Proof of Harten’s Lemma re the Convergence of TVNI Finite-Difference Schemes

An academic paper with a rather unusual history is that of the Israeli mathematician Ami Harten’s “High Resolution Schemes for Hyperbolic Conservation Laws.”  First published in the Journal of Computation Physics in 1983, it was republished in 1997 in the same journal, and is often cited with the later date.

At the time of republishing Peter Lax, who earlier had saved the computer from the hippie radicals, made the following statement about this paper in an introduction:

This paper was a landmark; it introduced a new design principle—total variation diminishing schemes—that led, in Harten’s hands, and subsequently in the hands of others, to an efficient, robust, highly accurate class of schemes for shock capturing free of oscillations. The citation index lists 429 references to it, not only in journals of numerical analysis and computational fluid dynamics, but also in journals devoted to mechanical engineering, astronautics, astrophysics, geophysics, nuclear science and technology, spacecraft and rockets, plasma physics, sound and vibration, aerothermodynamics, hydraulics, turbo and jet engines, and computer vision and imaging.

One point in the paper was a lemma concerning the convergence of TVNI (total variation nonincreasing) finite difference schemes.  Concerning the name of these schemes, Lax points out the following:

Harten originally called his schemes variation diminishing, abbreviated TVD; when Osher pointed out the usual meaning of these initials, the name was switched to total variation nonincreasing (TVNI), but was eventually settled on the more euphonious TVD.

The following is an expansion of Harten’s proof of the lemma.

Schemes which are total variation nonincreasing (TVNI) can be characterized as follows:
\sum_{j=-\infty}^{\infty}\left|u_{{j+1,n+1}}-u_{{j,n+1}}\right|\leq\sum_{j=-\infty}^{\infty}\left|\left(u_{{j+1,n}}-u_{{j,n}}\right)\right|

We can thus define
TV\left(u^{n}\right)=\sum_{j=-\infty}^{\infty}\left|\left(u_{{j+1,n}}-u_{{j,n}}\right)\right|

TV\left(u^{n+1}\right)=\sum_{j=-\infty}^{\infty}\left|u_{{j+1,n+1}}-u_{{j,n+1}}\right|

and substituting
TV\left(u^{n+1}\right)\leq TV\left(u^{n}\right)

Consider the general expression
u_{{j,n+1}}=u_{{j,n}}-C_{{j-1,n}}\left(u_{{j,n}}-u_{{j-1,n}}\right)+D_{{j,n}}\left(u_{{j+1,n}}-u_{{j,n}}\right)

where
C_{{j-1,n}}\geq 0

D_{{j,n}}\geq 0 C_{{j-1,n}}+D_{{j,n}}\leq 1

We should observe that our ultimate goal is to sum these values from negative infinity to positive infinity; thus, we can shift the index at will. The inequalities will still hold but the specific location in space may change. It is also worth noting that the coefficients may themselves change at different points in space.

Let us consider the next spatial step, to wit
u_{{j+1,n+1}}=u_{{j+1,n}}-C_{{j,n}}\left(u_{{j+1,n}}-u_{{j,n}}\right)+D_{{j+1,n}}\left(u_{{j+2,n}}-u_{{j+1,n}}\right)

Subtracting the previous spatial step from this yields
u_{{j+1,n+1}}-u_{{j,n+1}}=u_{{j+1,n}}-C_{{j,n}}\left(u_{{j+1,n}}-u_{{j,n}}\right)+D_{{j+1,n}}\left(u_{{j+2,n}}-u_{{j+1,n}}\right)-u_{{j,n}}+C_{{j-1,n}}\left(u_{{j,n}}-u_{{j-1,n}}\right)-D_{{j,n}}\left(u_{{j+1,n}}-u_{{j,n}}\right)

Some rearranging yields
u_{{j+1,n+1}}-u_{{j,n+1}}=\left(u_{{j+1,n}}-u_{{j,n}}\right)\left(1-D_{{j,n}}-C_{{j,n}}\right)+C_{{j-1,n}}\left(u_{{j,n}}-u_{{j-1,n}}\right)+D_{{j+1,n}}\left(u_{{j+2,n}}-u_{{j+1,n}}\right)

Taking the absolute value of both sides, we have
\left|u_{{j+1,n+1}}-u_{{j,n+1}}\right|=\left|\left(u_{{j+1,n}}-u_{{j,n}}\right)\left(1-D_{{j,n}}-C_{{j,n}}\right)+C_{{j-1,n}}\left(u_{{j,n}}-u_{{j-1,n}}\right)+D_{{j+1,n}}\left(u_{{j+2,n}}-u_{{j+1,n}}\right)\right|

At this point we observe that
C_{{j-1,n}}\geq 0
 D_{{j+1,n}}\geq 0
D_{{j,n}}+C_{{j,n}}\leq 1
1-D_{{j,n}}-C_{{j,n}}\geq 0

We can thus limit the absolute values and write the expression as follows:
\left|u_{{j+1,n+1}}-u_{{j,n+1}}\right|\leq\left|\left(u_{{j+1,n}}-u_{{j,n}}\right)\right|\left(1-D_{{j,n}}-C_{{j,n}}\right)+\left|\left(u_{{j,n}}-u_{{j-1,n}}\right)\right|C_{{j-1,n}}+\left|\left(u_{{j+2,n}}-u_{{j+1,n}}\right)\right|D_{{j+1,n}}

Taking the summation for both sides,
\sum_{j=-\infty}^{\infty}\left|u_{{j+1,n+1}}-u_{{j,n+1}}\right|\leq \sum_{j=-\infty}^{\infty}\left|\left(u_{{j+1,n}}-u_{{j,n}}\right)\right|\left(1-D_{{j,n}}-C_{{j,n}}\right)+\sum_{j=-\infty}^{\infty}\left|\left(u_{{j,n}}-u_{{j-1,n}}\right)\right|C_{{j-1,n}}+\sum_{j=-\infty}^{\infty}\left|\left(u_{{j+2,n}}-u_{{j+1,n}}\right)\right|D_{{j+1,n}}

Since, as we observed before, we can shift the indices (as the “centre” of the system is arbitrary with infinite boundaries) we can rewrite the above as follows
\sum_{j=-\infty}^{\infty}\left|u_{{j+1,n+1}}-u_{{j,n+1}}\right| \leq \sum_{j=-\infty}^{\infty}\left|\left(u_{{j+1,n}}-u_{{j,n}}\right)\right|\left(1-D_{{j,n}}-C_{{j,n}}\right)+\sum_{j=-\infty}^{\infty}\left|\left(u_{{j+1,n}}-u_{{j,n}}\right)\right|C_{{j,n}}+\sum_{j=-\infty}^{\infty}\left|\left(u_{{j+1,n}}-u_{{j,n}}\right)\right|D_{{j,n}}

in which case
\sum_{j=-\infty}^{\infty}\left|u_{{j+1,n+1}}-u_{{j,n+1}}\right| \leq \sum_{j=-\infty}^{\infty}\left|\left(u_{{j+1,n}}-u_{{j,n}}\right)\right|

Substituting, we have at last
TV\left(u^{n+1}\right)\leq TV\left(u^{n}\right)

Citation: Harten, A. (1997) “High Resolution Schemes for Hyperbolic Conservation Laws.” Journal of Computational Physics, Vol. 135, pp. 260-278.

Leave a Reply