1. Quick check §
What is the optimal equal-prior success at θ=0?
Hint 1
The two states coincide.
Worked solution
It is 1/2; there is no evidence to distinguish the labels.
Quantum Information Theory · Chapter 10
Research Practicum
How do we turn a theorem or small calculation into a reproducible research investigation?
In this chapter · 3 sections
Reading guide · preparation and approach
Starting point: Chapters 1–9; basic Python is useful for reproducing numerical examples.
Anchor example: Three bounded projects: binary testing, finite-block compression, and noise correction.
Method: Begin with an analytic benchmark, vary one assumption, and separate evidence from proof.
Watch for: A parameter scan cannot establish a universal statement or a channel capacity.
This lesson includes explanations, worked examples, and problems with solutions. Advanced results whose proofs are outside the chapter are identified explicitly.
10.1 Project A: certify a measurement
10.1.1 A spectral benchmark before optimization
Take the equally likely states \(\rho=|0\rangle\langle0|\) and \(\sigma_\theta=|\psi_\theta\rangle\langle\psi_\theta|\), where \(|\psi_\theta\rangle=\cos\theta|0\rangle+\sin\theta|1\rangle\), with \(0\le\theta\le\pi/2\). The analytic benchmark from Chapter 1 is \(p_{\rm guess}=(1+\sin\theta)/2\). At zero angle the best guess is random; at a right angle it is perfect. A numerical routine that misses these endpoints is not ready for a harder instance.
For each angle, diagonalize \(\rho-\sigma_\theta\), construct the projector onto its positive eigenspace, and compute the success directly. Record Hermiticity, eigenvalue bounds, and completeness of the two effects. Then change the prior to 0.8 and use the weighted difference instead. The equal-prior formula is no longer the benchmark; the weighted trace-norm expression is.
A useful extension restricts the receiver to computational-basis measurements. Compare its performance with the unrestricted optimum and identify where the gap is largest. Deliver the state definitions, parameter grid, precision, explicit effects at representative points, and residuals against the analytic result. These computations reproduce a known theorem and explore a constrained variant; they are not evidence of a new universal discrimination formula.
Worked check 10.1. At θ=π/6 and equal priors, what values should the unrestricted and computational-basis tests produce?
Reasoned solution
The unrestricted optimum is (1+1/2)/2=0.75. Guess |0⟩ for outcome 0 and the other state for outcome 1; success is (1+sin²θ)/2=0.625.
10.2 Project B: finite-block memory
10.2.1 Compare a bound with an achieved construction
Study the source with spectrum \((1-p,p)\), initially fixing p=0.2. For each block size n and cutoff k, retain the span of computational strings with at most k ones. Its dimension is \(D_{n,k}=\sum_{j=0}^k\binom nj\), and retained weight is \(w_{n,k}=\sum_{j=0}^k\binom nj p^j(1-p)^{n-j}\). Report memory \(m=\lceil\log_2D_{n,k}\rceil\), rate m/n, and the certified entanglement-fidelity lower bound \(w_{n,k}^2\).
At n=10,k=3, the benchmark is D=176, m=8, and w≈0.879126. To target fidelity at least 0.99 using this bound, choose the smallest cutoff with \(w^2\ge0.99\). The condition is sufficient, not a proof that smaller memories cannot work. For p below one half, weight ordering matches individual-string probability ordering, but including entire weight shells may still use more dimension than a finer selection at a boundary.
Vary n before varying p, so that changes in source entropy do not obscure finite-block effects. Plot the achieved sufficient rate against the entropy threshold. Describe the ceiling overhead and the confidence criterion. Extend to correlated binary sources only after specifying a transition model; replacing the binomial distribution without deriving a new source model would invalidate the calculation.
Worked check 10.2. Why must a report label w² a lower bound instead of “the fidelity”?
Reasoned solution
The retained-subspace Kraus contribution gives w², while fallback branches can contribute additional overlap. The inequality certifies performance but does not generally determine the full channel’s exact fidelity.
10.3 Project C: specify the noise
10.3.1 An honest bridge to a research note
Use the repetition code as a controlled experiment. Under independent bit flips with probability p and ideal recovery, the logical error is \(3p^2-2p^3\). Verify the polynomial by enumerating the eight error patterns and comparing their decoded action on an arbitrary logical superposition. Then introduce independent phase noise with probability q on each physical qubit. Any odd number of Z errors causes logical Z, giving phase-error probability \(3q(1-q)^2+q^3\). This exposes a failure mode that a bit-flip-only benchmark hides.
Choose one project and write a short research note: state the operational question, give a mathematical model, reproduce its benchmark, change one assumption, and explain the resulting limitation or conjecture. Supply equations, executable code, dependency versions, and enough intermediate data for another reader to reproduce the result. Include analytic endpoint checks and a separate discussion of floating-point tolerance.
Assessment should reward a precise negative result as much as an attractive plot. A complete submission identifies resources and error criteria, reproduces a baseline within a stated tolerance, explains the modified model, and distinguishes proved statements from numerical observations. The supplied numerical verification script checks selected course examples using NumPy; it is a starting benchmark, not a replacement for your derivation. For further reading, return to the theorem references in the relevant chapter and follow their original proofs before proposing a generalization.
Worked check 10.3. At p=q=0.01, compare the bit-flip logical error and the phase-flip logical error.
Reasoned solution
The bit-flip logical error is 0.000298. The phase error is 3(0.01)(0.99)²+(0.01)³=0.029404. Redundancy suppresses the modeled X errors while increasing exposure to uncorrected phase noise.
Summary
- Benchmark a numerical method against a calculation whose answer is known.
- Report resources, tolerances, assumptions, and whether a quantity is a bound or an exact value.
- Research readiness includes recognizing what the evidence does not establish.
Questions and worked problems
These exercises draw on the whole chapter. Attempt each question before opening a hint, explain your reasoning, then compare it with the worked solution.
2. Quick check §
What is the optimal equal-prior success at θ=π/2?
Hint 1
The states are orthogonal.
Worked solution
It is one, attained by the computational-basis measurement.
3. Quick check §
Name one check on a numerical POVM.
Hint 1
Physical effects must give valid probabilities.
Worked solution
Check Hermiticity, nonnegative eigenvalues within tolerance, and that the effects sum to identity; checking only the objective is insufficient.
4. Worked problem §
Find the computational-basis success at θ=π/4.
Hint 1
Use the restricted-measurement formula.
Worked solution
It is (1+1/2)/2=0.75, below the unrestricted value approximately 0.853553.
5. Worked problem §
For n=4,k=1,p=0.2 compute dimension and retained mass.
Hint 1
Include weights zero and one.
Worked solution
D=1+4=5, requiring three qubits. The mass is 0.8⁴+4(0.2)(0.8³)=0.8192.
6. Worked problem §
Compute the resulting certified fidelity lower bound.
Hint 1
Square the retained mass from the previous problem.
Worked solution
0.8192²=0.67108864. A three-qubit memory in this construction does not certify near-perfect fidelity.
7. Worked problem §
Derive the odd-parity phase-error probability for three independent phase flips.
Hint 1
Add the exactly-one and exactly-three cases.
Worked solution
It is 3q(1−q)²+q³=3q−6q²+4q³. The bit-flip syndrome does not detect these logical phase flips.
8. Misconception §
Do repeated optimizer runs prove global optimality?
Hint 1
Look for a bound that covers all feasible points.
Worked solution
No. Multiple runs can find the same local solution. A matching dual certificate, analytic bound, or applicable global theorem is needed.
9. Misconception §
Does an exhaustive parameter grid prove a continuous inequality?
Hint 1
A grid omits points between samples.
Worked solution
No. It is evidence on tested points. Rigorous interpolation bounds, interval methods, or an analytic argument are required to cover a continuum.
10. Extension §
Specify a reproducible extension of one project.
Hint 1
Change one assumption and identify the new benchmark.
Worked solution
For example, vary the prior in Project A. Record priors, angles, density matrices, numerical precision, and the constructed effects. Compare success against the weighted trace-norm formula and retain feasibility residuals. This is a verifiable extension without claiming a new theorem.