In the field of Quantum Machine Learning (QML), quantum kernel methods play an important role both for their effectiveness and for their mathematical formulation, which makes them a benchmark for QML analysis. However, these methods are affected by the phenomenon of exponential concentration, whereby the effectiveness of these approaches in learning problems decreases exponentially as the size of the problem increases. The two most commonly used methodologies are Quantum Fidelity Kernels and Projected Quantum Kernels (PQKs). Recent studies show that Quantum Fidelity Kernels do not allow for quantum advantage due to exponential concentration, while PQKs mitigate this phenomenon and are able to offer an advantage over classical models, at the cost of a reduction in expressiveness. A recent study proposes the Quantum Fisher Kernel (QFK), specifically designed to mitigate exponential concentration. The study theorizes QFK mathematically and simulates on synthetic datasets to support the proposed theory. This thesis demonstrates, through both ideal and non-ideal simulations, the effectiveness of QFK both in the context of classification tasks and in its ability to actually mitigate the exponential concentration problem, showing its applicability on current quantum computers with NISQ resources. Furthermore, the thesis enriches the methodology studied by placing it within the paradigm of quantum metric learning; the results obtained suggest that the computational load of the original implementation can be reduced by optimizing an alternative version of QFK. Finally, the research concludes with an experiment that demonstrates the competitiveness of QFKs compared to PQKs in obtaining the potential quantum advantage. This implies that it is possible to achieve a high quality kernel method through a purely quantum method as opposed to PQKs. This is a novel result that goes against what has been observed in studies regarding quantum kernel methods, which suggested a high level of difficulty.
Nel campo del Quantum Machine Learning (QML), i quantum kernel methods assumono un ruolo rilevante sia per la loro efficacia e sia per la loro formulazione matematica, che li rende un punto di riferimento per un’analisi relativa al QML. Tuttavia, tali metodi risultano afflitti dal fenomeno dell’exponential concentration, per cui, al crescere della dimensione del problema, la loro efficacia nei problemi di apprendimento diminuisce esponenzialmente. Le due metodologie prevalentemente utilizzate sono i Quantum Fidelity Kernels e i Projected Quantum Kernels (PQKs). Studi recenti mostrano che i Quantum Fidelity Kernels non consentono vantaggio quantistico a causa dell’exponential concentration, mentre i PQKs mitigano tale fenomeno e sono in grado di offrire un vantaggio rispetto i modelli classici, al costo di una riduzione dell’espressività offerta. Uno studio recente propone il Quantum Fisher Kernel (QFK), progettato nello specifico per mitigare l’exponential concentration. In tale studio il QFK viene teorizzato matematicamente e simulato su dataset sintetici a supporto della teoria proposta. La tesi, mediante simulazioni ideali e non ideali, dimostra l’effettiva efficacia di QFK sia nel contesto di compiti di classificazione e sia nell’effettiva mitigazione dell’exponential concentration, mostrandone l'applicabilità su attuali quantum computer con risorse NISQ. Inoltre il lavoro arricchisce la metodologia studiata inserendola all'interno del paradigma del quantum metric learning; i risultati ottenuti suggeriscono che il carico computazionale richiesto dall'implementazione originale può essere ridotto ottimizzando una versione alternativa di QFK. Infine, la ricerca si conclude con la realizzazione di un esperimento che dimostra la competitività di QFK rispetto i PQKs nell’ottenimento del potenziale vantaggio quantistico. Questo implica che è possibile ottenere un metodo kernel di alta qualità attraverso un procedimento puramente quantistico, senza l'uso dei PQKs. Questo è un risultato innovativo ed in controtendenza rispetto a quanto osservato negli studi nell'ambito dei quantum kernel methods che ne suggerivano l'elevata difficoltà.
Exponential concentration mitigation in Quantum Kernel methods: towards scalable quantum learning
LACAGNINA, MARCO
2024/2025
Abstract
In the field of Quantum Machine Learning (QML), quantum kernel methods play an important role both for their effectiveness and for their mathematical formulation, which makes them a benchmark for QML analysis. However, these methods are affected by the phenomenon of exponential concentration, whereby the effectiveness of these approaches in learning problems decreases exponentially as the size of the problem increases. The two most commonly used methodologies are Quantum Fidelity Kernels and Projected Quantum Kernels (PQKs). Recent studies show that Quantum Fidelity Kernels do not allow for quantum advantage due to exponential concentration, while PQKs mitigate this phenomenon and are able to offer an advantage over classical models, at the cost of a reduction in expressiveness. A recent study proposes the Quantum Fisher Kernel (QFK), specifically designed to mitigate exponential concentration. The study theorizes QFK mathematically and simulates on synthetic datasets to support the proposed theory. This thesis demonstrates, through both ideal and non-ideal simulations, the effectiveness of QFK both in the context of classification tasks and in its ability to actually mitigate the exponential concentration problem, showing its applicability on current quantum computers with NISQ resources. Furthermore, the thesis enriches the methodology studied by placing it within the paradigm of quantum metric learning; the results obtained suggest that the computational load of the original implementation can be reduced by optimizing an alternative version of QFK. Finally, the research concludes with an experiment that demonstrates the competitiveness of QFKs compared to PQKs in obtaining the potential quantum advantage. This implies that it is possible to achieve a high quality kernel method through a purely quantum method as opposed to PQKs. This is a novel result that goes against what has been observed in studies regarding quantum kernel methods, which suggested a high level of difficulty.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_03_Lacagnina_Tesi_01.pdf
accessibile in internet per tutti a partire dal 02/03/2027
Descrizione: Testo tesi
Dimensione
4.93 MB
Formato
Adobe PDF
|
4.93 MB | Adobe PDF | Visualizza/Apri |
|
2026_03_Lacagnina_Executive_Summary_02.pdf
accessibile in internet per tutti a partire dal 02/03/2027
Descrizione: Executive Summary
Dimensione
1.09 MB
Formato
Adobe PDF
|
1.09 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/253301