Differential constraint method for PDEs.

The method of differential constraints originated with Nikolai Yanenko in the 1960s as a systematic approach to constructing exact solutions of nonlinear partial differential equations by supplementing them with compatible overdetermined differential systems. This line of work was later consolidated and significantly expanded in the influential 1984 monograph by Yanenko, Sidorov, and Shapeev, which provided a rigorous formulation of the method and demonstrated its effectiveness, particularly in gas dynamics. Independently, during the 1980s, Peter J. Olver and Philip Rosenau developed closely related ideas under the name of “side conditions,” framing differential constraints as a unifying perspective that encompasses classical techniques such as separation of variables and symmetry reduction.

Geometric framework for DEs

To fix ideas (see this and, in particular, this), consider $n$ independent variables $x=(x^1,\dots ,x^n)$ and $m$ dependent variables $u=(u^1,\dots ,u^m)$. Recall:

• Jet Space $J^k(E)$: The manifold containing $(x,u)$ and all derivatives up to order $k$. Its dimension grows combinatorially.
• System of PDEs: Defined by a system of algebraic equations on the jet coordinates: ${\mathrm{\Delta }}_{\nu }(x,u^{(k)})=0$.
• The Manifold $\mathcal{E}$: These equations define a submanifold $\mathcal{E}\subset J^k(E)$.
  • Prolongations: A solution is a map $u=f(x)$. Its prolongation is the section of the jet bundle $j^{k}f:{\mathbb{R}}^n\to J^k$.
  • Geometric Definition: The graph of the prolonged solution must lie entirely within the submanifold $\mathcal{E}$.
• Vessiot distribution $\mathcal{V}$: The contact distribution establishes which submanifolds qualify as prolongations. We restrict that distribution on $\mathcal{E}$ and obtain the Vessiot distribution $\mathcal{V}$. $\mathcal{V}$ typically has dimension greater than $n$. This "excess dimension" reflects the fact that the general solution depends on arbitrary functions (infinite-dimensional solution space). Because the dimension of $\mathcal{V}$ is greater than $n$, there are many directions in which we can build an integral manifold.

Excess dimension and arbitrary functions

A solution is a geometric object—an $n$-dimensional integral manifold living inside $J^k$. At every point of that manifold, the tangent plane must lie inside $\mathcal{V}$: those are the "allowed directions" (they simultaneously satisfy the equation and the chain-rule contact conditions).

The dimension of $\mathcal{V}$ relative to $n$ tells us how constrained the construction is:

• $\dim (\mathcal{V})=n$ (no excess): The allowed space of directions is exactly as wide as the surface being built. There is only one $n$-dimensional plane to choose at each point—the surface is uniquely tracked. This is the situation for fully determined ODE systems where initial data at a point fixes the solution everywhere.
• $\dim (\mathcal{V})>n$ (excess dimension $e=\dim (\mathcal{V})-n$): The allowed space is strictly wider than $n$. At every point you must choose which $n$-dimensional sub-plane inside $\mathcal{V}$ to continue along. Because this choice is made at every continuous point of the domain, a smooth family of such choices is exactly a function. Each unit of excess dimension therefore corresponds to one arbitrary function in the general solution.

This is the geometric content of the Cartan–Kähler theorem: the Cartan characters $s_1,s_2,\dots$ measure precisely this excess dimension at successive stages of the integral element, telling you how many arbitrary functions (and of how many variables) the general solution requires. The differential constraint method exploits this picture directly: appending new equations shrinks $\mathcal{E}$, which shrinks $\mathcal{V}$, reducing the excess until $\dim (\mathcal{V})=n$ and the solution is pinned down.

The Strategy of Differential Constraints

To isolate specific solutions, we employ the strategy of Differential Constraints. We append $q$ new differential equations (constraints) to the original system.

This lowers the dimension of the manifold $\mathcal{E}$ and consequently lowers the dimension of the Vessiot distribution $\mathcal{V}$.
The Goal: We want to add enough constraints so that the dimension of the restricted Vessiot distribution becomes exactly $n$ (the number of independent variables).
Suppose we have reduced the distribution to rank $n$, the vector fields spanning this $n$-dimensional distribution are denoted as ${\mathbf{v}}_1,\dots ,{\mathbf{v}}_n$ (these act as the total derivatives $D_{x^i}$). For a solution submanifold to exist, these vector fields must commute (form an involutive distribution):

In the language of differential geometry:

• Failure: If $[{\mathbf{v}}_i,{\mathbf{v}}_j]\ne 0$, the distribution is non-involutive. The system is overdetermined and inconsistent. The curvature of the distribution is non-zero. Geometrically, if you integrate along $x$ then $t$, you get a different result than $t$ then $x$.
• Result: No solutions exist for this specific choice of constraints.

If the added constraints are carefully chosen such that the commutator conditions vanish ($[{\mathbf{v}}_i,{\mathbf{v}}_j]=0$), the distribution is Frobenius Integrable.

• Finite Type: The system is now of Finite Type.
• Result: The solution space is finite-dimensional. The value of the solution at a single point (along with finite derivatives) determines the solution everywhere.