7

There seem to be a lot of posts on the subject of heat equations with weird boundary conditions, but after a brief perusal of these I couldn't find quite what I'm looking for. The following 1D problem has arisen in my research, and I'm not sure how best to solve it:

$$\frac{\partial u}{\partial t} = \frac{\partial^2 u}{\partial x^2}; \\u(0,t) = f(t); \\ \frac{\partial u}{\partial x}(0,t) = g(t)$$

on the interval $(x,t)\in[0,\infty)\times (-\infty,\infty)$, where $f$ and $g$ are some specified smooth functions.

First of all: is this problem even well-posed? I don't know much existence/uniqueness theory for PDEs so I'm not sure.

Secondly, how can I go about solving this equation analytically? The unusual, nonhomogeneous B.C.'s are really tripping me up.


EDIT: Here's what I've tried. Letting $\tilde{u}(x,\omega)=\int dt \, e^{-i\omega t} u(x,t),$ we can transform the original equation into a 2nd order ODE in $x$ with Cauchy boundary conditions and solution

$$\tilde{u}(x,\omega) = \tilde{f}(\omega) \cosh\left((1-i)\sqrt{\frac{\omega}{2}}x\right)+\frac{1+i}{\sqrt{2\omega}}\tilde{g}(\omega) \sinh\left((1-i)\sqrt{\frac{\omega}{2}}x\right),$$

where $\tilde{f}$ and $\tilde{g}$ are of course the Fourier transforms of $f$ and $g$ respectively.

Formally, then, \begin{align}u(x,t) &= \int \frac{d\omega}{2\pi} \, e^{i \omega t} \left[ \tilde{f}(\omega) \cosh\left((1-i)\sqrt{\frac{\omega}{2}}x\right)+\frac{1+i}{\sqrt{2\omega}}\tilde{g}(\omega) \sinh\left((1-i)\sqrt{\frac{\omega}{2}}x\right) \right] \\ &=\int \frac{d \omega}{2 \pi} \int d\tau \, e^{i\omega(t-\tau)}\left[f(\tau) \cosh\left((1-i)\sqrt{\frac{\omega}{2}}x\right)+\frac{1+i}{\sqrt{2\omega}}g(\tau) \sinh\left((1-i)\sqrt{\frac{\omega}{2}}x\right)\right] \\ &= \int_{-\infty}^\infty d\tau \, \left(F(x,t-\tau)f(\tau)+G(x,t-\tau)g(\tau)\right), \end{align} where $$F(x,t):= \int_{-\infty}^{\infty} \frac{d\omega}{2\pi} \, e^{i \omega t} \cosh\left((1-i)\sqrt{\frac{\omega}{2}}x\right)$$

and

$$G(x,t):= \int_{-\infty}^{\infty} \frac{d\omega}{2\pi} \, e^{i \omega t} \frac{1+i}{\sqrt{2\omega}} \sinh\left((1-i)\sqrt{\frac{\omega} {2}}x\right). $$

I am getting discouraged, however, attempting to evaluate the kernels $F$ and $G$, which besides simply being challenging (for me anyway) to work with, appear to suffer from divergences.

Have I done something wrong? Are there closed-form, or at least nicer, expressions for $F$ and $G$? Does this approach even work?

  • 1
    Regarding existence: if we look at it from a numerical point of view with $C = \frac{\Delta t}{\Delta x^2}$ and $u_{i,j} = u(i\Delta x, j\Delta t)$ then discretizing $u_t = u_{xx}$ to second order leads to the recursion $$u_{i+1,j} = \frac{u_{i,j+1} - u_{i,j} + C[2u_{i,j} - u_{i-1,j}]}{C}$$ which can in principle be propagated to give us $u_{i,j}$ for all $i,j$ as the BC gives us $u_{0,j}$ and $u_{1,j}$. This is by far no proof of existence, but it suggest there might be just enough information in the BC to allow us to compute a solution at least locally around $x=0$ for all $t$. – Winther Jan 24 '17 at 00:11
  • I think this problem is ill poised. There can not be simultaneously 2 different boundary conditions for heat equation at $x=0$. The origin of the problem might be in some incorrect physical assumptions in the model. – Cave Johnson Jan 25 '17 at 13:14
  • @Nemo "There can not be simultaneously 2 different boundary conditions for heat equation at $x=0$." care to explain why not or share a reference? – mikefallopian Jan 25 '17 at 20:11
  • @mikefallopian is anything wrong with my answer? Are there any countering objections? – Cave Johnson Jan 26 '17 at 11:53
  • @Nemo That might be, however I’m afraid the proof below is not a valid one. Note that a system can seem to be overdetermined and still have a solution. To show that it’s ill defined in your approach you have to show that no matter what initial condition $v_0(x)$ you choose then you cannot satisfy the second boundary condition and I don't see that. – Winther Jan 26 '17 at 13:47
  • I have not looked properly at this so I might be missing something obvious here. If we try to take the Fourier transform wrt $t$ it does allow us to formally write down an expression for the solution $$u(x,t) = \frac{1}{2\pi}\int_{-\infty}^\infty\left[\frac{\hat{g}(k)}{\sqrt{ik}}\sin\left(\sqrt{ki} x\right) + \hat{f}(k)\cos\left(\sqrt{ki} x\right)\right]e^{-ikt},{\rm d}k$$ assuming all the Fourier-transforms involved exist (and if the square roots are well defined). For $u$ to exist it would require $\hat{f}$ and $\hat{g}$ to decay fast enough as to ensure convergence of the integral above. – Winther Jan 26 '17 at 13:58
  • @Winther, thank you very much for pointing out a possible issue in the answer. It has been deleted. However, even if $\hat{f},\hat{g}$ decay very fast, so that integral you mention converges, the solution will tend to infinity for large $x$ making it unphysical. Also in the last edit I have deleted the additional condition $\lim_{x\to+\infty}u(x,t)=0$ absent in OP's version of the problem. I think if this additional condition is imposed, then the problem will be ill posed. In either case OP's problem does not correspond to physical situation. – Cave Johnson Jan 26 '17 at 14:19
  • @Nemo Yes that is a true. I'm also not convinced that this problem makes sense from a physical point of view. – Winther Jan 26 '17 at 14:25
  • Physically, at least in terms of temperature, you’re specifying both the temperature and, in effect, the heat flux at $x=0$. It seems odd the initial state doesn’t explicitly enter the problem. – A rural reader Mar 23 '24 at 18:51
  • Both the question and the answer concentrate too much on the "heat equation". Look at your problem from purely mathematical point of view (to make it easier, replace $t$ with, e.g., $y$). Then you have a perfectly find problem for a second order PDE or, even better, for a system of two first order PDEs. For this system the Cauchy-Kovalevsky theorem can be applied that guarantees that if both $f$ and $g$ are analytic then there exists a unique analytic solution to your problem, which can be found as a convergent series (not sure how fast), so the problem is well posed in this case. – Artem Jul 31 '24 at 22:49

2 Answers2

-1

Approach $1$: separation of variables

Case $1$: $\text{Re}(t)\geq0$

Let $u(x,t)=X(x)T(t)$ ,

Then $X(x)T'(t)=X''(x)T(t)$

$\dfrac{T'(t)}{T(t)}=\dfrac{X''(x)}{X(x)}=-s^2$

$\begin{cases}\dfrac{T'(t)}{T(t)}=-s^2\\X''(x)+s^2X(x)=0\end{cases}$

$\begin{cases}T(t)=c_3(s^2)e^{-ts^2}\\X(x)=\begin{cases}c_1(s^2)\sin xs+c_2(s^2)\cos xs&\text{when}~s\neq0\\c_1x+c_2&\text{when}~s=0\end{cases}\end{cases}$

$\therefore u(x,t)=C_1x+C_2+\int_0^\infty C_3(s^2)e^{-ts^2}\sin xs~ds+\int_0^\infty C_4(s^2)e^{-ts^2}\cos xs~ds$

$\dfrac{\partial u(x,t)}{\partial x}=C_1+\int_0^\infty sC_3(s^2)e^{-ts^2}\cos xs~ds-\int_0^\infty sC_4(s^2)e^{-ts^2}\sin xs~ds$

$\dfrac{\partial u(x,t)}{\partial x}(x=0)=g(t)$ :

$C_1+\int_0^\infty sC_3(s^2)e^{-ts^2}~ds=g(t)$

$\int_0^\infty\dfrac{C_3(s^2)e^{-ts^2}}{2}d(s^2)=g(t)-C_1$

$\int_0^\infty\dfrac{C_3(s)e^{-ts}}{2}ds=g(t)-C_1$

$\mathcal{L}_{s\to t}\biggl\{\dfrac{C_3(s)}{2}\biggr\}=g(t)-C_1$

$C_3(s)=2\mathcal{L}^{-1}_{t\to s}\{g(t)\}-2C_1\delta(s)$

$\therefore u(x,t)=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(t)\}e^{-ts^2}\sin xs~ds-2C_1\int_0^\infty\delta(s^2)e^{-ts^2}\sin xs~ds+\int_0^\infty C_4(s^2)e^{-ts^2}\cos xs~ds=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(t)\}e^{-ts^2}\sin xs~ds-C_1\int_0^\infty\dfrac{\delta(s^2)e^{-ts^2}\sin xs}{s}d(s^2)+\int_0^\infty C_4(s^2)e^{-ts^2}\cos xs~ds=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(t)\}e^{-ts^2}\sin xs~ds-C_1\int_0^\infty\dfrac{\delta(s)e^{-ts}\sin x\sqrt{s}}{\sqrt{s}}ds+\int_0^\infty C_4(s^2)e^{-ts^2}\cos xs~ds=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(t)\}e^{-ts^2}\sin xs~ds-C_1\lim\limits_{s\to 0}\dfrac{e^{-ts}\sin x\sqrt{s}}{\sqrt{s}}+\int_0^\infty C_4(s^2)e^{-ts^2}\cos xs~ds=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(t)\}e^{-ts^2}\sin xs~ds-C_1\lim\limits_{s\to 0}\dfrac{\dfrac{xe^{-ts}\cos x\sqrt{s}}{2\sqrt{s}}-te^{-ts}\sin x\sqrt{s}}{\dfrac{1}{2\sqrt{s}}}+\int_0^\infty C_4(s^2)e^{-ts^2}\cos xs~ds=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(t)\}e^{-ts^2}\sin xs~ds-C_1\lim\limits_{s\to 0}(xe^{-ts}\cos x\sqrt{s}-2t\sqrt{s}e^{-ts}\sin x\sqrt{s})+\int_0^\infty C_4(s^2)e^{-ts^2}\cos xs~ds=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(t)\}e^{-ts^2}\sin xs~ds-C_1x+\int_0^\infty C_4(s^2)e^{-ts^2}\cos xs~ds=C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(t)\}e^{-ts^2}\sin xs~ds+\int_0^\infty C_4(s^2)e^{-ts^2}\cos xs~ds$

$u(0,t)=f(t)$ :

$C_2+\int_0^\infty C_4(s^2)e^{-ts^2}~ds=f(t)$

$\int_0^\infty\dfrac{C_4(s^2)e^{-ts^2}}{2s}d(s^2)=f(t)-C_2$

$\int_0^\infty\dfrac{C_4(s)e^{-ts}}{2\sqrt{s}}ds=f(t)-C_2$

$\mathcal{L}_{s\to t}\biggl\{\dfrac{C_4(s)}{2\sqrt{s}}\biggr\}=f(t)-C_2$

$C_4(s)=2\sqrt{s}\mathcal{L}^{-1}_{t\to s}\{f(t)\}-C_2\delta(\sqrt{s})$

$\therefore u(x,t)=C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(t)\}e^{-ts^2}\sin xs~ds+2\int_0^\infty s\mathcal{L}^{-1}_{t\to s^2}\{f(t)\}e^{-ts^2}\cos xs~ds-C_2\int_0^\infty\delta(s)e^{-ts^2}\cos xs~ds=C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(t)\}e^{-ts^2}\sin xs~ds+2\int_0^\infty s\mathcal{L}^{-1}_{t\to s^2}\{f(t)\}e^{-ts^2}\cos xs~ds-C_2=2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(t)\}e^{-ts^2}\sin xs~ds+2\int_0^\infty s\mathcal{L}^{-1}_{t\to s^2}\{f(t)\}e^{-ts^2}\cos xs~ds$

Case $2$: $\text{Re}(t)\leq0$

Let $u(x,t)=X(x)T(t)$ ,

Then $X(x)T'(t)=X''(x)T(t)$

$\dfrac{T'(t)}{T(t)}=\dfrac{X''(x)}{X(x)}=s^2$

$\begin{cases}\dfrac{T'(t)}{T(t)}=s^2\\X''(x)-s^2X(x)=0\end{cases}$

$\begin{cases}T(t)=c_3(s^2)e^{ts^2}\\X(x)=\begin{cases}c_1(s^2)\sinh xs+c_2(s^2)\cosh xs&\text{when}~s\neq0\\c_1x+c_2&\text{when}~s=0\end{cases}\end{cases}$

$\therefore u(x,t)=C_1x+C_2+\int_0^\infty C_3(s^2)e^{ts^2}\sinh xs~ds+\int_0^\infty C_4(s^2)e^{ts^2}\cosh xs~ds$

$\dfrac{\partial u(x,t)}{\partial x}=C_1+\int_0^\infty sC_3(s^2)e^{ts^2}\cosh xs~ds+\int_0^\infty sC_4(s^2)e^{ts^2}\sinh xs~ds$

$\dfrac{\partial u(x,t)}{\partial x}(x=0)=g(t)$ :

$C_1+\int_0^\infty sC_3(s^2)e^{ts^2}~ds=g(t)$

$\int_0^\infty\dfrac{C_3(s^2)e^{ts^2}}{2}d(s^2)=g(t)-C_1$

$\int_0^\infty\dfrac{C_3(s)e^{ts}}{2}ds=g(t)-C_1$

$\int_0^\infty\dfrac{C_3(s)e^{-ts}}{2}ds=g(-t)-C_1$

$\mathcal{L}_{s\to t}\biggl\{\dfrac{C_3(s)}{2}\biggr\}=g(-t)-C_1$

$C_3(s)=2\mathcal{L}^{-1}_{t\to s}\{g(-t)\}-2C_1\delta(s)$

$\therefore u(x,t)=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(-t)\}e^{ts^2}\sinh xs~ds-2C_1\int_0^\infty\delta(s^2)e^{ts^2}\sinh xs~ds+\int_0^\infty C_4(s^2)e^{ts^2}\cosh xs~ds=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(-t)\}e^{ts^2}\sinh xs~ds-C_1\int_0^\infty\dfrac{\delta(s^2)e^{ts^2}\sinh xs}{s}d(s^2)+\int_0^\infty C_4(s^2)e^{ts^2}\cosh xs~ds=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(-t)\}e^{ts^2}\sinh xs~ds-C_1\int_0^\infty\dfrac{\delta(s)e^{ts}\sinh x\sqrt{s}}{\sqrt{s}}ds+\int_0^\infty C_4(s^2)e^{ts^2}\cosh xs~ds=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(-t)\}e^{ts^2}\sinh xs~ds-C_1\lim\limits_{s\to 0}\dfrac{e^{ts}\sinh x\sqrt{s}}{\sqrt{s}}+\int_0^\infty C_4(s^2)e^{ts^2}\cosh xs~ds=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(-t)\}e^{ts^2}\sinh xs~ds-C_1\lim\limits_{s\to 0}\dfrac{\dfrac{xe^{ts}\cosh x\sqrt{s}}{2\sqrt{s}}+te^{ts}\sinh x\sqrt{s}}{\dfrac{1}{2\sqrt{s}}}+\int_0^\infty C_4(s^2)e^{ts^2}\cosh xs~ds=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(-t)\}e^{ts^2}\sinh xs~ds-C_1\lim\limits_{s\to 0}(xe^{ts}\cosh x\sqrt{s}+2t\sqrt{s}e^{ts}\sinh x\sqrt{s})+\int_0^\infty C_4(s^2)e^{ts^2}\cosh xs~ds=C_1x+C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(-t)\}e^{ts^2}\sinh xs~ds-C_1x+\int_0^\infty C_4(s^2)e^{ts^2}\cosh xs~ds=C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(-t)\}e^{ts^2}\sinh xs~ds+\int_0^\infty C_4(s^2)e^{ts^2}\cosh xs~ds$

$u(0,t)=f(t)$ :

$C_2+\int_0^\infty C_4(s^2)e^{ts^2}~ds=f(t)$

$\int_0^\infty\dfrac{C_4(s^2)e^{ts^2}}{2s}d(s^2)=f(t)-C_2$

$\int_0^\infty\dfrac{C_4(s)e^{ts}}{2\sqrt{s}}ds=f(t)-C_2$

$\int_0^\infty\dfrac{C_4(s)e^{-ts}}{2\sqrt{s}}ds=f(-t)-C_2$

$\mathcal{L}_{s\to t}\biggl\{\dfrac{C_4(s)}{2\sqrt{s}}\biggr\}=f(-t)-C_2$

$C_4(s)=2\sqrt{s}\mathcal{L}^{-1}_{t\to s}\{f(-t)\}-C_2\delta(\sqrt{s})$

$\therefore u(x,t)=C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(-t)\}e^{ts^2}\sinh xs~ds+2\int_0^\infty s\mathcal{L}^{-1}_{t\to s^2}\{f(-t)\}e^{ts^2}\cosh xs~ds-C_2\int_0^\infty\delta(s)e^{ts^2}\cosh xs~ds=C_2+2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(-t)\}e^{ts^2}\sinh xs~ds+2\int_0^\infty s\mathcal{L}^{-1}_{t\to s^2}\{f(-t)\}e^{ts^2}\cosh xs~ds-C_2=2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(-t)\}e^{ts^2}\sinh xs~ds+2\int_0^\infty s\mathcal{L}^{-1}_{t\to s^2}\{f(-t)\}e^{ts^2}\cosh xs~ds$

Hence $u(x,t)=\begin{cases}2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(t)\}e^{-ts^2}\sin xs~ds+2\int_0^\infty s\mathcal{L}^{-1}_{t\to s^2}\{f(t)\}e^{-ts^2}\cos xs~ds&\text{when Re}(t)\geq0\\2\int_0^\infty\mathcal{L}^{-1}_{t\to s^2}\{g(-t)\}e^{ts^2}\sinh xs~ds+2\int_0^\infty s\mathcal{L}^{-1}_{t\to s^2}\{f(-t)\}e^{ts^2}\cosh xs~ds&\text{when Re}(t)\leq0\end{cases}$

Approach $2$: power series method

Similar to diffusion equation, inhomogenous boundary conditions (the subtraction method):

Let $u(x,t)=\sum\limits_{n=0}^\infty\dfrac{x^n}{n!}\dfrac{\partial^nu(0,t)}{\partial x^n}$ ,

Then $u(x,t)=\sum\limits_{n=0}^\infty\dfrac{x^{2n}}{(2n)!}\dfrac{\partial^{2n}u(0,t)}{\partial x^{2n}}+\sum\limits_{n=0}^\infty\dfrac{x^{2n+1}}{(2n+1)!}\dfrac{\partial^{2n+1}u(0,t)}{\partial x^{2n+1}}=\sum\limits_{n=0}^\infty\dfrac{x^{2n}}{(2n)!}\dfrac{\partial^nu(0,t)}{\partial t^n}+\sum\limits_{n=0}^\infty\dfrac{x^{2n+1}}{(2n+1)!}\dfrac{\partial^nu_x(0,t)}{\partial t^n}=\sum\limits_{n=0}^\infty\dfrac{x^{2n}}{(2n)!}\dfrac{\partial^nf(t)}{\partial t^n}+\sum\limits_{n=0}^\infty\dfrac{x^{2n+1}}{(2n+1)!}\dfrac{\partial^ng(t)}{\partial t^n}$

doraemonpaul
  • 16,488
-1

The diffusion equation is invariant wrt. to the gauge of the zero of the field. A boundary condition with $u(0,t))=c$ is equivalent to $(u-c)(t)=0$ that mimics reflection of the solution function around the point $(0,c)$. The underlying proces can be implemented physically by a mirrored solution of Brownian antiparticles that annihilate in pairs when they meet at the boundary, a particle carrying substance u, the antiparticle 2c -u with mean c.

On the other hand, a boundary value of the derivative is an imprinted current. In order not ot blow up, the total current must be zero or positive outward over the boundary.

A mixed boundary condition is a convex combination of absorption and prescribed source gradient at each boundary segment.

Fixing a start distribution at $t=0$ and one of the three types of boundary conditions (Dirchlet, Neumann, mixed - Newton/Robin) - depending on the points of the boundary, determine the unique time forward solution; any additional incompatible boundary condition makes the problem ill posed.

Numerically on a lattice, one cannot describe boundary value funcrtions with its tangent derivative implicitly given, and the normal derivative without destroying ellipticity of the solutiion at the boundary.

Ellipticity means nothing more than the fact, the value at time $t+dt$ at $x$ is a weighted mean over a small neighborhood at time $t$ .

Roland F
  • 5,122