schemes, perturbation arguments, and numerical methods stable on bounded time intervals. Classical and mild solutions coincide when enough regularity is present. This principle lets analysts solve a problem first in the mild sense and then upgrade the solution under stronger assumptions. [quotetheorem:7984] This theorem explains why the domain $D(A)$ is not merely technical. It is the class of initial states for which the infinitesimal equation can be read directly in the Banach space. ## Beyond and Connected Topics The abstract Cauchy problem can be viewed as the operator-theoretic skeleton of an evolution equation. In a concrete PDE, the choice of the [Banach space](/page/Banach%20Space) $X$ determines the regularity in which the solution is measured, and the choice of $D(A)$ encodes boundary conditions and differentiability. It is closely related to [bounded linear operators](/page/Bounded%20Linear%20Operator), but its main applications require unbounded operators. The theory therefore depends on closed operators, dense domains, resolvent sets, and spectral estimates. In PDE, many choices of $A$ come from elliptic operators on [Sobolev spaces](/page/Sobolev%20Space). For parabolic equations, the associated semigroup often smooths the initial data. For wave equations, the problem is commonly rewritten as a first-order system on a product space. The abstract Cauchy problem is also a gateway to weak and distributional methods. If a classical derivative in time is unavailable, the mild formulation still gives an evolution in $X$, and further weak formulations can be developed by testing against elements of a [dual space](/page/Dual%20Space). The same framework extends in several directions: nonlinear abstract Cauchy problems replace $Au$ by a nonlinear operator $F(u)$; non-autonomous problems replace $A$ by a family $A(t)$; stochastic evolution equations add noise terms. The linear autonomous case remains the reference point because it has the most complete connection with semigroup theory. ## References Hille and Phillips, *Functional Analysis and Semi-Groups* (1957). Pazy, *Semigroups of Linear Operators and Applications to Partial Differential Equations* (1983). Engel and Nagel, *One-Parameter Semigroups for Linear Evolution Equations* (2000). Evans, *Partial Differential Equations* (2010). [Banach Space](/page/Banach%20Space). [Sobolev Space](/page/Sobolev%20Space).