Faithful Decoding

Faithful Decoding: Exact Dimensionality Reduction for High-Dimensional Equilibrium Systems

In the world of complex mathematical modeling—specifically when solving for equilibrium in economic or financial systems—we often face a brutal trade-off. You can have a high-dimensional model that captures every nuance of the system. Or, you can compress it into a low-dimensional version that actually finishes running. Traditionally, this compression is "lossy." This means you sacrifice precision to gain speed.

Current state-of-the-art methods, such as state aggregation or orthogonal projection, attempt to simplify these massive systems. They group similar states together or project them onto a lower-dimensional subspace. While these techniques are indispensable for making high-dimensional problems tractable, they inherently degrade information. You aren't just simplifying the math. You are fundamentally altering the solution.

This paper proposes a way out of that compromise. By treating these problems as dynamical systems and exploiting a mathematical property called "strong semiconjugacy," the authors demonstrate a new way to transform problems. You can turn a high-dimensional problem into a much smaller one. Then, you can "decode" the answer back to the original system with zero loss of accuracy.

The cost of losing information

When we deal with high-dimensional equilibrium objects—things like wealth distributions or firm size across millions of possible states—the computational burden scales poorly. The standard response is discretization and state aggregation. This involves approximating a continuous or high-dimensional space with a coarser, manageable grid.

The problem is that these are approximation methods. In many sensitive applications, particularly in finance or stochastic optimization, these "small" errors aggregate. If you are using stochastic approximation (a class of algorithms used to find roots of functions by using noisy samples), the errors become even more insidious. For instance, when trying to solve models involving heterogeneous beliefs, practitioners often encounter a "max-of-sample-means" issue. Because the model requires taking the maximum over several noisy estimates, the resulting estimate is systematically biased upward. As shown in, attempting to apply standard stochastic approximation directly to the high-dimensional operator $\hat{S}$ leads to a solution that drifts significantly above the true equilibrium.

Strong semiconjugacy as an architectural primitive

The authors' approach avoids the bias and the inaccuracy by reframing the problem through a specialized version of topological conjugacy. Topological conjugacy is a property where two systems are essentially identical under a continuous, reversible mapping. Instead of approximating the operator $S$ (the high-dimensional system), they construct a pair of maps: an encoder ($E$) and a decoder ($D$).

The architecture relies on a specific structural requirement. The high-dimensional operator $S$ and the low-dimensional operator $T$ must satisfy $S = D \circ E$ and $T = E \circ D$. This is what the authors call "strong semiconjugacy." It is a significant tightening of the standard mathematical definition of semiconjugacy. Standard semiconjugacy usually only requires surjective mappings (mappings where every element in the target is covered). By enforcing this dual-composition structure, the authors ensure that the fixed points (the steady-state solutions) of the two systems are in a one-to-one correspondence.

The workflow for a practitioner looks like this: 1. Encode: Map the high-dimensional state space into a lower-dimensional space $W$ using the encoder $E$. 2. Solve: Perform the heavy lifting—iterative solvers or stochastic approximation—entirely within the low-dimensional system $(W, T)$. 3. Decode: Once the low-dimensional fixed point $\bar{w}$ is found, apply the decoder $D$ to map it back to the exact solution $\bar{v}$ for the original high-dimensional system.

Crucially, the authors prove that this doesn't just preserve the solution. It also preserves the stability. If the simplified system is stable, the decoded high-dimensional system is guaranteed to be stable as well.

Massive speedups without the drift

The empirical results suggest that this isn't just a theoretical curiosity. It is a massive performance lever. In a real option problem—a model used to value the flexibility to delay investment decisions—the authors compare direct iteration on the high-dimensional operator $S$ against iteration on the low-dimensional operator $T$.

The paper reports that the speedup grows roughly quartically with the size of the transient state grid. Specifically, at a grid size of $m=9$, the authors demonstrate speed gains of up to 70,000 times .

This represents a massive reduction in the time required to reach a solution. Even more importantly, the accuracy remains intact. The authors validate the "faithfulness" of the decoding by measuring the sup-norm discrepancy ($\delta$) between the decoded low-dimensional solution and the direct high-dimensional solution. Across various grid sizes, they report $\delta < 9.9 \times 10^{-7}$. This confirms that the results match within the limits of floating-point error.

Furthermore, the method solves the bias problem mentioned earlier. In the asset pricing application with heterogeneous beliefs, applying the ROOT-SA algorithm (a variance-reduced stochastic approximation method) to the transformed operator $T$ results in an approximation that closely tracks the true equilibrium price function .

This stands in stark contrast to the biased, upward-drifting paths seen when applying the same logic to the original operator .

Where the framework hits a wall

Despite the impressive benchmarks, there are clear boundaries to where this method is applicable.

First, the "search space" for applicable problems is narrower than that of traditional lossy methods. You cannot simply throw an arbitrary high-dimensional system into an encoder-decoder pipeline. The system must possess the specific structural properties required for strong semiconjugacy. If your system doesn't admit a decomposition where $S = D \circ E$ and $T = E \circ D$, this entire framework is off the table.

Second, the complexity of finding the encoder and decoder themselves is not addressed. For a practitioner, the "hidden cost" is the engineering effort required to derive these maps. While the paper provides examples for real options and asset pricing, it does not provide a general-purpose heuristic for discovering these transformations in a brand-new domain. If you cannot analytically identify the $E$ and $D$ maps, you cannot ship the solution.

The verdict

Is this worth a prototype? If you are working on high-dimensional equilibrium models—particularly those involving persistent and transient state components (like macro-econometric models or multi-asset pricing engines)—the answer is a resounding yes.

The ability to achieve a 70,000x speedup while maintaining exactness is a tier of performance gain that justifies significant upfront engineering. More importantly, the method provides a mathematically sound way to use stochastic approximation without falling into the trap of systematic upward bias. However, do not expect this to be a "plug-and-play" library. This is a structural tool. You should only move toward this approach if you can prove your system's operators satisfy the strong semiconjugacy conditions. If you can, the efficiency gains are too large to ignore.

Figures from the paper