Perché il meccanismo "Phase Kickback" funziona con l'algoritmo di stima della fase quantistica?


13

Probabilmente ho letto il capitolo La trasformata quantistica di Fourier e le sue applicazioni da Nielsen e Chuang (edizione del 10 ° anniversario) un paio di volte prima e questo ha dato questa cosa per scontata, ma oggi, quando l'ho guardata di nuovo, non lo fa ' Non mi sembra affatto ovvio!

Ecco lo schema circuitale per l'algoritmo di stima di fase:

inserisci qui la descrizione dell'immagine

Il primo registro con qubit è presumibilmente il "registro di controllo". Se uno dei qubit nel primo registro è nello stato | 1 cancello unitaria controllata corrispondente viene applicata al secondo registro . Se è in uno stato | 0 allora non ottiene applicato al secondo registro . Se è in una sovrapposizione dei due stati | 0 e | 1 t|1|0|0|1l'azione del corrispondente unitario sul secondo registro può essere determinata dalla "linearità". Si noti che tutte le porte agiscono solo sul secondo registro e nessuna sul primo registro. Il primo registro dovrebbe essere solo un controllo .

Tuttavia, mostrano che lo stato finale del primo registro è :

12t/2(|0+exp(2πi2t1φ)|1)(|0+exp(2πi2t2φ)|1)...(|0+exp(2πi20φ)|1)

Sono sorpreso dal motivo per cui riteniamo che ci sia stato un cambiamento nello stato del primo registro di qubit, dopo l'azione delle porte Hadamard. Lo stato finale del primo registro avrebbe dovuto essere

(|0+|12)t

non è vero? Dico questo perché il primo registro dovrebbe essere solo un controllo. Non capisco come o perché lo stato del primo registro dovrebbe cambiare quando agisce come controllo.

Inizialmente pensavo che considerare i fattori esponenziali come parte del primo stato qubit del registro fosse solo una convenienza matematica, ma poi non aveva senso. Lo stato di un qubit o un sistema di qubit non dovrebbe dipendere da ciò che è matematicamente conveniente per noi!

Quindi, qualcuno potrebbe spiegare perché cambia esattamente lo stato del primo registro dei qubit, anche quando agisce semplicemente come "controllo" per il secondo registro? È solo una comodità matematica o c'è qualcosa di più profondo?


Non una risposta, ma: cosa significherebbe per essere una "convenienza matematica", se non rappresentasse un vero cambiamento nello stato? O la matematica descrive accuratamente come cambiano gli stati quantistici, oppure no. In caso contrario, hai problemi più grandi di questo esempio. Se supponete che la matematica descriva accuratamente la fisica, allora la rappresentazione matematica non è solo conveniente: gli stati dei fili di "controllo" (citazioni di paura avanti) in realtà cambiano in questa subroutine. Va bene essere confusi sul perché, ma prima devi accettare che cambiano.
Niel de Beaudrap,

La matematica è esattamente quella spiegata in questa risposta: quantumcomputing.stackexchange.com/a/1791/1837 ma quella situazione è più semplice e forse più facile da capire
DaftWullie

@NieldeBeaudrap Bene, la mia domanda è precisamente "perché" cambia
Sanchayan Dutta,

@DaftWullie La matematica non sembra difficile. Facciamo solo un semplice esempio di gate controllato . Se il registro di controllo è nello stato | 1 poi viene applicata a | u dare exp ( 2 π i 2 0 φ ) | u . Ma stanno considerando che il fattore esponenziale di exp ( 2 π i 2 0 ϕ ) sia un fattore del qubit di controllo nel primo registro, ovvero exp ( 2 π iU20|1|uexp(2πi20ϕ)|uexp(2πi20ϕ) e non del secondo registro. La mia domanda è: perché sì? exp(2πi20ϕ)
Sanchayan Dutta,

cc @NieldeBeaudrap ^
Sanchayan Dutta

Risposte:


9

Immagina di avere un autovettore di U . Se hai uno stato come | 1 | u e di applicare i dati di prevenzione U ad esso, si esce e i & Phi; | 1 |uU|1|uU . La fase non è collegata a un registro specifico, è solo un fattore moltiplicativo globale.eiϕ|1|u

Ora usiamo una sovrapposizione sul primo registro: È possibile riscrivere questo come ( | 0 + e i & Phi; | 1 ) | u

(|0+|1)|u|0|u+eiϕ|1|u
(|0+eiϕ|1)|u
quindi appare sul primo registro, anche se è stato creato in qualche modo sul secondo registro. (Naturalmente questa interpretazione non è del tutto vera perché è stata creata da un gate a due qubit che agisce su entrambi i qubit).

Questo passaggio è al centro di molti algoritmi quantistici.

Perché non scriviamo ed appena sostengono che non è separabile?|Ψ=|0|u+|1(eiϕ|u)

Non si può semplicemente rivendicarlo, ma è necessario dimostrarlo matematicamente. Ad esempio, si può prendere la traccia parziale rispetto al secondo qubit, Per prendere la traccia parziale, scegliamo una base per riassumere. Per semplicità, scegliamo { | u , | u } doveu | u = 0

TrB(|ΨΨ|AB)=TrB(|00||uu|+|10|eiϕ|uu|+|01||uu|eiϕ+|11|eiϕ|uu|eiϕ)
{|u,|u}u|u=0 e . Quindi si ottiene Tr B ( | ΨΨ | A B )= | 00 | + e i ϕ | 11 | + e - i ϕ | 01 | +u|(eiϕ|u=eiϕ Questo è il grado 1 (e puoi vedere che la fase è apparsa sul primo registro), quindi lo stato non è impigliato. È separabile.
TrB(|ΨΨ|AB)=|00|+eiϕ|11|+eiϕ|01|+|11|

|0(|u)+|1(eiϕ|u)eiϕ

Come definisci "impigliato"? Secondo qualsiasi definizione, questo non è impigliato. Prova a prendere la traccia parziale, per esempio. Inoltre, suppongo che in genere non hai problemi a rimuovere una fase globale da un'intera espressione, rispetto a tenere quella fase su componenti diversi?
DaftWullie,

A(|0)A(eiθ|0)B(|0)A(eiθ|0)Beiθ(|0)A(|0)B(eiθ|0)A|0B(|0)A(eiθ|0)B

Immagino di avere un problema con il passaggio a "fasi globali" del genere. Non ci avevo mai pensato prima.
Sanchayan Dutta,

Non c'è differenza fisica . Pensaci in questo modo: quale esperimento faresti per distinguere i due? Se c'è una differenza fisica, ci deve essere un modo per distinguerli.
DaftWullie,

9

Una prima osservazione

Questo stesso fenomeno di "controllo" induce a cambiare stati in alcune circostanze anche con porte NON-controllate; in effetti, questa è l'intera base della stima degli autovalori. Quindi non solo è possibile, è un fatto importante riguardo al calcolo quantico che è possibile. Ha persino un nome: un "kick kick", in cui i qubit di controllo (o più in generale, un registro di controllo) subiscono fasi relative come risultato dell'azione attraverso qualche operazione su un registro target.

Il motivo per cui questo accade

Perché dovrebbe essere così? Fondamentalmente si riduce al fatto che la base standard non è in realtà così importante come a volte la descriviamo come tale.

Versione breve. Solo gli stati base standard sui qubit di controllo non sono interessati. Se il qubit di controllo è in uno stato che non lo è uno stato base standard, in linea di principio può essere modificato.

Versione più lunga -

Considera la sfera Bloch. È, alla fine, una sfera - perfettamente simmetrica, senza che un punto sia più speciale di nessun altro e nessun asse più speciale di qualsiasi altro. In particolare, la base standard non è particolarmente speciale.

|00[1000],|01[0100],|10[0010],|11[0001]
CNOT[1000010000010010].

{0,1}

{0,1}

|++[1000],|+[0100],|+[0010],|[0001].
|0012[1111],|0112[1111],|1012[1111],|1112[1111].
Again, we're using the column vectors on the right only to represent the states on the left. But this change in representation will affect how we want to represent the CNOT gate.

A sharp-eyed reader may notice that the vectors which I have written on the right just above are the columns of the usual matrix representation of HH. There is a good reason for this: what this change of representation amounts to is a change of reference frame in which to describe the states of the two qubits. In order to describe |++=[1000], |+=[0100], and so forth, we have changed our frame of reference for each qubit by a rotation which is the same as the usual matrix representation of the Hadamard operator — because that same operator interchanges the X and Z observables, by conjugation.

This same frame of reference will apply to how we represent the CNOT operation, so in this shifted representation, we would have

CNOT14[1111111111111111][1000010000010010][1111111111111111]=[1000000100100100]
which — remembering that the columns now represent X eigenstates — means that the CNOT performs the transformation
CNOT|++=|++,CNOT|+=|,CNOT|+=|+,CNOT|=|+.
Notice here that it is only the first, 'control' qubits whose state changes; the target is left unchanged.

Now, I could have shown this same fact a lot more quickly without all of this talk about changes in reference frame. In introductory courses in quantum computation in computer science, a similar phenomenon might be described without ever mentioning the words 'reference frame'. But I wanted to give you more than a mere calculation. I wanted to draw attention to the fact that a CNOT is in principle not just a matrix; that the standard basis is not a special basis; and that when you strip these things away, it becomes clear that the operation realised by the CNOT clearly has the potential to affects the state of the control qubit, even if the CNOT is the only thing you are doing to your qubits.

The very idea that there is a 'control' qubit is one centered on the standard basis, and embeds a prejudice about the states of the qubits that invites us to think of the operation as one-sided. But as a physicist, you should be deeply suspicious of one-sided operations. For every action there is an equal and opposite reaction; and here the apparent one-sidedness of the CNOT on standard basis states is belied by the fact that, for X eigenbasis states, it is the 'target' which unilaterally determines a possible change of state of the 'control'.

You wondered whether there was something at play which was only a mathematical convenience, involving a choice of notation. In fact, there is: the way in which we write our states with an emphasis on the standard basis, which may lead you to develop a non-mathematical intuition of the operation only in terms of the standard basis. But change the representation, and that non-mathematical intuition goes away.

The same thing which I have sketched for the effect of CNOT on X-eigenbasis states, is also going on in phase estimation, only with a different transformation than CNOT. The 'phase' stored in the 'target' qubit is kicked up to the 'control' qubit, because the target is in an eigenstate of an operation which is being coherently controlled by the first qubit. On the computer science side of quantum computation, it is one of the most celebrated phenomena in the field. It forces us to confront the fact that the standard basis is only special in that it is the one we prefer to describe our data with — but not in how the physics itself behaves.


-1

Great question.
I once asked this too, but it is not just a matter of mathematical convenience.
The controlled-U is an "entangling" gate.
Once there's entanglement, you cannot separate the state into "first register" and "second register".
Only think of these registers separately at the beginning, or when there's no entanglement. After there's entanglement, your best bet is to work through the mathematics (matrix multiplications) thoroughly, and you will indeed get the state given by Nielsen and Chuang.


Trying to upvote the question but need to wait until I have 15 reputation.

I can't see any entanglement. The output seems to be separable between the two registers. 12t/2(|0+exp(2πi2t1φ)|1)(|0+exp(2πi2t2φ)|1)...(|0+exp(2πi20φ)|1) is state of the first register whereas |u is the state of the second register.
Sanchayan Dutta

1
@Blue I do not write it as a full answer because I myself find it difficult to internalize the concept in my mind, anyway this is due to the "Phase Kick-Back" phenomenon, and it is actually also due to the fact that control and target are somewhat entangled. Try and give a read to section 2.2 of Mosca's PhD thesis, it's the best explanation I have found so far.
FSic

@F.Siciliano Okay, thank you. I'll give it a read
Sanchayan Dutta
Utilizzando il nostro sito, riconosci di aver letto e compreso le nostre Informativa sui cookie e Informativa sulla privacy.
Licensed under cc by-sa 3.0 with attribution required.