Nominal vs as-built lens performance: what MTF spec really tells you

You spec a lens from its datasheet MTF curve, build the station, and the images come in softer than the curve promised. Nothing is broken: the curve is a nominal prediction for a lens built perfectly, and every production lens carries manufacturing tolerances that spend some of that contrast. This guide separates nominal from as-built lens performance, hands you a planning margin to derate published curves, and shows how to verify a lens at your own working distance, f-number and field points. You will leave with a worked contrast budget, a bench-check list and a clear pass or fail rule for ordering.

What does nominal lens performance actually promise?

Nominal performance is the ray-traced prediction for a lens built exactly to its design: every element perfectly centered, spaced and made of the specified glass, so a datasheet MTF curve is a probability-free idealization, not a production promise.

A designer traces rays through the prescription — every surface radius, thickness, glass index and the mechanical stop — and computes MTF, distortion and spot size for any working distance and f-number requested. Run the model twice and the curve comes back identical, which is what makes nominal data the right tool for ranking designs: it compares optics on equal terms, with no build noise mixed in. That is the reading the guide to lens performance curves teaches, and it is why shortlisting machine vision lenses by nominal numbers is sound practice. Distortion conventions get the same treatment in geometric versus TV distortion.

What the nominal model cannot tell you is whether the factory can hold the prescription: no tolerance bands, no yield estimate, no unit-to-unit spread. Treating the printed curve as a per-unit promise is the exact mistake this guide fixes.

What changes when a lens is built to tolerance?

As-built performance is the statistical prediction for lenses built with real tolerances: element positions, surface shapes, thicknesses and glass indices all vary within their bands, and a tolerance model for a typical multi-element vision lens tracks on the order of 100–400 such parameters, depending on element count and barrel design.

Production holds each parameter inside a band rather than at its design value. An element sits a few hundredths of a millimetre off axis (decenter), leans a fraction of a degree (tilt), runs a few hundredths of a millimetre thick, and its glass melt refracts slightly differently from the catalogue value. As-built analysis samples those bands at random, ray-traces the assembled lens, scores MTF at the specified frequencies and field points, and repeats the exercise thousands of times. The output is not one curve but a distribution — a probability that a randomly picked unit meets the requirement — and designers trade yield against cost by tightening or loosening each band. Design teams can also bank margin up front: one published tolerance-equivalence method simply designs at an enlarged field angle, so the built lens keeps headroom in hand.

The mount belongs to the same story. Flange focal distance carries its own tolerance, which is why the C-mount flange tolerance and shim fix exists — alignment error does not stop at the last glass surface.

How much contrast does manufacturing tolerance really cost?

Plan on 80–90% of the nominal contrast at your working spatial frequency as a production budget, because tolerance-driven losses of 10–20% are the realistic span for a well-made machine vision lens, not a defective one.

Treat the span as a budgeting margin rather than a property of any one lens. Simple few-element designs cluster near the top of the band; long, many-element designs drift toward the bottom, and corner field points sit below the axis because tilt-driven aberrations grow with field height. The loss comes from two mechanisms: lateral misalignments, which add aberrations no refocus can remove, and focus shifts, which your depth of field absorbs up to a limit. The MTF contrast at Nyquist guide reads curves at the sensor limit; apply this margin on top of it.

Use k = 0.9 on-axis for a short, few-element design and k = 0.8 at the corners of a complex one. The exact probability for a specific series is a Monte-Carlo result over its tolerance file, so on a close decision the toleranced number replaces this rule of thumb. The margin sits on top of application mismatch, which comes next — do not spend it twice.

Why does your bench measurement disagree with the datasheet curve?

A datasheet curve is computed at one design working distance, one f-number, one wavelength set and defined field points, so your bench image differs because your optics, sensor, lighting and alignment sit elsewhere, before any tolerance is applied.

The hard ceiling comes first. Diffraction cuts a lens off completely at the spatial frequency ν_c = 1 / (λ × N), with λ the wavelength in mm and N the f-number:

Four variables then move the curve under you:

• Working distance. Focus and magnification both follow distance: the field of view versus working distance relation quantifies magnification, and focus error costs MTF directly until it exhausts your depth of field.
• f-number. At λ = 0.55 µm, an F2.8 optic cuts off at 1 / (0.00055 mm × 2.8) ≈ 649 lp/mm; the same optic at F8 cuts off at 1 / (0.00055 mm × 8) ≈ 227 lp/mm. That 2.9× swing dwarfs any tolerance effect. The focus side of the same knob is in the aperture, f-number and depth of field guide, and the blur spot itself is the Airy disk.
• Wavelength. Nominal curves are computed over a design band, while a narrow red LED focuses at its own plane; a curve read at 550 nm can mislead a 660 nm station by the chromatic focal shift.
• Field point. Curves are printed on-axis and at defined field heights, commonly 0.5 and 1.0 of image height, and tangential and sagittal contrast separate away from the axis. A defect at 0.8 field needs the curve at 0.8 field.

Match the curve to your station first; only then argue about tolerances. The sensor sets which frequency matters, and the lp/mm to pixel pitch method makes that conversion.

How do you budget MTF for an inspection?

Budget from the sensor up: convert pixel pitch to a spatial frequency, read nominal MTF there, multiply by 0.8–0.9, and require the result to clear your algorithm's contrast threshold with margin.

Worked example — a 2/3″ 5 MP sensor with 3.45 µm pixels, 660 nm red light, lens at F2.8:

1. Sensor Nyquist frequency: f_N = 1 / (2 × 0.00345 mm) = 144.9, call it 145 lp/mm. A two-pixel feature lives at 100 lp/mm, so read the curve there.
2. Nominal contrast at 100 lp/mm on-axis: an illustrative curve value of 0.55 modulation.
3. Derate for tolerance: 0.55 × 0.85 = 0.47 planned as-built contrast.
4. Compare with the tool threshold: subpixel edge work needs 0.30. Headroom = (0.47 − 0.30) / 0.30 = 0.57 — a 57% margin, so order.
5. Counter-case: nominal 0.35 at 100 lp/mm derates to 0.35 × 0.85 ≈ 0.30, zero headroom. Reject, and recover with a higher-contrast series, larger features, or coarser pixels that drop the working frequency.
6. Diffraction sanity check: at F2.8 the cutoff is ≈ 649 lp/mm, so 100 lp/mm sits at 15% of cutoff, well below the collapse region. At F8 the cutoff is ≈ 227 lp/mm and 100 lp/mm sits at 44% of cutoff, where the curve is already falling — read the F8 curve, not the F2.8 one.

Check pixel-to-feature matching directly in the camera resolution calculator:

The pass rule that falls out: order when derated contrast clears your threshold by at least 20%, redesign — optics, feature size or pixels — when it does not.

Does decenter hurt a lens more than a thickness error?

Element decenter and tilt usually cost more image quality than thickness or radius errors, because a tilted or off-center element breaks the symmetry and adds coma and astigmatism across the field, while a thickness error mostly shifts focus, which you can refocus out.

Defocus is the forgiving error. A thickness or spacing slip moves best focus along the axis, and the focus ring finds it again; spacing errors can even be shimmed out, as the lens spacer and shim guide shows. Lateral errors are not forgiving: a decentered element bends the wavefront asymmetrically, and the result is one-sided blur — the coma and astigmatism signature — that survives every refocusing attempt. Blur that remains after careful refocus can equally be spherical aberration driven to the edge of its band; the diagnosis differs, the remedy is the same: swap the unit.

Tolerance is only the starting alignment. Vibration, thermal cycling and a worn focus lock all move a barrel away from its as-built state — the same mechanics that blur lenses on mobile robots. Where the station rides a moving axis, an anti-vibration barrel holds element alignment mechanically. The LS-FFL-0475 is a 36.61 mm, 2/3″ C-mount anti-vibration lens.

Do telecentric measurement lenses carry the same risk?

Yes — telecentric measurement lenses carry the same MTF tolerances plus two more: magnification and telecentricity drift with the same element tilts, so the scale you calibrate can differ from the nominal scale printed in the datasheet.

Telecentric designs earn their keep by holding magnification constant across a depth range, which is what the entocentric versus telecentric comparison shows. Tolerance then adds three drifts to the MTF story:

• Magnification. For illustration, a ±0.2% magnification tolerance on a nominal 1× lens turns a 20.00 mm gauge into 19.96–20.04 mm — a ±40 µm band, about 12 pixels of uncertainty at 3.45 µm pixels. Never ship a tolerance-critical measurement on nominal magnification: calibrate each lens's scale against a certified standard, and re-check after any mechanical disturbance.
• Telecentricity. Element tilts angle the chief rays, so a part-height difference reads as a size difference — the mechanism behind the concentricity errors measured at different heights. Object-space pixel size inherits the same drift, which the object-space pixel size calculation converts to microns.
• Field of view. The nominal telecentric field of view calculation sizes the field the same idealized way, and the calibration reference pose guide shows where residual scale error surfaces in real measurements.

The LS-TCL-2118 is a 1× object-side telecentric lens, C-mount, with 108 mm working distance for 2/3″ sensors.

Longer standoffs move to lower magnification over larger formats: 0.166× over a 1″ sensor at 258 mm working distance, or 0.287× bi-telecentric over APS-H at 288 mm — both telecentric lenses reward per-unit calibration.

When should you stop trusting the nominal curve?

Stop treating the nominal curve as the answer when your derated contrast lands within roughly 20% of the threshold your software needs, because then normal production spread can flip individual lenses between pass and fail.

Keep using nominal data for ranking and shortlisting — comparison is what it is good at. Distrust it at three moments:

• Thin margins. A derated value of 0.36 against a 0.30 threshold gives (0.36 − 0.30) / 0.30 = 20% headroom; one weak unit in the order turns a passing station into a failing one. Change the design instead — a higher-contrast series at your frequency, larger features, or a sensor whose pixels move the feature to a lower frequency. The feature-size-to-MTF selection guide runs the same logic from the feature side, and running a lens off its design magnification costs contrast for reasons no tolerance budget covers (why MTF falls off design magnification).
• High pixel counts. Small pixels push the working frequency up the curve, where absolute contrast is already low, so the same 10–20% tolerance loss costs proportionally more detections. The LS-FFL-0567 is a 16 mm C-mount lens for 1.1″ sensors, specified at 20 MP with an F2.8–22 iris and less than 0.3% distortion.

• Substitutions. A replacement with an identical nominal curve can still be a different tolerance build, and catalogue entries marked as end of life soon invite such swaps. Bench a sample of the new part at your working distance before it enters production.

Large formats raise the stakes on both frequency and alignment. The LS-FFL-1676 is a 60 mm fixed focal length lens for 65 MP sensors, M58 mount, with 214–1536 mm working distance.

What do you check before you order?

Check four things before you order: the spatial frequency your sensor needs, nominal contrast there at your f-number and wavelength, the derated 80–90% value against your software threshold, and the WD and field points where each curve is defined.

Run this list for every lens on the bill of materials:

• Frequency from pixels. Compute f_N = 1 / (2 × pixel pitch) and read curves there, not at comfortable round numbers. Contrast at the wrong frequency proves nothing about your station.
• Curve conditions. Confirm the f-number, wavelength band, working distance and field heights the datasheet used. Each one moves the numbers, as the sections above show.
• Derated margin. Multiply nominal by 0.8 at corners and 0.9 on axis, and require at least 20% headroom over your tool threshold. That margin is what unit-to-unit spread spends.
• Mechanical plan. Decide how each unit gets focused at your working distance and, for measurement, how its scale gets calibrated — shims, spacers or a focus mechanism. As-built focus and scale are per-unit properties.
• Sample check. Bench the actual model at your station — slanted edge or resolution target at your WD, iris and lighting — before the order closes. It collapses every unknown above into one measurement.

Shortlist in the lens finder, configure a candidate at the lens configurator step, and find the rest of the selection maths on the lenses topic page. If a margin lands thin, send the curve readings through the contact page and we will size the alternative series with you.