Differentiation schemes

Differentiation Schemes

Goal

Provide a reliable workflow to select a differentiation scheme, generate stencils, and assess accuracy for simulation discretization.

Requirements

• Python 3.10+
• NumPy (for stencil computations)
• No heavy dependencies

Inputs to Gather

Input	Description	Example
Derivative order	First, second, etc.	1 or 2
Target accuracy	Order of truncation error	2 or 4
Grid type	Uniform, nonuniform	uniform
Boundary type	Periodic, Dirichlet, Neumann	periodic
Smoothness	Smooth or discontinuous	smooth

Decision Guidance

Scheme Selection Flowchart

Is the field smooth?
├── YES → Is domain periodic?
│   ├── YES → Use central differences or spectral
│   └── NO → Use central interior + one-sided at boundaries
└── NO → Are there shocks/discontinuities?
    ├── YES → Use upwind, TVD, or WENO
    └── NO → Use central with limiters

Quick Reference

Situation	Recommended Scheme
Smooth, periodic	Central, spectral
Smooth, bounded	Central + one-sided BCs
Advection-dominated	Upwind
Shocks/fronts	TVD, WENO
High accuracy needed	Compact (Padé), spectral

Script Outputs (JSON Fields)

Script	Key Outputs
scripts/stencil_generator.py	offsets, coefficients, order, accuracy, scheme
scripts/scheme_selector.py	recommended, alternatives, notes
scripts/truncation_error.py	error_scale, order, reduction_if_halved

Workflow

1. Identify requirements - derivative order, accuracy, smoothness
2. Select scheme - Run scripts/scheme_selector.py
3. Generate stencils - Run scripts/stencil_generator.py
4. Estimate error - Run scripts/truncation_error.py
5. Validate - Test with manufactured solutions or grid refinement

Conversational Workflow Example

User: I need to discretize a second derivative for a diffusion equation on a uniform grid. I want 4th-order accuracy.

Agent workflow:

1. Select appropriate scheme (boundary type was not stated; if the domain is bounded rather than periodic, add --boundary to surface one-sided/ghost-cell guidance, or ask the user):

   python3 scripts/scheme_selector.py --smooth --order 2 --accuracy 4 --json
   

2. Generate the stencil:

   python3 scripts/stencil_generator.py --order 2 --accuracy 4 --scheme central --json
   

3. Result: 5-point stencil with coefficients [-1/12, 4/3, -5/2, 4/3, -1/12] / dx².

Pre-Discretization Checklist

• Confirm derivative order and target accuracy
• Choose scheme appropriate to smoothness and boundaries
• Generate and inspect stencils at boundaries
• Estimate truncation error vs physics scales
• Verify with grid refinement study

CLI Examples

# Select scheme for smooth periodic problem
python3 scripts/scheme_selector.py --smooth --periodic --order 1 --accuracy 4 --json

# Generate central difference stencil for first derivative
python3 scripts/stencil_generator.py --order 1 --accuracy 2 --scheme central --json

# Generate 4th-order second derivative stencil
python3 scripts/stencil_generator.py --order 2 --accuracy 4 --scheme central --json

# Estimate truncation error
python3 scripts/truncation_error.py --dx 0.01 --accuracy 2 --scale 1.0 --json

Error Handling

Error	Cause	Resolution
order must be positive	Invalid derivative order	Use 1, 2, 3, ... (max 6)
order must be <= 6	Derivative order too large	Use 1–6
accuracy must be even for central	Odd accuracy requested for central scheme	Use 2, 4, 6, ...
scheme must be central, forward, or backward	Invalid --scheme value	Use central, forward, or backward

Interpretation Guidance

Stencil Properties

Property	Meaning
Symmetric offsets	Central scheme (no directional bias)
Asymmetric offsets	One-sided or upwind scheme
More points	Higher accuracy but wider stencil

Truncation Error Scaling

Accuracy Order	Error Scales As	Refinement Factor
2nd order	O(dx²)	2× refinement → 4× error reduction
4th order	O(dx⁴)	2× refinement → 16× error reduction
6th order	O(dx⁶)	2× refinement → 64× error reduction

Common Stencils

Derivative	Accuracy	Points	Coefficients (× 1/dx or 1/dx²)
1st	2	3	[-1/2, 0, 1/2]
1st	4	5	[1/12, -2/3, 0, 2/3, -1/12]
2nd	2	3	[1, -2, 1]
2nd	4	5	[-1/12, 4/3, -5/2, 4/3, -1/12]

Verification checklist

Before trusting a generated stencil or accepting a scheme recommendation, record concrete evidence for each item below:

• Ran stencil_generator.py --json and confirmed results.accuracy matches the requested order AND that len(results.offsets) equals the expected stencil width (e.g. 5 points for a 4th-order central second derivative); for a central scheme also verified the offsets are symmetric about 0.
• Sanity-checked the returned coefficients against references/stencil_catalog.md: confirmed they sum to ~0 (consistency: the operator annihilates a constant) and reproduce a known catalog stencil for at least one standard case (e.g. 2nd-order d²/dx² gives [1, -2, 1]/dx²).
• For a central scheme, confirmed --accuracy is even (odd values exit 2 with accuracy must be even for central); recorded the actual exit code rather than assuming the requested order was achieved.
• Recorded the error_scale, order, and reduction_if_halved from truncation_error.py and confirmed reduction_if_halved == 2**accuracy, then compared error_scale against the smallest physical feature size (dx/L_feature from references/error_guidance.md) to confirm the grid actually resolves the physics.
• Ran an independent grid-refinement / manufactured-solution study on >=3 grids (the scripts do NOT do this) and confirmed the observed order p_obs = log(e_h/e_{h/2})/log(2) is within ~10% of the formal accuracy before quoting that order.
• For a bounded (non-periodic) domain, confirmed boundary stencils were generated/selected explicitly (--scheme forward|backward or --boundary guidance) per references/boundary_handling.md, since the interior stencil alone does not define the scheme order at the boundary.
• For non-smooth fields (shocks/fronts), confirmed scheme_selector.py did NOT recommend high-order central FD and that a limiter/WENO/upwind path was chosen instead.

Common pitfalls & rationalizations

Tempting shortcut	Why it's wrong / what to do
"The stencil generator returned coefficients, so the scheme is the order I asked for."	The accuracy field just echoes your request; it is not measured. Verify the achieved order with a grid-refinement study and confirm the coefficients match a catalog stencil and sum to ~0.
"I asked for 4th order on a central scheme with --accuracy 3, it'll just round up."	It will not — central schemes reject odd accuracy with accuracy must be even for central (exit 2). Pass an even accuracy; an odd request is an error, not a silent upgrade.
"Higher accuracy order always means lower error here."	truncation_error.py reports asymptotic scaling (scale * dx**accuracy); for a coarse grid or under-resolved feature the higher-order term need not dominate, and roundoff (O(ε/dx^p)) can win on very fine grids. Compare error_scale to the feature size, do not assume monotone improvement.
"Two grids agree closely, so it's converged."	Two grids cannot estimate observed order or confirm the asymptotic range. Use >=3 grids and compute p_obs before claiming the formal order (see references/error_guidance.md).
"The interior stencil is 4th order, so my whole solve is 4th order."	The generator emits interior stencils only; boundary closures often limit the global order. Generate one-sided/ghost-cell stencils explicitly and verify the boundary does not drop the observed order (references/boundary_handling.md).
"The field has a shock but a wide central stencil is more accurate, so use it."	High-order central FD oscillates (Gibbs) at discontinuities. scheme_selector.py recommends FV with limiter/WENO or upwind for --discontinuous; follow it rather than maximizing formal order.
"Custom --offsets let me build any stencil I want."	Offsets must be distinct, length-capped (51), and number more than the derivative order, or the script exits 2. A valid run still does not guarantee the intended accuracy — verify the coefficients and observed order.

Security

Input Validation

• --order (derivative order) is validated as a positive integer with an upper bound (order <= 6)
• --accuracy is validated as a positive integer (<= 8), and additionally must be even for central schemes
• --scheme is validated against a fixed allowlist (central, forward, backward)
• --offsets (custom stencil) is length-capped (max 51), parsed as distinct integers, and must exceed the derivative order
• --dx and --scale are validated as finite, non-negative numbers (--dx strictly positive)
• No user-supplied strings are interpolated into code paths or shell commands

File Access

• Scripts read no external files; all inputs are provided via CLI arguments
• Scripts write only to stdout (JSON output); no files are created unless the agent explicitly uses the Write tool

Tool Restrictions

• Read: Used to inspect script source, references, and user configuration files
• Bash: Used to execute the three Python scripts (stencil_generator.py, scheme_selector.py, truncation_error.py) with explicit argument lists
• Write: Used to save generated stencil coefficients or scheme recommendations; writes are scoped to the user's working directory
• Grep/Glob: Used to locate relevant files and search references

Safety Measures

• No eval(), exec(), or dynamic code generation
• All subprocess calls use explicit argument lists (no shell=True)
• Stencil computation uses only small, bounded arrays (derivative order capped at 6, accuracy at 8, and custom offset lists capped at 51 points)
• All output is deterministic JSON with no shell-interpretable content

Limitations

• Boundary handling: Stencil generator provides interior stencils; boundaries need special treatment
• Nonuniform grids: Standard stencils assume uniform spacing
• Spectral: Not covered by stencil generator

References

• references/stencil_catalog.md - Common stencils
• references/boundary_handling.md - One-sided schemes
• references/scheme_selection.md - FD/FV/spectral comparison
• references/error_guidance.md - Truncation error scaling

Version History

• v1.2.2 (2026-06-24): Added a Verification checklist (evidence-based, tied to script JSON outputs and the references) and a Common pitfalls & rationalizations table.
• v1.2.0 (2026-06-23): Enforced even-accuracy and order upper-bound validation, corrected Security/error-handling/output docs to match scripts, fixed CLI/eval examples, hardened input validation
• v1.1.0 (2024-12-24): Enhanced documentation, decision guidance, examples
• v1.0.0: Initial release with 3 differentiation scripts