ODE: Separation of Variables for PDEs

The method of separation of variables is a cornerstone technique for solving linear partial differential equations (PDEs) that arise in countless engineering contexts, from heat transfer and wave propagation to quantum mechanics and electrostatics. Its power lies in transforming a seemingly intractable PDE into a system of simpler, solvable ordinary differential equations (ODEs), allowing you to construct exact solutions to complex physical problems governed by specific boundary and initial conditions. Mastering this method is essential for any engineer or physicist who needs to move beyond numerical simulations to gain deep analytical insight into a system's behavior.

The Core Idea: Assuming a Product Solution

The entire method rests on a single, critical assumption: that the solution to the PDE can be written as a product solution of functions, each depending on only one of the independent variables. For a function $u(x,t)$ describing, say, temperature along a rod over time, we assume: $$u(x,t)=X(x)T(t)$$ Here, $X(x)$ is a function solely of the spatial variable $x$, and $T(t)$ is a function solely of the time variable $t$. This assumption is not always valid—the PDE and its boundary conditions must be homogeneous and linear—but when it is, it unlocks a powerful simplification. The genius of the approach is that by substituting this product form back into the PDE, the dependencies on $x$ and $t$ can be completely isolated.

Separating Variables and Introducing the Separation Constant

The next step is to substitute the product solution into the original PDE. Consider the one-dimensional heat equation: $\frac{\partial u}{\partial t}=\alpha \frac{{\partial }^{2}u}{\partial x^2}$, where $\alpha$ is the thermal diffusivity. Substituting $u=X(x)T(t)$ yields: $$X(x)\frac{dT}{dt}=\alpha T(t)\frac{d^{2}X}{d{x}^2}$$ Now, we separate variables by dividing both sides by the product $\alpha X(x)T(t)$, provided it is not zero: $$\frac{1}{\alpha T}\frac{dT}{dt}=\frac{1}{X}\frac{d^{2}X}{d{x}^2}$$ The left side is now a function of $t$ only, and the right side is a function of $x$ only. For this equation to hold for all independent values of $x$ and $t$, both sides must be equal to the same constant. This introduces the critical separation constant, often denoted by $-\lambda$ (the negative sign is conventional and simplifies subsequent steps): $$\frac{1}{\alpha T}\frac{dT}{dt}=\frac{1}{X}\frac{d^{2}X}{d{x}^2}=-\lambda$$ This single act of separation decomposes one PDE into two independent ODEs:

1. Temporal ODE: $\frac{dT}{dt}+\alpha \lambda T=0$
2. Spatial ODE (often called the eigenvalue problem): $\frac{d^{2}X}{d{x}^2}+\lambda X=0$

Solving the Resulting Eigenvalue Problem

The spatial ODE, $X^{\prime \prime }+\lambda X=0$, is not an initial value problem; it is a boundary value problem. Its solutions, and the allowable values of the separation constant $\lambda$, are completely determined by the physical boundary conditions imposed on the system (e.g., $u(0,t)=0$ and $u(L,t)=0$ for a rod with fixed temperatures at its ends). This equation, together with its homogeneous boundary conditions, forms an eigenvalue problem.

The nature of the solutions depends on the sign of $\lambda$:

• $\lambda >0$: Solutions are sinusoidal: $X(x)=A\cos (\sqrt{\lambda }x)+B\sin (\sqrt{\lambda }x)$.
• $\lambda =0$: Solution is linear: $X(x)=Ax+B$.
• $\lambda <0$: Solutions are exponential (hyperbolic), which typically do not satisfy homogeneous boundary conditions on finite domains for problems like the heat or wave equation.

Applying the boundary conditions forces $\lambda$ to take on a specific, discrete set of positive values. For the boundary conditions $X(0)=0$ and $X(L)=0$, we find that $B=0$ from the first condition, and $\sin (\sqrt{\lambda }L)=0$ from the second. This leads to the eigenvalues ${\lambda }_n=(\frac{n\pi }{L}{)}^2$ and corresponding eigenfunctions $X_n(x)=\sin (\frac{n\pi x}{L})$, for $n=1,2,3,...$.

Superposition of Separated Solutions

For each eigenvalue ${\lambda }_n$, we solve the corresponding temporal ODE. For the heat equation example, $\frac{d{T}_n}{dt}+\alpha {\lambda }_n{T}_n=0$ gives $T_n(t)=C_n{e}^{-\alpha {\lambda }_{n}t}$. Therefore, for each $n$, we have a valid product solution: $$u_n(x,t)=X_n(x)T_n(t)=\left[\sin \left(\frac{n\pi x}{L}\right)\right]\left[C_n{e}^{-\alpha (n\pi /L{)}^{2}t}\right]$$ Because our original PDE is linear and homogeneous, the principle of superposition applies. This means the sum of any number of these individual solutions is also a solution. Thus, the general solution is an infinite series: $$u(x,t)=\sum _{n=1}^{\infty }{u}_n(x,t)=\sum _{n=1}^{\infty }{C}_n\sin \left(\frac{n\pi x}{L}\right)e^{-\alpha (n\pi /L{)}^{2}t}$$

Applying Boundary and Initial Conditions

The final step is to make this general solution satisfy the specific initial condition of the problem, $u(x,0)=f(x)$, which describes the initial temperature distribution. Substituting $t=0$ into our series gives: $$f(x)=u(x,0)=\sum _{n=1}^{\infty }{C}_n\sin \left(\frac{n\pi x}{L}\right)$$ This is precisely a Fourier sine series expansion of the function $f(x)$. The coefficients $C_n$ are no longer arbitrary; they are determined by the standard Fourier coefficient formula: $$C_n=\frac{2}{L}{\int }_0^{L}f(x)\sin \left(\frac{n\pi x}{L}\right)dx$$ Evaluating this integral (for a given $f(x)$) completes the solution. The process for other canonical equations like the wave equation or Laplace's equation follows an identical logical workflow: assume product form, separate variables, solve the eigenvalue problem from the spatial ODE with boundary conditions, solve the temporal ODE, superpose, and finally fit the initial conditions.

Common Pitfalls

1. Misapplying the product assumption to inhomogeneous conditions: The method requires homogeneous boundary conditions to generate a proper eigenvalue problem. A condition like $u(0,t)=50$ is inhomogeneous. A common correction is to use a technique like shifting the data—defining a new variable $v(x,t)=u(x,t)-w(x)$, where $w(x)$ is a steady-state solution chosen to absorb the inhomogeneity, leaving homogeneous conditions for $v$.

1. Incorrectly handling the separation constant: Forgetting the negative sign convention ($-\lambda$) or incorrectly analyzing the cases for $\lambda$ (positive, zero, negative) based on the boundary conditions is a frequent error. Always let the physics of the boundary conditions dictate which case yields non-trivial solutions. On a finite domain with homogeneous Dirichlet or Neumann conditions, $\lambda >0$ is almost always the relevant case.

1. Forgetting the principle of superposition for the general solution: After finding the family of product solutions $u_n(x,t)$, the most general solution is their sum (a series), not just one term. Stopping at a single term like $u_1(x,t)$ only satisfies the PDE and boundary conditions, but it cannot satisfy an arbitrary initial condition $f(x)$.

1. Mismatching eigenfunctions and Fourier series type: The boundary conditions determine the orthogonal eigenfunction family (sines, cosines, or both). Using the wrong Fourier coefficient formula to solve for $C_n$ will yield an incorrect solution. For Dirichlet conditions ($u=0$) at both ends, you use a sine series. For Neumann conditions ($\partial u/\partial x=0$) at both ends, you use a cosine series. Mixed conditions lead to a mixed series.

Summary

• The method of separation of variables reduces solvable linear PDEs to systems of ODEs by assuming the solution is a product of single-variable functions.
• The key step introduces a separation constant (often $-\lambda$), which splits the PDE into separate temporal and spatial ODEs.
• The spatial ODE, combined with homogeneous boundary conditions, forms an eigenvalue problem that determines the permissible values of $\lambda$ (eigenvalues) and corresponding spatial modes (eigenfunctions).
• The principle of superposition allows the general solution to be constructed as an infinite series sum of all product solutions.
• The final, specific solution is obtained by using the initial condition to determine the series coefficients, typically via an orthogonal expansion like a Fourier series.