Quantum Source Coding and Data Compression

Introduction

Data compression removes redundancy so that information can be represented with fewer physical resources. In classical information theory, a source produces symbols according to a probability distribution, and long sequences are compressed by exploiting the fact that most sequences belong to a highly probable typical set. Quantum source coding addresses an analogous problem: a source produces quantum states, and the goal is to store or transmit many independently prepared systems using fewer qubits while preserving the ability to reconstruct the source states with high fidelity.

The central result is Schumacher compression. It states that an independently and identically distributed quantum source with density operator ρ can be compressed asymptotically to approximately S(ρ) qubits per signal, where S(ρ) is the von Neumann entropy. This result is quantum information theory’s counterpart to Shannon’s noiseless coding theorem.

Quantum compression is not simply classical compression applied to measurement outcomes. Measurement may destroy superposition and phase information. Instead, compression must preserve the quantum state itself, including its coherence and its entanglement with external systems.

Learning Objectives

By the end of this chapter, a learner should be able to:

1. Explain typical sets and typical subspaces.
2. Calculate the von Neumann entropy of simple quantum ensembles.
3. Describe ensemble coding and the Schumacher compression protocol.
4. Compare classical Shannon coding with quantum source coding.
5. Explain fidelity, asymptotic rates, and the meaning of compression limits.
6. Identify practical limitations in implementing quantum compression.
7. Apply the theory to a small quantum communication scenario.

1. Classical Source Coding as the Starting Point

A classical memoryless source emits symbols X from an alphabet 𝒳 according to probabilities p(x). For n independent emissions, a sequence xⁿ = (x₁, …, xₙ) has probability

The source entropy is

where H(X) is measured in bits per symbol. The noiseless coding theorem says that, for long sequences, reliable lossless compression is possible at any average rate greater than H(X) bits per symbol, while a rate below H(X) is impossible in the asymptotic limit under the usual source assumptions.

A sequence is strongly typical when the relative frequency of each symbol is close to its probability. The weak typicality condition can be expressed through the sequence probability:

For sufficiently large n, the typical set contains almost all probability, approximately 2ⁿᴴ⁽ˣ⁾ sequences, and each typical sequence has probability approximately 2⁻ⁿᴴ⁽ˣ⁾. Thus, the encoder need only identify one of roughly 2ⁿᴴ⁽ˣ⁾ likely sequences, requiring approximately nH(X) bits.

This argument is an asymptotic statement. It does not mean that every finite data block can be compressed to exactly its entropy, nor does it imply that atypical sequences can be ignored in applications where errors are unacceptable.

2. Quantum Sources and Ensembles

A quantum source is commonly described by an ensemble

where:

• x labels a possible source preparation;
• p(x) is the probability of that preparation;
• ρ_x is the density operator describing the emitted quantum state.

The average or ensemble state is

A density operator is positive semidefinite and has unit trace: ρ ≥ 0 and Tr(ρ) = 1. Pure states have the form ρ = |ψ⟩⟨ψ|, while mixed states represent classical uncertainty or entanglement with an unobserved system.

An important distinction is that an ensemble and its average density operator are not identical descriptions. Different preparation procedures can produce the same ρ. If a receiver has no access to the preparation label x, all operational predictions concerning measurements on the emitted system are determined by ρ. A compression protocol designed to preserve the quantum system must therefore preserve the relevant density operator, and often must preserve its correlations with a reference system.

For n independent emissions, the average state is

The tensor product assumption means that the source is memoryless: each emission is statistically and quantum mechanically independent of the others. Sources with memory require more advanced models and may have different compression rates.

3. Von Neumann Entropy

The quantum analogue of Shannon entropy is the von Neumann entropy:

If ρ has spectral decomposition

then

The eigenvalues λᵢ are non-negative and sum to one. The entropy is measured in qubits per source state when it determines a quantum compression rate.

Important special cases are:

A pure state has zero von Neumann entropy even if its state vector is unknown to an observer. This does not mean that the state can always be replaced by no physical resource: the sender and receiver must share the description or preparation structure needed to reconstruct it. Schumacher compression concerns an ensemble and its average state, and its operational rate is determined by S(ρ), not merely by whether individual emissions are pure.

Worked Example 1: Entropy of a Qubit Source

Suppose a source emits |0⟩ with probability 0.75 and |1⟩ with probability 0.25. The average state is diagonal:

Therefore,

The asymptotic quantum compression rate is approximately 0.811 qubits per emitted state, provided the source is memoryless and the fidelity requirement is asymptotically close to one.

4. Typical Subspaces

For a classical source, typical sequences contain almost all probability. For a quantum source, the corresponding object is the typical subspace of ρ^{⊗n}.

Let the eigenvalues of ρ be λ₁, …, λ_d, with eigenvectors |i⟩. The eigenvectors of ρ^{⊗n} are tensor products |i₁⟩⊗…⊗|iₙ⟩, and their eigenvalues are products λ_{i₁}…λ_{iₙ}. The quantum typical subspace is spanned by tensor-product eigenvectors whose eigenvalue products satisfy an approximate entropy condition:

For any fixed tolerance δ > 0, the typical projector Πₙ,δ has the following asymptotic properties as n becomes large:

1. The probability of the typical subspace approaches one:

1. Its dimension is approximately exponential in nS(ρ):

1. Within the typical subspace, the eigenvalues of ρ^{⊗n} are approximately 2⁻ⁿˢ⁽ρ⁾.

These statements are asymptotic and depend on the source being independently and identically distributed. The dimension result is the key compression insight: although n qubits span a Hilbert space of dimension 2ⁿ, almost all probability lies in a subspace of dimension close to 2ⁿˢ⁽ρ⁾. Representing that subspace requires approximately nS(ρ) qubits.

Figure: Schumacher compression retains the high-probability typical subspace and encodes its effective dimension using approximately nS(ρ) qubits.

5. Schumacher Compression

Schumacher compression is a quantum source-coding protocol with three conceptual stages.

5.1 Encoding

For a block of n source systems, the encoder identifies the typical subspace associated with ρ^{⊗n}. The state is decomposed into a typical component and an atypical component. The typical component is mapped isometrically into a smaller register of approximately nS(ρ) qubits. Additional degrees of freedom may be fixed to a standard state or discarded, depending on the implementation.

The encoding must be a valid quantum operation. It cannot simply select an unknown eigenvector or measure the source label, because such actions may disturb superpositions. In an ideal asymptotic description, the encoder performs a coherent projection or an equivalent unitary embedding.

5.2 Transmission or storage

The compressed register is transmitted or stored. If the source state lies in the typical subspace, the number of qubits required is approximately log₂(dim Πₙ,δ), which is close to nS(ρ).

5.3 Decoding

The receiver applies the inverse isometry, reconstructing a state in the original n-system Hilbert space. The decoded state is close to the original state with high probability. The protocol does not promise exact equality for every finite block; it promises that the average error tends to zero as n increases.

A commonly used error measure is fidelity. For a pure state |ψ⟩ and reconstructed state σ, fidelity is

For two mixed states ρ and σ, one standard convention is

The convention should always be stated because some literature uses the square root of this quantity. High fidelity means that the reconstructed state is operationally difficult to distinguish from the intended state in many applications.

The asymptotic theorem can be stated informally as follows:

• If the compression rate R satisfies R > S(ρ), there exist encoding and decoding procedures whose fidelity approaches one as n → ∞.
• If R < S(ρ), asymptotically faithful compression is impossible for the general memoryless source model.

Here R is measured in qubits per source state. The strict inequalities express asymptotic achievability and converse results; finite-block operation requires a margin, and the required margin depends on error tolerance and block length.

6. Ensemble Coding and the Role of Nonorthogonal States

Classical symbols can often be copied and inspected without disturbance. Quantum states behave differently. Nonorthogonal states cannot be perfectly distinguished. If |ψ⟩ and |φ⟩ are nonorthogonal, then

and no measurement can identify them perfectly in a single use.

An ensemble coder must therefore preserve the quantum alternatives coherently. Consider an ensemble of nonorthogonal states {|ψₓ⟩, p(x)}. The average state is

The relevant compression rate is S(ρ), not automatically H(X). In general, H(X) may be larger than, equal to, or conceptually different from S(ρ), because the classical label may contain preparation information that is not accessible as a perfectly distinguishable quantum degree of freedom.

For orthogonal states |x⟩ with probabilities p(x), the density operator is diagonal in the source basis, and

Thus, quantum source coding reduces to classical source coding in the orthogonal case. For nonorthogonal states, the quantum entropy captures the effective dimension of the ensemble’s state space.

The Holevo quantity is useful for interpreting ensemble information:

For pure signal states, S(ρₓ) = 0, so χ(𝓔) = S(ρ). The Holevo quantity bounds the classical information that can be extracted about x by measurement. It should not be confused with a direct statement that every quantum source can be replaced by χ classical bits without loss. Schumacher compression preserves quantum states, whereas Holevo theory concerns accessible classical information.

7. Classical and Quantum Source Coding Compared

The comparison must be interpreted carefully. A classical description of a quantum state may require many real numbers and may be impossible to transmit exactly with finite classical communication. Quantum compression is instead a physical compression of a quantum ensemble. It preserves the states generated by that ensemble, not necessarily an arbitrary state chosen outside the model.

8. Compression Limits and Assumptions

The rate S(ρ) is a fundamental limit under a specific task definition. It is not a universal statement that every quantum object can be compressed to its entropy. The following assumptions are important:

1. Memoryless source: The source emits ρ^{⊗n}. Correlated or nonstationary sources require other entropy rates.
2. Known source model: The encoder and decoder know, or can estimate, the relevant density operator.
3. Asymptotically large blocks: The theorem describes the limit n → ∞.
4. Vanishing error: Fidelity is required to approach one.
5. No extra side information: If the decoder has correlated classical or quantum side information, the achievable rate can change.
6. Preservation of quantum structure: Compression must retain coherence and relevant entanglement.

A source with entropy S(ρ) = 0 is supported on a one-dimensional subspace when its average state is pure. Such a source can be represented with zero asymptotic qubits per signal under the model, although initialization, synchronization, classical metadata, and physical control may still require resources.

At the other extreme, a maximally mixed d-dimensional source has S(ρ) = log₂d. It offers no asymptotic compression advantage: the typical subspace is essentially the entire dⁿ-dimensional space. For qubits, the maximally mixed state has S(ρ) = 1, so n independent maximally mixed qubits generally require n qubits for faithful universal preservation.

9. Finite-Block and Practical Considerations

Real systems use finite blocks. The approximation “approximately nS(ρ) qubits” hides several costs:

• Finite-block overhead: A tolerance δ enlarges the code space beyond the ideal entropy rate.
• Projection failure: Atypical components have nonzero probability for finite n.
• Source estimation: The density operator may be unknown or drifting.
• Circuit complexity: Constructing typical-subspace projectors can be computationally difficult.
• Noise: Storage and transmission errors may destroy the compressed state.
• Entanglement preservation: The source may be entangled with a reference system, requiring a stronger fidelity criterion.
• Latency: Block coding requires collecting many states before compression.
• Hardware constraints: Current quantum memories have limited coherence times and qubit counts.

Trace distance is another useful error measure. For states ρ and σ, it is

where ||A||₁ = Tr√(A†A) is the trace norm. Trace distance lies between 0 and 1 and has an operational interpretation as distinguishability. Fidelity and trace distance are related by standard inequalities, so a high-fidelity reconstruction generally corresponds to small distinguishability, although the exact bound depends on the convention and whether states are pure or mixed.

In engineering practice, one should report the block length, target fidelity or trace-distance error, estimated source state, code rate, and computational resources. Reporting only the entropy rate can make a theoretical result appear more practical than it is.

10. Applied Case: Compressing a Biased Qubit Source

Background

Consider a quantum sensor that produces one qubit per measurement. After calibration, the source is modeled as

The states are orthogonal in this illustrative model, but the physical data are retained as quantum states rather than measured immediately.

Problem

A storage unit records blocks of n = 1000 qubits. The design team wants to reduce the storage requirement while maintaining high average fidelity. What is the theoretical asymptotic rate, and what practical issues must be considered?

Analysis

The eigenvalues are 0.9 and 0.1. Hence,

The asymptotic ideal is therefore approximately 0.469 qubits per source state. For 1000 states, the leading-order compressed size is approximately

This is not a guaranteed finite-block code size. A real code must include a margin for atypical sequences, target fidelity, implementation overhead, and possibly error correction.

Approach and solution

The encoder estimates or uses the calibrated density operator, constructs an approximation to the typical subspace, and maps that subspace into a compressed register. The decoder applies the inverse map. If the source model is accurate and the block is sufficiently large, most of the source probability lies in the retained subspace.

Because the source states are orthogonal, the entropy agrees with the Shannon entropy of the classical bias. If the sensor instead emitted nonorthogonal pure states with the same preparation probabilities, the average density operator would need to be recomputed; the correct quantum rate would be S(ρ), not automatically 0.469 qubits.

Outcome

The theoretical saving is substantial: roughly 531 qubits per 1000 source systems at leading order. However, the actual device may save less because finite-block coding, control circuitry, noise protection, and source-estimation costs consume resources.

Learning points

1. Entropy gives an asymptotic benchmark, not a complete hardware specification.
2. Orthogonal quantum ensembles reproduce the classical entropy result.
3. Nonorthogonal signals require density-operator analysis.
4. Fidelity and block length must accompany any claim of compression.

11. Broader Applications

Quantum source coding is relevant to:

• quantum network nodes that buffer generated states;
• quantum sensor data streams;
• distributed quantum computing and memory management;
• quantum repeaters and entanglement distribution;
• compression of laboratory tomography data, where the stored object is a classical estimate rather than the unknown state itself;
• quantum simulation, in which low-entropy states may occupy a small effective subspace.

Compression can also be combined with error correction. These are different functions: source coding removes redundancy in the source, while channel coding adds redundancy to protect against noise. A complete system may first compress a source and then encode the compressed qubits using a quantum error-correcting code. The combined design must account for both the source entropy and the physical channel’s noise characteristics.

12. Contemporary Directions and Governance Considerations

Current research investigates finite-block quantum compression, sources with memory, compression with side information, and resource trade-offs involving classical communication and shared entanglement. Practical algorithms seek approximate typical-subspace transformations that can be implemented with realistic circuits rather than ideal projectors.

Quantum data governance also matters. Compression can change where quantum information is stored, who controls it, and how long it remains recoverable. In sensitive applications, designers should document source assumptions, reconstruction fidelity, failure probabilities, and whether discarded atypical components may contain information. A compression protocol should never discard states merely because they are statistically unlikely when legal, safety, or scientific requirements demand worst-case preservation.

Best-Practice Checklist

When designing or evaluating a quantum compression system:

1. Specify the source ensemble and average density operator.
2. Check whether the source is memoryless or correlated.
3. Diagonalize or otherwise characterize ρ to determine S(ρ).
4. Define the error metric and target fidelity.
5. Distinguish asymptotic rate from finite-block code size.
6. Preserve coherence and relevant reference-system correlations.
7. Include source-estimation, circuit, memory, and error-correction overheads.
8. Test robustness under source drift and hardware noise.
9. Report qubits per source state together with block length and error.
10. Do not replace quantum data by measurement outcomes unless measurement is part of the explicitly defined task.

Chapter Summary

Classical source coding compresses typical symbol sequences at a rate determined by Shannon entropy H(X). Quantum source coding replaces typical sets with typical subspaces and Shannon entropy with von Neumann entropy S(ρ). For a memoryless quantum source described by ρ, almost all of ρ^{⊗n} lies in a subspace of dimension approximately 2ⁿˢ⁽ρ⁾. Schumacher compression therefore permits asymptotically faithful storage or transmission at approximately S(ρ) qubits per source state.

The result depends on important assumptions: independent emissions, a known source model, large block lengths, and an error criterion such as fidelity approaching one. Orthogonal ensembles reproduce classical source coding, while nonorthogonal ensembles require coherent preservation of quantum states and cannot generally be treated as ordinary classical symbols. Entropy is an asymptotic limit, not a complete practical specification; finite-block effects, circuit complexity, noise, side information, and entanglement can significantly affect implementation.

Key Terms

Quantum source: A process that emits quantum states according to a specified statistical model.

Ensemble: A set {p(x), ρₓ} of states and their preparation probabilities.

Density operator: A positive, unit-trace operator representing a quantum state.

Average state: The density operator ρ = ∑ₓ p(x)ρₓ of an ensemble.

Memoryless source: A source whose n-use state is ρ^{⊗n}.

Shannon entropy: H(X) = −∑ₓ p(x)log₂p(x), the classical uncertainty in bits.

Von Neumann entropy: S(ρ) = −Tr(ρlog₂ρ), the quantum entropy in qubits.

Typical set: The high-probability set of classical sequences with approximately equal probability scale.

Typical subspace: The high-probability eigenspace of ρ^{⊗n} with dimension approximately 2ⁿˢ⁽ρ⁾.

Typical projector: The projector onto a chosen quantum typical subspace.

Schumacher compression: Asymptotically faithful quantum source coding at rate S(ρ).

Fidelity: A measure of similarity between intended and reconstructed quantum states.

Trace distance: D(ρ,σ) = ½||ρ−σ||₁, an operational distinguishability measure.

Nonorthogonal states: States whose inner product is neither zero nor one in magnitude and therefore cannot be perfectly distinguished.

Holevo quantity: χ = S(ρ) − ∑ₓp(x)S(ρₓ), an upper bound on accessible classical information.

Asymptotic rate: Resource consumption per source symbol in the limit of large block length.

Memoryless quantum source: Another term for an independently and identically distributed quantum source.

Entanglement fidelity: A fidelity criterion that tests preservation of a state’s correlations with a reference system.

Finite-block overhead: Additional resources required because a practical code uses a finite, rather than infinite, block.

Review Questions

Short Answer Questions

1. What is the von Neumann entropy of a pure quantum state?
2. Why is the typical subspace important in Schumacher compression?
3. State the asymptotic Schumacher compression rate for a memoryless source ρ.
4. Why can nonorthogonal quantum states not be treated as ordinary classical symbols?
5. Name two practical factors that make finite-block compression differ from the ideal entropy rate.

Long Answer Questions

1. Explain the analogy between the classical typical set and the quantum typical subspace.
2. Describe the encoding, transmission, and decoding stages of Schumacher compression.
3. Compare Shannon source coding and quantum source coding, including their resources and error criteria.
4. Explain why the average density operator, rather than only the preparation labels, determines the quantum compression rate.
5. Discuss the assumptions and limitations underlying the claim that a source can be compressed to S(ρ) qubits per state.

Analytical / Application Questions

1. A qubit source has eigenvalues 0.8 and 0.2. Calculate S(ρ) and estimate the leading-order number of qubits required for 10,000 independent emissions.
2. Two preparation procedures produce the same average density operator but use different labels and probabilities. Explain whether their Schumacher compression rates differ and why.
3. Design a checklist for deciding whether a proposed quantum compression experiment has demonstrated genuine compression rather than merely measured and classically stored the source data.

Multiple Choice Questions with Answers

Question 1

Which quantity determines the asymptotic Schumacher compression rate for a memoryless quantum source?

A. The number of possible preparation labels only

B. The von Neumann entropy S(ρ)

C. The number of measurements performed

D. The physical temperature of the quantum processor

Correct Answer: B. The von Neumann entropy S(ρ)

Question 2

What is the approximate dimension of the typical subspace for n independent copies of ρ?

A. nS(ρ)

B. 2ⁿ

C. 2ⁿˢ⁽ρ⁾

D. S(ρ)/n

Correct Answer: C. 2ⁿˢ⁽ρ⁾

Question 3

For an ensemble of mutually orthogonal pure states with probabilities p(x), which relationship holds?

A. S(ρ) = 0 always

B. S(ρ) = H(X)

C. S(ρ) = 2H(X)

D. S(ρ) is undefined

Correct Answer: B. S(ρ) = H(X)

Question 4

Why is measuring every quantum source state before compression generally unsuitable?

A. Measurement always increases the number of qubits

B. Measurement may destroy coherence and quantum correlations

C. Measurement changes classical probabilities into amplitudes

D. Measurement is possible only for maximally mixed states

Correct Answer: B. Measurement may destroy coherence and quantum correlations

Question 5

A maximally mixed qubit has eigenvalues 1/2 and 1/2. What is its von Neumann entropy?

A. 0 qubits

B. 0.5 qubits

C. 1 qubit

D. 2 qubits

Correct Answer: C. 1 qubit

Question 6

Which condition is required for the standard memoryless Schumacher theorem?

A. The source emits correlated states with arbitrary memory

B. The n-use source state is ρ^{⊗n}

C. Every source state must be orthogonal

D. The decoder must measure the source label

Correct Answer: B. The n-use source state is ρ^{⊗n}

Question 7

For ρ = 0.75|0⟩⟨0| + 0.25|1⟩⟨1|, what is S(ρ) approximately?

A. 0.189 bits

B. 0.500 bits

C. 0.811 bits

D. 1.500 bits

Correct Answer: C. 0.811 bits

Question 8

What does high reconstruction fidelity indicate in a quantum compression protocol?

A. The source labels were recovered with certainty

B. The reconstructed state is close to the intended quantum state

C. The compressed register contains no noise under all conditions

D. The source had zero entropy

Correct Answer: B. The reconstructed state is close to the intended quantum state

Question 9

Which statement about asymptotic compression is correct?

A. Every finite block can always be compressed to exactly nS(ρ) qubits

B. Rates below S(ρ) are always achievable with zero error

C. Rates above S(ρ) can achieve fidelity approaching one under the theorem’s assumptions

D. The rate is independent of the source density operator

Correct Answer: C. Rates above S(ρ) can achieve fidelity approaching one under the theorem’s assumptions

Question 10

For a 1000-state source with asymptotic entropy 0.469 qubits per state, what is the leading-order compressed size?

A. 46.9 qubits

B. 469 qubits

C. 1000.469 qubits

D. 1469 qubits

Correct Answer: B. 469 qubits