⚛️ quantum physics

Assessing Projected Quantum Kernels for the Classification of IoT Data

This paper evaluates the effectiveness of the Projected Quantum Kernel (PQK) for classifying IoT-generated occupancy data, demonstrating that it achieves performance comparable to classical methods when using a compatible, feature-ready dataset and a shallow quantum encoding circuit.

Original authors: Francesco D'Amore, Luca Mariani, Carlo Mastroianni, Francesco Plastina, Luca Salatino, Jacopo Settino, Andrea Vinci

Published 2026-02-10

📖 3 min read🧠 Deep dive

CC BY 4.0

Original authors: Francesco D'Amore, Luca Mariani, Carlo Mastroianni, Francesco Plastina, Luca Salatino, Jacopo Settino, Andrea Vinci

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to recognize whether an office is "busy" or "empty" just by looking at sensor data (like temperature, light, and CO2 levels).

Usually, we use Classical Computers—the super-fast, reliable machines we use every day. But scientists are now trying to use Quantum Computers, which work on a completely different, much more complex set of rules.

This paper explores a specific way to use these quantum machines to solve this problem. Here is the breakdown of how they did it, using some everyday analogies.

1. The Problem: The "Language Barrier"

Think of your sensor data (temperature, light, etc.) as a book written in English. A classical computer reads English perfectly. However, a quantum computer speaks "Quantum-ese."

To make the quantum computer understand the data, you have to translate the English book into Quantum-ese. This is called "Encoding."

The researchers found a catch: if you make the translation too complicated (using very long, complex "sentences" or deep circuits), the quantum computer actually gets confused and performs worse. It’s like trying to translate a simple grocery list into Shakespearean English—it’s overkill and just makes things messy. They discovered that simple, short translations (shallow circuits) actually worked best.

2. The Solution: The "Quantum Projector" (PQK)

Quantum computers operate in a "Hilbert Space," which is a mathematical dimension so massive it’s almost impossible to visualize. Imagine trying to organize a massive, infinite library. It’s too much information to handle at once.

To solve this, the researchers used a method called the Projected Quantum Kernel (PQK).

The Analogy: Imagine you have a 3D object, like a complex sculpture. It’s hard to describe every tiny curve and shadow in words. But, if you shine a flashlight on that sculpture, it casts a 2D shadow on the wall. That shadow is a simplified version of the object that still tells you a lot about its shape.

The PQK acts like that flashlight. It takes the incredibly complex "quantum sculpture" of the data and projects it back down into a "classical shadow" that a standard computer can easily understand and use to make decisions.

3. The Results: A New Contender

The researchers compared three "players" in this game:

The Old Pro (Classical SVM): The reliable, standard way we do things now.
The Quantum Specialist (QK): A pure quantum approach.
The Hybrid (PQK): The "flashlight" method described above.

What they found:

The Hybrid is a heavyweight: The PQK (the hybrid method) was able to match—and sometimes even beat—the "Old Pro" (the classical computer).
Efficiency: The hybrid method was "leaner." It needed fewer "support vectors" (think of these as the key landmarks used to draw a map) to get the job done.
The "Happy Accident" of Noise: In quantum computing, "noise" (random errors) is usually the enemy. But the researchers found that a little bit of "shot noise" (randomness from not measuring enough times) actually helped the model learn better. It’s like how a little bit of "white noise" in a room can sometimes help you focus by masking distracting sounds.

Summary in a Nutshell

The researchers proved that we don't need to build massive, overly complex quantum "engines" to solve real-world problems like monitoring office occupancy. By using a clever "flashlight" technique to simplify quantum information, we can create smart, efficient models that are just as good as the classical ones we've used for decades, and potentially even better.

Technical Summary: Assessing Projected Quantum Kernels for the Classification of IoT Data

1. Problem Statement

The integration of Quantum Machine Learning (QML) into real-world applications faces two primary hurdles:

Dataset Compatibility: Most QML research relies on benchmark datasets (like MNIST) that require significant preprocessing (e.g., PCA or Autoencoders) to reduce dimensionality for current NISQ-era hardware. This preprocessing can distort the data's inherent nature and obscure potential quantum advantages.
The Dimensionality Challenge: While quantum computers can map data into exponentially large Hilbert spaces, this vast dimensionality can hinder the learning process and increase computational complexity.

The authors aim to evaluate how Projected Quantum Kernels (PQK) perform when applied directly to real-world, high-dimensional Internet of Things (IoT) sensor data without the need for dimensionality reduction.

2. Methodology

The study utilizes a binary classification task: predicting office room occupancy (occupied vs. non-occupied) based on environmental parameters (illuminance, blinds, lamps, humidity, $CO_2$ , and temperature).

A. Data Pipeline:
The researchers used the COGITO-IoT dataset (2,865 records). Unlike traditional approaches, this data was injected directly into quantum circuits, mapping each of the six features to a specific qubit.

B. Quantum Encoding Strategies:
The authors compared several feature maps (quantum circuits used to encode classical data):

Shallow/Simple: RotX (single-qubit rotations) and a custom 3D circuit (rotations around $X, Y,$ and $Z$ axes).
Entangled/Complex: 3D-CNOT (adding circular CNOT gates), IQP (Instantaneous Quantum Polynomial-time), Trotterized (simulating Hamiltonian evolution), and ZZFeatureMap.

C. Kernel Approaches:

Quantum Kernel (QK): Computes similarity via the direct overlap of quantum states ( $|\langle\phi(\mathbf{x}')|\phi(\mathbf{x})\rangle|^2$ ).
Projected Quantum Kernel (PQK): Projects the high-dimensional quantum state back into a classical feature space using a set of Pauli measurements ( $X, Y, Z$ ) on single qubits ( $m=1$ ), qubit pairs ( $m=2$ ), or both. This creates a classical representation that is then used in a Support Vector Machine (SVM).

D. Experimental Setup:
Experiments were conducted using a noiseless statevector simulator. The authors performed grid searches for hyperparameters ( $C$ and $\gamma$ ) and used 10-fold cross-validation. They also investigated the impact of shot noise (finite-sampling noise) by varying the number of circuit evaluations.

3. Key Contributions

Direct IoT-to-Quantum Mapping: Demonstrated a workflow where real-world IoT data is used without dimensionality reduction, providing a more "honest" assessment of QML performance.
Complexity Analysis: Provided a theoretical comparison showing that PQK has a lower quantum computational complexity—scaling linearly with the number of records $\mathcal{O}(N)$ —compared to QK, which scales quadratically $\mathcal{O}(N^2)$ .
Empirical Validation of Shallow Circuits: Challenged the intuition that deeper, more complex circuits are always better, showing that shallow architectures can be more effective for specific classification tasks.

4. Results

Performance Parity: PQK achieved accuracy and F1-scores comparable to, and in some cases superior to, the classical SVM-RBF model. Specifically, the 3D-CNOT encoding with $m=2$ measurements was highly competitive.
PQK vs. QK: For shallow feature maps, PQK significantly outperformed QK (statistically significant via paired t-tests). However, this advantage vanished with deeper, more complex feature maps.
Model Compactness: PQK produced more compact models, achieving up to an 18% reduction in the number of support vectors compared to classical SVM, suggesting better generalization potential.
The "Noise Benefit": Interestingly, the study found that a moderate amount of shot noise (low number of shots) can actually improve model performance, likely by helping the optimizer escape local minima or saddle points.

5. Significance

This paper is significant because it moves QML research away from "toy" datasets toward practical, real-world sensor data. It provides a roadmap for designing efficient QML pipelines by highlighting that encoding design and measurement strategy are often more critical than circuit depth. The findings suggest that PQK is a computationally efficient and effective candidate for edge computing and IoT applications where resource constraints (both in data dimensionality and quantum gate depth) are paramount.