%\documentstyle[aps,epsfig,pra]{revtex}
%\documentclass[preprint,pra,aps]{revtex4}
\documentclass[twocolumn,prl,aps]{revtex4}
%
%\documentclass[pra,aps]{revtex}
%\usepackage{epsfig,wick}
\usepackage[matrix,frame,arrow]{xypic}
\usepackage{amsfonts,amsmath,amssymb}
\usepackage{graphicx,epsfig}
\input{Qcircuit}
\newcommand\beq{\begin{equation}}
\newcommand\eeq{\end{equation}}
\newcommand\bea{\begin{eqnarray}}
\newcommand\eea{\end{eqnarray}}
\def\eps{\epsilon}
\newcommand{\QED}{\Box}
\newcommand{\ba}{\begin{array}}
\newcommand{\ea}{\end{array}}
\newtheorem{theo}{Theorem}
\newtheorem{defi}{Definition}
\newtheorem{lem}{Lemma}
\newtheorem{exam}{Example}
\newtheorem{prop}{Property}
\newtheorem{propo}{Proposition}
\newtheorem{cor}{Corollary}
\newtheorem{conj}{Conjecture}
\renewcommand{\>}{\rangle}
\newcommand{\<}{\langle}
\begin{document}
\noindent {\bf \em Supplementary Information} \vskip 0.2cm
{\it The need for ancilla verification}---In the theory of quantum
fault tolerance there are two
%PA3: general-> well-known
well-known methods for performing fault-tolerant error correction.
The first, due to Shor \cite{shorver} and DiVincenzo and Shor
\cite{shordiv}, makes use of ancillae encoded in the classical
repetition code informally called ``cat'' states. The second,
which is due to Steane \cite{Steane97} and applies to the
Calderbank-Shor-Steane (CSS) class of quantum codes, uses as
ancillae logical $|0\rangle$ and $|+\rangle$ states encoded in the
same code as the data. In both methods, the preparation of these
ancillae by directly executing their encoding circuit is not in
general a fault-tolerant procedure, and extra steps are needed
before the ancillae are considered pure enough to be coupled to
the data. This purification procedure, also known as
\emph{verification}, is performed by running parity checks on the
ancilla using additional ancillary verifier states.
Ancilla verification is typically done non-deterministically by simply rejecting ancillae that fail the parity checks of verification and starting anew. This procedure however is disadvantageous when measurement is slow since the verified ancilla qubits will have to be stored for long times in memory before measurements that are part of verification finish. Alternatively, verification can be done deterministically by correcting errors in the verified ancilla and thus avoiding post-selection.
%PA2: added next comment
(Note that the deterministic protocol in Fig.$\,$4 in \cite{SDT}
that does not correct errors in the ancillae is unusable when
measurement is slow: In this protocol, an element of
non-determinism exists in {\em which} ancillae are to be coupled
to the data depending on the verification measurement outcomes.)
%
The price for a deterministic verification procedure is a penalty in efficiency: A larger number of verifier
qubits and verification operations are needed compared to the
non-deterministic procedures (e.g., compare the deterministic
protocol in Fig.$\,$9 in \cite{preskill:faulttol} with the
non-deterministic one in Fig.$\,$11 in \cite{AGP}). And this
increase in qubit and operation overhead translates into a decrease in the accuracy threshold as compared to
non-deterministic verification---this is the reason why
non-deterministic verification has been used hitherto in almost
all threshold studies.
We will here describe how ancilla verification can be avoided
altogether and replaced by suitable decoding of the encoded
ancillae {\em after} these interact with the data during error
correction. We find that not only does this method improve the
efficiency of present schemes that use verification, but it also
leads---in the examples of interest we have studied in detail---to
an improvement in the accuracy threshold as compared even to
non-deterministic verification procedures.
%PA3: Added this new paragraph by combining pieces scattered later in the discussion
We will first discuss the case of distance-3 quantum codes
correcting arbitrary errors on any one qubit in the code block. In
the end we will discuss the generalization to codes of higher
distance. Throughout, we use the shorthand notation $X\equiv
\sigma_{\rm x}$ and $Z\equiv \sigma_{\rm z}$ and we also use
superscripts to denote the qubit on which an operator acts (e.g.,
$Z^{(i)}$ denotes a Pauli $\sigma_{\rm z}$ operator acting on the
$i$-th qubit).
{\it Ancilla decoding instead of verification}---The intuition
leading to this method is guided by a few basic observations on
how fault-tolerant error correction is implemented.
The first observation is that ancilla verification checks
particular error patterns and does not need to protect against
arbitrary multi-qubit errors. In particular, we are concerned with
faults that appear in first order in the encoding circuit and
propagate to cause multi-qubit errors at the output of the ancilla
encoder that our distance-3 code cannot correct. Verification
works by exactly trying to detect (or correct, in deterministic
schemes) these particular errors appearing in first order, with
the additional condition that the verification step should not,
also in first order, introduce multi-qubit errors in the ancilla
that is being verified.
The second observation is that the ancilla after interacting with
the data block remains, ideally, in a {\em known} logical state.
For example, after the cat state $|\bar{+}\rangle_{\rm rep} \equiv
(|00\dots 0\rangle + |11\dots 1\rangle) / \sqrt{2}$ interacts with
the data, it remains encoded in either $|\bar{+}\rangle_{\rm rep}$
or $|\bar{-}\rangle_{\rm rep}$ (the latter differs by the sign
having flipped from $+$ to $-$ in the above superposition).
Similarly in Steane's error correction method, an ancilla prepared
in the logical $|0\rangle$ state remains in the same state after
interacting with the data. This is more information than just
knowing that the ancilla state is
%PA2: rephrased the following
protected by a code---we have additional information about what the state should be.
The idea of the new method is to couple the encoded ancilla to the data without attempting any verification, followed by a procedure that is run afterwards on it and allows us to learn and \emph{invert} possible multi-qubit errors produced by the encoder and having propagated to the data. This procedure could be implemented in various ways, but it seems that the most efficient (and eye-pleasing) one is via a suitable decoder. A schematic of this method is given in Fig.$\,$\ref{ECwoVER}.
%
\begin{figure}[htb]
\begin{center}
\includegraphics[width=7.1cm,trim=0 0 0 0]{ECwoVER}
\caption{\label{ECwoVER}
\footnotesize{
Schematic of fault-tolerant error correction without ancilla verification for distance-3 codes. The ancilla is encoded by starting with ancillary qubits in a reference state (e.g., $|0\rangle^{\otimes k}$), the state to be encoded ($|\psi_{\rm enc}\rangle$), and running the unitary encoding circuit ($\mathcal{D}^{-1}$). After interacting with the data block (schematically shown as a {\sc cnot}), the ancilla is decoded using a unitary decoding circuit ($\mathcal{D}$). The final measurements yield syndrome information ($\{x_i\}$) and the eigenvalue of a logical operator ($\bar{L}$).
}
}
\end{center}
\end{figure}
To prove that this error-correction (EC) method is fault tolerant,
%(for a suitable choice of the decoding circuit)
we need to show that properties $0',0-2$ in \cite{AGP} are satisfied.
Properties $0$ (EC without faults takes an arbitrary input to the
code space) and $1$ (EC without faults corrects single-qubit
errors in the data block) are true since, without faults,
verification is superfluous and the measurement outcomes after
decoding will give the necessary syndrome information to perform
ideal error correction. Properties $0'$ (EC with one fault always
produces an output deviating by at most a weight-one operator from
the code space) and $2$ (EC with one fault produces a block with
at most one error if the input has no errors) require proof. As
with the known error-correction methods that use ancilla
verification, showing that property $0'$ holds is made easier
after understanding why property $2$ is satisfied. And as we will
see, satisfying property $2$ will determine how our decoding
circuit must be designed.
What we need to show is that a single fault in the EC circuitry cannot cause more than one error in the corrected data block (which for property $2$ initially has no errors). Consider first the case where the fault occurs during the interaction with the data. Since this interaction is done with transversal gates, no gate operates on more than one qubit in the same block. Therefore a single fault cannot cause more than one error in the data. An error can also appear in the ancilla block, but this does not prevent successful correction: If cat states are used the syndrome extraction is repeated, or in Steane's method the errors in data and ancilla blocks occur always in the same qubit in the code block and a single syndrome extraction can be trusted.
The second case is when the fault occurs inside the ancilla
encoder. In general the encoding circuit has no special structure
%DDV appear in
and more than one errors can appear in the produced encoded
ancilla. The encoder can however be designed so that no logical
error appears at the output due to a single fault inside it and,
furthermore, the possible error patterns caused by all such first
order fault events are known. Of interest are multi-qubit errors which
may propagate to the data via the ancilla-data interaction.
However, letting ancilla and data interact without verification as
in Fig.$\,$\ref{ECwoVER} is permissible since the information at
the output of the following decoder will allow us to diagnose
whether such error propagation did occur and invert it. To
understand how this is possible, it is useful to think of the
decoding circuit as performing error correction while
simultaneously decoding its input block to a qubit. Since in the
case considered the decoder is fault-free, both processes are
executed ideally. The decoder then performs the mapping
%
\begin{equation}
\label{prop1}
E_i|\bar{0} \rangle \rightarrow |i\rangle \otimes |0\rangle \, ; \hspace{0.2cm}
E_i|\bar{1} \rangle \rightarrow |i\rangle \otimes |1\rangle \, ,
\end{equation}
%
\noindent where $\{E_i\}$ is the set of all single-qubit Pauli errors that our distance-3 code corrects, $|\bar{0} \rangle$ ($|\bar{1} \rangle$) is the ideal logical $|0 \rangle$ ($|1 \rangle$) state, and $i$ is syndrome information that, after decoding, indicates the error $E_i$ (if the code is not {\em perfect} and the basis $\{E_i|\bar{0} \rangle, E_i|\bar{1} \rangle \}$ does not span the whole Hilbert space,
%PA2: rephrased the following
the action of the unitary decoder can be extended to include some non-correctable errors).
Now, if we decode a $n$-qubit cat state, then measuring the qubits carrying the syndrome will allow us to learn all eigenvalues of the $Z^{(i)}Z^{(j)}$ code stabilizers for $1\leq i=j-1 \leq n-1$. Hence we can diagnose whether multi-qubit $X$ errors were produced in the encoder and can invert them by updating the Pauli frame of the data block. In Steane's method, the syndrome information will diagnose any $X$ (resp. $Z$) errors in the logical $|0\rangle$ (resp. $|+\rangle$) ancilla that propagate to the data. In addition, measuring the decoded qubit (the second tensor-product factor in Eq.$\,$(\ref{prop1}); denoted by $\bar{L}$ in Fig.$\,$\ref{ECwoVER}) will reveal the eigenvalue of the {\em logical} $Z$ (resp. $X$) operator. This resolves the ambiguity about the error causing a particular syndrome since, for $E_i$ and $E_j$ to have the same syndrome, $E_i^{\dagger} E_j = \bar{O}$ where $\bar{O}$ is either trivial (equal to the identity or $\bar{L}$) or anti-commutes with $\bar{L}$. Thus, either $E_i$ and $E_j$ are equivalent up to an element of the ancilla stabilizer, or lead to orthogonal decoded states which allows distinguishing them. Again, whenever multi-qubit errors are detected, the appropriate Pauli-frame change is done on the data block to invert them.
The final case to consider is when a single fault occurs in the
decoding circuit. In the case discussed above we were concerned
with correcting multi-qubit errors caused by a single fault in the
encoder and subsequently propagating to the data. But we also need
to guarantee that such a corrective step is not mistakenly taken
due to a single fault inside the decoder (which can cause no
errors to the data block). Thus, to satisfy property $2$, we must
ensure that no single fault inside the decoding circuit can give
the same syndrome
%PA2: added the following in parenthesis
(including the eigenvalue of the logical operator $\bar{L}$) as
any of the multi-qubit errors which a single fault inside the
encoder can produce. Constructing the decoding circuit to meet
this condition must be done with care given the chosen ancilla
encoder.
In the examples section that follows such decoder designs are
shown for some very frequently used cases: fault-tolerant
measurements using four- and seven-qubit cat states and Steane's
error-correction method for the [[7,1,3]] code. We conjecture that
an appropriate decoding circuit can be found for any distance-3
code. In any case, a less efficient but general solution is always
possible: We can further encode the ancilla after interacting with
the data in the two-bit classical repetition code and then decode
the two sub-blocks separately as shown in Fig.$\,$\ref{ECwoVER2}.
In this circuit, when the syndrome for $X$ errors at the two
%DDV agrees -> agree
sub-blocks agree, then we can be confident that, to first order,
any detected multi-qubit $X$ error has occurred during ancilla
encoding and has propagated to the data. Otherwise, if the
syndromes for $X$ errors disagree, we conclude that, again to
%DDV moved again
first order, a fault has happened in one of the two decoded
sub-blocks and no multi-qubit $X$ error has propagated to the
data. The syndrome for $Z$ errors (revealing $Z$ errors initially
in the input data block) can be obtained by taking the {\em
parity} of the syndromes at the two sub-blocks.
\begin{figure}[htb]
\begin{center} \vskip 0.2cm
\includegraphics[width=8.5cm,trim=0 0 0 0]{ECwoVER2}
\caption{\label{ECwoVER2}
\footnotesize{
To ensure that a single fault during ancilla decoding cannot be mistaken for a fault inside the ancilla encoder leading to the same syndrome, the ancilla is encoded in the two-bit classical repetition code and then the two sub-blocks are decoded separately. If the syndrome in {\em both} sub-blocks indicates a multi-qubit $X$ error then we are confident a fault in the ancilla encoder occurred and propagated to the data. Otherwise the fault must have occurred during the decoding procedure.
}
}
\end{center}
\end{figure}
For such designs that satisfy property $2$,
%PA2: rephrased "it is now easy to show"
let us now discuss property $0'$.
%holds as well.
%
%PA2: took out "The important observation is"
We observe that errors that may propagate from the encoder to the data block
(e.g., $X$ errors when ancilla and data interact via a {\sc cnot}
as in Fig.$\,$\ref{ECwoVER}) are decoupled from errors that
propagate the other way around ($Z$ errors). Hence detecting and
inverting errors caused in the encoder does not interfere with the
way errors initially in the data are treated by EC. Therefore, the
%DDV reword
success of our method for dealing with single faults inside the EC
circuitry is independent of whether the data block starts in the
code space or not. This implies that EC methods which satisfy
property $0'$ when ancilla verification is used (e.g., Steane's method as discussed in \cite{AGP}), will
still satisfy it if decoding of the ancilla is performed instead as described above.
For non-perfect codes we need to worry also about replacing
verification against multi-qubit errors that do {\em not}
propagate from the ancilla to the data (e.g., $Z$ errors in
Fig.$\,$\ref{ECwoVER}). (Recall that a code is called {\em
perfect} if all possible syndromes point to correctable errors, as
is e.g. the case for the
%PA2: took out [[5,1,3]] since it has not been used in threshold calculations
Steane [[7,1,3]] and Golay [[23,1,7]] codes.)
%
Verification against such errors can be easily avoided by repeating the syndrome extraction: The circuit in Fig.$\,$\ref{ECwoVER} or \ref{ECwoVER2} can be repeated three times and the syndrome for $Z$ errors in the data must only be trusted if at least two of the syndromes for $Z$ errors agree.
%DDV add parenthetical
(The same applies for the $X$ error correction.)
Finally, one technical point must be addressed. With a single syndrome extraction as in
Fig.$\,$\ref{ECwoVER}, property $2$ is satisfied separately for
$X$ and $Z$ errors. That is, a single fault inside the ancilla
encoder may lead to a single $X$ and a single $Z$ error in the
output data block with the two errors acting on {\em different}
qubits inside the block. This is not a problem as long as the
subsequent logical operation does not mix $X$ and $Z$ errors, as
e.g. is the case for the logical {\sc cnot} which are transversal
for CSS codes. If however a logical $S$ gate is applied next then
we must enforce property $2$ in a stricter sense and prevent $X$
and $Z$ errors acting on different qubits at the output of EC.
This can be achieved by extracting the syndrome a second time by
running the EC circuit again. This modification will have no
effect on the accuracy threshold since the
%DDV reword
logical $S$ gate can be handled via injection by
teleportation~\cite{SDT}.
%PA2: move next one paragraph earlier
%The second technical point depends on whether the code is perfect or not. For perfect codes (i.e., codes for which any possible syndrome points to a correctable error, as is e.g. the case for the Steane [[7,1,3]] and Golay [[23,1,7]] codes) no modification is necessary. If the code is not perfect then the EC circuit in Fig.$\,$\ref{ECwoVER} or \ref{ECwoVER2} must be used three times and the syndrome for $Z$ errors in the data must only be trusted if at least two of the syndromes for $Z$ errors agree. (The same applies for the $X$ error correction.)
A similar ancilla decoding technique can replace verification when
EC uses a quantum code that corrects $t>1$ errors. Now, properties
0 -- 3 in \S 10 in \cite{AGP} are sufficient to guarantee fault
tolerance and our decoding circuit must be designed appropriately.
Most importantly, we must ensure that, with $k\leq t$ faults
inside EC, errors acting on more than $k$ qubits that may
propagate from the ancilla to the data can be diagnosed by the
subsequent ancilla decoding and inverted. This can be accomplished
by encoding the ancilla into a $t$-bit classical repetition code
before decoding each sub-block separately similar to
Fig.$\,$\ref{ECwoVER2}. For example, for the [[23,1,7]] Golay code
that corrects $t=3$ errors, ancilla decoding can replace
verification if we encode the ancilla after interaction with the
data into the $3$-bit classical repetition code
\cite{Reichardt-private}.
\noindent {\it Some examples}---We will now give some examples of this method in use. Let us begin with fault-tolerant syndrome measurement using four-qubit cat states, which is e.g. useful for EC with the [[5,1,3]] or [[7,1,3]] codes. The circuit for encoding, verifying and interacting these cat states with the data was shown in Fig.$\,$\ref{f1}. An alternative circuit that performs the same measurement without ancilla verification was shown in Fig.$\,$\ref{f2}.
The circuit in Fig.$\,$\ref{f2} is fault tolerant because the
error $X^{(1)}X^{(2)}$ (or its equivalent $X^{(3)}X^{(4)}$) which
may appear in the encoder in first order will be detected as it
leads to all three measurements of $Z$ after decoding giving
outcome $-1$. In addition, no single fault inside the decoding
circuit can flip all three $Z$-measurement outcomes, something
%DDV to
which would lead us to mistakenly cause a weight-two error in the
data by applying an unnecessary correction.
The second example is fault-tolerant measurement using seven-qubit cat states, which is e.g. needed for logical measurements that prepare ancilla needed for universality in the [[7,1,3]] code (see Figs.$\,$13 and 14 in \cite{AGP}). The circuit for encoding and verifying these cat states (Fig.$\,$14 in \cite{AGP}) is shown in Fig.$\,$\ref{7cat}.
%
\begin{figure}[t]
\begin{center}
\includegraphics[width=4.8cm,trim=0 0 0 0]{7cat}
\caption{\label{7cat}
\footnotesize{
Encoding and verification of a $7$-qubit cat state. Similar to Fig. 1 of the main text, a verifier qubit is used to measure the parity $Z^{(1)}Z^{(7)}$ on the ancilla. Any of the high-weight $X$ errors ($X^{(1)}X^{(2)}$, $X^{(6)}X^{(7)}$, $X^{(1)}X^{(2)}X^{(3)}$, or $X^{(5)}X^{(6)}X^{(7)}$) created in the encoder in first order is thus detected before the ancilla interacts with the data.
}
}
\end{center}
\end{figure}
The verification can again be omitted if, after interaction with
the data, the cat state is decoded with the circuit shown in
Fig.$\,$\ref{7cat_dec}. To understand how the outcomes of the
final measurements of $Z$ allow us to diagnose any multi-qubit $X$
errors having resulted from a single fault in the ancilla encoder, we
follow the propagation of these errors through the decoding
circuit:
%
\begin{equation}
\label{prop}
\begin{array}{rcl}
X^{(1)}X^{(2)} & \rightarrow & X^{(1)}X^{(2)}X^{(6)} \\
X^{(6)}X^{(7)} & \rightarrow & X^{(6)}X^{(7)} \\
X^{(1)}X^{(2)}X^{(3)} & \rightarrow & X^{(1)}X^{(2)}X^{(3)}X^{(6)}X^{(7)} \\
X^{(5)}X^{(6)}X^{(7)} & \rightarrow & X^{(2)}X^{(4)}X^{(6)}X^{(7)} \,
\end{array}
\end{equation}
%
An $X$ error appearing on the right-hand side of Eq.$\,$(\ref{prop}) will result in the measurement of $Z$ on the corresponding qubit giving an outcome $-1$. We note that different initial errors propagate to different final error patterns and, hence, distinct measurement outcomes which allows distinguishing them. In addition, it is straightforward to see that no single fault inside the decoder can lead to any of the final error patterns in Eq.$\,$(\ref{prop}). So no fault inside the decoder can make us mistakenly apply a multi-qubit correction to the data.
%
\begin{figure}[h]
\begin{center}
\includegraphics[width=4cm,trim=0 0 0 0]{7cat_dec}
\caption{\label{7cat_dec}
\footnotesize{
The decoding circuit that replaces verification for the $7$-qubit cat state prepared as in Fig.$\,$\ref{7cat}. The outcome of the measurement of $X$ on the fifth qubit gives the eigenvalue of the measured operator on the data. As in Fig.$\,$\ref{f2}, all measurements of $Z$ give ideally outcome $+1$.
}
}
\end{center}
\end{figure}
Our third example is fault-tolerant EC using Steane's method for the [[7,1,3]] code. In this method, logical $|0\rangle$ and $|+\rangle$ states are sequentially coupled to the data block to extract the syndrome information. In Steane's original scheme each ancilla is verified by either comparing two independently-encoded ancilla copies or by measuring suitable parities with extra verifier qubits (for the [[7,1,3]] code the measurement of a single parity is sufficient; see \cite{SDT}). In our variant of this method no verification is performed and after encoding the ancilla is allowed to interact with the data. The decoding circuit that is applied next to the ancilla is identical with the encoding circuit. A schematic of this EC method is shown in Fig.$\,$\ref{Steane_wover}, where the encoder ($\mathcal{D}^{-1}_{\rm |0\rangle}$) and decoder ($\mathcal{D}_{\rm |0\rangle}$) of the logical $|0\rangle$ state are shown separately in Fig.$\,$\ref{0enc} and Fig.$\,$\ref{0dec}, respectively.
%
\begin{figure}[t]
\begin{center}
\includegraphics[width=5.2cm,trim=0 0 0 0]{SteaneEC_wover}
\caption{\label{Steane_wover}
\footnotesize{
Steane's EC without ancilla verification for the [[7,1,3]]. The encoding circuit for the $|0\rangle$ state ($\mathcal{D}^{-1}_{\rm |0\rangle}$) is shown separately in Fig.$\,$\ref{0enc}, and the corresponding decoder, which here includes the final measurements, ($\mathcal{D}_{\rm |0\rangle}$) in Fig.$\,$\ref{0dec}. The encoder and decoder for the $|+\rangle$ state are identical with the direction of the {\sc cnot} gates reversed and with qubit preparations and measurements performed in the conjugate bases.
}
}
\end{center}
\end{figure}
To show the tolerance of this design to single faults, we first list the possible multi-qubit $X$ errors produced in first order in the encoder $\mathcal{D}^{-1}_{\rm |0\rangle}$: $X^{(1)}X^{(7)}$, $X^{(2)}X^{(3)}$, and $X^{(4)}X^{(5)}$. Propagating them through the decoding circuit of Fig.$\,$\ref{0dec} we obtain
%
\begin{equation}
\label{prop2}
\begin{array}{rcl}
X^{(1)}X^{(7)} & \rightarrow & (X^{(1)})X^{(3)}X^{(5)} \\
X^{(2)}X^{(3)} & \rightarrow & (X^{(2)})X^{(6)}X^{(7)} \hspace{0.3cm} , \\
X^{(4)}X^{(5)} & \rightarrow & (X^{(4)})X^{(6)}X^{(7)}
\end{array}
\end{equation}
%
\noindent where we have put in parenthesis trivial errors acting on qubits subsequently measured in the $X$ eigenbasis. As seen in Eq.$\,$(\ref{prop2}), the first weight-two error gives a distinct syndrome from the other two, which have identical syndromes (they both flip the eigenvalues of the measurements of $Z$ on the sixth and seventh qubit). This is however to be expected, because their product $X^{(2)}X^{(3)}X^{(4)}X^{(5)}$ is in the code stabilizer and so the same recovery operator can be applied for both.
Finally, it can be easily checked that the decoder is designed so that a single fault inside it cannot lead to any of the final error patterns in Eq.$\,$(\ref{prop2}). Note that running the encoding circuit backwards would not have provided a decoding circuit with this property: Indeed, we can see that if e.g. the {\sc cnot} gate from the second to the third qubit is applied in the first time step of the decoder, then it will not be possible to distinguish whether the error $X^{(2)}X^{(3)}$ was produced by a fault in the encoder or by this decoder {\sc cnot} gate failing.
\begin{figure}[t]
\begin{center}
\includegraphics[width=4.7cm,trim=0 0 0 0]{0enc}
\caption{\label{0enc}
\footnotesize{
Our encoder for the logical $|0\rangle$ state in the [[7,1,3]] code. Single fault events in this circuit can lead to weight-two $X$ errors ($X^{(1)}X^{(7)}$, $X^{(2)}X^{(3)}$, or $X^{(4)}X^{(5)}$), which will propagate to the data block through the transversal {\sc cnot} gates of Fig.$\,$\ref{Steane_wover}.
}
}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=5.1cm,trim=0 0 0 0]{0dec}
\caption{\label{0dec}
\footnotesize{
Our decoder for the logical $|0\rangle$ state in the [[7,1,3]] code corresponding to the encoder of Fig.$\,$\ref{0enc}. The measurements of $X$ yield the syndrome for $Z$ errors in the data block, while all measurements of $Z$ give ideally outcome $+1$.
}
}
\end{center}
\end{figure}
\noindent {\it Conclusion}---Avoiding ancilla verification is
helpful when measurements are slow since measurement outcomes need
not be available immediately. Furthermore, the EC circuits become
more efficient in both the number of qubits and operations compared to the EC circuits that use ancilla
verification. For example, using this new method to perform EC
inside the [[7,1,3]] {\sc cnot}-exRec in \cite{AGP} decreases the total
number of locations from $575$ to $351$ and the number of
%DDV by
ancillary qubits by half. Counting malignant pairs in the new
circuit nearly doubles the accuracy threshold lower
bound found in \cite{AGP} which changes from $2.73\times 10^{-5}$ to
$5.36\times 10^{-5}$. This method may also prove beneficial for
fault tolerance with geometric locality constraints since the
reduction in the number of qubits will result in a smaller unit
cell and so shorter movement. Finally, we would like to comment
that it would be interesting to investigate whether a similar
ancilla decoding technique can replace verification in the
teleported error-correction method discussed in \cite{knill:ftqc}.
\begin{thebibliography}{10}
%PA: commented out titles
\bibitem{shorver}
P.~W.~Shor.
%\newblock Fault-Tolerant Quantum Computation.
\newblock p. 56 in {\em 37th Annual Symposium on Foundations of Computer
Science} (FOCS '96), 1996.
\bibitem{shordiv}
D.~P.~DiVincenzo and P.~W.~Shor.
%\newblock Fault-tolerant error correction with efficient quantum codes.
\newblock {\em Phys. Rev. Lett.} 77, 3260--3263, 1996.
\bibitem{Steane97}
A.~Steane.
%\newblock Active stabilisation, quantum computation and quantum state synthesis.
\newblock {\em Phys. Rev. Lett.} 78(2252), 1997.
\bibitem{SDT}
K.~M.~Svore, D.~P.~DiVincenzo, and B.~M.~Terhal.
%\newblock Noise Threshold for a Fault-Tolerant Two-Dimensional Lattice.
\newblock{quant-ph/0604090}.
\bibitem{preskill:faulttol}
J.~Preskill.
%\newblock Fault-tolerant quantum computation,
\newblock pp. 213--269 in {\em Introduction to Quantum Computation}, eds. H.-K.
Lo, S. Popescu and T.P. Spiller (1998, World Scientific, Singapore).
%\newblock{quant-ph/9712048}.
\bibitem{AGP}
P.~Aliferis, D.~Gottesman, and J.~Preskill.
%\newblock Quantum accuracy threshold for concatenated distance-3 codes.
\newblock {\em Quant. Inf. Comput.} 6({2}), 97--165, 2006.
%\newblock{quant-ph/0504218}.
\bibitem{Reichardt-private}
B. Reichardt, private communication.
\bibitem{knill:ftqc}
E.~Knill.
%\newblock Quantum computing with realistically noisy devices.
\newblock {\em Nature}, 434:39--44, 2005.
\end{thebibliography}
\end{document}