.png)
1 month ago
4
A few days ago, we announced the results of an experiment we carried out together with a team of researchers from HSBC Bank.
Here’s how the market reacted:
Press enter or click to view image in full size
And honestly, it could have been even more 🙂
So what was this all about? In a nutshell, the experiment consisted of taking a dataset that’s usually fed into machine-learning models and transforming it using a quantum algorithm developed by IBM researchers. That algorithm belongs to the family of Quantum Machine Learning techniques known as the Projected Quantum Feature Map (PQFM).
Press enter or click to view image in full size
By the way, the experiment run in our ibm_torino Heron processor (133 qubits) using my beloved Estimator primitive. Stay tuned, because what is coming is even better than that 😉
Projected Quantum Feature Map (PQFM)
In Machine Learning, it’s very common to transform the data before training the models. The idea is to highlight properties and relationships that otherwise would be very hard to detect. When those properties remain hidden, the models perform poorly — their predictions are fuzzy or flat-out wrong. That’s exactly what HSBC (and anyone doing corporate bond trading) was experiencing.
The core idea of our experiment was simple: transform those datasets using a quantum algorithm (and a quantum computer). The PQFM algorithm, when applied to classical data, generates many more features and relationships than classical transformations can. And it does so by exploiting the strange, rich properties of quantum mechanics.
There are several ways to implement PQFM, but in this case, we used a Heisenberg-type ansatz circuit.
What’s an ansatz circuit? Think of it as a parameterized, reusable quantum circuit. The gates form a kind of “template,” and the parameters are the data itself — encoded into the angles of the gates. The purpose is straightforward: turn classical data into quantumly enriched classical data.
In a Heisenberg-type circuit, the dataset’s properties are encoded as angles of single-qubit gates Rx(angle), Ry(angle), and Rz(angle). Then, entangling gates like RXX, RZZ, and RYY are applied, weaving the qubits together. This is where the magic happens.
If my intuition is right, what happens here is that when we entangle the qubits carrying encoded data, we create relationships that simply cannot be expressed in classical terms. The raw data doesn’t change, but new correlations emerge — correlations that machine-learning models can actually understand and exploit.
Here’s the fascinating twist: in this experiment, we discovered that quantum noise makes those relationships even richer. Yes, the very thing we usually fight against turned out to be helpful this time. We still don’t know why. It was a complete surprise. Somehow, the noise smooths and normalizes these nonlinear relationships in a way that improves prediction power. The result? A 34% boost in model accuracy. That’s huge. That’s relevant. But…
Between Skepticism and Opportunity
I’m not a scientist by training (though definitely by vocation 😉), but I’ve been working alongside scientists long enough to know this outcome left many with mixed feelings. Science needs explanations. Surprises are welcome, sure, but they must eventually be understood. So I get why some in the community view these results with skepticism — or even dismiss the experiment altogether.
There’s still so much to learn about quantum mechanics, and by extension, quantum computing. What’s undeniable, though, is that this result is both impressive and relevant. The Heisenberg circuit could be applied to many use cases beyond bond trading: basically, any dataset that’s noisy, sparse, nonlinear, and high-dimensional.
Press enter or click to view image in full size
It would be foolish not to try. For comparison: in AI, there are countless behaviors that we don’t fully understand — yet no one doubts the value they bring to industry and society.
Promising Progress
This might be one of the most meaningful demonstrations yet of quantum computers operating at a utility scale (>100 qubits), with enormous potential for both science and industry.
What remains to be seen is whether this result was just a fortunate coincidence of uncontrolled factors — or whether we’ve cracked open a door to a new path of practical significance and scientific understanding.
