Sparse Estimation Backend Crosswalk

Visual: sparse backend decision map from Jacobian structure to factorization, Schur, marginalization, covariance, and PCG.

Related docs

• Nonlinear Solver Diagnostics Crosswalk
• Solver Selection and Convergence Diagnosis
• Cholesky, LDLT, and Normal Equations
• QR, SVD, and Rank-Revealing Solvers
• Eigenvalues, Hessian Conditioning, and Observability
• Sparse Matrices, Fill-In, and Ordering
• Schur Complement, Marginalization, and PCG
• Square-Root Information and Covariance Recovery

Backend decision matrix

Sparse Estimation Backend Crosswalk is the routing page for rank and nullspaces, sparsity, ordering, and fill-in, factorization, Schur complement for solving, marginalization prior construction, covariance recovery, and PCG stagnates diagnostics.

Backend	SPD assumptions	Rank robustness	Sparsity/fill behavior	Covariance recovery	Runtime/memory	Diagnostic value
Cholesky	Requires symmetric positive definite normal equations or damped approximation.	Poor when rank deficient; failure is useful.	Very fast when ordering controls fill.	Selected covariance possible from factors if rank valid.	Excellent for well-posed sparse systems.	Factor failure exposes SPD, gauge, and conditioning issues.
LDLT	Handles symmetric indefinite systems better than Cholesky depending on pivoting.	More diagnostic for indefiniteness, still not full rank-revealing.	Fill depends on ordering and pivoting.	Useful for debugging signs and pivots.	Slightly higher overhead than Cholesky.	Pivots indicate indefiniteness or weak modes.
QR	Does not require normal-equation SPD.	More robust than normal equations.	Can create more fill than Cholesky but avoids squaring condition number.	Square-root covariance paths are natural.	More expensive but safer for debug and some production problems.	Reveals rank and conditioning better than normal equations.
SVD	No SPD assumption.	Strongest rank and nullspace diagnostic.	Usually dense or small-case tool.	Direct pseudoinverse and covariance diagnostics under threshold.	Expensive, often for reduced snapshots.	Best for explaining weak modes and threshold sensitivity.
Schur	Reduced system should be SPD after valid elimination and damping.	Depends on invertible eliminated blocks.	Can reduce problem size but create dense reduced blocks.	Camera/pose covariance possible, but fill and rank matter.	Excellent for bundle adjustment structure; risky for dense separators.	Separates eliminated-variable issues from kept-variable solve.
PCG	Requires symmetric positive definite matrix-vector operator.	Weak for rank diagnosis unless monitored carefully.	Avoids explicit factor fill; preconditioner controls convergence.	Not a covariance recovery method by itself.	Low memory, scalable if preconditioned well.	Residual norms and stagnation expose conditioning or operator bugs.

Rank/nullspaces/conditioning

Rank and nullspaces answer whether the local linear system has enough independent information. Gauge freedom is an expected nullspace caused by arbitrary global coordinates such as pose-graph translation and yaw. Rank deficiency outside the expected gauge usually means missing constraints, bad geometry, or invalid elimination.

Conditioning is different. A weak mode can be technically observable while still producing fragile steps and misleading covariance. Rank thresholds, damping, and priors can make a system appear full rank while covariance remains nonsensical. The diagnostic sequence is: expected gauge dimension, singular spectrum, weak-mode vectors, condition estimate, then physical interpretation.

Sparsity/ordering/fill-in

Sparsity is the reason estimation backends are tractable. Each residual touches a small number of variables, so J, H, and square-root factors have exploitable block structure. Ordering decides which variables are eliminated first. Fill-in is the new nonzero structure created by elimination. A poor ordering or a dense marginalization prior can turn a sparse SLAM graph into a memory-heavy dense solve.

Inspect symbolic nonzeros, numeric nonzeros, elimination tree depth, separator size, and block structure. If runtime/memory grows suddenly after graph growth, open the fill report before changing nonlinear methods.

Schur/marginalization/covariance/PCG

Schur complement for solving eliminates variables inside one linear solve and then back-substitutes their updates. Marginalization prior construction removes variables from the active estimator and stores their information as a new prior over remaining separator variables. The algebra can look similar, but the lifecycle is different: solving is temporary, marginalization changes the future problem.

Square-root information stores normalized residual and factor structure without forming a dense covariance inverse. Covariance recovery extracts selected marginal covariance blocks from the local information system or square-root factor. Marginal covariance integrates out other variables; conditional covariance holds them fixed. The inverse of an information diagonal block is not generally a marginal covariance.

PCG requires an SPD operator, a preconditioner, and telemetry: unpreconditioned residual norm, preconditioned residual norm, iteration count, stopping tolerance, symmetry test for matrix-vector products, and nonlinear progress. Stagnation symptoms include residual norms flattening, iteration limits on many nonlinear steps, and accepted nonlinear progress disappearing even though linear work is high.

Concept cards

Rank deficiency

Field	Explanation
What it means here	The linearized system lacks independent constraints in one or more directions.
Math object	Rank of J, R, or H; zero singular values.
Effect on the solve	Makes updates non-unique and covariance undefined without gauge handling.
What it solves	Diagnoses missing information and singular backends.
What it does not solve	It does not choose the physical anchor policy.
Minimal example	Relative pose graph with no fixed world pose.
Failure symptoms	Cholesky fails, covariance huge, solution drifts in global frame.
Diagnostic artifact	Singular values and nullspace basis.
Normal vs abnormal artifact	Normal nullspace matches expected gauge; abnormal nullspace contains constrained states.
First debugging move	Count expected gauge freedoms and compare with numeric rank.
Do not confuse with	Poor conditioning.
Read next	Eigenvalues, Hessian Conditioning, and Observability.

Nullspace

Field	Explanation
What it means here	Directions in state perturbation space that do not change residuals locally.
Math object	Vectors n such that J n = 0 or H n = 0.
Effect on the solve	Allows arbitrary motion unless constrained or projected out.
What it solves	Explains gauge freedoms and unobservable directions.
What it does not solve	It does not imply every weak mode is exactly unobservable.
Minimal example	Global yaw in a planar relative pose graph.
Failure symptoms	State changes with no cost change, singular covariance, anchor sensitivity.
Diagnostic artifact	Nullspace vectors visualized as state perturbations.
Normal vs abnormal artifact	Normal vectors match known symmetries; abnormal vectors reveal missing factors.
First debugging move	Apply a small nullspace perturbation and verify residual change.
Do not confuse with	Eigenvector of a small but nonzero eigenvalue.
Read next	QR, SVD, and Rank-Revealing Solvers.

Gauge freedom

Field	Explanation
What it means here	Arbitrary coordinates not determined by relative measurements.
Math object	Known nullspace basis or anchor constraint.
Effect on the solve	Requires fixing, projecting, or respecting gauge in covariance interpretation.
What it solves	Separates coordinate arbitrariness from physical uncertainty.
What it does not solve	It does not add real sensor information.
Minimal example	Fixing the first pose in SLAM.
Failure symptoms	Different absolute maps with same relative residual cost.
Diagnostic artifact	Gauge dimension and anchor sensitivity test.
Normal vs abnormal artifact	Normal gauge affects only arbitrary frame; abnormal anchor changes relative geometry.
First debugging move	Remove the anchor and inspect expected singular modes.
Do not confuse with	Prior knowledge.
Read next	SLAM/VIO Observability, FEJ, and Nullspace Consistency.

Condition number

Field	Explanation
What it means here	Ratio between strongest and weakest constrained directions.
Math object	kappa(A) = sigma_max / sigma_min.
Effect on the solve	Controls numerical sensitivity and PCG convergence.
What it solves	Quantifies weak-mode fragility.
What it does not solve	It does not identify wrong residual semantics by itself.
Minimal example	Low-parallax triangulation with tiny depth information.
Failure symptoms	Slow PCG, noisy covariance, unstable small pivots.
Diagnostic artifact	Spectrum, condition estimate, weak eigenvectors.
Normal vs abnormal artifact	Normal spectrum has expected weak directions; abnormal spectrum spans many orders unexpectedly.
First debugging move	Check scaling and whitening before changing factorization.
Do not confuse with	Exact rank deficiency.
Read next	Eigenvalues, Hessian Conditioning, and Observability.

Sparsity

Field	Explanation
What it means here	Most Jacobian or Hessian blocks are zero because factors touch few variables.
Math object	Block sparse J, H, or factor graph adjacency.
Effect on the solve	Enables real-time memory and runtime scaling.
What it solves	Avoids dense linear algebra on large estimation problems.
What it does not solve	It does not guarantee low fill after elimination.
Minimal example	Odometry factor touches only two neighboring poses.
Failure symptoms	Unexpected dense rows, memory growth, slow symbolic phase.
Diagnostic artifact	Sparsity plot and block adjacency graph.
Normal vs abnormal artifact	Normal pattern follows factor topology; abnormal pattern has broad accidental coupling.
First debugging move	Print each residual block's parameter dependencies.
Do not confuse with	Fill-in.
Read next	Sparse Matrices, Fill-In, and Ordering.

Fill-in

Field	Explanation
What it means here	New nonzeros created during elimination or factorization.
Math object	Nonzero pattern of factor L, R, or reduced system.
Effect on the solve	Increases memory, runtime, and covariance recovery cost.
What it solves	It is not a solver; it explains factorization growth.
What it does not solve	It does not indicate model accuracy.
Minimal example	Eliminating a pose connects all neighboring landmarks or poses.
Failure symptoms	Runtime/memory explodes after graph growth.
Diagnostic artifact	Symbolic fill report and elimination tree.
Normal vs abnormal artifact	Normal fill grows predictably; abnormal fill jumps after ordering or marginalization.
First debugging move	Compare fill under alternative orderings on the same graph.
Do not confuse with	Original sparsity.
Read next	Sparse Matrices, Fill-In, and Ordering.

Ordering

Field	Explanation
What it means here	Variable elimination sequence used by the sparse backend.
Math object	Permutation matrix or ordered variable list.
Effect on the solve	Controls fill, separator size, and Schur structure.
What it solves	Reduces runtime and memory by exploiting graph topology.
What it does not solve	It does not repair rank deficiency.
Minimal example	Eliminating landmarks before cameras in bundle adjustment.
Failure symptoms	Direct solve becomes dense, Schur reduced matrix too large.
Diagnostic artifact	Ordering report and fill comparison.
Normal vs abnormal artifact	Normal ordering respects block structure; abnormal ordering creates large cliques.
First debugging move	Run AMD/COLAMD or domain-specific ordering and compare nonzeros.
Do not confuse with	Solver method choice.
Read next	Sparse Matrices, Fill-In, and Ordering.

Cholesky

Field	Explanation
What it means here	Factorization for SPD normal equations or information matrices.
Math object	H = L L^T or R^T R.
Effect on the solve	Fast direct solve when assumptions hold.
What it solves	Efficient sparse SPD systems.
What it does not solve	It does not handle rank deficiency robustly.
Minimal example	Damped pose-graph normal equations after gauge anchor.
Failure symptoms	Non-positive pivot, factorization abort, huge condition warning.
Diagnostic artifact	Pivot log and factor nonzeros.
Normal vs abnormal artifact	Normal pivots positive and stable; abnormal pivots vanish or turn negative.
First debugging move	Check SPD assumptions, gauge, and whitening.
Do not confuse with	QR on the original least-squares system.
Read next	Cholesky, LDLT, and Normal Equations.

LDLT

Field	Explanation
What it means here	Symmetric factorization that exposes diagonal or block pivots.
Math object	P^T H P = L D L^T.
Effect on the solve	Can diagnose indefiniteness and handle some symmetric systems better than Cholesky.
What it solves	Symmetric systems where pivot information matters.
What it does not solve	It is not a full rank-revealing replacement for SVD.
Minimal example	Debugging an indefinite Hessian approximation.
Failure symptoms	Negative or tiny pivots, pivoting instability.
Diagnostic artifact	D pivots and permutation.
Normal vs abnormal artifact	Normal pivots match expected definiteness; abnormal pivots reveal sign or rank problems.
First debugging move	Inspect pivot sequence and compare with damped Cholesky.
Do not confuse with	LM damping.
Read next	Cholesky, LDLT, and Normal Equations.

QR

Field	Explanation
What it means here	Least-squares factorization that avoids forming normal equations.
Math object	J = Q R.
Effect on the solve	Improves numerical behavior for ill-conditioned least squares.
What it solves	More robust linearized least-squares solves.
What it does not solve	It does not make a wrong residual correct.
Minimal example	Debugging calibration Jacobians without squaring condition number.
Failure symptoms	Higher runtime but stable answer compared with Cholesky failure.
Diagnostic artifact	R diagonal and rank estimate.
Normal vs abnormal artifact	Normal R has clear rank; abnormal has tiny diagonals in expected constrained modes.
First debugging move	Compare QR step with normal-equation step on a small case.
Do not confuse with	SVD rank thresholding.
Read next	QR, SVD, and Rank-Revealing Solvers.

SVD

Field	Explanation
What it means here	Factorization that exposes singular values and vectors directly.
Math object	J = U Sigma V^T.
Effect on the solve	Gives best local rank and nullspace diagnosis.
What it solves	Small or reduced rank-revealing debug problems.
What it does not solve	It is usually too expensive for full production sparse graphs.
Minimal example	Explaining weak yaw/scale mode in a calibration snapshot.
Failure symptoms	Singular values collapse, threshold changes covariance dramatically.
Diagnostic artifact	Singular spectrum and right singular vectors.
Normal vs abnormal artifact	Normal tiny singular vectors match known gauge; abnormal vectors reveal missing excitation.
First debugging move	Visualize the weakest singular vector as a state perturbation.
Do not confuse with	Eigenanalysis of a damped Hessian without interpreting damping.
Read next	QR, SVD, and Rank-Revealing Solvers.

Schur complement

Field	Explanation
What it means here	Algebraic elimination of a variable block to solve a reduced system.
Math object	S = B - E C^-1 E^T.
Effect on the solve	Reduces system size and exploits conditional independence.
What it solves	Landmark-heavy bundle adjustment and related block systems.
What it does not solve	It does not mean permanent marginalization by itself.
Minimal example	Eliminate points, solve camera system, back-substitute point updates.
Failure symptoms	Singular eliminated block, dense reduced matrix, wrong back-substitution.
Diagnostic artifact	Block ranks, Schur nonzeros, reduced residual.
Normal vs abnormal artifact	Normal eliminated blocks are invertible; abnormal blocks need damping or removal.
First debugging move	Validate C^-1 block solves on representative landmarks.
Do not confuse with	Marginalization prior.
Read next	Schur Complement, Marginalization, and PCG.

Marginalization prior

Field	Explanation
What it means here	Prior factor created after removing variables from the active estimator.
Math object	Schur-reduced information and right-hand side over separator variables.
Effect on the solve	Preserves old information but can introduce dense couplings and stale linearization.
What it solves	Keeps fixed-lag smoothing bounded.
What it does not solve	It does not remain exact after large future relinearization changes.
Minimal example	Marginalize old poses and keep a prior on the newest boundary poses.
Failure symptoms	Dense prior explosion, overconfident estimator, loop closure fights prior.
Diagnostic artifact	Prior residual, prior matrix sparsity, linearization point.
Normal vs abnormal artifact	Normal prior separator is controlled; abnormal prior grows dense and stale.
First debugging move	Track separator size and prior nonzeros at every marginalization.
Do not confuse with	Schur complement for a temporary solve.
Read next	Out-of-Sequence Measurements and Fixed-Lag Smoothing.

Square-root information

Field	Explanation
What it means here	Factor form of information used to whiten residuals or represent solved factors.
Math object	L where L^T L = Sigma^-1, or QR factor R.
Effect on the solve	Avoids explicit dense inverses and keeps normalized residuals.
What it solves	Stable weighting and selected covariance workflows.
What it does not solve	It does not avoid rank analysis.
Minimal example	Premultiply pose residual by inverse square-root covariance.
Failure symptoms	Double whitening, bad scale, covariance mismatch.
Diagnostic artifact	L, R, and whitened residual checks.
Normal vs abnormal artifact	Normal factor reconstructs information; abnormal factor has wrong units or order.
First debugging move	Verify L^T L against intended information.
Do not confuse with	Covariance square root.
Read next	Square-Root Information and Covariance Recovery.

Covariance recovery

Field	Explanation
What it means here	Extracting selected local uncertainty from information or square-root factors.
Math object	Blocks of H^-1, marginal covariance, or pseudoinverse under rank handling.
Effect on the solve	Used after solving for diagnostics, gating, and integrity checks.
What it solves	Provides local uncertainty when gauge and rank assumptions are explicit.
What it does not solve	It does not validate global consistency.
Minimal example	Recover camera pose marginal covariance after sparse BA.
Failure symptoms	Negative variance, overconfident weak mode, dense inverse memory blowup.
Diagnostic artifact	Selected covariance block, rank threshold, factor backsolve trace.
Normal vs abnormal artifact	Normal block is positive and gauge-aware; abnormal block changes wildly with anchor.
First debugging move	Compare selected covariance against SVD on a small snapshot.
Do not confuse with	Measurement covariance.
Read next	Square-Root Information and Covariance Recovery.

PCG

Field	Explanation
What it means here	Iterative method for large SPD systems using matrix-vector products.
Math object	Preconditioned conjugate gradient iterations for A x = b.
Effect on the solve	Reduces memory but makes convergence depend on conditioning and preconditioning.
What it solves	Large sparse or implicit Schur systems where direct factorization is too expensive.
What it does not solve	It does not work on indefinite or nonsymmetric operators.
Minimal example	Iterative Schur solve for large bundle adjustment.
Failure symptoms	PCG stagnates, reaches iteration limit, nonlinear cost stops improving.
Diagnostic artifact	Unpreconditioned and preconditioned residual norms, iteration count, tolerance.
Normal vs abnormal artifact	Normal residuals decrease with accepted nonlinear progress; abnormal residuals flatten or oscillate.
First debugging move	Test symmetry and compare to a direct solve on a small case.
Do not confuse with	Schur complement itself.
Read next	Schur Complement, Marginalization, and PCG.

Preconditioner

Field	Explanation
What it means here	Approximation that improves PCG convergence by reshaping the linear system.
Math object	Matrix or operator M approximating A or A^-1.
Effect on the solve	Changes iteration count and residual decay.
What it solves	Reduces effective condition number for iterative solves.
What it does not solve	It does not fix an invalid SPD assumption.
Minimal example	Block-Jacobi preconditioner for pose blocks.
Failure symptoms	Many PCG iterations, residual plateaus, nonlinear budget wasted.
Diagnostic artifact	Preconditioned residual norm and iteration comparison.
Normal vs abnormal artifact	Normal preconditioner improves decay; abnormal preconditioner has little effect or breaks symmetry.
First debugging move	Compare identity, diagonal, and block preconditioners on the same snapshot.
Do not confuse with	Whitening or LM damping.
Read next	Schur Complement, Marginalization, and PCG.

Sources

• Ceres Solver, "Solving Non-linear Least Squares": https://ceres-solver.readthedocs.io/latest/nnls_solving.html
• Davis, "Direct Methods for Sparse Linear Systems": https://people.engr.tamu.edu/davis/publications.html
• Saad, "Iterative Methods for Sparse Linear Systems" homepage: https://www-users.cse.umn.edu/~saad/books.html
• Dellaert and Kaess, "Square Root SAM": https://www.cc.gatech.edu/~dellaert/pubs/Dellaert06ijrr.pdf
• Triggs, McLauchlan, Hartley, and Fitzgibbon, "Bundle Adjustment - A Modern Synthesis": https://www.cs.jhu.edu/~misha/ReadingSeminar/Papers/Triggs00.pdf