Chapter 40 — Synthesis Quiz
Twelve questions that span the whole book and this chapter's forward look. These are synthesis questions: most ask you to connect an advanced field back to an idea you already own. Try each before opening the answer.
1. A tensor is to a multilinear map what a matrix is to a ______. Fill in the blank and explain the analogy.
Answer
A **linear map** (Chapter 7). A matrix is the coordinate representation of a linear map; a tensor (as a multidimensional array) is the coordinate representation of a *multilinear* map — a function linear in each of several arguments. Pick a basis and the abstract map becomes a concrete array of components, in both cases.2. True or false: matrix multiplication is a special case of tensor contraction. Justify.
Answer
**True.** Contraction means summing the product over a shared index. Matrix multiplication is $(AB)_{ik}=\sum_j a_{ij}b_{jk}$ — exactly a contraction over the shared index $j$, written `np.einsum('ij,jk->ik', A, B)`. The dot product, the trace, and $A\mathbf{x}$ are also contractions. Tensors just let you contract objects with more slots.3. What is the single property that makes a complete inner product space a Hilbert space, and what does that property guarantee about an infinite Fourier series?
Answer
**Completeness** — every Cauchy sequence converges to a limit *inside the space*. For a Fourier series (an infinite sum of orthonormal basis projections), completeness guarantees the infinite sum actually converges to a function in the space, with no "energy" leaking into a missing limit (Parseval's identity, §40.3). Finite-dimensional spaces are automatically complete, which is why this book never needed the word.4. Why is the plain transpose $A^{\mathsf{T}}$ the wrong adjoint for quantum gates, and what replaces it?
Answer
Quantum states live over the **complex** numbers, where the length-preserving condition is $U^{*}U=I$ using the **conjugate (Hermitian) transpose** $A^{*}$ (Chapter 27), not $U^{\mathsf{T}}U=I$. For a complex gate like the phase gate $\mathrm{diag}(1,i)$, the plain transpose fails to give the identity while the conjugate transpose succeeds. (For real gates like the Hadamard the two coincide.)5. PageRank, spectral clustering, and PCA were presented as three different applications. What single linear-algebra idea unifies all three?
Answer
**Eigenvectors / eigenvalues** (Theme 6). PageRank is the dominant eigenvector of a link matrix (Chapter 29); spectral clustering uses the Fiedler eigenvector of a graph Laplacian (§40.4); PCA uses the top eigenvectors of a covariance matrix (Chapter 32). All three "read a matrix through its eigenvalues" to reveal hidden structure — networks, clusters, or directions of variance.6. The chapter says the derivative "is a linear map." Explain what this means for the multivariable case, and name the matrix involved.
Answer
The best local linear approximation to a map $\mathbf{f}:\mathbb{R}^n\to\mathbb{R}^m$ near a point is multiplication by a matrix — the **Jacobian** of partial derivatives. Single-variable: the derivative is multiplication by a number (the tangent slope). Multivariable: multiplication by the Jacobian (the tangent map). This is the deep link to [multivariable calculus](/calculus/multivariable) — calculus *linearizes*, and linear algebra solves the linearized problem.7. Which of the four fundamental subspaces is the home of "what a linear system $A\mathbf{x}=\mathbf{b}$ can reach," and how does this idea survive into infinite-dimensional operator theory?
Answer
The **column space** $C(A)$ (Chapter 13): $A\mathbf{x}=\mathbf{b}$ is solvable exactly when $\mathbf{b}\in C(A)$. In functional analysis the same question becomes the *range* of a bounded operator, and the kernel/null space becomes the operator's null space (Chapter 35). The four-subspace orienting questions — what can it reach, what does it kill — transfer directly to infinite dimensions (Theme 5).8. The randomized SVD recovered the exact top singular values of a $200\times 200$ matrix using only ~15 random probe vectors. Why did so few probes suffice?
Answer
Because the matrix was **effectively low rank** — it really had only ~10 meaningful directions (Chapter 31's lesson). A handful of random projections almost surely span the dominant subspace, so the exact SVD of the small sketch recovers the top singular values. The method exploits low rank; on a genuinely full-rank matrix it would only approximate.9. Restate Theme 4 ("the most applied branch of pure mathematics") and give two applications of the same SVD from different chapters.
Answer
Theme 4: the same tool, learned once, solves problems across many fields. The SVD (Chapter 30) compresses an image (Chapter 31), reduces a dataset via PCA (Chapter 32), powers matrix-factorization recommenders (Chapter 33), denoises signals, and underlies randomized methods (§40.4). One decomposition, a dozen applications — that universality is Theme 4 made concrete.10. A quantum gate is a unitary matrix, a quantum computation is a product of unitaries, and measurement projects the state. Which earlier chapters does each of these three facts come from?
Answer
**Unitary gate** — orthogonal/unitary matrices preserve length (Chapter 21). **Product of unitaries** — matrix multiplication as composition of transformations (Chapter 8); a product of unitaries is unitary. **Measurement projects** — orthogonal projection (Chapter 19), and the measured values are eigenvalues of a Hermitian observable (Chapter 27). The entire linear-algebraic skeleton of quantum computing is Parts IV–V.11. The Hessian's eigenvalues were said to control how fast an optimization converges. In one sentence, what does a large condition number of the Hessian mean geometrically, and what do practitioners do about it?
Answer
A large condition number (ratio of largest to smallest Hessian eigenvalue, Chapter 38) means the minimization surface is a long, narrow valley, so gradient descent zig-zags slowly; practitioners **precondition** — change variables (a change of basis, Chapter 16) to make the eigenvalues more equal and round out the valley. Even training speed is, underneath, a statement about eigenvalues.12. The threshold concept of this chapter is that "linear algebra is not a course you finish; it is a language you learn to read everywhere." What is the practical test, six months from now, of whether this book worked?