What “Numerical Data” Means Here

In the highlighted phrase, “numerical data” means the quantitative input values that a quantum algorithm is supposed to store, access, estimate, or compute with. These are not just abstract symbols in a paper. They are numbers that must be made physically available to the algorithm in some representation.1

In the parent document, examples include real numbers such as

\[ x = 0.734918\ldots, \]

fixed-point bit strings, phase angles, probabilities, amplitudes, matrix entries, payoff values, model parameters, correlations, and discretized samples. In short, “numerical data” refers to the numbers that carry the problem instance.

For a data-intensive quantum algorithm, the important question is not only “how many qubits and gates does the circuit use?” but also “where are the numbers, and how accurately can the algorithm use them?”

Numerical data as problem information

A quantum algorithm usually has two kinds of structure. One part is the general procedure: the gates, measurements, and subroutines that define the method. Another part is the particular data for the problem instance. For example, in a financial risk calculation, the general procedure might be amplitude estimation, while the numerical data might include a loss distribution, asset correlations, payoff functions, and thresholds. In a quantum linear-algebra algorithm, the procedure might manipulate a block-encoding, while the numerical data may be the entries or singular-value structure of a matrix [Gilyén et al. 2019].

The parent document is concerned with this second part: the input numbers. These numbers are often high precision. If a computation needs a value to 40 or 64 bits of precision, then the algorithm must somehow access distinctions on the scale of

\[ 2^{-40} \quad\text{or}\quad 2^{-64}. \]

That is a very fine distinction. The document’s point is that one cannot simply hide those bits inside a continuous quantum parameter and then pretend that the algorithm has obtained them for free.

So “numerical data” here does not mean merely “data written using digits.” It means operationally relevant numerical information: numerical values whose small differences may change the output of the computation.

Digital numerical data

The most direct way to represent a number is digitally. For example, an \(n\)-bit approximation to a number \(x \in [0,1)\) can be written as

\[ x \approx 0.b_1b_2\cdots b_n, \]

where each \(b_j\) is a bit, either \(0\) or \(1\). A classical computer stores such information using memory bits. A quantum computer can also store classical bit strings reversibly in computational-basis states. An \(r\)-qubit register can represent one of \(2^r\) basis labels,

\[ |0\rangle, |1\rangle, \ldots, |2^r-1\rangle. \]

If those qubits are used to store an ordinary binary label, then they are acting as digital qubit memory. This is what the parent document calls “exact digital qubits”: qubits used to hold discrete, addressable information.

This kind of storage is “expensive” in memory because each bit of precision takes physical logical storage. But it is also clear what has been stored. If the algorithm has a register containing the first 40 bits of \(x\), then those 40 bits are explicitly available as digital information, subject of course to the usual constraints of reversible computation and measurement in quantum information [Nielsen and Chuang 2010].

Numerical analysis has long emphasized that real numbers are not stored exactly in ordinary computation; instead, finite representations such as floating-point numbers store approximations [Goldberg 1991; Higham 2002]. The parent document is importing that concern into quantum algorithm design: precision has to live somewhere.

Phase-encoded numerical data

Quantum systems also allow continuous parameters. A simple example is the one-qubit rotation

\[ R_z(\theta) = \begin{pmatrix} e^{-i\theta/2} & 0 \\ 0 & e^{i\theta/2} \end{pmatrix}. \]

Here \(\theta\) is a real-valued angle in the mathematical description of the operation. Because \(\theta\) can be written with arbitrarily many digits on paper, it is tempting to say: instead of storing many binary bits of a number, encode the number as an angle.

For instance, one might set

\[ \theta = 2\pi x, \]

so that the number \(x\) is represented by a phase rotation. This is a kind of continuous phase encoding.

But this is exactly where the phrase “numerical data were encoded too cheaply” becomes important. If the algorithm is merely given access to a unitary operation depending on \(\theta\), it has not automatically received all the binary digits of \(\theta\). It has received the ability to interact with a physical transformation. To extract or use fine-grained information about \(\theta\), the algorithm must distinguish nearby possible rotations.

For example, suppose two candidate data values differ by

\[ \Delta x = 2^{-40}. \]

Then their corresponding phase angles differ by

\[ \Delta \theta = 2\pi \Delta x = 2\pi 2^{-40}. \]

Those two rotations are extremely close. A single use of the rotation will generally not make them reliably distinguishable. The algorithm may need many queries, coherent repetitions, or estimation steps before that tiny phase difference becomes visible in measurement statistics. This is the link to quantum detection and estimation theory: distinguishability is not just a matter of two mathematical parameters being unequal, but of whether measurements can tell the corresponding physical states or channels apart [Helstrom 1976; Holevo 1982].

Why “numerical” matters, not just “data”

The word “numerical” is doing important work. Some data are labels: “choose branch 3,” “apply this gate,” or “look up item \(j\).” Those may require only discrete identification. Numerical data, by contrast, often have a metric structure. The difference between \(0.7349\) and \(0.7350\) may be small but meaningful. A simulation, optimization problem, risk estimate, or matrix computation may require values within a specified additive error.

That means the algorithm may be sensitive not only to which value is stored, but also to how close values are. If a task requires resolution \(\varepsilon\), then values separated by roughly \(\varepsilon\) may need to be distinguished. For phase-encoded data, resolving such small separations is a metrological problem.

A simplified model helps. Suppose the data value \(x\) appears through a unitary

\[ U_x = e^{-i x G}, \]

where \(G\) is a Hermitian generator. If an algorithm uses \(U_x\) repeatedly, then in favorable coherent settings the distinguishability between \(x\) and \(x+\Delta x\) can grow with the number of uses \(q\). Roughly speaking, one expects the accumulated phase sensitivity to involve a product like

\[ q\,\Delta x\,\|G\|, \]

where \(\|G\|\) measures the scale of the generator. Under bounded-generator assumptions, resolving \(\Delta x\) therefore requires \(q\) to grow as \(\Delta x\) shrinks. This is the intuition behind Heisenberg-limited phase scaling in ideal quantum metrology [Giovannetti, Lloyd, and Maccone 2006; Giovannetti, Lloyd, and Maccone 2011].

The parent document is not claiming that every phase encoding is bad. Rather, it is warning that phase-encoded numerical precision is not automatically usable. The data may be present in the mathematical parameter, but not yet available as operational information.

What “encoded too cheaply” means

In this context, “too cheaply” means that the representation appears to save memory by putting many bits of numerical precision into a small number of quantum degrees of freedom, such as one or a few continuous phases, without counting the cost of recovering or exploiting that precision.

For example, imagine that a computation needs a 64-bit number. A fully digital representation might store all 64 bits. A hybrid representation might store only 40 bits digitally and place the remaining 24 bits into a phase. On paper, this looks like a saving of 24 bits of digital memory. But if the algorithm later needs those 24 bits, it must resolve phase differences of order

\[ 2^{-24}. \]

In the parent document’s idealized bounded-generator picture, that can imply query costs on the order of \(2^{24}\) for a single phase channel, up to constants and modeling details. The saved memory has reappeared as execution cost.

So the highlighted phrase points to a tradeoff:

\[ \text{fewer stored digital bits} \quad\text{may require}\quad \text{more queries to extract equivalent precision}. \]

That is the main conceptual role of “numerical data” in the introduction. The book is not merely about storing arbitrary information in qubits. It is about storing and using high-precision numbers, and about how the cost of precision moves between memory and time.

The central interpretation

The safest reading is this:

In the parent document, “numerical data” means high-precision quantitative input values that a quantum algorithm must use accurately enough for its task. These values may be stored as digital bit strings, encoded into phase rotations, placed in oracles, embedded in state-preparation routines, or represented through block-encodings. The document’s concern is that if such numerical information is represented with too little exact digital memory, then the missing precision may have to be recovered through additional quantum queries or estimation effort.

This interpretation is directly supported by the surrounding text. The later claims about precise lower bounds, such as a scaling of the form

\[ q \gtrsim 2^{(n-r)/m}, \]

depend on assumptions that the parent document says will be formalized later: bounded generators, calibrated channels, constant success probability, and tasks sensitive to the displaced bits. But the meaning of “numerical data” in the highlighted sentence is already clear: it is the problem’s quantitative content, especially the digits or precision that the algorithm cannot ignore.

References

Brassard, Gilles, Peter Høyer, Michele Mosca, and Alain Tapp. 2002. “Quantum Amplitude Amplification and Estimation.” In Quantum Computation and Information, Contemporary Mathematics 305, 53–74. American Mathematical Society.

Gilyén, András, Yuan Su, Guang Hao Low, and Nathan Wiebe. 2019. “Quantum Singular Value Transformation and Beyond: Exponential Improvements for Quantum Matrix Arithmetics.” In Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing, 193–204.

Giovannetti, Vittorio, Seth Lloyd, and Lorenzo Maccone. 2006. “Quantum Metrology.” Physical Review Letters 96: 010401.

Giovannetti, Vittorio, Seth Lloyd, and Lorenzo Maccone. 2011. “Advances in Quantum Metrology.” Nature Photonics 5: 222–229.

Goldberg, David. 1991. “What Every Computer Scientist Should Know About Floating-Point Arithmetic.” ACM Computing Surveys 23 (1): 5–48. DOI: 10.1145/103162.103163.

Helstrom, Carl W. 1976. Quantum Detection and Estimation Theory. Academic Press.

Higham, Nicholas J. 2002. Accuracy and Stability of Numerical Algorithms. 2nd ed. Society for Industrial and Applied Mathematics.

Holevo, Alexander S. 1982. Probabilistic and Statistical Aspects of Quantum Theory. North-Holland.

Nielsen, Michael A., and Isaac L. Chuang. 2010. Quantum Computation and Quantum Information: 10th Anniversary Edition. Cambridge University Press.