Linear Algebra Interview Questions
Linear algebra questions in quant interviews cluster into five recurring families: eigenvalue and spectral-structure puzzles, covariance and correlation-matrix constraints, PCA and factor structure, least-squares and projection geometry, and numerical linear algebra at scale.
The common thread: stop grinding matrix entries and ask what the matrix does to space — which directions it stretches, and which it kills.
Last updated 2 July 2026 · sub-areas, difficulty mix and firm attributions on this page are compiled directly from the 60 linear-algebra problems in the QuantVault bank; firm tags are candidate-reported, not employer-verified, and no individual author is named in our source data.
- Core sub-areas
- Eigenvalues & spectral structure · covariance & correlation matrices · PCA & factor models · least squares & projections · numerical linear algebra
- Typical difficulty
- Skews medium — 9 easy, 37 medium and 14 hard across the 60-problem set
- Who leans on it
- Quant-research loops above all — 53 of the 60 problems carry a quant-researcher tag, with candidate reports from Two Sigma, Citadel, Jump Trading, Renaissance and others
- Practice pool
- 60 problems in the bank · 8 free to open with the full worked solution
Where linear algebra shows up: mostly in quant-research phone screens and onsites, where a covariance matrix or a regression is the natural excuse to test whether you understand the objects behind the models — trading tracks tend to see the lighter, faster spectral puzzles. For how it sits inside specific funnels, see the Two Sigma interview questions guide and the Citadel interview questions guide.
The surfaceThe linear algebra sub-areas quant interviews test
Five families cover essentially every linear-algebra question in the bank. Each row pairs a sub-area with the recurring question shape and a representative type — a flavor drawn from real problems, not a leaked wording — so you can see where the difficulty actually lives.
| Sub-area | Recurring shape | Representative type (flavor, not a real question) |
|---|---|---|
| Eigenvalues & spectral structure | Read the spectrum off structure — rank, symmetry, trace, determinant — instead of the characteristic polynomial | A rank-1, all-ones or otherwise patterned matrix whose full set of eigenvalues follows from one inner product plus a trace check. |
| Covariance & correlation matrices | Translate “is this matrix valid?” into positive semidefiniteness, then test it on eigenvalues | Deciding which pairwise correlations three or four assets can simultaneously have before the matrix stops being PSD. |
| PCA, SVD & factor models | Interpret the top eigenpair as variance explained and factor exposure | Explaining what the leading principal component of a return covariance matrix captures — and using the spectral decomposition to simulate or whiten. |
| Least squares & projection geometry | Treat regression as an orthogonal projection onto a column space | A Gram-Schmidt or vector-projection setup that collapses to two lines once you write the residual as orthogonal to the span. |
| Numerical linear algebra at scale | Reorganize the computation around small Gram matrices, factorizations and memory limits | Fitting a regression when the design matrix cannot fit in memory, by streaming the sufficient statistics chunk by chunk. |
What's confirmed vs. what varies: the sub-areas and the difficulty mix above come straight from the problem set, so they are stable. Which firm asks which flavor is candidate-reported through problem tags — treat the firm attributions as directional, and expect research tracks to weight the PCA, factor-model and numerical families more heavily than trading tracks.
The patternsSignature linear algebra question patterns
Three moves generate most correct answers in this topic. Each worked box below is a 60–90 second micro-example in the interview's actual cadence — the reasoning template, applied to a generic setup rather than any firm's wording.
Spectral shortcuts — rank and trace before polynomials
Takeaway: never expand a characteristic polynomial you can avoid — count the zero eigenvalues with rank, then let trace (and determinant) pin down whatever is left.
Shape. What are the eigenvalues of \(A = uv^{\top}\) for vectors \(u, v \in \mathbb{R}^n\) with \(v^{\top}u \neq 0\)?
1. Count zeros with rank. \(A\) has rank 1, so its null space has dimension \(n-1\): the eigenvalue \(0\) appears with multiplicity \(n-1\), before any computation.
2. Find the invariant direction. \(Au = u\,(v^{\top}u)\), so \(u\) is an eigenvector with eigenvalue \(v^{\top}u\) — a single inner product finishes the spectrum.
3. Audit with trace. \(\operatorname{tr}(A) = v^{\top}u\) must equal the eigenvalue sum, and it does. The move: rank for the zeros, one special direction for the rest — the all-ones matrix is just the case \(u = v = \mathbf{1}\), with spectrum \(n, 0, \dots, 0\).
PSD constraints — when is a correlation matrix even possible?
Takeaway: a proposed correlation matrix is legitimate exactly when it is positive semidefinite — and for symmetric patterned matrices you check that on eigenvalues, not on determinants of minors.
Shape. Four random variables each have pairwise correlation \(\rho\). What values can \(\rho\) take?
1. Write the matrix structurally. The correlation matrix is \(R = (1-\rho)I + \rho\,\mathbf{1}\mathbf{1}^{\top}\) — identity plus a rank-1 piece.
2. Read off the spectrum. Along \(\mathbf{1}\) the eigenvalue is \(1 + 3\rho\); on the orthogonal complement it is \(1 - \rho\), with multiplicity 3.
3. Impose PSD. \(1 + 3\rho \ge 0\) and \(1 - \rho \ge 0\) give \(-\tfrac13 \le \rho \le 1\). The move: for \(n\) variables the bound is \(\rho \ge -\tfrac{1}{n-1}\) — mutual negative correlation runs out of room as the universe grows.
Compute around the constraint — normal equations at scale
Takeaway: when the data is too big, shrink the algebra rather than the model — OLS only ever needs the small \(p \times p\) Gram matrix, which you can accumulate one chunk at a time.
Shape. Fit \(\beta = (X^{\top}X)^{-1}X^{\top}y\) when \(X\) is \(n \times p\) with enormous \(n\) and small \(p\), and \(X\) does not fit in memory.
1. Rewrite the sufficient statistics. Split \(X\) into row-chunks \(X_i\): then \(X^{\top}X = \sum_i X_i^{\top}X_i\) and \(X^{\top}y = \sum_i X_i^{\top}y_i\) — one pass over the data in \(O(p^2)\) memory.
2. Solve the small system. A Cholesky factorization of the accumulated \(p \times p\) system finishes in negligible time compared to the data pass.
3. Flag the conditioning cost. Forming \(X^{\top}X\) squares the condition number, so mention QR or SVD as the stable alternative when the design is near-collinear. The move: name the sufficient statistics, stream them, and volunteer the numerical caveat unprompted — that last line is what interviewers listen for.
Free practiceLinear algebra practice questions by difficulty
All 8 free problems from the 60-problem linear-algebra set, grouped by difficulty — every link opens the full worked solution, so you can check your reasoning line by line. One honesty note: the free spread is 2 easy, 4 medium and 2 hard; the bank's other 52 problems, including most of the hard tail, sit in the paid tier. For depth past this page, work through the full 60-problem linear algebra set in the problem bank.
Easy warm-ups
- Eigenvalues of a Rank-1 MatrixSpectral structureeasyfree
- Eigenvalue Interpretation in PCAPCA & factorseasyfree
Medium — the interview's center of gravity
- Eigenvalues of the All-Ones MatrixSpectral structuremediumfree
- Vector Projection and Gram-Schmidt OrthogonalizationProjection geometrymediumfree
- Sampling Multivariate Normals via Spectral DecompositionCovariance & spectramediumfree
- Computing OLS When the Design Matrix Is Too Large for MemoryNumerical linear algebramediumfree
Hard — the separators
- Correlation Range with Four VariablesPSD constraintshardfree
- Maximum Ones in a Nilpotent Binary MatrixSpectral structurehardfree
The planHow to prepare for linear algebra questions
Five techniques to make reflexive, in the order they pay off. Each maps onto a sub-area from the table above, so you can drill it immediately after reading.
- Make rank-and-trace your first move on eigenvalue puzzles. Before touching a characteristic polynomial, count the zero eigenvalues from rank and constrain the rest with trace and determinant — rank-1, all-ones and nilpotent matrices all fall to this in under a minute.
- Internalize PSD as the master constraint. Every “can these correlations coexist?” question is a positive-semidefiniteness check; for patterned matrices, decompose into identity-plus-low-rank and read the eigenvalues directly.
- Learn the spectral picture of PCA cold. The top eigenvector of a covariance matrix maximizes explained variance, the eigenvalue is that variance, and \(\Sigma = Q\Lambda Q^{\top}\) doubles as your recipe for simulating or whitening correlated normals.
- Treat regression as projection. Writing the fitted values as the orthogonal projection of \(y\) onto the column space of \(X\) turns Gram-Schmidt, residual-orthogonality and omitted-variable questions into pictures instead of algebra.
- Practice the scale reflex. When a matrix is described as too large, reach for the small Gram matrix, a streaming accumulation and a factorization — and volunteer the conditioning caveat. Then run the free practice set and take the interactive linear algebra playlist on a clock.
FAQLinear algebra interview questions — frequently asked
How important is linear algebra for quant interviews?
Core for quant-research roles and a steady presence everywhere else: covariance, PCA and regression questions are all linear algebra in disguise. In this set, 53 of the 60 problems carry a quant-researcher tag, so if you are interviewing for QR you should treat it as first-tier material.
What linear algebra topics should I focus on?
Eigenvalues and spectral structure dominate — they appear in 47 of the 60 problems in this set — followed by covariance and correlation-matrix reasoning and PCA. Round it out with least-squares-as-projection and one honest pass at numerical linear algebra: Gram-matrix accumulation, factorizations and conditioning.
How hard are linear algebra interview questions?
Mostly medium: of the 60 problems in this set, 9 are easy, 37 are medium and 14 are hard. The difficulty rarely comes from long computation — it comes from choosing the right decomposition or invariant, after which the algebra usually collapses to a few lines.
Are these real quant interview questions?
They are representative, not verbatim. The problems are curated from our bank to match the linear-algebra shapes candidates report, rewritten for clarity with worked solutions we author ourselves — we never claim any wording is a leaked question.