Damping in perpetual demand lending pools
A note on what the γ-dial adds to the single-pool-vs-two-pools question that Chitra, Diamandis, Sheng, Sterle, and Yusubov pose in DeFi.
Chitra et al.’s Perpetual Demand Lending Pools reaches, in Appendix B, for exactly the object this site is about: the Schur complement of a partitioned covariance. But they reach for it in its undamped form, and they use it to answer a binary question — merge two pools or keep them apart. This note works out what the damped form would add: a continuous architecture dial whose setting is fixed by how well the cross-pool covariance can actually be estimated — which, in DeFi, is the whole problem.
The setup, in their notation
A PDLP holds a portfolio of risky assets with covariance $\Sigma$. To ask whether two pools should be one, Appendix B splits the assets into a pool-1 block and a pool-2 block with no assets in common, and writes
$A$ is pool 1’s own covariance, $C$ is pool 2’s, and $B$ is the cross-pool block — how the two pools’ assets co-move. The mean-variance delta hedge for a merged pool uses the conditional covariances, the Schur complements
while an isolated pool hedges against $A$ (or $C$) alone, as if the other pool did not exist. Claim B.1 then gives a spectral sufficient condition — $\sigma_{\max}(A) < \sigma_{\min}(\Sigma/A)$ — under which the single merged pool delta-hedges better than the two isolated ones, and the authors note these conditions are “similar to hierarchical risk-parity methods” ([Cot24], [LdP16]).
The decision is the two endpoints of the γ-bridge
Read the two options through the bridge and they are not two architectures but the two ends of one. The isolated pool keeps only its own block $A$; the merged pool conditions fully on the other through $A - B C^{-1}B^{\mathsf T}$. Interpolating the coupling gives a one-parameter family of partially merged pools:
At $\gamma=0$ the pools are split — no cross-pool hedging, robust but blind to real diversification. At $\gamma=1$ the pools are fully merged — maximal diversification credit, but the hedge now leans on every entry of $B$ being correct. Chitra et al. test which of these two endpoints is better; they never sit at an interior point. This is precisely the $\gamma\in\{0,1\}$ binary that the literature map marks for KFAC and for HRP-vs-min-variance — the same dial, left at its stops.
Why the interior is the right place to be here
The binary choice assumes $B$ is known. In a PDLP it is the opposite of known. The pools are disjoint by construction, so $B$ is the cross-block covariance — the entries with the fewest joint observations and the most estimation noise. On-chain return histories are short, pools rebalance fast, and the asset count can rival or exceed the number of return samples. Trusting $B$ completely (merge) over-fits noise; discarding it (split) throws away signal. The right answer is to trust it in proportion to how well it is measured — which is exactly what $\gamma$ encodes.
The reliability that sets $\gamma$ is a James–Stein / Wiener ratio of signal to signal-plus-noise. In the two-block case with conditional correlation $\rho$ estimated from $n$ observations,
Strong, well-sampled cross-pool coupling pushes $\gamma^\star \to 1$ and the merge deserves trust; weak or undersampled coupling pushes $\gamma^\star \to 0$ and keeping the pools apart was right. The undersampled DeFi regime — small $n$ — pulls $\gamma^\star$ toward $0$: the theory predicts pools should usually stay mostly split, merging only as the cross-coupling proves both strong and stable. That is a graded, testable prediction the threshold test in Claim B.1 cannot make — it can only flip from “split” to “merge” the instant the spectral condition is met.
What it would change operationally
Concretely, a damped PDLP delta-hedges pool 1 against $A(\gamma^\star)$ rather than against either $A$ or $\Sigma/C$ — a soft sharing of risk across pools, with the sharing fraction a function of data, not a governance flag. This speaks directly to the paper’s closing observation that GMX V2’s dynamic-pricing upgrade “does not take into account pool asset covariance”: a target-weight mechanism that sets inter-pool collateral sharing in proportion to $\gamma^\star$ would put exactly that covariance information back, at the reliability-weighted amount, and would degrade gracefully as histories shorten instead of flipping discretely.
It is the same damping as everywhere else
None of this is new machinery. $A - \gamma B C^{-1}B^{\mathsf T}$ is the augmented sub-covariance of Schur Complementary Allocation, and the identity that the damping a portfolio needs when assets outnumber returns is, term for term, the damping a spatial model needs when stations outnumber observations is the subject of Two Sides of Schur Damping. A perpetual demand lending pool is just an unusually sharp instance of the outnumbered regime — short histories, fast pools — which is why the interior of the dial, rather than its endpoints, is where its architecture question is most likely to live.