A regulator wants to fix a market with thousands of products. To do that she needs to know how every price affects every demand. What she actually has is this: an estimate so noisy that no single entry can be trusted.

The paper's claim is that when the market has enough large scale structure, she can still raise total surplus, with probability close to one, in every state the model allows.

The way in is to stop caring about individual entries. A few large scale patterns survive the noise. She aims her subsidies along those patterns and ignores everything else.

Here is the model. There are n firms. Each sells one product and sets its own price. Some pairs of products are substitutes, like two phones. Some are complements, like a phone and its earbuds.

All of that lives in one matrix, D. Entry i j says how demand for product i moves when the price of product j moves. Positive means substitutes. Negative means complements.

The regulator's tool is a vector, sigma. It is a small per unit subsidy, or tax, on each firm. Firms re-price, demand shifts, and surplus changes. Everything in the paper is first order around today's equilibrium.

If you think in control terms, this is a plant sitting at an operating point. Sigma is a small input, and D is the Jacobian.

The firms' first order conditions give a linear response. In units where every own price effect is minus one, the price change solves I minus D, times p dot, equals minus sigma.

Quantities follow as D times the price change. So the welfare effect runs through the inverse of I minus D. The true inverse is tame, because every eigenvalue of I minus D is at least one. The trouble is the estimate. Noise of size root n pushes eigenvalues of the estimated matrix across zero, so the plugged in inverse can be close to singular. With a noisy D she cannot even be sure of the sign.

Before any noise enters, one accounting identity constrains everything. The change in consumer surplus, plus half the change in producer surplus, equals the money spent.

A single monopolist sits here. A subsidy cuts the price by half its size, so consumers get fifty cents of each dollar, and the firm gets the full dollar.

Slide up the line until consumers get nothing, and producers get two dollars per dollar spent. Net of the spending, that is one dollar of new surplus per dollar. With consumers held harmless, that is the ceiling, and the theorem gets within epsilon of it.

Now the key move. D is symmetric, so it has an orthonormal basis of eigenvectors. Think of each eigenvector as a bundle of products.

Subsidize in the shape of one bundle, and only that bundle's price and quantity respond. The n coupled markets become n separate lanes.

In each lane the response is a scalar. Every eigenvalue is negative or zero, so from here on lambda means its size. The bundle's price falls by one over one plus lambda. Its quantity rises by lambda over one plus lambda.

Inside one lane, the whole market reduces to a supply and demand picture. The firms' pricing conditions give this line: the quantity change equals the price change plus the subsidy.

Households give the other line: the quantity change equals minus lambda times the price change. When lambda is small that line is flat, and the two lines meet far to the left. The subsidy shows up almost entirely as a price cut.

At lambda equal to one they meet in the middle. Half goes to price and half to quantity. That is the monopoly case.

When lambda is large the demand line is steep on these axes, which an economist would call very elastic demand, and the crossing point slides to the top. The price barely moves, and the quantity absorbs almost the entire subsidy. Firms sell more units at a positive markup, and that is where the new surplus comes from.

So where does a subsidy dollar go? It depends only on the eigenvalue of the lane you aim at. Consumers get one over one plus lambda, and that share falls as lambda grows.

Producers get two lambda over one plus lambda, which climbs toward two. The dashed line is the net new surplus.

At lambda equal to one you recover the monopoly split. At large lambda you are at the corner of the budget line: consumers unchanged, two dollars to producers, one dollar of net surplus. So the whole problem becomes: find a lane with a large eigenvalue.

Which lanes can she find? Look at the spectrum. If her errors are independent across entries, average zero, and bounded, the noise matrix has operator norm of order root n. That is the width of this gray bulk.

Now suppose complementarities of fixed strength run across many product categories. Then the top eigenvalue of D grows like n itself, and it sits far outside the bulk.

Perturbation theory does the rest. Eigenvalues move by at most the norm of the noise. And the Davis Kahan theorem says that eigenvectors attached to well separated eigenvalues barely rotate. The big lanes survive.

The running example makes this concrete. Products come in categories. Inside a category they are substitutes: one racket or another. Across categories they are complements: rackets and shoes.

In the symmetric version of this model, the top eigenvector is exactly the all ones vector. Every cross category link pushes that direction the same way, so its eigenvalue adds up across the whole market and grows like n.

Substitutes cannot play this role. A short argument from our survey, which the paper does not give, shows why. Write D as minus the identity plus a nonnegative matrix. Negative semidefiniteness caps that matrix's top eigenvalue at one, and Perron Frobenius then caps its bottom eigenvalue too. Every eigenvalue of D is stuck between minus two and zero, at any market size. So a huge lane needs complements at scale. The paper calls its full set of conditions significant structure.

Here is the rule as an algorithm. Take the noisy matrix and rescale it so the diagonal is minus one.

Compute its eigenvectors, and keep every one whose eigenvalue, in size, clears three quarters of a threshold. The threshold comes with the assumed environment. The rule does not estimate it from data. Call the span of the kept eigenvectors L hat.

Project the observed quantity vector onto L hat. Notice what this avoids. Individual eigenvectors may be badly estimated, and each comes with an arbitrary sign. A projection onto their span depends on neither.

Then divide by the squared length of that projection and multiply by the budget. By construction, the estimated spending is exactly on target.

One more point. Her quantity data can be as noisy as the signal itself. It does not matter, because the span she projects onto has far fewer than n dimensions, and independent noise spread over n directions almost vanishes there.

Theorem one says that under significant structure, for a large enough market, this rule delivers three things at once, with probability as close to one as you like, in every state of the world.

Consumer plus producer surplus rise by almost two dollars per dollar spent.

Consumer surplus barely moves, up or down, and that holds household by household.

And realized spending lands on the target. The guarantee is frequentist. There is no prior over markets. The only randomness is the noise in her data. One observation of ours: a subsidy proportional to the quantities already raises welfare a little in every state. What the theorem buys is the rate, the neutrality for consumers, and the budget.

How large is large enough? The paper's simulation answers with a picture. The horizontal axis is the size of the top eigenvalue, measured in units of root n.

Below one, the recovered eigenvector is close to random. The median gain is near zero, and close to half of the runs lose welfare. This is the spiked random matrix transition, a name the paper never uses. It sits at half the norm of the noise, well before the Davis Kahan bound says anything.

Well above it, alignment is perfect, and every dollar buys a dollar of surplus.

In between is the dangerous band, and this reading of the figure is ours. The median outcome is already positive, but losses are common. In the paper's figure, at least one run in ten still loses welfare when the top eigenvalue is twice root n, and one in twenty still loses at three and a half times root n. So the practical question is whether you can tell, from data, which side you are on. The paper proposes a check for that, and the second deep dive examines it.

Before going deeper, two questions. Pause after each one and answer it in your head. First. She aims a subsidy along a lane whose eigenvalue has size nine. Of each dollar spent, how much reaches consumers, and how much net surplus is created?

The answer. Consumers get one over one plus nine, which is ten cents. Producers get a dollar eighty. Net new surplus is ninety cents. And the budget identity checks: ten cents plus half of a dollar eighty is one dollar.

Second question. Her quantity data are as noisy as the quantities themselves. Why does the rule still spend almost exactly its budget?

The answer. The rule only uses the projection of the observed quantities onto the kept lanes. That space has far fewer than n dimensions, and independent noise spread over n directions has almost no energy inside it. The signal, by assumption, keeps a fixed share of its length there.

Three limits to keep in mind. Everything is first order, so these are small interventions. The gains go to producers, and a separate result constructs markets where consumers cannot be robustly guaranteed any share. And without significant structure, there are markets where no rule can guarantee more than a negligible return on the money spent. One remedy for the first point comes from the paper itself: the authority can claw the producers' gain back with fixed transfers.

If you want to go deeper, the next stop is the proof. The interesting part is the case where the large eigenvalues are close to each other. Then no single eigenvector can be recovered, and the argument works with the whole subspace.

After that, read the impossibility result and the data diagnostic. The notes page has each of these laid out with page links.