All interactions are built upon the ability of a user to target the element they want to interact with, regardless of the input modality.
To enable a user to work with an experience successfully, we need alignment between the system's calculated understanding of their intent, and the user's actual intent. To the degree that the system interprets the user's intended actions correctly, satisfaction increases and performance improves.
The gaze vector has been shown repeatedly to be usable for fine targeting, but often works best for gross targeting (acquiring somewhat larger targets). Minimum target sizes of 1-1.5 degrees should allow successful user actions in most scenarios, though targets of 3 degrees often allow for greater speed. Note that the size that the user targets is effectively a 2d area even for 3d elements - whichever projection is facing them should be the targetable area. Providing some salient cue that an element is "active" (that the user is targeting it) is extremely helpful - this can include treatments like visible "mouseover" effects, audio highlights or clicks, or clear alignment of a cursor with an element.
Users will often fail to find very-high or very-low UI elements, focusing most of their attention on areas around their main focus (usually roughly eye level). Placing most targets in some reasonable band around eye level can help. Given the tendency for users to focus on a relatively small visual area at any time (the attentional 'cone' of vision is roughly 10 degrees), grouping UI elements together to the degree that they're related conceptually can leverage attention-chaining behaviors from item to item as a user moves their gaze through an area.
If user intent to target something can be determined (or approximated closely enough), it can be very helpful to accept "near miss" attempts at interaction as though they were targeted correctly. There are a handful of successful methods that can be incorporated in HoloLens experiences:
This should be turned on most/all of the time. This technique removes the natural head/neck jitters that users may have. Also movement due to looking/speaking behaviors.
These work best in areas with sparse interactive content. If there is a high probability that you can determine what a user was attempting to interact with, you can supplement their targeting abilities by simply assuming some level of intent.
This mechanism is useful in tasks requiring speed. When a user is moving through a series of targeting/activation maneuvers at speed, it can be useful to assume some intent and allow missed steps to act upon targets which the user had in focus slightly before or slightly after the tap (50ms before/after was effective in early testing).
This mechanism is useful for pathing movements, reducing the slight jitter/wobble due to natural head movement characteristics. When smoothing over pathing motions, smooth by size/distance of movements rather than over time
This mechanism can be thought of as a more general version of "Closest link" algorithms - drawing a cursor toward a target, or simply increasing hitboxes (whether visibly or not) as users approach likely targets, using some knowledge of the interactive layout to better approach user intent. This can be particularly powerful for small targets.
When determining which nearby interactive elements to give focus to, provide a bias to the element that is currently focused. This will help reduce erratic focus switching behaviours when floating at a midpoint between two elements with natural noise.