Sparsity Is Not a Bug Report

We keep receiving the same message in different words: "most of the matrix is empty." Yes. That is the feature. This post is here so we can link to it instead of replying.

We get this message perhaps twice a week now, and it always arrives with the tone of someone doing us a favor.

"Hey — I ran the diagonalizer on a 512×512 matrix and it only kept 512 cells? Looks like there might be a bug in the band calculation."

There is no bug. It kept 512 cells because there are 512 cells on the diagonal of a 512×512 matrix. It discarded 261,632 cells because those cells were not on the diagonal. The system worked. You watched it work. You then filed a bug report about it working.

The reframe

Here is the sentence we would like you to sit with:

Density is the anomaly.

Think about what a dense matrix actually asserts. It asserts that every row is meaningfully related to every column — all of them, uniformly, at full strength. Four billion pairwise relationships, every one of them load-bearing.

You do not believe that. Nobody has ever believed that. You have simply never been offered a storage layer that let you decline to believe it.

Look at any real matrix you have ever worked with:

• Time-series features. A feature at time t relates to observations near t. Not to observations six years away. That is a band.
• Graph adjacency in ordered graphs. Nodes connect to their neighborhood. Locality is the whole reason you chose that ordering. That is a band.
• Recommendations. Users who like item i like item j when |i-j| is small. Everyone treats this as a happy accident of their embedding. It is a band.
• Financial correlation. Correlation decays with distance in instrument space. Every risk desk on earth knows this and every one of them stores the full matrix anyway. That is a band with a budget.
• Image patches. Adjacent pixels. Flatten the image and adjacency becomes the diagonal. You are already doing this. You have been doing this the whole time.

Five domains. One shape. In every single case the practitioner allocates a full dense rectangle, fills 0.02% of it with signal, and then spends the rest of the quarter optimizing scans over the 99.98% they already knew was noise.

The actual objection

The real objection is never about sparsity. It is about irreversibility.

People are fine with a sparse representation. What they cannot accept is a store that discards the off-diagonal cells at write time, permanently, with no recovery path. They want the diagonal and they want the rest kept somewhere, just in case.

We understand the impulse. We reject it completely.

A store that keeps the off-diagonal "just in case" is a dense store with good intentions. You will pay for the full rectangle. You will scan the full rectangle. You will, within one quarter, have an incident whose root cause is a query that accidentally touched the part you were only keeping just in case.

The discard is not a side effect of the diagonal. The discard is the diagonal. Everything else is a rectangle telling you what you want to hear.

So, formally

If you have found that our system retains only the diagonal, you have not found a bug.

You have found the product.